Developing Medical Decision Support Tools with Logistic Regression
Introduction
Machine Learning Methods
Linear regression
“least squares” method
Logistic Regression
Sigmoid Activation Function
CHD Versus Age: Sigmoid Distribution
Logistic Regression Formula
Maximum Likelihood Estimation
PPT Slide
Univariate Ranges
One Dimensional Decision Problem
Use of the ROC Curve to Optimize Threshold
Dichotomous Data
A Linear (2D) Decision Boundary
Curved Decision Surfaces
Curved Decision Boundary
An Over-fitted Model
Likelihood Ratios: Edinburgh Data
Likelihood Ratios : Sheffield Data
Optimizing Performance on Unseen Data
Validation Set
Jack Knifing
Example Models: Chest Pain
Comparison of Different Model Types
ROC Areas for Small and Large LR Models
Performance of Small and Large LR Models Sheffield Test
Chest Pain Data: Model Comparisons
Neural Net, Edinburgh Data (700 training, 553 test)
Low Blood Pressure Prediction, Pfizer Data
Benefit of Simple Models: Predicting WHF with ECG Data
Linear and Non-Linear Models (Predicting Low BP)
Stepwise LR: Low BP Medications
Low BP Meds: Results
LR Benefits for Developing Prediction Models
Advantages of Multi-layer Neural Network Programs
Suggested Model Building Strategy
Further Reading
Email: hamish@mit.edu
Home Page: http://medg.lcs.mit.edu/people/hamish/index~3.htm
Other information: Medical Computing Class