Title: Actionable machine learning in genetics, personalized medicine, and health
Advisor: Su-In Lee
Supervisory Committee: Su-In Lee (Chair), Meliha Yetisgen (GSR, Biomedical Informatics), Ali Shojaie (Biostatistics, Public Health), and Walter Ruzzo
Abstract: Modern machine learning is changing how we approach biology, medicine, and health. However, a key challenge is to build models that produce actionable insights rather than opaque predictions. To support this we propose four methods that learn interpretable, and hence actionable, machine learning models:
1) ChromNet learns the network structure of interactions among proteins in the human cell nucleus, and has provided effective visualization of these results to over 5,000 users so far. This allows us to understand which proteins collaborate to manage our DNA.
2) SHAP unifies several methods designed to provide additive explanations of any machine learning model. It uses an axiom-based approach to justify why a single explanation can be applied to models of any type.
3) Prescience learns an ensemble model for the prediction of hypoxia in the operating room and then uses a SHAP-based approach to explain the prediction. This demonstrates the use of a complex model while still allowing a doctor to understand what causes a patient's risk.
4) Study Flow will enable us to combine multiple models from different studies while retaining the relevant causal assumptions. By encoding the causal assumptions of each model, we can automated large parts of the systematic review process.