COMMITTEE CHAIR: Dr. Lin Li
COMMITTEE CO-CHAIR: Dr. Lijun Qian

TITLE: A FRAMEWORK FOR EXPLAINABLE PTSD SEVERITY CLASSIFICATION AND AUTOMATED SOAP NOTE GENERATION

ABSTRACT: Post-traumatic stress disorder (PTSD) remains underdiagnosed in military populations due to stigma, variability in symptom presentation, and challenges associated with accurately identifying symptom severity. Concurrently, clinicians face increasing documentation burdens that reduce time available for direct patient care. This study presents a two-phase explainable clinical decision support framework designed to support both PTSD severity assessment and automated clinical documentation. In Phase I, supervised machine learning models were applied to classify PTSD severity using symptom based representations aligned with DSM-5, CAPS-5, and PCL-5 frameworks. Rather than predicting prior PTSD diagnosis, participants were categorized into four clinically informed severity levels (Minimal, Mild, Moderate, and Severe). The learning models were evaluated using both domain-aggregated symptom and individual symptom level feature representations. Experimental results demonstrated that domain-aggregated symptom representations improved both classification performance and interpretability. Explainable artificial intelligence (XAI) techniques, including SHAP and LIME analysis, were conducted to improve inference transparency by identifying influential symptom domain and visualizing model reasoning. Phase II evaluated automated SOAP (Subjective, Objective, Assessment, and Plan) note generation from unstructured mental health narratives using small language models (SLMs). Five instruction-tuned models for locally deployable SOAP-note generation using zero-shot prompting, few-shot prompting, and QLoRA supervised fine-tuning were evaluated using a curated dataset of 1,487 narrative SOAP note pairs, while examining whether medical pretraining consistently benefits the task. The results indicate that task-adapted, quantized small language models can support in-house SOAP drafting, but the external semantic understanding gaps caution against autonomous use. Together, these findings demonstrate how explainable machine learning and small language models can support clinician-centered PTSD workflows by combining transparent severity assessment with structured clinical documentation while maintaining clinician oversight and professional judgment.

Keywords: PTSD, Machine Learning, Severity Classification, Explainable Artificial Intelligence, SOAP Note Generation, Small Language Models, Supervised Fine-Tuning.

Room Location: S.R. Collins Building, Room 111