Early Identification of End‑of‑Life Risk at Emergency Department Triage Using Routine Electronic Health Record Data: Development, Validation, and Subgroup

Abstract

Singapore’s rapid population aging has increased emergency department (ED) visits by older adults with multimorbidity. Early end-of-life (EoL) identification at ED presentation may enable timely goals-of-care discussions and palliative referral, yet pragmatic tools are limited. We conducted a retrospective cohort study using electronic health records (EHR) from Singapore General Hospital ED (2008-2020). We defined EoL patients as those died within 30 days after the index ED visit (excluding trauma presentations and/or deaths during the ED encounter). We retained the most recent visit per patient, yielding 550,897 Singaporean adults; 27,837 (5.1%) being EoL patients. Using routinely available ED-arrival predictors (demographics, comorbidities, prior 1-month healthcare utilization, shift time, triage acuity, and visit year), we developed three models: LASSO, XGBoost, and AutoScore and conducted temporal validation (2008-2016 training period, 2017-2020 validation period). We evaluated discrimination (AUC) and calibration (O/E ratio) for each model overall and across demographic subgroups (i.e. age, gender, ethnicity). As ED deployment requires minimizing false positives, we also evaluated specificity. LASSO and XGBoost showed excellent AUC( 0.937 and 0.938), while AutoScore had a lower AUC (0.776) but greater interpretability. Recalibration improved all models’ O/E ratios to 1.00 (CI: 0.978-1.022). Across all demographic subgroups, XGBoost had the most stable AUC and LASSO demonstrated more balanced specificity. Nevertheless, performance declined in the oldest-old across all models. In summary, an EHR-based prediction model using ED arrival data can reliably identify potential EoL patients. Future work should externally validate the model for generalizability and transportability, ensure ongoing monitoring for bias, and conduct periodic recalibration.