The basic goal of my research is understanding reasoning by developing computational analogies for facets of reasoning required in human problem solving. The forum for this interest is the area of Artificial Intelligence called ``Expert Systems''. Specifically, I want to understand and develop mechanisms for the kinds of reasoning that the present generation of expert systems is incapable of handling. There are two sides to the research: the kinds of reasoning and the domain. I am working on problems such as reasoning about time, about causation, and about the implications of mechanisms, but reality is maintained by developing solutions to clinically relevant problems in cardiology in collaboration with expert cardiologists.
My primary commitment is the development of a program to help physicians reason about the diagnosis and management of patients in heart failure. The program will assess the clinical information entered as evidence about the physiological state of the patient. Then it will provide the user with assessments of the possible causes and complicating factors (both ruled in and yet to be ruled out), about what information would clarify the diagnosis, about possible therapies to correct the problems, about possible effects of therapies in the current patient state, and about the interpretation of additional findings. This project is being developed in collaboration with three cardiologists at Tufts-New England Medical Center. A second project is the development of a general knowledge representation for medicine by representing the knowledge pertinent to coronary artery disease in a uniform representation language usable by many kinds of programs. This project is part of a larger project involving my colleagues at MIT and a group of physicians at T-NEMC. A third project is a program to give advice in the Cardiac Intensive Care Unit on the management of ventricular arrhythmias. This program will take heart beat classifications from an arrhythmia monitor attached to the patient, information about the current drug regimen, and clinical information about the patient, and provide assessments and advice for adjusting the anti-arrhythmia therapy. This project is a collaboration between our group at MIT, physicians and researchers at University Hospital, and engineers at Hewlett-Packard.
These projects provide an ample supply of reasoning problems from which to develop ideas and mechanisms. I have looked at issues of reasoning about time in a couple of contexts. We are developing a mechanism for reasoning in an environment where data arrives over time, but not necessarily in time order, in conjunction with the arrhythmia advisor. The important idea is the separation of reasoning into static contexts defined by the data trends and into reasoning about trends across time. Time also plays a role in reasoning about diagnosis em the possible sequences of events leading to the present state. Driven by the causal relationships in the heart failure domain, I have developed a strategy for reasoning about possible diagnoses even when causation takes place over a variety of times. For example, it is possible to have pedal edema caused by high blood volume present at the same time as low blood volume caused by aggressive diuretic therapy. Accounting for such possibilities requires explicit representation of the times required for causation. In the heart failure domain I have also developed several ideas for reasoning about causation in a physiological model including explicit representation of the justifications for conclusions to allow appropriate problem formation when contradictions arise. Most recently I have been investigating mechanisms for using quantitative and ordering information in reasoning about complicated causal physiological models. The technique I am developing makes use of the ideas of signal flow analysis to capture the abstractions of the feedback loop and the causal pathway to extend the kinds of reasoning that have been explored in AI as qualitative reasoning. This new technique holds the promise of providing the right abstractions for reasoning about the dominance of mechanisms in a complicated model. There are many other significant problems in these domains that I would like to explore, time permitting. The exploration of reasoning is just beginning.