A Multimodal Transformer-Based Framework for Early Prediction of In-Hospital Cardiac Arrest Using Sparse Electronic Health Records

Background and Objective: In-hospital cardiac arrest (IHCA) remains a leading cause of preventable mortality, often preceded by subtle physiological changes unrecognized by standard early warning scores. We developed and externally validated CardioAlert, a multimodal transformer-based framework that integrates sparse electronic health record (EHR) data with unstructured nursing notes to predict IHCA up to ۲۴ hours in advance. Methods: This retrospective multicenter study used admissions from two hospitals (development cohort: ۱۵۲,۳۸۶; external validation: ۴۸,۷۱۲). The model fused ۳۸ irregularly sampled dynamic variables (vital signs, labs), ۱۲ static features, and clinical text embeddings from BioBERT via a cross-modal attention mechanism with time-aware masking. Performance was compared against NEWS۲, gradient boosting (XGBoost), LSTM, GRU-D, and a text-free transformer. Primary metric: area under the receiver operating characteristic curve (AUROC). Results: CardioAlert achieved an AUROC of ۰.۹۱۷ (۹۵% CI ۰.۹۰۵–۰.۹۲۹) in external validation, significantly outperforming NEWS۲ (۰.۷۴۲, p<۰.۰۰۱), XGBoost (۰.۸۳۱), LSTM (۰.۸۵۴), GRU-D (۰.۸۶۷), and the text-free transformer (۰.۸۷۲). At a ۱۰% alert rate, sensitivity was ۰.۸۲۲ with a positive predictive value of ۰.۳۱۶, providing a median early warning of ۱۸.۲ hours. The addition of note embeddings improved AUROC by ۰.۰۴۵. Influential features included respiratory rate trends, altered consciousness descriptions, and rising lactate. Conclusion: CardioAlert, a multimodal transformer fusing sparse EHR and clinical notes, accurately predicts IHCA, enabling preemptive interventions.