Explaining Machine Learning Models for Early Detection of Pregnancy Risk

Abstract

Care management programs for high-risk pregnancies aim to detect pregnant women with pregnancy risk factors early so they can receive proper care or preventative treatment. To detect these women, pregnant members are first detected, then they are checked for high risk diagnosis codes or fed into a risk prediction algorithm. Members predicted to be most at risk are outreached and provided guidance on how to manage or monitor symptoms.

In this thesis, we work with the high risk pregnancy care management team at Independence Blue Cross to (1) build a pregnancy identification algorithm to detect pregnant women earlier in their pregnancy, (2) model impactable pregnancy risk factors, and (3) explain these models’ predictions. We introduce a new framework for thinking about explainability methods in healthcare – working in assumptions about a prior understanding a clinician may have about the patient and working with high dimensional, redundant data – and we conduct a user study to examine deployability and impact of these algorithms.