Transport-based transfer learning on Electronic Health Records: application to detection of treatment disparities.

Abstract

Objectives: Electronic Health Records (EHRs) sampled from different populations can introduce unwanted biases, limit individual-level data sharing, and make the data and fitted model hardly transferable across different population groups. In this context, our main goal is to design an effective method to transfer knowledge between population groups, with computable guarantees for suitability, and that can be applied to quantify treatment disparities.

Materials and methods: For a model trained in an embedded feature space of one subgroup, our proposed framework, Optimal Transport-based Transfer Learning for EHRs (OTTEHR), combines feature embedding of the data and unbalanced optimal transport (OT) for domain adaptation to another population group. To test our method, we processed and divided the MIMIC-III and MIMIC-IV databases into multiple population groups using ICD codes and multiple labels.

Results: We derive a theoretical bound for the generalization error of our method, and interpret it in terms of the Wasserstein distance, unbalancedness between the source and target domains, and labeling divergence, which can be used as a guide for assessing the suitability of binary classification and regression tasks. In general, our method achieves better accuracy and computational efficiency compared with standard and machine learning transfer learning methods on various tasks. Upon testing our method for populations with different insurance plans, we detect various levels of disparities in hospital duration stay between groups.

Discussion and conclusion: By leveraging tools from OT theory, our proposed framework allows to compare statistical models on EHR data between different population groups. As a potential application for clinical decision making, we quantify treatment disparities between different population groups. Future directions include applying OTTEHR to broader regression and classification tasks and extending the method to semi-supervised learning.