Weak-Supervision Expectation-Maximization Framework for Identifying Decisional Vulnerability in Older Emergency...

Background and Objectives: Decision-making capacity is essential for informed consent, yet its assessment in emergency departments (EDs) is often subjective and inconsistently documented. Older adults are particularly vulnerable to impaired capacity during acute illness. We aimed to develop a scalable, electronic health record (EHR)-based approach to support early identification of older ED patients at risk of decisional vulnerability using the Medical Information Mart for Intensive Care (MIMIC)-IV database. Methods: We conducted a retrospective cohort study of 51,195 ED patients aged ≥65 years. Clinicians manually reviewed 2000 patients using a conservative consensus protocol to establish a consensus-derived proxy for decisional vulnerability. Such an approach yielded a definitive reference subset (capacity vs. no capacity) and an "uncertain" category when consensus was not achieved. We developed a weak-supervision expectation-maximization (EM) label model that combined multiple noisy labeling functions derived from triage vital signs, acuity measures, arrival mode, and large language model (LLM)-classified chief complaints to estimate the probabilistic risk of impaired capacity. Model discrimination and calibration were assessed on an independent holdout subset of definitive reference labels using receiver operating characteristic area under the curve (ROC-AUC), precision-recall area under the curve (PR-AUC), calibration plots, and Brier score. To support clinically conservative use, operating thresholds were a priori constrained to limit automated flagging to ≤15% of patients, with the remaining ones deferred for clinician review. Results: On the definitive holdout set, the weak-supervision model achieved an ROC-AUC of approximately 0.855 and a PR-AUC of approximately 0.837. Calibration assessment demonstrated residual miscalibration in the generative posterior, which improved after a lightweight discriminative refinement step (logistic regression trained on EM-derived probabilistic labels), reducing the Brier score to approximately 0.224 on holdout evaluation. Under the prespecified operational constraint (≤15% auto-flagged), the model functioned as a conservative, selective alerting strategy, achieving high specificity and positive predictive value while identifying only a minority of patients with decisional vulnerability. Conclusions: This study demonstrates a methodological proof of concept for using weak supervision to model a retrospectively defined proxy for decisional vulnerability from routinely collected ED EHR data. The framework is intended to support conservative, triage-oriented prioritization. Further prospective validation, external testing, and workflow governance are needed before clinical implementation.