Dissertations and Theses (Open Access)
Graduation Date
Summer 2025
Degree Name
Doctor of Philosophy (PhD)
School Name
McWilliams School of Biomedical Informatics at UTHealth Houston
Advisory Committee
Hongfang Liu, PhD,
Abstract
Suicide remains a leading cause of preventable death, with particularly high burden in underserved populations treated in safety-net psychiatric hospitals. Early detection is challenged not by the absence of data, but by the informational opacity of psychiatric clinical narratives which is unstructured, subjective, and heterogeneous records that resist traditional analytic methods. This article-based dissertation addresses this challenge by developing an integrated MINDFUL framework that combines clinically validated annotation, advanced machine learning, and explainable AI to enable accurate, interpretable suicide phenotyping in high-acuity psychiatric settings. The first study created a gold-standard annotated corpus of psychiatric notes labeled for suicidal ideation, suicide attempt, exposure to suicide, and non-suicidal self-injury. Developed with multidisciplinary input, the annotation process achieved near-perfect agreement (Cohen’s Kappa = 0.95) and revealed key psychosocial correlates and documentation challenges. Baseline modeling demonstrated feasibility for AI-assisted risk identification (micro-F1 = 0.70). The second study evaluated transformer-based language models for multi-label suicide phenotyping. RoBERTa fine-tuned as a single multi-label classifier outperformed all alternatives (accuracy = 0.88, F1 = 0.81), confirming that domain-relevant pretraining and unified modeling improve performance and efficiency in psychiatric NLP tasks. The third study addressed explainability using reasoning large language models with in context learning. These models matched or exceeded a fine-tuned GPT-3.5 baseline (best accuracy = 0.94, F1 = 0.90) while generating clinically coherent justifications aligned with psychiatric reasoning. Together, the three studies show that validated NLP can extract clinically relevant suicide phenotypes from unstructured psychiatric notes with both accuracy and interpretability. The MINDFUL framework offers a pathway for transforming complex clinical narratives into actionable insights, advancing both methodological innovation and suicide prevention in clinical practice
Recommended Citation
Li, Zehan, "MINDFUL: Machine Intelligence Towards Neuropsychiatric Data For Unified Learning" (2025). Dissertations and Theses (Open Access). 92.
https://digitalcommons.library.tmc.edu/uthshis_dissertations/92
Keywords
Machine Intelligence, Neuropsychiatric Data, Suicide Phenotyping, Natural Language Processing, Large Language Models, Explainable AI, Clinical Notes, Psychiatric Informatics, Multi-label Classification, Electronic Health Records