‹ -*- Text -*- ‹ TAR's Master's Thesis Text Source File. ‹ File formatter used is R. ‹ ‹ WARNING: PARTS OF THE ORIGNIAL TEXT USED IN THIS THESIS HAVE NOT ‹ BEEN RECOVERED. THERE ARE GAPS. Two chapters and all ‹ of the case reports have not been found. ‹ ‹ *** INCOMPLETE *** INCOMPLETE *** INCOMPLETE *** INCOMPLETE *** ‹ .dv press ‹ .fo 0 timesroman10 .fo 1 timesroman10B .fo 2 timesroman10I .fo 3 timesroman14B .fo 4 timesroman8 .fo 5 gacha9 .fo 6 gacha9b .fo 7 gacha9i .fo 8 timesroman12bh .fo 9 timesroman6 .fo A hippo8 .fo B hippo10 .fo E math10 .fo F symbol10 .nr normal_lisp_font 5 .nr bold_lisp_font 6 .nr italic_lisp_font 7 .nr small_font 4 .nr tiny_font 9 .nr text_font 0 .nr fnfont 4 .nr footnote_font 4 .sr bullet_font 8 .sr math_font E .sr symbol_font F .sr greek_font B .sr small_greek_font A .def_font_macro small \\small_font! .def_font_macro text \\text_font! .def_font_macro tiny \\tiny_font! ‹ ‹ Initialize registers for macro package: ‹ .sr figure_name_form Figure \figure_number_form .nr tty_table_of_contents 1 .nr section_font 1 ‹ bold font. .so r;r macros .‹ .‹ SMALL_FONT and TINY_FONT are used to set up sub- and superscripting; .‹ GREEK_FONT and SMALL_GREEK_FONT are string registers that hold the .‹ font names for the respective fonts. If they are defined, then some .‹ lowercase greek letters will be available as string register contents .‹ (prefixed with an "s" for small): .‹ ALPHA, BETA, CHI, LAMBDA, MU, PI, TAU, THETA .‹ BULLET_FONT must be an "xxxH" font which contains the bullet character .‹ as "7" (otherwise a bold em-dash will be used). .‹ .sr mit Massachusetts Institute of Technology .sr en ‹ en-dash .sr em ‹ em-dash (followed by a word break) .sr minus ‹ minus sign .nd small_font -1 .nd tiny_font -1 .nd bold_font 1 .sd my_list_left_margin 450m .sd my_list_right_margin 450m .sd bullet_font bold_font .sr _bull_f_comp bold_font .nr bullet_font_defined 0 ‹ 0 if no special bullet font defined. .sc bullet_font_defined bullet_font _bull_f_comp .de set ' nr tab_pos_\0 hpos ' em .de tab ' if \. ' br ' en \(\tab_pos_\0!m) ' em .de para . br . if ls<150 . sp . ne 2l . ef . ne 1.5l . en . ti +5 . em .de _int_title_par ‹ Args: Center (1=yes), Text . br . if ls<200 . sp 1.2l . en . ne 3l . fs bold_font . if \0 \\1 . ef \1 . en . fs . if ls<150 . sp . en . ti +5 .em .de subtitle ._int_title_par 0 "\0" . em .de csubtitle ._int_title_par 1 "\0" . em .de bspar . para . fs bold_font \0  ' fs ' em .de ref_begin . br . ev ref_env . nr ref_num 0 . in 5 . set . fi b . ls 1 . em .de ref_end . br . ev . xe ref_env . em .de nref . sp . ne 4 . ti -5 [\+ref_num]\tab!\ . em .de draft . ls 1.6 . margin 1i 1i 2i 2i . sr left_footing sdate . sr center_footing \1*** DRAFT ***\* . sr right_footing time . nr print_footings 1 . em .de cin . be indented_block . nv save_indent \0 . in \0 . ir \0 . em .de cout . in -\save_indent . ir -\save_indent . en ‹ indented_block . em .de int_list . begin standard_list . sv list_start \0 . sv list_left_margin my_list_left_margin . sv list_right_margin my_list_right_margin . nv st_list_indent 4 . if nargs==2 . nr st_list_indent \1 . en . ilist \st_list_indent . em .de nlist ‹ NUMBERED LISTS . int_list "\\\list_count. " ‹ (ends with a tab) . em .de rlist ‹ ROMAN NUMBERED LISTS . int_list "\\\:list_count. " 6 ‹ (ends with a tab) . em .de alist ‹ LETTERED LISTS . int_list "\\\;list_count. " ‹ (ends with a tab) . em .if bullet_font_defined ‹ BULLET LISTS . de blist . int_list "bullet_font!7* " ‹ [bullet] (ends with a tab) . em . ef . de blist . int_list "bullet_font!en!* " ‹ [dash] (ends with a tab) . em . en .de elist . end_list . en ‹ standard_list . em .if small_font>-1 . if tiny_font>-1 . de sup . if font==small_font \tiny(\0) . ef 1 \small(\0) . en . em . de sub . if font==small_font \tiny(\0) . ef 1 \small(\0) . en . em . ef . de sup \small(\0) . em . de sub \small(\0) . em . en .en .if ?greek_font . sr alpha greek_font!a* . sr beta greek_font!b* . sr chi greek_font!x* . sr lambda greek_font!l* . sr mu greek_font!m* . sr pi greek_font!p* . sr tau greek_font!t* . sr theta greek_font!q* . en .if ?small_greek_font . sr salpha small_greek_font!a* . sr sbeta small_greek_font!b* . sr schi small_greek_font!x* . sr slambda small_greek_font!l* . sr smu small_greek_font!m* . sr spi small_greek_font!p* . sr stau small_greek_font!t* . sr stheta small_greek_font!q* . en ‹ Defines the string registers with substitutions for math. ‹ Requires the following fonts in the input file: Math, Symbol, Splunk. ‹ The symbols for the font numbers should be in the string registers ‹ math_font, symbol_font and splunk_font. The math_font and ‹ symbol_font must be defined to use this file. The splunk_font is ‹ optional. It is only needed if LSIG or LPI are used. The string ‹ registers must be initialized before loading this file. ‹ ‹ -- Invoked by  ‹ .sr TRI \math_font!2*‹ regular triangle .sr COMP \math_font!o*‹ functional composition .sr PRIME \math_font!(*‹ prime mark (') .sr SSIG \math_font!*‹ small summation (sigma) .sr SPI \math_font!*‹ small product (pi) .sr LSIG \splunk_font!$*‹ large summation (sigma) .sr LPI \splunk_font!)*‹ large product (pi) .sr MULT \math_font!X*‹ large "X" .sr DIV \math_font!%*‹ division sign (-:-) .sr OSUM \math_font!4*‹ orthogonal sum (+ in a circle) .sr ADDSUB \math_font!+*‹ plus or minus (symbol) .sr EQUIV symbol_font!*‹ equivalence (three horizontal bars) .sr APPROX 0~(+0.1c)~(-0.1c)*‹ approximate (two ~'s) .sr BOTTOM symbol_font!*‹ bottom symbol .sr INTER \math_font!y*‹ set intersection .sr UNION \math_font!U*‹ set union .sr SUBSET \math_font!G*‹ proper subset .sr ESUBSET \math_font!I*‹ subset (with equality) .sr NSUBSET \math_font!H*‹ not subset .sr SUPSET \math_font!J*‹ proper superset .sr ESUPSET \math_font!L*‹ superset (with equality) .sr NSUPSET \math_font!K*‹ not superset .sr NULL \math_font!O*‹ null set .sr MEM \math_font!B*‹ set membership .sr NMEM \math_font!N*‹ not a member (set) .sr HASMEM \math_font!,*‹ has as a member (mem backwards) .sr NEQ \math_font!=*‹ not equal .sr LEQ \math_font!<*‹ less than or equal .sr GEQ \math_font!>*‹ greater than or equal .sr ALL \math_font!A*‹ for all (upside down A) .sr EXISTS \math_font!E*‹ there exists (backwards e) .sr DUPARR symbol_font!**‹ double shaft arrow upwards .sr DDNARR symbol_font!+*‹ double shaft arrow downward .sr DLARR symbol_font!(*‹ double shaft arrow left .sr DRARR symbol_font!)*‹ double shaft arrow right .sr DDARR symbol_font!,*‹ double shaft arrow both directions .sr UPARR symbol_font!"*‹ single shaft arrow upwards .sr DNARR symbol_font!#*‹ single shaft arrow downward .sr LARR symbol_font!7*‹ single shaft arrow left .sr RARR symbol_font!!*‹ single shaft arrow right .sr DARR symbol_font!$*‹ single shaft arrow both directions .sr DOT symbol_font!*‹ raised dot, as for multiplication .sr DBLBR symbol_font!l*‹ double left bracket .sr DBRBR symbol_font!m*‹ double right bracket .de script ‹ Script for large letters symbol_font!\0* .em .de program . nf l . begin program . nv line_space_save ls/100 . nv normal_font normal_lisp_font . nv bold_font bold_lisp_font . nv italic_font italic_lisp_font . ls 1 . fs \normal_lisp_font . em .de end_program . sp . ls line_space_save . fs . en‹ program . fi b . em .def_font_macro lisp \\normal_lisp_font! .def_font_macro blisp \\bold_lisp_font! .def_font_macro ilisp \\italic_lisp_font! .smargins .75i .75i 1i 1i .5i .5i ‹ key: space .tr ~ .rs .sr dun ~mu!g/ml‹ Drug UNits (ug/ml) .begin foo ‹ -*- Text -*- . nr ref_indent 0 . nr ref_rindent 0 . so ml:medg;ref > . en ‹ foo .ref_subtitle "General Medical References" .ref alluisi Earl A. Alluisi, ``Attention and Vigilance as Mechanisms of Response,'' ital(Acquisition of Skill), edited by Edward Bilodeau, New York, Academic Press, pp. 201en!213, 1966. .em .ref bigger J. Thomas Bigger, Jr. and Francis M. Weld, ``Analysis of Prognostic Significance of Ventricular Arrhythmias after Myocardial Infarction: Shortcomings of Lown Grading System,'' ital(British Heart Journal), bold(45):717en!724, 1981. .em .ref cdp Coronary Drug Project, ``Prognostic Importance of Premature Beats Following Myocardial Infarction,'' ital(Journal of the AMA), bold(233):1116en!1124, 1973. .em .ref davison Richard Davison, Michele Parker, and Arthur Atkinson, ``Excessive Serum Lidocaine Levels during Maintenance Infusions: Mechanisms and Prevention,'' ital(American Heart Journal), bold(104):203en!208, August 1982. .em .ref Lown0 Bernard Lown and Marshall Wolf, ``Approaches to Sudden Death from Coronary Heart Disease,'' ital(Circulation), bold(44):130en!140, July 1971. .em .ref lown_ed Bernard Lown, ``Lidocaine: Antiarrhythmic Panacea or Cardiac Cosmetic Agent?'', (editorial), ital(Journal of the AMA), bold(246):2482en!2483, November 27, 1981. .em .ref Lown1 Bernard Lown, ``Sudden Cardiac Death em 1978,'' ital(Circulation), bold(60):1593en!1599, December 1979. .em .ref macdonald Clement J. McDonald, ``Protocol-Based Computer Reminders, the Quality of Care and the Non-Perfectability of Man,'' ital(New England Journal) ital(of Medicine), bold(295):1351en!1355, December 9, 1976. .em .ref osborn Harold Osborn, ``Lidocaine Prophylaxis after Acute MI,'' ital(Hospital Physician), bold(2):108en!111+, 1982. .em .ref pfeifer Henry Pfeifer, David Greenblatt and Jan Koch-Weser, ``Clinical Use and Toxicity of Intravenous Lidocaine: A Report from the Boston Collaborative Drug Surveillance Program'', ital(American Heart Journal), bold(92):168en!173, August 1976. .em .ref ruberman W. Ruberman, E. Weinblatt, ital(et al.), ``Ventricular Premature Beats and Mortality after Myocardial Infarction,'' ital(New England Journal) ital(of Medicine), bold(279):750, 1977. .em .ref schulze R. A. Schulze, H. W. Strauss, and B. Pitt, ``Sudden Death in the Year Following Myocardial Infarction: Relation to Ventricular Premature Contractions in the Late Hospital Phase and Left Ventricular Ejection Fraction,'' ital(American Journal of Medicine), bold(62):192en!199, 1977. .em .ref stedman ital(Stedman's Medical Dictionary), 22nd ed., The Williams & Wilkins Company, Baltimore, 1972. .em .ref steel Knight Steel, Paul Gertman, ital(et al.), ``Iatrogenic Illness on a General Medical Service at a University Hospital,'' ital(New England) ital(Journal of Medicine), bold(304):638en!642, March 1981. .em .ref_subtitle "Pharmacokinetic References" .ref greenblatt_study D. R. Abernethy and D. J. Greenblatt, ``Impaired Clearance of Lidocaine in Elderly Male Patients,'' to appear in the ital(Journal of) ital(Cardiovascular Pharmacology). .em .ref arzbaecher R. Arzbaecher and J. Poyezdala, ``An Interactive System for On-Line Planning and Delivery of Anti-Arrhythmia Therapy in Coronary Care,'' ital(Computers in Cardiology), 1982. .em .ref bauer Larry Bauer, Todd Brown, ital(et al.), ``Influence of Long-Term Infusions on Lidocaine Kinetics,'' ital(Clinical Pharmacology and) ital(Therapeutics), bold(31):433en!437, April 1982. .em .ref buckman Kevin Buckman, Keith Claiborne, ital(et al.), ``Lidocaine Efficacy and Toxicity Assessed by a New, Rapid Method,'' ital(Clinical Pharmacology) ital(and Therapeutics), bold(28):177en!181, 1980. .em .ref collins1 Steve Collins and Robert Arzbaecher, ``Computer Control of Cardiac Arrhythmia'', ital(Proceedings of the IEEE Frontiers of Engineering) ital(in Health Care Conference), 276en!279, 1979. .em .ref collins2 Steve Collins and Robert Arzbaecher, ``Feedback Control in the Management of Cardiac Arrhythmias,'' ital(ISA Transactions), bold(18):95en!100, 1979. .em .ref crone J. Crone, J. Belic, ital(et al.), ``A Programmable Infusion Pump Controller,'' ital(30th ACEMB), p. 95, November 1977. .em .ref greenblatt David Greenblatt, Victoria Bolognini, ital(et al.), ``Pharmacokinetic Approach to the Clinical Use of Lidocaine Intravenously,'' ital(Journal of the American Medical Association), bold(236):273en!277, July 19, 1976. .em .ref jelliffe1 Roger Jelliffe, John Rodman, ital(et al.), ``Clinical Studies with Computer-Assisted Lidocaine Infusion Regimens,'' ital(29th ACEMB), p. 308 November 1976. ‹ital(Abstracts of the 49th Scientific Sessions), 1976. .em .ref jelliffe3 Roger Jelliffe, ``Computers in Cardiovascular Therapeutics,'' in Lee D. Cady, ed., ital(Computer Techniques in Cardiology), Marcel Dekker, Inc., pp. 261en!326, 1979. .em .ref jelliffe2 R. Jelliffe, D. D'Argenio, ital(et al.), ``A Time-Shared Computer Program for Adaptive Control of Lidocaine Therapy Using an Optimal Strategy for Obtaining Serum Concentrations,'' ital(Computer) ital(Applications in Medical Care), IEEE, 975en!981, 1980. .em .ref lelorier Jacques LeLorier, Denis Grenon, ital(et al.), ``Pharmacokinetics of Lidocaine after Prolonged Intravenous Infusions in Uncomplicated Myocardial Infarction,'' ital(Annals of Internal Medicine), bold(87):700en!702, 1977. .em .ref lima John Lima, David Conti, ital(et al.), ``Clinical Pharmacokinetics of Procainamide Infusions in Relation to Acetylator Phenotype,'' ital(Journal of Pharmacokinetics and Biopharmaceutics), bold(7):69en!85, 1979. .em .ref manion Carl Manion, David Lalka, ital(et al.), ``Absorption Kinetics of Procainamide in Humans,'' ital(Journal of Pharmaceutical Sciences), bold(66):981en!984, July 1977. .em .ref narang Prem Narang, Joseph Adir, ital(et al.), ``Pharmacokinetics of Bretylium in Man after Intravenous Administration,'' ital(Journal of) ital(Pharmacokinetics and Biopharmaceutics), bold(8):363en!372, 1980. .em .ref nation R.L. Nation, E.J. Triggs and M. Selig, ``Lignocaine Kinetics in Cardiac Patients and Aged Subjects,'' ital(British Journal of Clinical) ital(Pharmacology), bold(4):439en!448, 1977. .em .ref ochs Hermann Ochs, Gerhard Carstens and David Greenblatt, ``Reductions in Lidocaine Clearance During Continuous Infusion and by Coadministration of Propranolol,'' ital(New England Journal of Medicine), bold(303):373en!377, Aug. 1980. .em .ref peck Carl C. Peck, ital(Bedside Techniques for Optimizing Drug Dosage), ‹Uniform Services University of the Health Sciences, Instruction Manual, 1982. .em .ref rowland Malcolm Rowland, Pate Thomson, ital(et al.), ``Disposition Kinetics of Lidocaine in Normal Subjects,'' ital(Annals New York Academy of) ital(Sciences), bold(179):383en!398, 1971. .em .ref sheiner1 Lewis B. Sheiner, Stuart Beal, ital(et al.), ``Forecasting Individual Pharmacokinetics,'' ital(Clinical Pharmacology and Therapeutics), bold(26):294en!305, September 1979. .em .ref thomson Pate Thomson, Kenneth Melmon, ital(et al.) ``Lidocaine Pharmacokinetics in Advanced Heart Failure, Liver Disease, and Renal Failure in Humans,'' ital(Annals of Internal Medicine), bold(78):499en!508, 1973. .em .ref ueda Clarence Ueda, David Hirschfeld, ital(et al.) ``Disposition Kinetics of Quinidine,'' ital(Clinical Pharmacology and Therapeutics), bold(19):30en!36, 1975. .em .ref wagner John Wagner, ital(Fundamentals of Clinical Pharmacokinetics), Drug Intelligence Publications, 1975. .em .ref woosley Raymond Woosley and David Shand, ``Pharmacokinetics of Antiarrhythmic Drugs,'' ital(The American Journal of Cardiology), bold(41):986en!995, May 1978. .em .ref_subtitle "Computer Science References" .ref allen1 James F. Allen, ital(A General Model of Action and Time), Dept. of Computer Science, University of Rochester TR 97, November 1981. .em .ref allen2 James F. Allen, ital(Maintaining Knowledge about Temporal Intervals), Dept. of Computer Science, University of Rochester TR 86, January 1981. .em .ref bruce Bertram C. Bruce, ``A Model for Temporal References and its Application in a Question Answering Program,'' ital(Artificial Intelligence 3), 1972. .em .ref clemmer T. P. Clemmer, R. M. Gardner, and J. F. Orme, Jr., ``Computer Support in Critical Care Medicine,'' ital(Proceedings, 1980 Symposium on Computer) ital(Applications in Medical Care), pp. 1557en!1561, IEEE, 1980. .em .ref dambrosia James M. Dambrosia and Jonas H. Ellenberg, ``Statistical Considerations for a Medical Data Base,'' ital(Biometrics), bold(36):323en!332, June 1980. .em .ref fagan Lawrence M. Fagan, ital(VM: Representing Time-Dependent Relations in a) ital(Medical Setting), Ph.D. Thesis, Department of Computer Science, Stanford University, June 1980. .em .ref gorry G. Anthony Gorry, Howard Silverman and Stephen Pauker, ``Capturing Clinical Expertise: A Computer Program that Considers Clinical Responses to Digitalis,'' ital(The American Journal of Medicine), bold(64):452en!460, March 1978. .em .ref hp_monitor Hewlett-Packard, ital(A Compendium on Automated Arrhythmia Detection), a Hewlett-Packard publication, 1976. .em .ref kahn Kenneth M. Kahn and G. Anthony Gorry, ``Mechanizing Temporal Knowledge,'' ital(Artificial Intelligence 9), 1977. .em .ref long William J. Long and Thomas A. Russ, ``A Control Structure for Time Dependent Reasoning,'' to appear in the ital(Proceedings of) ital(IJCAIen!83), 1983. .em .ref mcdermott Drew V. McDermott, ``A Temporal Logic for Reasoning About Processes and Plans,'' Computer Science Department, Yale University, RR 196, 1981. .em .ref arry_monitor Kenneth Ripley and Alan Murray, eds., ital(Introduction to Automated) ital(Arrhythmia Detection), IEEE Computer Society Press, 1980. .em .ref arry_detection1 Donald Romhilt, Saul Bloomfield, ital(et al.), ``Unreliability of Conventional Electrocardiographic Monitoring for Arrhythmia Detection in Coronary Care Units,'' ital(The American Journal of Cardiology), bold(31):457en!461, April 1973. .em .ref russ Thomas A. Russ, ``A Knowledge-Based Approach to Ventricular Arrhythmia Management,'' in ital(Proceedings of the International Conference on) ital(Cybernetics and Society), IEEE Systems, Man and Cybernetics Society, pp.~10en!14, 1982. .em .ref shortliffe1 Edward H. Shortliffe, ital(Computer Based Medical Consultations:) ital(MYCIN), Elsevier North Holland, Inc., 1976. .em .ref shortliffe2 Edward H. Shortliffe, Bruce G. Buchanan, and Edward A. Feigenbaum, ``Knowledge Engineering for Medical Decision Making: A Review of Computer-Based Clinical Decision Aids,'' ital(Proceedings of the) ital(IEEE), bold(67):1207en!1224, September 1979. .em .ref silverman Howard Silverman, ital(A Digitalis Therapy Advisor), mit Laboratory for Computer Science, MIT/LCS/TRen!143, December 1974. .em .ref singer Daniel Singer, Albert Mulley, ital(et al.), ``The Course of Patients with Suspected Myocardial Infarction. Prediction of Complications Using a Computer-Based Databank,'' ital(Proceedings, 1980 Symposium on) ital(Computer Applications in Medical Care), pp. 1590en!1593, IEEE, 1980. .em .ref swartout1 William R. Swartout, ital(A Digitalis Therapy Advisor with) ital(Explanations), mit Laboratory for Computer Science, MIT/LCS/TRen!176, February 1977. .em .ref swartout2 William R. Swartout, ital(Producing Explanations and Justifications of) ital(Expert Consulting Systems), mit Laboratory for Computer Science, MIT/LCS/TRen!251, January 1981. .em .ref psz1 Peter Szolovits, ital(Artificial Intelligence and Clinical Problem) ital(Solving), mit Laboratory for Computer Science MIT/LCS/TMen!140, September 1979. .em .ref psz_book Peter Szolovits (ed.), ital(Artificial Intelligence in Medicine), AAAS Selected Symposium Series, Westview Press, 1982. .em .ref psz2 Peter Szolovits and Stephen G. Pauker, ``Computers and Clinical Decision Making: Whether, How, and For Whom?,'' ital(Proceedings of the IEEE), bold(67):1224en!1226, September 1979. .em .ref arry_detection2 N. J. Vetter and D. G. Julian, ``Comparison of Arrhythmia Computer and Conventional Monitoring in Coronary-Care Unit,'' ital(The Lancet), pp.~1151en!1154, May 24, 1975. .em .ref hp_eval J. Y. Wang, M. N. Shaya, ital(et al.), ``The Design and Evaluation of a Real-Time Arrhythmia Monitoring Algorithm,'' ital(Computers in) ital(Cardiology), 1982. .em .if draft . smargins .75i .75i 1.5i 1i .5i .5i .‹ key: . ls 1 .ef . smargins 1.25i .65i 1.5i 1i .5i .5i . ls 1.7 .en . bp 2 .rs .ls 1 .nf c big(Ventricular Arrhythmia Management:) .sp .15i big(A Knowledge-Based Approach) .sp .2i by .sp .1i Thomas~~Anton~~Russ .sp .35i Submitted to the Department of Electrical Engineering and Computer Science on May 12, 1982 in partial fulfillment of the requirements for the Degree of Master of Science .fi b .sp .25i .ls 1.5 .csubtitle "Abstract" This thesis presents a prototype algorithm for the management of patients with ventricular arrhythmias (abnormal heartbeats). The algorithm formalizes and abstracts the therapeutic strategy used by an expert cardiologist in a Cardiac Intensive Care Unit. Formal mathematical models are incorporated where they exist. For example, mathematical models of drug distribution in the body (pharmacokinetic models) are used. Where formal models are not available, such as in the treatment management area, clinical experience is used as the guideline. A second area where medical expertise is needed is to combine data from different sources. For these reasons an expert systems approach is taken. .para The algorithm implicitly generates expectations about the effects of therapy and compares them with the clinical state of the patient. The difference between the expected results of therapy and the patient's actual response drives the reasoning process. This feedback system allows the program to tailor its recommendations to individual variations in therapeutic response. .para This thesis describes the design of the therapy management algorithm. Its capabilities and limitations are discussed and its performance is illustrated in a series of sample cases obtained from clinical records. .ls 1 .vp 8.5i Thesis supervisor: Peter Szolovits .br Title: Associate Professor of Computer Science and Engineering .sp 1.5 This research was supported (in part) by the National Institutes of Health Grant No.~1~P01~LM~03374en!04 from the National Library of Medicine, (in part) by BRSG~S07~RR~07047en!17, awarded by the Biomedical Research Support Grant, Division of Research Resources, National Institutes of Health, and (in part) by the Whitaker Health Sciences Fund. .ls 1.7 . bp .rs .sp 3 big(Acknowledgements) .para I would like to thank the members of the Arrhythmia Advisor project for the support and encouragement during the development and testing of this algorithm: Maureen Cochran, Rob Friedman, David Greenblatt, Bill Hardy, Mike Klein, Buck Locke, Bill Long, and Peter Szolovits (who was also my thesis supervisor). .para I would especially like to express appreciation to Dr. Klein for volunteering to have his brain picked to form the basis of the expert system. He is the source of the medical knowledge that went into the Advisor algorithm. Any errors of medicine found in this thesis, however, are solely my responsibility and are due to my incomplete grasp of the application domain. .para Bill Long has been a great help in the research and writing of this thesis. He was always willing to discuss technical and philosophical issues and provided invaluable guidance and encouragement. In addition to giving technical assistance, he has also been a good friend. Without his help this thesis would not have been possible. .para Finally I would like to thank the other members of the Medical Decision Making Group for their support, in particular Glenn Burke, who kept ML alive, and Phyllis Koton, who unflaggingly sent me messages over the Arpanet telling me to get to work. . bp . en . begin_table_of_contents 3 .chapter "Introduction" .para The increasing sophistication of medicine is leading to problems in cardiac intensive care. As will be argued in greater detail in the next chapter, clinicians are being confronted with an increasing volume of information that has a bearing on the treatment of critically ill cardiac patients. There is a need to coordinate data from various sources and to provide decision support aids in order to allow the physicians to make better use of the medical knowledge relevant in this domain. .para Progress in the state of the art of expert systems in medicine~[shortliffe2,~psz_book] gives cause to believe that these techniques can be successfully applied to the domain of cardiac intensive care. The methodology followed in this research was to take the actions of a clinical expert in the field as the basis for the decision making algorithm. In those areas where formal mathematical models exist, they were used to supplement the basic algorithm. The experience of a senior cardiologist was again used to determine how to combine information from various sources and provide advice for the management of ventricular arrhythmias. .sect "Overview of the Arrhythmia Advisor Project" .para The interest of the MIT Clinical Decision Making Group in cardiac care begins with the development of a Digitalis Advisor~[silverman] in the early 1970s. This was an expert system to give advice about the use of the drug digitalis in the treatment of cardiac problems. This system was oriented around a specific therapy and therefore suffered from a lack of flexibility. Because of the development of other cardiac drugs, the use of digitalis has declined in recent years. This led to an examination of the broader problem of cardiac care itself. At the present time, the research group is pursuing two projects in the area of cardiology. One is the Arrhythmia Advisor, of which the program described in this thesis is a part, and the other is research on the modeling of congestive heart failure. .para The Arrhythmia Advisor project involves workers from a number of institutions. The participants from the Clinical Decision Making at MIT are Dr. William Long, Prof. Peter Szolovits and the author. Medical expertise is provided by collaborating researchers affiliated with the Boston University School of Medicine and University Hospital. They are Dr. Robert Friedman and Dr. William Hardy of the Medical Information Systems Unit, and the domain experts Dr. Michael Klein, head of the Cardiac Care Unit (CCU), and Maureen Cochran, R.N., head CCU nurse. Mr. W. Buck Locke from Hewlett-Packard Corporation provides liaison for the automatic arrhythmia monitoring equipment used in the CCU. Dr. David Greenblatt from the Tuftsen!New England Medical Center is the project's pharmacokinetics consultant. .sect "Goals and Limits of this Thesis" .para This thesis project is intended to be an experimental effort. The program as described implements a preliminary design which will be extensively revised in the course of the next few years as part of an on-going research effort. One purpose of this experiment is to identify the areas of reasoning that present a special challenge. To this end, an effort has been made to use a simple structure for the program and to avoid the construction of special mechanisms. This has resulted in a flowchart style implementation of the algorithm and is responsible for the implicit rather than explicit inclusion of expectations and assumptions. By testing the limits of simple mechanisms, it is possible to identify the areas where more sophisticated reasoning and system support structures are needed. These areas will be discussed in the concluding chapter of the thesis. .para The major part of the thesis research was the elucidation and refinement of an algorithm for reasoning about therapy management. This algorithm covers the initial recommendation of therapy based on pharmacokinetic principles and the evaluation of the effectiveness of the therapy. Using the evaluation, the program can recommend adjustments in drug regimens or changes of therapy. In particular, the program is capable of recognizing the failure of its therapeutic suggestions, and thus will exhibit limited knowledge of its own boundaries. .para This thesis consists of the development of the initial therapy evaluation and management algorithm. A moderately complex pharmacokinetic model for lidocaine and a simple model for procainamide have been implemented. A simple explanation capability is provided, although this was not the focus of much effort. The justifications are simply a review of the reasoning steps executed. Arrhythmia analysis is performed by code implemented by Dr. William Long. It is assumed that all patients to be treated have a single cause of their arrhythmia (myocardial infarction) and that they will be treated with one anti-arrhythmic drug at a time. The selection of a treatment strategy is not part of the thesis, since the order in which the drugs are tried is fixed. .sect "Organisation of the Thesis" .para The thesis will begin with an overview of the medical domain in which the program operates (Chapter~2). This will describe the problem and identify the factors which go into the decision making process. Other research relevant to the medical problem and computer science solutions will be discussed (Chapter~3). Next, the development and organization of the Arrhythmia Advisor will be discussed (Chapter~4), followed by an example of the system in use (Chapter~5). Following the example will be a preliminary, informal evaluation of the program's performance (Chapter~6) and a discussion of the areas where further research is needed (Chapter~7). .chapter "Medical Aspects of Ventricular Arrhythmia Management" .para This chapter contains a discussion of the medical issues relevant to the management of ventricular arrhythmias. It provides a background for the domain in which the thesis program operates. The decisions regarding therapy and the factors entering into these decisions will be presented. A glossary of medical terms can be found in Appendix :glossary!. Much of the information concerning the treatment practices and principles followed in cardiac care comes from consultations with Dr. Michael Klein, Chief of the Cardiac Care Unit at Boston Universityem!University Hospital, a collaborating researcher working with the Clinical Decision Making Group at MIT. Dr. Klein is the expert whose patient management strategies form the basis of the prototype advisor program. .para The treatment of serious ventricular arrhythmias takes place in a ital(Cardiac Care Unit (CCU)), an intensive care station that specializes in the treatment of heart patients. A common manifestation of the disease processes in these patients is the existence of ital(ventricular) ital(premature beats (VPBs)) or other forms of ital(ventricular) ital(arrhythmia). .sect "The Cardiac Care Unit" .para The cardiac care unit was instituted to facilitate special observation of patients at high risk of death due to heart problems. The key aspects of therapy management in a CCU consist of assessing the severity of a patient's problem; selecting an appropriate therapy, usually the administration of an anti-arrhythmic drug; the adjustment of a therapeutic intervention to the specific requirements of the individual patient; and evaluating the success of the therapy in controlling the problem. The selection of a therapy is based in part on the diagnosis of the underlying disease process and in part on an assessment of the danger posed by the VPBs. In many cases, however, the underlying problem cannot be treated directly. Instead an attempt is made to prevent the occurrence of life-threatening arrhythmias such as ital(ventricular tachycardia (VT)) or ital(ventricular fibrillation) ital((VF)). This course of action assumes the existence of indicator or warning arrhythmias, that is, patterns of VPBs that indicate a predisposition to or danger of the development of VF, and seeks to suppress these VPBs~[osborn]. .subsection "Overview" .para The cardiac care unit is a special intensive care unit where extra equipment and specially trained staff are available to handle critical cardiac patients. The environment is technology intensive.‹ and a technically high level of life support is provided. Patients enter a CCU when there is reason to believe that they are at high risk of developing cardiac complications. Often these patients have just suffered an ital(acute myocardial infarction) ital((MI)) (heart attack) or have ital(congestive) ital(heart failure (CHF)). Other patients have chronic problems such as insufficient cardiac blood flow (ital(ischemia)) manifesting itself as chest pain (ital(angina)). In other cases the heart rhythm is intrinsically irregular (ital(primary) ital(arrhythmia)). Adverse reactions to drugs can also cause cardiac problems that would be treated in the CCU. .para One characteristic common to all patients is that they are in danger of developing life-threatening rhythm disturbances such as ventricular tachycardia or ventricular fibrillation. The patients are placed in the CCU so that they may be more closely observed than would be possible on a general medical ward. .para Since one of the manifestations of poor cardiac function is the presence of abnormal beats known as ventricular premature beats, they are an important source of information about the patient. The normal beating of the heart is coordinated through electric discharge starting at a certain location in the heart known as the sinus node. As the electric discharge is conducted through the heart, the muscle cells contract. It is the electric potential of this contraction that is recorded on an ital(electrocardiogram (EKG)). The EKG thus gives a picture of the electrical state of the heart. If the conduction is smooth, a coordinated pumping occurs. If the conduction is disrupted due to disease, injury or lack of oxygen then abnormal conduction may result. This is the cause of VPBs. .para These beats come in several varieties. The simplest ones are beats that are early and arise in the ventricle instead of at the sinus node. A more complex case is when the abnormal beats have more than one shape on the EKG, presumably due to the existence of more than one abnormal conduction pathway or source for the differently shaped VPBs. Two abnormal beats following one another are called ital(couplets) and more than two are a ital(run). The final variety are exceptionally early VPBs. They are known as R-on-T VPBs.foot The name comes from the EKG waveform characteristic of heart beats. The largest spike is called the ``R~wave'', and the peak that represents the electrical recovery of the heart is known as the ``T~wave.'' An ``R-on-T VPB'' is one where the start of the beat occurs early, before the heart has completed its electrical recovery from the previous beat. On the EKG the R~wave of the early beat occurs during the T~wave of the normal beat (See also Appendix :glossary). .efoot A system for classifying arrhythmias according to malignancy was developed by Lown~[lown1]. The Lown system characterizes the danger of the VPBs as increasing in the order that the types are listed above. A study by Bigger~[bigger] confirms the basic principle, but indicates that frequency of VPBs are also important. His data also shows that the prognostic significance of the R-on-T phenomena may be overrated. .para Since the therapy itself comes with risks (outlined below), it is not general practice to treat patients prophylactically unless there is reason to believe that they are in immediate danger of developing VT or VF. The VPBs serve as a warning signal to identify the patients at riskM~[cdp,~osborn,~ruberman]. The patients who exhibit dangerous so-called ital(warning arrhythmias) are the ones treated. .subsection "Types of Decisions" The types of decisions that must be made can be divided into two major categories: which patients to treat and how to manage the therapy. The decision to treat is based on the perception of the relative dangers of the disease and the therapy. The management encompasses the selection of an appropriate treatment strategy and the proper execution of the strategy. For the purposes of this thesis, the strategy will be taken as a given. The current implementation of the Advisor is concerned only with the management of the therapy proper. .para Therapy management must deal with three broad types of outcome: success, insufficient effect, and therapeutic failure. An additional complication is the possibility of adverse side-effects caused by the therapy itself. .subsection "Sources of Information" .para The information that clinicians use to evaluate a patient's condition and choose a therapy comes from many sources. A diagrammatic representation is given in Figure current_figure. General medical knowledge is the basis of the decision and serves as the framework within which the data about a specific patient are evaluated. Part of this general knowledge also concerns the action and disposition of drugs on the body. Other information comes from the patient's medical history, laboratory tests and other measurements such as electrocardiograms, and clinical observations. Some of these sources are discussed in further detail below. .figure "Factors Affecting the Therapy Decision" <== .‹ THIS FILE IS MISSING!! (pp. 29--42 of the thesis) . so tar;thprog > .chapter "An Example" .nr immediate_figure 1 .para An example showing the potential usefulness of such a knowledge-based system follows. The data is derived from an actual case treated at the University Hospital CCU in 1982. .foot In the course of this research, appropriate safeguards for all patient-derived data are being taken. The guidelines used by University Hospital are being followed. ``Fred Smith'' is a name invented by the author. .efoot The case was retrospectively run through the Arrhythmia Advisor. The actual therapy given was entered as a correction to the program's recommendations so that the consultation could proceed beyond the initial session. At each decision point a comparison can be made between the program's recommendations and the action of the CCU staff. .para In the examples the user's input is in italics. .figure "Initial Information for the Patient Specific Model" .program Welcome to the Arrhythmia Therapy Advisor. What is the patient's name? ital(Fred Smith) What is the patient's sex? ital(m) [MALE] How old is the patient? ital(55) What is his weight (kg)? ital(161 lb) How tall is the patient (cm)? ital(5'10) Please characterize the patient's renal function: ital(n) [NORMAL] And his hepatic function? ital(n) [NORMAL] Is heart failure present? ital(y) [YES] What is the cardiac output (L/min)? ital(u) [UNKNOWN] .end_program .finish_figure .para The first set of questions is used to set up the patient specific model (Figure~last_figure). The questions about renal and hepatic function are needed because they are known to affect the pharmacokinetics of some of the drugs used in cardiac therapy. Note that the program can deal with unknown information. The input is made as flexible as possible without a full natural language interfaceem!abbreviations are accepted and values with units are automatically converted into the internal program representation. .nr ex2_figure current_figure .figure "Initial Therapy Recommendation" .program From the 60 minutes of information there were MULTIFORM VPBs at 13.0/minute. There were also long runs with runs of all kinds totalling 0.667 per minute. Computing therapy recommendation . . . Since the arrhythmia state is not good, I recommend lidocaine treatment. I suggest increasing the dose: The old therapy is: IV -- None. Bolus - None. The therapy I recommend adds IV -- 1.00 mg/min at 0:00 Bolus - 50.00 mg at 0:00 50.00 mg at 0:10 Therapy level is 2; Estimated target serum level is 2.20 .end_program .finish_figure .para An assessment of the state of the patient's arrhythmia is made based on data from the arrhythmia monitoring equipment (Figure~last_figure). Since this data is not supplied by the user, it is displayed to give the background for the treatment advice that is to follow. The therapy advice is presented to the user along with a one-line summary. The summary of the recommended treatment is presented in terms of the abstract therapy level and the expected steady state (target) drug concentration. .figure "Explanation Facility" .program Is this acceptable? ital(ex) [EXPLAIN] What should I explain? ital(vpb) [VPB-EVAL] 0:00 -- From the 60 minutes of information there were MULTIFORM VPBs at 13.0/minute. There were also long runs with runs of all kinds totalling 0.667 per minute. The initial arrhythmia state is very fast uniform runs. The very fast uniform runs constitute an urgent risk to the patient. What should I explain? ital(d) [DOSE] 0:00 -- The dosage goal level is 2.84 ug/ml with the range between 1.64 and 4.04. The IV rate is set to 1.00 mg/min based on parameters v1 = 21.91 L (56% of normal) and Kel = 0.021 (109%) At time 0:00 the concentration estimate is 0.00 ug/ml, therefore a bolus of 50.0 mg is suggested. At time 0:10 the concentration estimate is 1.57 ug/ml, therefore a bolus of 50.0 mg is suggested. Is the recommendation acceptable? ital(y) [YES] I recommend consulting again at time 0:15 .end_program .finish_figure .para The initiative remains with the user. He is asked to concur with the advice given by the program. If desired, he can examine a trace of the reasoning steps used to arrive at a decision (Figure~last_figure). The dosage explanation identifies the criteria used to calculate the dosage recommendation. The IV rate is explained in terms of the parameters of the underlying pharmacokinetic model. The modification of the pharmacokinetic parameters (56% and 109%) is explained in Figure~current_figure. .para The reader should note the discrepancy between the dosage goal level of 2.84dun! asserted in Figure~last_figure and the expected target serum level from Figure~ex2_figure of 2.20dun!. The difference is explained by the desire to have a ``reasonable'' value for the IV infusion. An exact match would have required an IV rate of 1.31~mg/min, which is impractical in a clinical setting. Reasonable units for the various drugs were determined through interviews with Dr. Klein and Ms. Cochran concerning the current practice in the CCU. The unit sizes considered reasonable are parameters passed to the dosage calculation routines and can thus be easily modified. .nr expk_figure current_figure .figure "Modification of the Pharmacokinetic Model for Heart Failure" .program What now? ital(ex) [EXPLAIN] What should I explain? ital(model) Modify parameters K12 K21 Kel V1 of lidocaine by 1.23 0.971 1.09 0.566 for heart failure. Lidocaine: Alpha half-life = 0:07 Beta half-life = 1:55 V1 = 21.91 l. Total body clearance = 453.82 ml/min .end_program .finish_figure .para Figure~last_figure explains the changes that were made to the parameters of the pharmacokinetic model in response to the heart failure that was reported for the patient. The changes are reported in terms of the changes to the parameters used in the pharmacokinetic simulation (useful for program debugging), and also in terms of the half-life, initial volume of distribution for the patient, and the clearance. The latter values are more useful for intuitive clinical judgment and are included for human engineering reasons. The pharmacokinetic parameters are explained in Appendix~:pk_model. .figure "Entering the Actual Therapy" .program What now? ital(u) [UPDATE-PATIENT] What is the current time (zero based)? ital(20) Was the IV rate changed to 1.00 mg/min at 0:00? ital(n) [NO] Was a bolus of 50.00 mg given at 0:00? ital(y) [YES] Was a bolus of 50.00 mg given at 0:10? ital(y) [YES] Were any other lidocaine treatments made? ital(y) [YES] Please enter the treatments below ('Return' alone ends input). The last update was at 0:00. It is now 0:20. What is the iv dose [mg/min] and time [min]? ital(2 mg/min at 0:10) What is the iv dose [mg/min] and time [min]? What is the bolus dose [mg] and time [min]? From the 12 minutes of information there were UNIFORM VPBs at 1.92/minute. There were also pairs with runs of all kinds totalling 0.0833 per minute. Is the patient toxic? ital(n) [NO] Computing therapy recommendation ... Since an adequate effect has been achieved, I suggest maintaining the current therapy Is this acceptable? ital(y) [YES] I recommend consulting again at time 0:50 .end_program .finish_figure .para Although the system gives a recommendation for the next consultation time, the user is free to ask for advice whenever he wishes. The purpose of the session shown in Figure~last_figure is to tell the Advisor what was actually done (the IV was started 10 minutes later and at a higher rate than the program recommended). The final decision about whether to follow the advice given by the Arrhythmia Advisor remains with the physician. In order to assure that the program's view of the therapeutic intervention corresponds to reality, it is necessary to confirm the expectations. .para Since the patient's state appears to have improved, the program sees no need to modify current therapy. The program uses the information about the actual therapy to modify its view of what the drug level is likely to be. As can be seen in Figure~current_figure, the abstract therapy level is now at 3. .figure "The Arrhythmia Returns" .program What now? ital(u) [UPDATE-PATIENT] What is the current time (zero-based)? ital(1:20) Were any other lidocaine treatments made? ital(n) [NO] From the 60 minutes of information there were MULTIFORM VPBs at 3.83/minute. There were also short runs with runs of all kinds totalling 0.25 per minute. Is the patient toxic? ital(n) [NO] Computing therapy recommendation ... The drug level is low, therefore I recommend a bolus. I suggest increasing the dose: The old therapy is: IV -- 2.00 mg/min at 0:10 Bolus - 50.00 mg at 0:00 50.00 mg at 0:10 The therapy I recommend adds IV -- None. Bolus - 50.00 mg at 1:20 Therapy level is 3; Estimated target serum level is 4.41 Is this acceptable? ital(y) [YES] I recommend consulting again at time 1:50 .end_program .finish_figure .para The patient developed another series of runs, one of the warning arrhythmias. Since this is an urgent condition, some intervention is deemed necessary. Using the pharmacokinetic data from the model, the program calculates that the patient is in the slowly increasing phase (see Figure~drug_level!). Based on this information, the program recommends a bolus to immediately increase the blood level. A change in IV rate is not indicated because the current drug level is not close to the eventual steady state level. .figure "The Patient Becomes Toxic" .program What now? ital(u) [UPDATE-PATIENT] What is the current time (zero-based)? ital(4:20) Was a bolus of 50.00 mg given at 1:20? ital(n) [NO] Were any other lidocaine treatments made? ital(y) [YES] Please enter the treatments below ('Return' alone ends input). The last update was at 1:20. It is now 4:20. What is the iv dose [mg/min] and time [min]? ital(3 at 1:20) Note: The estimated steady state drug concentration of 6.61 ug/ml will exceed the patient's toxic limit of 6.0 ug/ml. What is the iv dose [mg/min] and time [min]? What is the bolus dose [mg] and time [min]? From the 60 minutes of information there were UNIFORM VPBs at 0.05/minute. Is the patient toxic? ital(y) [YES] How long ago did the signs of toxicity first appear? ital(0) Computing therapy recommendation ... The patient is toxic, but the arrhythmia has been controlled. I suggest stopping treatment for now and adjusting the toxic limit. I suggest stopping therapy until toxicity disappears. I estimate that the patient will be below the blood concentration level when toxic signs first occurred by about 4:26. The patient's toxic level has been reduced to 4.73 ug/ml .end_program .finish_figure .para As Figure~last_figure shows, the actual therapeutic action taken was to increase the infusion rate. This does not have an immediate impact on the drug level. Instead, it raises the steady state value, an effect which is not seen for hours. (Recall from Figure~expk_figure that the beta half-life is almost two hours.) In the case of the example patient, this leads to a toxic reaction three hours after the increase in the infusion rate. The program would not have made a recommendation of such a high rate because the steady state drug concentration is calculated to exceed the default toxic limit of 6.0dun!. If the user were entering a treatment plan rather than reporting past treatment, a warning would be printed and the user asked to confirm the dosage. .para An explanation for the staff behavior lies in an insufficient understanding of the pharmacokinetic principles underlying the therapy. One of the effects of congestive heart failure is to reduce the volume of distribution of the drug, thus raising the concentration for a given dose. Although the staff apparently took this effect into consideration when administering the initial bolus doses (50~mg is less than the normal starting bolus of lidocaine in the CCU), they did not subsequently apply the reasoning to the IV doses. The program applies knowledge at all appropriate points. .para Another noteworthy point is the use of the toxic response information together with the pharmacokinetic data to modify the patient's individual toxic level. The new level of 4.73dun! replaces the default value of 6.0dun!. The new toxic level is set to half of a therapy step increment below the estimated concentration at which toxic signs were noticed. The reduction is designed to provide a margin of safety for future drug therapy. In this case, the drug level estimate at toxicity was 5.42dun!. It was reduced by half of a therapeutic step (0.71dun!) to yield the new toxic level. Although data on this point is scarce, Dr. Klein's clinical experience suggests that the toxic concentration does not vary much between hospital visits. .para One final point: At the time the patient is reported to be toxic, the computer's estimate of the drug level was 5.42dun!, slightly below the default toxic limit. This means that it would have been possible for the program to make a recommendation which would have exceeded the limit for this particular patient. This underscores the importance of using the clinical state of the patient as the major feedback variable in managing therapy. A knowledge-based approach allows information that cannot be mathematically captured to influence the decision making process. .para This example shows that by considering and consequently applying all relevant knowledge, improvements in clinical practice in a CCU are possible. .nr immediate_figure 0 .chapter "Evaluation of Program Performance" .nr eval_chapter chapter .para The program has the goal of providing therapeutic management advice at the level of competence of the house staff at a teaching hospital. It is hoped that the quality of therapy will eventually be better, but the initial goal is to at least match the current therapy. .para There are two parts to the evaluation. First, the accuracy of the pharmacokinetic model for lidocaine was evaluated. This evaluation made use of raw data from a pharmacokinetic study undertaken by Dr. David Greenblatt from Tuftsen!New England Medical Center~[greenblatt_study]. This provides a picture of the level of accuracy that can be expected from the model. The second part of the evaluation is of the advice generated by the therapy management algorithm. The program's performance was evaluated by collecting several typical cases from the University Hospital CCU during 1982 and early 1983. These cases were selected by Dr. Klein or Ms. Cochran as cases where anti-arrhythmic drugs covered by the initial design of the program were used. The cases were retrospectively run through the advisor. .para The quality of the advice given was judged by Dr. Klein and Ms. Cochran. This evaluation indicates how well the algorithm that is currently implemented models the expert decision-making process. In addition to providing a benchmark for progress toward the goal stated above, the failings of the algorithm will serve to show the areas that need more refinement. In keeping with the experimental nature of this thesis, it is hoped that areas for further research can be identified through an analysis of the shortcomings of the algorithm. .para The use of review by Dr. Klein (or another cardiac care expert) is a necessary component of the evaluation process because there is no ital(gold standard) against which program performance can be measured. Unlike the pharmacokinetic models, where the actual measured drug levels can be used to evaluate the accuracy of the model, the therapy management problem does not have a single widely accepted criterion for measuring success. The methodology generally followed in medicine is to engage in randomized clinical trials with a comparison of the outcomes. This is neither feasible nor ethical for the testing of a program still under development. For a more extensive discussion of the evaluation issue see~[psz2]. .sect "Evaluation of the Pharmacokinetic Model" .para The pharmacokinetic model was evaluated by taking data from Dr. Greenblatt's study. The study consisted of giving each of the test subjects a single 25~mg bolus of lidocaine and then drawing blood samples and analyzing the plasma concentrations. Involved in the study were six groups: young men and women, elderly men and women, and obese men and women. The study was conducted to determine whether these characteristics had effects on the pharmacokinetics of lidocaine. As of this time the analysis is not complete. .subsection "General Accuracy of the Model" .para In order to test the accuracy of the program's model, the study subjects were run on the simulation of the pharmacokinetic model using the parameters listed in Appendix :pk_model. The results for the two control groups and the entire study population are summarized in table :current_table. .begin_table "Accuracy of the Model" .ta 1.25i 2.5i 3.25i 4i 4.75i Group Model Prediction Number Mean Standard Prediction Ranges in Range Error Deviation Below Actual .sp .1i small_font [ng/ml] ~~~~[% of Prediction]~~~~ [number~~(percent)]* .program 1Young Men:* ~~~~<~~10 21 50.5% 23.6 ~2~~~~~(9%) ~10~-~~50 27 59.6% 14.0 ~0~~~~~(0%) ~50~-~100 22 54.3% 13.2 ~0~~~~~(0%) 100~-~200 27 48.3% 15.3 ~0~~~~~(0%) ~~~~>~200 25 42.2% 17.6 ~0~~~~~(0%) 1Young Women:* ~~~~<~~10 ~3 41.6% 15.2 ~1~~~~(33%) ~10~-~~50 10 51.5% 21.5 ~1~~~~(10%) ~50~-~100 ~7 64.4% 14.1 ~0~~~~~(0%) 100~-~200 16 60.4% 14.3 ~0~~~~~(0%) ~~~~>~200 ~9 49.0% ~9.9 ~0~~~~~(0%) 1All Subjects:* ~~~~<~~10 53 69.9% 89.2 20~~~~(37%) ~10~-~~50 75 50.9% 23.0 11~~~~(14%) ~50~-~100 75 45.8% 22.1 ~5~~~~~(6%) 100~-~200 72 48.6% 19.9 ~3~~~~~(4%) ~~~~>~200 59 41.5% 16.5 ~2~~~~~(3%) .end_program .rtabs .finish_table .para The accuracy is broken down into categories depending on the concentration predicted by the model. The mean error is expressed as a percentage of the model's prediction. The error is the absolute percentage error. It was calculated as follows: .sp 1 Error = 100~~~mult!~~~div("| predicted minus actual |" "predicted") .sp 1 Absolute differences were used to calculate the error in order to measure the degree to which an individual prediction could be trusted, rather than to measure how close to the average of the cases the prediction lay. The number of cases listed under ``Prediction Below Actual'' are the instances in which the model predicted concentrations that were lower than the actual measured concentrations. .para In general, the model predictions had a mean error less than 50% of the predicted value. This is partially explained by the inherent inaccuracies in using population data to predict individual kinetics. A further explanation is that this was a study conducted using small boluses. Note that the drug levels in this data are significantly below the generally accepted therapeutic range of 1.5en!5.0~smu!g/ml (1500en!5000~ng/ml). There is also evidence that the concentrations achieved with a steady infusion are higher than those predicted using parameters based on the kinetics of single bolus studies~[bauer!,~lelorier!,~ochs,~rowland!]. Since the error in most cases was on the high side, part of the deviation from the predictions can be explained by this mechanism. .para One should note that the greatest inaccuracies occur at the lowest concentrations. Unfortunately these are also the concentrations of the samples farthest in time from the administration of the bolus. It is thus not clear whether the inaccuracies are an artifact of the very low concentrations or whether it is a more serious problem reflecting errors in the clearance parameters. The results do however, underscore the need for a corrective mechanism to modify the therapy based on the clinical response, since the drug level predictions can not be relied upon. .subsection "Intrinsic Variation in Single Patients" .para Two of the subjects in the study were tested more than once. This provides an opportunity to examine the stability of the measurements within one patient. A summary of results is presented in Table :current_table. .begin_table "Variation in a Single Patient" .ta 1i 2i 3i 4i 5i Subject Time Model Actual Actual Actual (Weight) From Bolus Prediction Test 1 Test 2 Test 3 small_font! [hours:minutes] [ng/ml] [ng/ml] [ng/ml] [ng/ml]* .sp .program norm(DA) 0:15 283.6 ~99.4 224.8 233.0 (165 lbs.) 1:00 123.4 ~59.1 ~90.0 ~75.4 2:00 ~82.2 ~33.9 ~45.6 ~54.7 4:00 ~37.9 ~16.5 ~21.2 ~23.4 8:00 ~~8.0 ~~5.1 ~--~~ ~~7.3 norm(RSTH) 0:15 290.6 138.2 247.5 (161 lbs.) 1:00 126.5 ~68.9 ~88.8 2:00 ~84.2 ~48.1 ~61.2 4:00 ~38.8 ~20.3 ~22.1 8:00 ~~8.2 ~~4.1 ~~8.3 .end_program .rtabs .finish_table .para The first of the subjects was tested on three occasions, the other on two occasions. As is evident from the table, there can be significant variations between the plasma levels in the same patient. This means that even though a given dose is found to be effective in a patient, it is not guaranteed that the same dose will be effective at a later time. It would be useful to know whether the large variation within a single patient is also present with longer term infusions or whether it is peculiar to a bolus situation. .para Again basic reliability problems with the data from population based models manifest themselves. One consequence of this is that an adaptive feedback mechanism that can take advantage of any laboratory measurements of drug levels that are available should be included. This would be a useful extension of the present model. The literature indicates that methods for utilizing this information exist~[jelliffe2,~sheiner1]. .sect "Basis for Therapy Advice Evaluation" .para The methodology that is followed is that of running the test cases on the computer and then having Dr. Klein and Ms. Cochran evaluate the quality of the advice. This provides a subjective expert opinion about the validity of the management suggestions. Using expert opinion as the standard against which the performance is measured provides insight not only into the quality of the advice, but also into the success in modeling the expert decision maker himself. .para The nature of the evaluation problem, in particular the inability to determine what would have happened if the computer-generated advice were followed when the clinical staff actually did something else, restricts the possibilities for formal evaluation. The standard medical practice of randomized clinical trials is inappropriate for early testing of a prototype program. Practical and ethical reasons prevent this approach. Unlike for pharmacokinetic models, there is no ital(gold standard) that can be used to judge the quality of the advice. .sect "Performance on Test Cases" .para The test cases were selected by Ms. Cochran and Dr. Klein from the patients that were admitted to the CCU. Eight test cases were used. The format of the presentation is a summary of the case, including a history of the therapy actual administered. This is followed by the recommendations that resulted when the case was retrospectively entered into the Advisor. The recommendations of the Arrhythmia Advisor are presented along with comments from Dr. Klein and Ms. Cochran. The Advisor produced a consultation summary for each case that was successfully entered. These summaries can be found in Appendix~:pat_rec. .para All times are expressed relative to the start of therapy. Thus ``1:20'' means one hour and twenty minutes after therapy began. .subtitle "Case 1" The patient, 55 years old and weighing 161~lbs. (73~kg), was admitted suffering from ventricular arrhythmias as a result of theophylline toxicity. He was in heart failure. (This is the case presented as an example in the previous chapter.) Immediately prior to the start of lidocaine therapy he developed VT and was defibrillated. .para The initial therapy given by the CCU staff was: .list A 50 mg bolus of lidocaine at time 0 and another 50 mg at 10 minutes. .br 2 mg/min IV infusion starting at time 0. .end_list At 1:20 into the therapy VPBs became frequent and another run was detected. The IV infusion was changed to 3~mg/min but no bolus was given. At 4:20 the patient developed signs of toxicity and the infusion was reduced to 2~mg/min. An hour later the patient was no longer toxic. .bspar "Results:" The therapy recommendation was to give 50~mg boluses of lidocaine initially and again 10 minutes later. This was the same as what was actually done. The Advisor, however, recommended a smaller IV infusion, namely 1~mg/min, because the patient was in heart failure and could be expected to have a reduced total body clearance for the drug. After consideration of this aspect, Dr. Klein concluded that the recommendation was reasonable. .para At 1:20 when the arrhythmic activity picked up again, the Advisor recommended that another 50~mg bolus be given and the IV rate left unchanged. This decision was based on the pharmacokinetic estimates that were made by the program. The estimated current blood concentration was 2.72dun with an expected steady state concentration of 4.41dun. A bolus would increase the immediate drug level without an increase in the steady state level. Dr. Klein found this conclusion to be reasonable. He also noted that the IV rate is often increased (as the CCU staff did in this case), and that this can cause trouble later in the therapy. .para At 4:20 the patient became toxic. The increase in the IV rate to 3~mg/min as was actually done turned out to be incorrect. The staff responded by reducing the IV rate. The Advisor suggested stopping the IV entirely and watching the patient. Dr. Klein found the program's advice superior to the actual clinical action. .subtitle "Case 2" The patient was admitted with coronary artery disease (CAD), a suspected MI and heart failure. He was 61 years old and weighed 174~lbs. (79~kg). On entry he was being treated simultaneously with lidocaine, procainamide and bretylium. .bspar "Results:" This case involved the use of more than one of the anti-arrhythmic agents at the same time. The management algorithm was designed around the assumption that the drugs would be tried sequentially. This assumption was such a central part of the design that it was not possible to enter the case into the advisor. No advice could be generated. .subtitle "Case 3" The patient was admitted with CAD and an acute MI. He is 34 years old and weighed 211~lbs. (96~kg). He received lidocaine and procainamide simultaneously. .bspar "Results:" See the results of Case 2. .subtitle "Case 4" The patient was admitted with MI and heart failure. She weighed 169~lbs. (77~kg) and was 57 years old. She had 9 VPBs over a period of twenty-four hours. She did not receive anti-arrhythmic therapy because the arrhythmic activity was insignificant. .bspar "Results:" The evaluation of the arrhythmia by Dr. Long's code indicated that the state of the arrhythmia was considered to be good. That being the case, the Advisor suggested not initiating any therapy. This was judged to be the correct action and corresponded to the actual clinical decision. .subtitle "Case 5" The patient was admitted with an acute MI. He was 38 and weighed 246~lbs (112~kg). He had a run of VPBs of length four in the hour before therapy was started. He had no other VPBs. On the next day, a single VPB was detected hours after therapy was stopped. The therapy he received was: .list A 75 mg bolus of lidocaine at time 0 and 50 mg at 25 minutes. .br A 2 mg/min IV infusion starting at time 2 minutes. .end_list No other VPBs were detected and the therapy remained unchanged. .bspar "Results:" The program gave a recommendation of a 200~mg bolus of lidocaine and the initiation of a 3~mg/min IV infusion. The program justified the large doses with the size of the patient (246~lbs.). Dr. Klein decided that the infusion dose was reasonable given the size of the patient, but noted that it would be rare in the University Hospital CCU to start a patient at that high a dose level. Ms. Cochran observed that other institutions have different approaches. .para After considering the weight factor and performing some mental arithmetic, Dr. Klein approved the bolus dose in principle. Ms. Cochran pointed out that the nursing staff does not like to administer more than 100~mg of lidocaine at one time. This is based on the observation of dizziness and other signs of lidocaine toxicity immediately following large boluses in some patients. It was suggested that the dosage calculation algorithm be modified to limit the size of bolus doses and to split large boluses into more smaller ones spread out over five or ten minutes. .para An explanation for this discrepancy could be found by considering the assumptions underlying the pharmacokinetic model used as input to the dosage calculator. It is assumed that any drug injected into the body is immediately distributed throughout the plasma compartment.foot For a description of the model and the compartments, see Appendix :pk_model. .efoot This mathematical assumption breaks down in the real world. The distribution of the drug in the blood takes a short but measurable amount of time. Although this is not important for describing the kinetics of the drug, it can result in momentary toxicity due to locally high concentrations from the as yet incompletely diluted bolus dose. .para At ten hours, the program considered the therapy a success because a good state had been maintained for eight hours. Both Dr. Klein and Ms. Cochran felt that this was too optimistic and suggested that twenty-four hours be a more reasonable time frame. This is because of evidence indicating that the first twenty-four hours following an MI are the most critical. .subtitle "Case 6" The patient was admitted with unstable angina and a diagnosis of MI. She was 52 and weighed an estimated 180~lbs. (82~kg). She showed signs of frequent VPBs and runs which were abolished by the following lidocaine therapy: .list An immediate 75~mg lidocaine bolus followed by 50~mg ten minutes later. .br An IV infusion at 2~mg/min starting immediately. .end_list .bspar "Results:" The program's recommendation was to start 2~mg/min infusion and give a 150~mg bolus. The infusion advice was judged to be correct, but the bolus should have been split into two smaller boluses as per the discussion above. .subtitle "Case 7" The patient was admitted in heart failure with an acute MI. He weighed 137~lbs. (62~kg) and was 66 years old. Therapy had already been started before we have arrhythmia data. The evaluation considers a decision to increase the infusion. For 14 hours before the time of interest for the evaluation he received a lidocaine infusion at the rate of 1~mg/min. This was sufficiently long for steady state concentrations to be achieved. Arrhythmia data was available for 7 hours prior to the decision to increase the infusion. At the time of interest (7:00 based on the start of the arrhythmia data), he received a 50~mg bolus and the infusion rate was increased to 2~mg/min in response to several episodes of VT. .bspar "Results:" The program's recommendation was the same as the action actually taken in the CCU and was judged to be correct. This demonstrates that the advisor has the capability to properly handle patients who present with drug treatment in progress. .subtitle "Case 8" The patient was 56 and weighed 172~lbs. (78~kg). He had unstable angina and tight aortic stenosis. Although this case was not one with a diagnosis of acute MI, it was included in the evaluation. Therapy was started in response to frequent uniform and multiform VPBs. The initial lidocaine therapy was: .list A 75~mg bolus followed by 50~mg after ten minutes. .br A 2~mg/min IV infusion. .end_list Over the next two hours the VPB frequency slowly declined. Three hours and ten minutes after therapy was started, there was an increase in VPBs to the original level. This led to an increase in the infusion rate to 3~mg/min. .bspar "Results:" The Advisor's initial recommendation was to treat with a 150~mg bolus and a 2~mg/min infusion of lidocaine. As in Case 6, the infusion was judged to be correct whereas the bolus should have been split into two smaller boluses. .para The Advisor would have been more aggressive in increasing the therapy, recommending another bolus of 50~mg after one hour since the arrhythmia didn't improve. Dr. Klein was not able to reach a definite conclusion about the validity of this advice because the information about the couplets currently being provided by the arrhythmia monitor lacked sufficient detail. An important consideration would be the rate of the couplets. (This problem will be corrected in the coming months.) The reason the decision remained equivocal was because the cause was not an acute MI. If the patient were being treated for arrhythmia secondary to an MI then an increase would have been appropriate. Recall that the prototype program operates under this assumption. .para At 3:10 the Advisor recommended increasing the infusion rate to 3~mg/min as was actually done, and giving a bolus of 100~mg. This was judged to be reasonable. It was again pointed out that the CCU staff often does not give a bolus when the IV infusion rate is increased, thus not achieving a quick increase in the blood concentration. .sect "Discussion" The discussion is divided into two parts. First, the ability of the program to give good advice is examined and the underlying principles are identified. In the second part, the failures are analyzed and some suggestions for improvements are made. .subsection "Success" .para The successes of the program in generating advice can be attributed to its ability to combine relevant information from several sources. One factor which resulted in performance superior to that of the attending staff was the strict adherence to pharmacokinetic principles. For example, in Case 1 the program did not recommend an increase in the infusion rate to solve an immediate problem. A change in the infusion takes hours to have an effect on the blood concentration of the drug and is thus not a suitable response to an acute problem. Furthermore, the effects of the heart failure were taken into consideration in the planning of both the bolus doses and the infusion rates. By using the data from the model, the program was able to conclude that the patient was still not near the steady state concentration that would have been achieved with the initial 2~mg/min infusion. Since the effect of this infusion could not be evaluated, it did not suggest an increase. The actual clinical decision to increase resulted in a toxic reaction by the patient. A study of lidocaine toxicity by Davison ital(et~al.)~[davison] found that 71% of the patients were in heart failure or shock.foot Heart failure and shock both result in a decrease in blood circulation and have similar effects on lidocaine clearance. In the current system implementation, shock is treated the same as heart failure. .efoot The failure to consider the pharmacokinetic effects of heart failure on lidocaine distribution appears to be a general phenomenon. .para The other situation where the advice was judged superior was also in Case~1. After toxicity developed, the program recommendation was that the infusion be temporarily terminated. This was Dr. Klein's preferred course of action. Since the algorithm consistently applies its decision criteria, and since that is the action that Dr. Klein had indicated was proper in this situation, it is hardly surprising that the algorithm performed correctly. .para The correct performance in Case~4 demonstrates that the program is capable of concluding that intervention is not necessary. It is reassuring to see that the management algorithm can also decide that the proper advice is not to treat. .para The acceptable performance with the infusion dose in Case~6 again stems from the consistent application of pharmacokinetic dosing principles. It should be noted, however, that the blind following of the pharmacokinetic principles itself is not in and of itself a guarantee of correct treatment decisions. The problem uncovered with the bolus dose is an example and is discussed below. .para One area where the program performance was systematically superior was in the recommendation of a bolus every time the infusion rate was increased. This is also a direct consequence of the consideration of the pharmacokinetic principles of therapy management. It is also an area where the clinical staff does not perform consistently. .para Although the program still has some shortcomings, it shows promise of producing reasonable advice for those cases which it can accept. It can thus serve a role as a mechanical ``second opinion'' and source of reminders about aspects of the decision process that are not always considered by the staff in the hectic environment of intensive care. .subsection "Reasons for Failures" .para This section analyzes the failures in the handling of the above cases and attempts to indicate possible solutions. .para Cases 2 and 3 were not handled at all. The existence of a treatment regimen with more than one anti-arrhythmic drug in use was not anticipated in the initial algorithm design. The assumption of only one drug in use at a time was so fundamental that it was not even possible to enter the cases into the program. It would be a simple matter of programming to modify the Advisor to allow multi-drug therapy regimens to be entered, but that would not solve the underlying problem. .para The basic problem with multi-drug regimens is that the management problem has not been solved. If more than one drug is in use the advisor, human or computer, is faced with the problem of assigning credit for the beneficial effects and blame for the toxic effects. To the extent that there exist criteria that allow one to distinguish between the effects due to each agent, the problem can be solved. In the case of toxic effects, that may often be the case, since different anti-arrhythmic drugs act differently. The credit assignment case remains difficult, since each of the drugs is being administered in an attempt to abolish the same problem. If the arrhythmia is not improving, then it is not clear which drug should be increased, since it is impossible to determine which of the drugs was responsible for the beneficial effects (if any) seen thus far. .para This is a difficult problem which may defy solution altogether. Dr. Klein does not feel that this type of regimen is effective because of the problems in deciding what to do in the face of therapeutic failure. It does not appear as if there is any expertise to draw on in this area. Before this can be incorporated into an expert system, there must be more medical research done in order to identify criteria for making the management decisions. .para Another point which arose during the evaluation was the problem of convincing the clinical staff that the higher doses recommended for heavier patients were valid. It was suggested that doses that exceeded an arbitrary ``normal'' size be accompanied by an explanation to the effect that the dose increase was related to the size of the patient. It was also pointed out by Ms. Cochran that the patients are not weighed immediately upon admission. This was felt to be of only minor concern since it is possible to estimate weight within acceptable limits. The actual weight can be entered later.foot This updating ability does not exist in the present version of the complete Arrhythmia Advisor, but it is present in the incomplete implementation of the next version. .efoot .para Case 5 uncovered two problems: the overly optimistic success criterion and the recommendation of large boluses. Changing the time required for successful suppression of VPBs from eight to twenty-four hours is not a trivial change. The twenty-four hour period is the time since the onset of the MI. During the first day, the patient is in the period of greatest cardiac instability. Taking this factor into account requires knowledge about the disease state itself. This is to be handled by a module of the Arrhythmia Advisor which has not yet been implemented. The determination of success depends on the interaction of information about the disease state and the condition of the arrhythmia. The problem of the large boluses (which also occurred in Cases 6 and 8) can be solved by modifying the dosage calculation algorithm described on page dosage_page. The algorithm could be easily modified by placing a maximum limit on the size of any single bolus that would be administered. Ms. Cochran suggests a limit of 100~mg for lidocaine and procainamide boluses. The ability to incorporate such empirically determined modifications of advice generated from purely mathematical data is one of the strengths of the expert systems technology. It can be exploited to make the mathematical models more useful in real-world settings. .chapter "Conclusion" .sect "Suggestions for Program Improvements" .para This section deals with improvements and extensions of the Arrhythmia Advisor. These are changes that will be relatively easy to implement and which can be accommodated within the design of the Advisor. The changes for the most part make use of existing techniques and knowledge. None of them should require a major research effort. .subsection "Pharmacokinetic Changes" .para In order to capture more of the therapy management, it will be necessary to increase the number of drugs which can be handled by the advisor. As indicated in Chapter 2, pharmacokinetic models for many of the anti-arrhythmic agents exist. If the strategies for administering them are the same as for lidocaine and procainamide, then they can be easily added by including the relevant parameters in the pharmacokinetic model module. To the extent that the management strategies differ, this will entail changes to the therapy advisor as well. The effectiveness of the Advisor will also be increased through the inclusion of other drugs that are used to treat cardiac patients as well. This will include diuretics, potassium supplements and drugs which affect the circulatory system and the contraction strength of the heart. Examples of these types of drugs are lasix, digitalis, dopamine, nitroglycerin, propranolol and nitroprusside. Since these drugs are a part of therapy process, they can be used to provide more complete patient care suggestions. In addition, it will make it possible to account for drug interactions. It is known, for example, that propranolol affects the pharmacokinetics of lidocaine~[ochs]. .para Improvements in the form and function of the pharmacokinetic models themselves are also possible. As alluded to in the previous chapter, it is possible to use laboratory drug concentration measurements to adjust the parameters of the model to a account for the individual variation. This is the approach taken by Jelliffe [jelliffe2]. Since the pharmacokinetic model in the Arrhythmia Advisor serves only as a data source, it can easily be replaced by a more sophisticated version without the need to reprogram the system. The model used in the current version was chosen to provide the minimum amount of sophistication needed to demonstrate that the system can use clinical information to modify the parameters of the model, and that the therapy management algorithm can interpret the data from the model in light of the clinical reaction. Since neither the MIT Laboratory for Computer Science nor the Boston University CCU is equipped to develop such models, it has been the intention to chose the best model published in the literature and not attempt to construct our own. The Arrhythmia Advisor design addressed the issues of what use to make of the data provided by such mathematical models. Nevertheless, an improvement in the accuracy of the predictions would enhance the performance of the Advisor. .para Therapy regimens other than a constant infusion with boluses could be added. The existence of pharmacokinetic models and programmable infusion pumps has made it possible to use the knowledge about the exponential behavior of the drug distribution phase to develop infusion regimens that compensate for the drug distribution by having initially high infusion rates that taper to the rate needed to achieve steady state. Programs to calculate the regimens and program the pumps are currently available~[crone!,~jelliffe1,~arzbaecher]. The use of exponential infusion regimens would also have the beneficial side-effect of making the management decision process simpler, since such regimens are capable of nearly instantaneously achieving and maintaining a given target concentration level. The management problem need then only concern itself with assessing the effectiveness of the chosen target level. The use of programmed pumps also allows the easy use of non-integer infusion rates without placing an unreasonable burden on the nursing staff. .subsection "Implementation of Modules" .para In order to expand the capabilities of the Arrhythmia Advisor by including more drugs, a therapy selection component must be added to the therapy management module. At the present time, the therapy selection is a fixed order list of anti-arrhythmic agents that are tried if the state of the patient's arrhythmia warrants treatment. An extension to handle a broader spectrum of cardiac therapy, not exclusively related to the arrhythmia itself, requires that the types of treatment be selected. In order to be fully effective, this requires that diagnostic and physiologic knowledge be incorporated into the Advisor. .para Part of the knowledge needed is to be encapsulated in the disease evaluation module. One example of the interaction of these two additions occurs in the case of arrhythmia caused by a low serum potassium concentration. This condition can be treated directly by administering potassium supplements. Using the diagnosis of arrhythmia partially caused by low potassium, the treatment selection algorithm would add potassium supplements to the treatment regimen. .subsection "Extension of Capabilities" .para The use of programmer provided explanations of the trace of the program execution meets the minimum requirement for program explanation. The explanations could be improved if the basis for the reasoning could be directly examined. One major fault is that the assumptions made in the design of the algorithm are no longer available to be examined at run time. Explanations could be improved by making the operating assumptions explicit and by keeping a record of the goals that are being pursued by individual sections of code. This would allow users to examine in detail the reasoning behind recommendations which appear to be unusual. A large improvement in the explanation capability will remain a research issue. One approach that has been demonstrated experimentally is to use automatic programming techniques to generate the actual advisor program from a high-level goal representation of the task~[swartout2]. .para Another way in which the capabilities of the system could be improved would be by the inclusion of multiple hypotheses. Since it is known that certain of the data are inaccurate, multiple hypotheses could be based on different assumptions about what the actual state of the patient is. For example, given information about the distribution of drug levels for normal patients, an average, a high and a low case can be considered as parallel hypotheses. Maintaining multiple hypotheses can provide the opportunity to perform a sensitivity analysis on the decision. The effect of different assumptions on the management advice could be calculated and the various possibilities presented to the user. For example, this process has the potential for reducing the number of alarms from the arrhythmia monitor. If the presence or absence of a particular arrhythmia would have no effect on the advice that would be generated, then it is not necessary to have the staff confirm the existence of that arrhythmia. This provides a higher level mechanism to compensate for inaccuracies at the signal processing stage of arrhythmia monitoring. It will be necessary to choose the assumptions to be investigated wisely, in order to avoid combinatorial explosion in the number of hypotheses under consideration. One solution is to apply the expert systems approach to this sub-problem as well. .para Other types of analysis are also possible. Some data exists concerning the danger represented by the various forms of ventricular arrhythmia. As more data is gathered about the statistical likelihood of toxicity at various drug concentration levels, the inputs to the risk-benefit decision could be made more mathematically rigorous. A formalization of the decision basis would extend the work of this thesis. .sect "Suggestions for Further Research" .para The suggestions for further research are divided into two sections depending on whether they are more germane to the medical domain or the computer science area. One by-product of the formalization process used in developing the expert algorithm was the identification of some areas in which medical knowledge is incomplete. The attempt to automate the therapy management also uncovered areas of computer science where further work is necessary. .subsection "Medical Research" .para This domain uses descriptive models as information sources for the control of drug therapy because more fundamental causal models of the drug action are not available. If such models were available, then the management algorithm would depend less upon the intuition developed by clinicians and could engage in reasoning at the causal level. This approach has found favor in the AI in medicine community. By understanding the mechanisms behind the action of the drugs, it should be possible to reason about treatment strategies from first principles. Being able to supply a justification for action that is founded in a deeper understanding of the disease processes involved is desirable. At the present time, much of the information is empirical. For example, lidocaine has been found to be effective in controlling many types of ventricular arrhythmia, but the exact mechanism by which this control is achieved remains to be explained. .para The determination of the endpoints of therapy and a description of what constitutes acceptable progress in the control of arrhythmias is another area where research is needed. At present, there is some controversy about what the best method of treating arrhythmias is~[lown_ed,~osborn]. An alternate approach to this problem is discussed in the computer science section where aids to support the specification of management strategies are considered. .para One type of case that cannot be handled by the present Arrhythmia Advisor is the treatment of patients with more than one anti-arrhythmic agent simultaneously. As indicated in the previous chapter, this is a difficult problem because beneficial effects cannot be attributed to any of the agents directly. If the mechanisms by which the drugs act can be better defined, then it may become possible to isolate the individual effects of multiple drugs used together. In order to handle such treatment regimens in a rational manner, however, some method must be discovered for attributing beneficial and toxic effects to the individual constituents of such a multi-drug regimen. This is in general a very difficult problem. .subsection "Computer Science" .subsubsection "Control of Reasoning in Time Dependent Domains" .para System support for the control of reasoning in time dependent domains is insufficient. Research in the area of temporal reasoning has focused on the problem of reasoning about time relations [allen1,~allen2,~bruce,~kahn,~mcdermott] rather than on providing a control system for reasoning modules that operate in domains where time is important. .para Two examples from the domain of ventricular arrhythmia management will serve to illustrate the problem. First, consider the integration of information from laboratory tests. At some time, samples are drawn and sent to the clinical laboratory. After processing, which typically takes several hours, the results are available to the CCU staff. In the meantime, therapy may have been started. Thus the test results do not describe the state of the patient at the time they become available to the decision makers, but rather at a time in the past, namely when the samples were drawn. In order to reason using this data, it is necessary to view the laboratory results as describing this past state and then consider the effects of any therapeutic interventions that may have changed that state. Concretely, assume that the laboratory results indicate that the serum concentration of potassium is low. This can cause VPBs. If the patient was suffering from severe arrhythmia, then some therapy would have been administered. An assessment of the current state of the patient must take into account both the alternate cause of the VPBs and any anti-arrhythmic therapy administered since the blood samples were drawn. Reasoning in a domain in which the state changes over time and not all information is immediately available requires that the control system used be able to account for the difference in the times at which information from several sources become available. Fagan indicates that this is an area that would be an extension of the work done in VM. .para The second example demonstrates the need to alter conclusions reached by earlier executions of modules. The pharmacokinetic model establishes expectations and plans therapy based on the information about the pharmacokinetics of the drug being used, modified by facts known about the patient. With lidocaine, for example, it is known that patients with liver disease will have higher concentrations than patients without liver disease. If it is subsequently discovered that a patient being treated has liver disease, then the drug level expectations generated in the past under the false assumption of no liver problems must be recomputed. Since this recomputation refers to the ital(past), however, no changes in therapy can be considered. Currently the Arrhythmia Advisor recomputes the drug levels by starting at time zero. This is a simple control system that assures that all relevant information about the patient is used to calculate the present state of the therapy even if some of the information had changed. It unfortunately suffers from a lack of efficiency, particularly when cases are followed over several days. It would be useful to have a control system which would allow the reexecution to be limited to the time periods in which a change would have an effect. .para Examination of these problems led to the establishment of criteria for a control system for reasoning processes in a time dependent domain. A control structure for use in such a domain should support: .nlist The ability to monitor the data dependencies of the reasoning modules and engage in daemon style invocation of modules whose input values have changed. This would allow changes in the input values to automatically propagate through all relevant reasoning modules. .next State variables that will allow a computation to be broken into a number of segments and continued. The relevant state variables must be saved from one execution until the continuation of the same execution later. This allows an incremental recomputation when data values change. It also supports the continuation of a process when data about a time further in the future is desired, rather than requiring the process to start from the beginning. .next A distinction between the future and the past when the modules are executing. Since two types of reasoning functions are being performed, namely conclusions or expectations are being generated and suggestions about possible interventions are being made, it is necessary to distinguish between the past and the future since past actions cannot be undone. When reasoning in the past, actions must be accepted as given and only conclusions can be altered. .elist .para At present there is no system support for the implementation of such reasoning. A report on the early stages of work being done by the author and Dr. Long in this area can be found in [long]. Dr. Long is responsible for the detailed design and coding of the control system. .para The new implementation of the Advisor will use this control structure. The pharmacokinetic model module has gone through initial testing and has demonstrated the feasibility of this type of programming. .subsubsection "Specification of Clinical Strategies" .para The major focus of research in this thesis was the development and implementation of a strategy for ventricular arrhythmia management based on the clinical strategies used by a cardiologist. The development of a general method for the specification of clinical strategies to support the generation of such algorithms appears to be an area worthy of further investigation. .para The fundamental characteristic of the clinical strategy is that it is based on mental models of processes at work in the patient. These models allow predictions of effects of therapy to be made and thus are crucial to planning therapeutic actions. Based on these predictions, expectations of improvements in a patient's clinical status can be made. The difference between the ideal expectations and the actual outcome is then used to refine the initial treatment plan. .para System support would be especially useful at a basic level since the detailed reasoning about therapy can be difficult and tedious. A large amount of time was spent hand-crafting the therapy advice algorithm described in this thesis. Most of the time was spent handling the difficulties associated with making decisions based on time and concentration pairs. Primitives that would make the description of the data easier would simplify the programming task. .para A mechanism to aid in the development and implementation of clinical treatment strategies should include the following features: .nlist It must be possible to specify models of processes which can then serve as data sources for the reasoning algorithms. .next The interface to the models should be simplified by providing abstraction mechanisms for reasoning about the model-produced data at an abstract level. An example would be support for the sub-therapeutic, therapeutic and toxic drug concentration concepts. .next The separation of general knowledge from patient specific information should be supported. .next Primitive constructs for the description of decisions and actions should be developed. For example, increasing drug doses and changing therapies are concepts that should be supported. .next The system should be easily extensible, so that domain specific reasoning strategies can be incorporated into the overall strategy specification system. .next In addition to a specification of the action or method to be used, the system should provide a mechanism for declaring the ital(intent) or purpose of a decision or action. .elist The result would be a formal language that could be used to describe clinical decision making strategies. The existence of such a language could be expected to provide the following benefits: .blist The construction and modification of therapy algorithms would become simpler. .next Having the intent of an action accessible to the program can serve as a check on the method being used. It also makes more meaningful explanations of a program's action possible, since the purpose of an action can also be reported. .next Since therapy management is an inexact art, differences in strategy exist. If an expert system is to find wide acceptance in such an environment, it should be easy to modify to meet the criteria for acceptable therapeutic advice in the community in which it will be used. By providing a high level method for specifying the actions, the therapy algorithm could be adjusted to reflect the local practices in the management of patients. .next The existence of a formal language for specifying clinical treatment plans could promote discussion of the strategies in the medical community. The language would provide a common basis for the comparison of different strategies, serving a function similar to that envisioned for ALGOL in the area of algorithm description. .elist .sect "Summary" .para In this thesis, the strategy used by an expert cardiologist was elucidated and served as the basis for a therapy management algorithm. His management strategy was modified to include information from a formal mathematical model which provides a more accurate picture of the time course of drug concentrations in patients than the simpler intuitive model generally used in clinical practice. The expert knowledge used in the Arrhythmia Advisor serves two purposes: first, it enables the program to perform reasoning in areas of the domain for which formal models do not exist, and second it provides a method for drawing conclusions based on data from multiple sources. Clinical expertise is needed to use the results of formal mathematical models in an actual clinical setting. For example, the proper interpretation of data from the pharmacokinetic model requires that the patient's state be considered, which requires expert knowledge about cardiology. .para The informal evaluation shows that such an approach is capable of producing reasonable advice when it operates within the limits of its area of expertise, and can on occasion produce therapeutic recommendations which are better than current clinical practice. This indicates that the approach to management that was taken in this thesis research has the potential for improving the status of medical care. Expert systems technology can produce decision support tools for an intensive care environment. .ls 1 .de gitem .sp .55 .ne 3 .ti -10 bold(\0~~em!)~~ .em .de aitem .sp .55 bold(\0)(\p_value!m)\1 .em .appendix "Glossary of Medical Terms" .achap "Abbreviations" .nr save_vpos vpos .nr p_value 1100 .in +.5i .aitem "AV" "atrio-ventricular" .aitem "CAD" "coronary artery disease" .aitem "CCU" "cardiac care unit" .aitem "CHF" "congestive heart failure" .aitem "CI" "cardiac index" .aitem "CNS" "central nervous system" .aitem "CO" "cardiac output" .aitem "EF" "ejection fraction" .in -.5i .nr end_vpos vpos .vp save_vpos!m .nr p_value 4100 .in +3.5i .ns .aitem "EKG" "electrocardiogram" .aitem "IV" "intravenous" .aitem "MI" "myocardial infarction" .aitem "PSM" "patient specific model" .aitem "SA" "sinoatrial" .aitem "VF" "ventricular fibrillation" .aitem "VPB" "ventricular premature beat" .aitem "VT" "ventricular tachycardia" .in -3.5i .vp end_vpos!m .sp 2 .achap "Glossary" .in +10 .gitem "acute" of sudden onset, happening quickly. .gitem "angina" [actually: ital(angina pectoris)] a squeezing chest pain due to ischemia (q.v.) of heart muscle. .gitem "aorta" The major artery leaving the heart and supplying blood to the body. .gitem "aortic" Referring to the aorta or to the valve connecting the heart to the aorta. .gitem "arrhythmia" irregularity in the heart rhythm. It can be caused by irregular conduction of electrical signals. One type of arrhythmia is an early beat that starts in the ventricle (q.v.). .gitem "atrium" The upper chamber of the heart. The atria receive blood coming into the heart. In the normal heart beat sequence, they contract before the ventricles (q.v.). .gitem "atrio-ventricular (AV) node" tissue that forms the electrical conducting pathway between the atria and the ventricles (qq.v.). In order to maintain separate contractions, the atria and ventricles are electrically insulated from one another. The AV node conducts the electrical signal initiating a heart beat. A delay is introduced, thus assuring proper sequencing of contractions. .gitem "bolus" a one time shot of drug. .gitem "cardiac care unit (CCU)" intensive care unit with a staff specially trained in cardiac care. Patients at high risk of heart complications are treated here. .gitem "cardiac" pertaining to the heart. .gitem "cardiac index (CI)" cardiac output (q.v.) normalized for body surface area, generally expressed in liters per minute per square meter (normal approx 3.5 l/mindot!msup(2)) .gitem "cardiac output (CO)" rate of blood flow out of the heart, generally expressed in liters per minute (normal approx 5 l/min). .gitem "chronic" long term, not sudden. .gitem "congestive heart failure (CHF)" condition in which the heart does not pump enough blood to meet the body's demand resulting in fluid buildup in the body. .gitem "coronary artery disease (CAD)" build up of fatty deposits in the arteries supplying the heart with blood, ``hardening of the arteries.'' .gitem "couplet" two ventricular premature beats (q.v.) in a row. .gitem "creatinine" protein excreted by the kidneys. The rate of creatinine clearance can be used to determine kidney function. .gitem "defibrillation" stopping fibrillation (q.v.) often by means of an electric shock. .gitem "diuretic" drug that causes the kidneys to remove water from the body. Often used to help reduce edema (q.v.). .gitem "ectopic" coming from an unusual place, used to describe the origin of abnormal heart beats. .gitem "edema" accumulation of fluid in the tissues. Pulmonary edema is fluid in the lungs, pedal edema is fluid in the feet and ankles. .gitem "ejection fraction (EF)" ratio of amount of blood pumped out per beat to the volume of the heart chamber. Expressed as a ratio or percentage (normal approx 65%). .gitem "electrocardiogram (EKG)" tracing of the voltage of the discharge of the heart muscles during contraction and the voltage of electrical recovery. .gitem "fibrillation" uncoordinated individual contractions of heart muscle cells without significant pumping effect. Often described as having the appearance of the motion of a bag of worms. .gitem "heart failure" see congestive heart failure. .gitem "hepatic" pertaining to the liver. .gitem "highly perfused organs" organs that have a large blood flow, generally the heart, brain, kidneys and liver. .gitem "intravenous (IV) infusion" introduction of fluid to the body by allowing it to flow into a vein. .gitem "ischemia" insufficient blood supply causing oxygen starvation of tissues. .gitem "multiform VPBs" VPBs (q.v.) which have more than one shape on an electrocardiogram. Thought to be due to more than one pacemaker or conduction path that causes the difference in the electrical discharge pattern. .gitem "myocardial infarction (MI)" death of heart muscle tissue due to occlusion of an artery. Commonly called a ``heart attack.'' .gitem "myocardium" the heart muscle. .gitem "P wave" initial wave on an EKG (q.v.) immediately preceding the contraction of the atria (cf. Figure current_figure below). .gitem "pacemaker" bold(a))~~origin of the electrical signal that causes a heart contraction. This is normally the function of the SA node, but other sources are possible.~~bold(b))~~a mechanical device that triggers the contraction. .gitem "pathology" study of disease processes. .gitem "pedal" pertaining to the feet. .gitem "pharmacodynamics" study of the relationship between drug concentration and drug effect. .gitem "pharmacokinetics" study of the time and dose dependent distribution of a drug in the body. .gitem "physiology" study of the function of the body. .gitem "primary arrhythmia" arrhythmia (q.v.) without an apparent cause. .gitem "prognosis" forecast of the outcome of a disease or therapy. .gitem "pulmonary" relating to the lungs. .gitem "Q wave" the first deflection of the EKG (q.v.) signal before the contraction of the ventricles (q.v.). Start of the QRS complex (q.v.)~~(cf. Figure current_figure below). .gitem "QRS complex" the major electrical signal on an EKG (q.v.) immediately preceding the contraction of the ventricles. The major deflection is the R wave (q.v.)~~(cf. Figure current_figure below). .gitem "QT interval" time from the Q wave until the T wave (qq.v.)~~(cf. Figure current_figure below). This is used as measure of the time required for the heart to recover electrically after a beat. This quantity is often adjusted for the heart rate, when it is called the corrected QT interval (KQT). Usually expressed in seconds (normal KQT approx 0.38) .gitem "R wave" the largest signal caused by the electric discharge of the heart prior to contraction of the ventricles (q.v.)~~(cf. Figure current_figure below). .gitem "R-on-T VPB" an especially early beat in which the early beat's R wave begins during the T wave (qq.v.) of the preceding beat. It can be categorized by computing the ratio of the time between beats (RRprime) to the QT interval (RRprime!/QT ratio). If this ratio is less than 0.9 it is definitely R on T, if greater than 1.0 then it is just an early beat, and otherwise it is borderline. .gitem "rales" gurgling sound caused by fluid in the lungs (pulmonary edema). .gitem "renal" pertaining to the kidneys. .gitem "run" more than two VPBs (q.v.) in a row. .gitem "sinoatrial (SA) node" the main pacemaker (q.v.) of the heart. It normally initiates the electrical signal which coordinates the contraction of the heart. .gitem "sinus node" see sinoatrial node. .gitem "stenosis" narrowing of a blood vessel or, more commonly, a heart valve. It is a condition which restricts blood flow and causes a pressure gradient. .gitem "sudden death" cardiac arrest due to arrhythmia. Sudden death today is a reversible phenomenon, since the heart can be restarted through appropriate resuscitation. .gitem "T wave" wave on the EKG (q.v.) recording which traces the electrical recovery of the heart. After the recovery, the heart is capable of conducting electrical signals and contracting again (cf. Figure current_figure below). .gitem "tachycardia" fast heart beat. (See also ventricular tachycardia). .gitem "therapeutic window" range of drug concentrations between the minimum effective dose and the minimum toxic dose. Drugs which have a ital(narrow) therapeutic window have only a small difference between ineffective, effective and poisonous drug concentrations. .gitem "torsade de pointes" description of a run of VPBs (qq.v.) in which the morphology of the individual beats changes. This is considered to be a very serious arrhythmia (q.v.). .gitem "ventricle" the lower chamber of the heart. The ventricles pump the blood to either the lungs or the body. .gitem "ventricular fibrillation (VF)" fibrillation of the ventricle (q.v.). A very serious arrhythmia (q.v.) which will lead to death if not stopped. .gitem "ventricular premature beats (VPB)" early heartbeats which originate in the ventricle as opposed to the sinus node (qq.v.). They are a subclass of arrhythmia (q.v.). .gitem "ventricular tachycardia (VT)" fast heart beat (usually greater than 120 beats per minute) with a ventricular origin (see VPB). Pumping efficiency is decreased. A dangerous arrhythmia (q.v.) which can degenerate into ventricular fibrillation (q.v.). .gitem "volume of distribution" a mathematical pharmacokinetic (q.v.) concept that describes the volume of a compartment of a pharmacokinetic model (see appendix :pk_model) into which a drug is distributed. This has an effect on the concentration of the drug (concentration = amount of drug div volume of distribution). .in -10 .sp .ne 15c .para Figure~current_figure shows an idealized EKG trace with the different waves, the QRS complex and the QT interval identified. .nr immediate_figure 1 .figure "An Idealized Electrocardiogram Trace" <== .bp bold(Case 4) .so tar;case4 > .bp bold(Case 5) .so tar;case5 > .bp bold(Case 6) .so tar;case6 > .bp bold(Case 7) .so tar;case7 > .bp bold(Case 8) .so tar;case8 > . insert_refs