Logo image
Counterfactual Optimization of Combinatorial Treatments Using Inverse Classification
Conference proceeding

Counterfactual Optimization of Combinatorial Treatments Using Inverse Classification

Sulyun Lee, W. Nick Street, Aaron Sporrer, Barry L. Carter and Linnea A. Polgreen
Proceedings (IEEE International Conference on Healthcare Informatics. Online), pp.1394-1396
06/01/2026
DOI: 10.1109/ICHI69079.2026.00217

View Online

Abstract

Optimizing multi-drug regimens from observational data is challenging due to combinatorial explosion: k binary treatments yield 2 k possible combinations. We present an inverse classification framework using neural networks that avoids explicit interaction modeling by learning a predictive function across the treatment space. Applied to 124,031 post-AMI Medicare patients optimizing 7 medication classes (128 combinations) to minimize 2-year mortality, the neural network achieved AUC 0.793. Individualized recommendations differed from observed treatments in 98% of cases, with 1.8% of patients showing predicted mortality reductions exceeding 10 percentage points. Predicted benefits were largest among vulnerable subgroups including older patients, females, minorities, and those with higher comorbidity burden.
Machine Learning Optimization Equations inverse classification Modeling myocardial infarction Printing Testing Training

Details

Metrics

1 Record Views
Logo image