Sy Hwang

Clinical informatics pipelines increasingly compute validated clinical endpoints from upstream NLP outputs. Even when the endpoint is defined by an established rubric, translating that rubric across representations—natural language instructions, program logic, and reference implementations—can introduce specification drift, where ostensibly equivalent calculators yield meaningfully different scores. We study this phenomenon for the Expanded Disability Status Scale (EDSS), a standard measure of disability in multiple sclerosis. Holding constant a shared set of functional-system (FS) subscores extracted by a large language model (LLM), we compare EDSS values computed across three representations of the same scoring rubric: prompt-executed natural language, LLM-generated code, and a canonical reference implementation. We characterize disagreement structure, distributional shifts, and clinically salient boundary flips, and we propose an audit workflow that treats endpoint computation as a first-class verification target in clinical NLP systems.

We built and evaluated a zero-shot LLM pipeline with automated, task-aware prompt optimization to extract radiology and symptom fields for gallstone phenotyping from de-identified EHR text. Across symptomatic, asymptomatic, and control cohorts, it performed reliably on high-signal binary fields and symptom flags but lagged on fine-grained stone burden and complications, establishing a practical baseline and motivating targeted refinements

Background: The use of artificial intelligence (AI) to analyze health care data has become common in behavioral health sciences. However, the lack of training opportunities for mental health professionals limits clinicians’ ability to adopt AI in clinical settings. AI education is essential for trainees, equipping them with the literacy needed to implement AI tools in practice, collaborate effectively with data scientists, and develop skills as interdisciplinary researchers with computing skills. Objective: As part of the Penn Innovation in Suicide Prevention Implementation Research Center, we developed, implemented, and evaluated a virtual workshop to educate psychiatry and psychology trainees on using AI for suicide prevention research. Methods: The workshop introduced trainees to natural language processing (NLP) concepts and Python coding skills using Jupyter notebooks within a secure Microsoft Azure Databricks cloud computing and analytics environment. We designed a 3-hour workshop that covered 4 key NLP topics: data characterization, data standardization, concept extraction, and statistical analysis. To demonstrate real-world applications, we processed chief complaints from electronic health records to compare the prevalence of suicide-related encounters across populations by race, ethnicity, and age. Training materials were developed based on standard NLP techniques and domain-specific tasks, such as preprocessing psychiatry-related acronyms. Two researchers drafted and demonstrated the code, incorporating feedback from the Methods Core of the Innovation in Suicide Prevention Implementation Research to refine the materials. To evaluate the effectiveness of the workshop, we used the Kirkpatrick program evaluation model, focusing on participants’ reactions (level 1) and learning outcomes (level 2). Confidence changes in knowledge and skills before and after the workshop were assessed using paired t tests, and open-ended questions were included to gather feedback for future improvements. Results: A total of 10 trainees participated in the workshop virtually, including residents, postdoctoral researchers, and graduate students from the psychiatry and psychology departments. The participants found the workshop helpful (mean 3.17 on a scale of 1‐4, SD 0.41). Their overall confidence in NLP knowledge significantly increased ( P =.002) from 1.35 (SD 0.47) to 2.79 (SD 0.46). Confidence in coding abilities also improved significantly ( P =.01), increasing from 1.33 (SD 0.60) to 2.25 (SD 0.42). Open-ended feedback suggested incorporating thematic analysis and exploring additional datasets for future workshops. Conclusions: This study illustrates the effectiveness of a tailored data science workshop for trainees in psychiatry and psychology, focusing on applying NLP techniques for suicide prevention research. The workshop significantly enhanced participants’ confidence in conducting data science research. Future workshops will cover additional topics of interest, such as working with large language models, thematic analysis, diverse datasets, and multifaceted outcomes. This includes examining how participants’ learning impacts their practice and research, as well as assessing knowledge and skills beyond self-reported confidence through methods such as case studies for deeper insights.

Background: The safety planning intervention (SPI) is a suicide prevention intervention that results in a written plan to help patients reduce suicide risk. High-quality safety plans—that is, those that are the most complete, personalized, and specific—are more effective in reducing suicide risk. Measuring SPI quality is labor-intensive, which means that clinicians rarely get specific, actionable feedback on their use of the SPI. Objective: This study aimed to develop the Safety Plan Fidelity Rater, an automated tool that assesses the quality of written safety plans leveraging 3 large language models (LLMs)—GPT-4, LLaMA 3, and o3-mini. Methods: Using 266 deidentified safety plans from outpatient mental health settings in New York, LLMs analyzed four key steps: warning signs, internal coping strategies, making the environment safe, and reasons for living. We compared the predictive performance of the three LLMs, optimizing scoring systems, prompts, and parameters. Results: Findings showed that LLaMA 3 and o3-mini outperformed GPT-4, with different step-specific scoring systems recommended based on weighted F ₁ -scores. Conclusions: These findings highlight LLMs’ potential to provide clinicians with timely and accurate feedback on safety plan quality, which could greatly improve its implementation in community practice.

This feasibility study assesses using Large Language Models (LLMs) to automate the extraction of Multiple Sclerosis (MS) progression metrics from unstructured clinical notes. While the Expanded Disability Status Scale (EDSS) is the standard for measuring disability, calculating it during routine care is burdensome, often leaving progression details locked in narrative text. Using OpenAI’s GPT-4.1 in a zero-shot setting, the study processed 288 progress notes from 239 patients. The model extracted scores for eight Functional Systems (FS) and derived a total EDSS score. Performance was evaluated against expert annotations using a strict and lenient agreement criteria. Results showed strong performance for FS extraction, with accuracy ranging from 0.782 to 0.971 across domains. While exact EDSS agreement was variable due to documentation ambiguity, applying clinically meaningful tolerance bands yielded high consistency, with most disagreements being minor adjacent errors unlikely to alter therapy. The study confirms that LLMs can feasibly derive detailed disability measures, potentially reducing documentation burden enabling the automated, population-wide tracking of MS progression over time.

Ascertainment of multiple sclerosis (MS) progression is important for informing clinical care decisions and supporting biomedical research. However, the details to infer a patient's MS progression status are locked within clinical notes. In this feasibility study, we assessed the feasibility of developing and validating a large language model (LLM)-based EDSS and FS classifier for ascertaining MS progression from clinical notes.

Early detection of cognitive impairment (CI) is critical for timely intervention in Alzheimer’s disease and AD-related dementias. To address this, we propose the Warning Assessment and Alerting Tool for Cognitive Health from Spontaneous Speech (WATCH-SS), a modular and explainable three-stage framework for detecting CI from a patient’s speech sample. The framework uses detectors for five linguistic and acoustic indicators of CI, aggregates their outputs into a set of clinically interpretable summary features, and uses a predictive model for CI classification. We consider multiple approaches to implementing these detectors that range from simple, computationally efficient methods suitable for real-time analysis to strong, resource-intensive methods, better for high accuracy offine analysis. On the DementiaBank ADReSS dataset, WATCH-SS achieved strong predictive performance (AUC = 80% on the test set). This work demonstrates that a modular, feature-based approach can achieve strong performance while providing a transparent diagnostic profile, representing a significant step towards a trustworthy and clinically-usable screening tool for primary care.

In-hospital cardiac arrest (IHCA) affects over 200,000 patients annually in the United States, yet its detection through manual chart review remains resource-intensive and often delayed. We evaluated the performance of four open-source large language models (LLMs) and GPT-4o in identifying IHCA cases from 2,674 clinical notes across five hospitals. While GPT-4o achieved the highest performance (F1-score: 0.90, recall: 0.97), several open-source models demonstrated comparable capabilities, suggesting their viability for clinical applications. Our systematic analysis of model outputs revealed that performance was strongly influenced by site-specific documentation practices, with inter-site agreement rates varying by over 20%. Through detailed error analysis, we identified key challenges including medical terminology hallucinations and structural inconsistencies in model reasoning. These findings establish a framework for implementing LLM-based IHCA detection systems while highlighting critical considerations for their clinical deployment.

The Safety Planning Intervention (SPI) produces a plan to help manage patients’ suicide risk. High-quality safety plans – that is, those with greater fidelity to the original program model – are more effective in reducing suicide risk. We developed the Safety Planning Intervention Fidelity Rater (SPIFR), an automated tool that assesses the quality of SPI using three large language models (LLMs)—GPT-4, LLaMA 3, and o3-mini. Using 266 deidentified SPI from outpatient mental health settings in New York, LLMs analyzed four key steps: warning signs, internal coping strategies, making environments safe, and reasons for living. We compared the predictive performance of the three LLMs, optimizing scoring systems, prompts, and parameters. Results showed that LLaMA 3 and o3-mini outperformed GPT-4, with different step-specific scoring systems recommended based on weighted F1-scores. These findings highlight LLMs’ potential to provide clinicians with timely and accurate feedback on SPI practices, enhancing this evidence-based suicide prevention strategy.