Logo image
Back
Explaining Fairness Enhancement with Collaborative Learning for Prediction of Treatment Completion in Patients with Substance Use Disorder
Conference paper

Explaining Fairness Enhancement with Collaborative Learning for Prediction of Treatment Completion in Patients with Substance Use Disorder

Mary M Lucas, Quyen M Ngo and Christopher C Yang
IEEE International Conference on Healthcare Informatics (ICHI)
12 Aug 2026
DOI: https://doi.org/10.1109/ICHI69079.2026.00016
Featured in Collection :   Drexel's Newest Publications

Abstract

Unequal performance of machine learning models across patient groups can lead to disparities in health outcomes, potentially leading to poor outcomes and patient harm. While many AI bias reduction methods exist, they often do not provide explanations or insights into how the model fairness is achieved. The objective of this work is to investigate how a collaborative learning approach can improve fairness enhancement in a machine learning model for predicting treatment completion for substance use disorder, and how these improvements can be explained. Using dimensionality reduction and clustering, we identify patient subgroups where the collaborative model is more successful at correcting errors made by the group-specific models. We then analyze these patient clusters to understand their feature distributions and treatment characteristics, finding that corrected cases often show distinct patterns when compared with the overall patient population. To explain individual predictions, we apply SHAP values and compare feature contributions across clusters, showing how the collaborative model leverages complex feature interactions to override dominant risk signals in corrected cases. Our findings demonstrate the value of localized interpretability in understanding model behavior and tradeoffs introduced by enhancing model fairness.

Metrics

1 Record Views

Details

Logo image