Structured-to-Narrative Representation Framework for Next-Day Discharge Prediction Using Structured EHR Data and Clinical Language Models

<title>Abstract</title> <p>Objective Accurate prediction of next-day hospital discharge after elective spine surgery is critical for optimizing inpatient operations and perioperative care coordination. This study introduces a structured-to-narrative modeling framework that transforms high-dimensional structured electronic health record data into clinically interpretable textual representations for discharge prediction. Materials and Methods We analyzed structured electronic health record data from 1,958 patients undergoing elective spine surgery, comprising approximately 600 binary and categorical variables spanning demographics, laboratory, diagnosis, and procedure data. We evaluated machine learning models and a fine-tuned domain-specific transformer trained exclusively on structured-to-narrative representations derived from these variables. Models were assessed under strict patient-level separation using five-fold cross-validation. Discrimination, calibration, and class-specific performance were evaluated using AUROC, PRAUC, Brier score, precision, recall, and F1-score, with interpretability examined using SHAP and narrative-level probability analysis. Results Ensemble machine learning models achieved strong discrimination and calibration, with a maximum AUROC of 0.94. The narrative-centered transformer achieved competitive performance with a mean AUROC of 0.84, PRAUC of 0.57, and Brier score of 0.17. Threshold optimization shifted the transformer toward a sensitivity-focused prediction behavior, achieving a mean recall of 0.75. SHAP analysis and narrative-level probability inspection demonstrated clinically coherent attribution aligned with procedural burden, age, and physiologic instability. Discussion While ensemble models achieved superior discrimination, the narrative-centered transformer provides a language-native, representation-driven alternative that supports clinically interpretable reasoning directly from structured data. Conclusion Structured-to-narrative modeling offers a scalable and interpretable pathway for integrating predictive modeling into real-world perioperative decision support.</p>