Reliable and Trustworthy AI for Evidence-based Clinical Decision Support in Cancer Care

Abstract

The integration of cutting-edge AI methods with real-world clinical data has moved from being a novelty to a necessity in oncology. However, the deployment of AI faces challenges, including the complexity of reliably modeling longitudinal Electronic Health Records (EHR) characterized by missing data and frequent patient drop-outs, patient heterogeneity which leads to disparities in AI performance, and the need for validating AI models' clinical benefits, especially in managing challenging cancer cases. This thesis presents research focused on addressing these challenges: developing a continuous time model-based time-to-event regression framework to improve the prediction of clinically meaningful patient outcomes from irregularly sampled EHR data; utilizing data and algorithm-driven approaches to mitigate AI performance disparity for predicting cancer-associated adverse events across diverse patient demographics; and developing an AI-based decision support tool that integrates genomics and clinical data for evidence-based cancer care, with a focus on improving management of difficult-to-treat cancer cases. This work contributes towards transforming cancer care through reliable and trustworthy AI-driven clinical decision support.