Leveraging large language models to predict hospital readmissions from free-text clinical notes at Jefferson Health

PI: Meng Zhao

Co-PI: Mooi Choo Chuah 

University: Lehigh University

Industry partner: Jefferson Health 

Jefferson Health is one of the largest and most highly ranked healthcare systems in Pennsylvania. Headquartered in Philadelphia, it serves as the clinical arm of Thomas Jefferson University. Following recent mergers with Einstein Healthcare Network and Lehigh Valley Health Network (LVHN), it now provides care across a massive footprint spanning from Philadelphia and its suburbs into the Lehigh Valley and Northeastern Pennsylvania. While Jefferson Health already captures rich patient data in their electronic health record (EHR) system, many of the strongest signals of readmission risk are embedded in free-text clinical notes including evolving symptoms, functional status, discharge readiness, and nuanced clinician assessments that do not reliably appear in structured EHR fields. Recently, large language models (LLMs) offer a powerful way to extract meaning from free-text clinical notes. LLMs can summarize, classify, and infer latent risk factors described in clinical notes. However, using LLMs directly on clinical notes requires strong privacy protection. Jefferson Health, like all covered entities, must ensure that protected health information (PHI) is not exposed in research, model development, or downstream analytics artifacts. Though de-identification tools exist, many are not optimized for local documentation conventions, nonstandard abbreviations, templated text, or the “long tail” of identifiers (e.g., family member names, rare locations, embedded phone numbers, etc.). A robust, locally validated de-identification capability is therefore a prerequisite to responsibly leveraging free-text notes for model development and eventually operational deployment.

This proposal addresses both needs: privacy-preserving text access and improved readmission prediction through two integrated aims: 1) Develop a de-identification tool to remove PHI from free-text clinical notes using a hybrid approach combining rule-based and statistical NLP methods, tune and validate on Jefferson Health clinical notes and 2) Develop and evaluate LLM models that use de-identified clinical notes to predict readmissions with strong calibration and interpretable outputs that are suitable for integration into clinical workflows.