Toward Trustworthy Machine Learning in Precision Health: Interpretable, Scalable, and Fair Methods for Clinical Prediction

Abstract

Artificial intelligence (AI) and machine learning (ML) play an increasingly important role in precision health, with applications in disease detection, risk stratification, treatment planning, and outcome prediction. Yet healthcare data remain challenging because they are often heterogeneous, imbalanced, large-scale, difficult to preprocess, and prone to limited transparency and subgroup disparities in performance. This dissertation addresses these challenges through studies focused on interpretability, fairness, scalability, and clinically meaningful prediction. The dissertation first reviews AI and ML applications in precision health. It then applies interpretable machine learning methods to maternal health problems, including cesarean delivery prediction among class III obese women undergoing labor induction and length-of-stay prediction among patients with preeclampsia. Next, it introduces a multilevel deep neural network framework for efficient learning on large, imbalanced datasets, together with a fairness-aware extension. It also develops an interpretable, fairness-aware survival modeling framework to identify childhood cancer survivors at risk of disengaging from long-term follow-up care. Finally, it evaluates language-model-based tabular representations for simplifying healthcare data preparation and enriching learned representations. Overall, this dissertation demonstrates that healthcare machine learning must extend beyond predictive accuracy to also support interpretability, fairness, scalability, and clinical relevance. This work contributes to the development of more robust, transparent, and equitable AI/ML methods for precision health.