Temporal Reasoning in Clinical Narratives: From Information Extraction to Early Disease Detection

Timely and accurate prediction of chronic disease risk, such as Type 2 Diabetes, requires models that can reason over complex, longitudinal clinical narratives. This thesis presents a unified framework for temporally grounded patient modeling that integrates structured event representations, fine-grained temporal reasoning, and scalable predictive architectures. Initial experiments show that concept-based models paired with visit-level temporal modeling outperform traditional baselines that ignore temporal dynamics, motivating a deeper exploration into concept and temporal relation extraction from clinical narratives. Towards this, I introduce GraphTREx, a state-of-the-art temporal relation extraction approach that achieves a 5% F1 improvement on the end-to-end temporal relation extraction task in the I2B2 2012 challenge with ~9% gains on long-distance relations. GraphTREx generalizes well, showing strong performance on the E3C corpus and robust out-of-domain results (without additional adaptation) on a newly annotated UI Health dataset with dense temporal relations, highlighting its portability. Building on these temporal graphs, I develop HiTGNN—a Hierarchical Temporal Graph Neural Network that integrates intra-document temporal relations, inter-visit dynamics, and external medical knowledge, enabling reasoning across both local event structures and longitudinal patient trajectories. I also introduce ReVeAL (Reasoning with Verifier Aided Labeling), an inference-time scaling framework where a smaller LLM validates predictions from a larger frozen LLM, inheriting interpretability and improving accuracy without full retraining. While our LLM-based approach offers explanations, HiTGNN outperforms it (and other LLM baselines) in accuracy and efficiency, especially over shorter prediction horizons, demonstrating the strength of lightweight, temporally grounded graph-based models, well-suited for low-resource, privacy-sensitive clinical settings. Rigorous dataset curation, ablations, and analyses reinforce robustness of these approaches. Together, these contributions establish a generalizable framework for clinical prediction by integrating concept abstraction, temporal structure, and semantic enrichment to support robust, privacy-conscious AI systems for real-world clinical decision support.