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

Area of Science:

• Geriatric Medicine
• Health Informatics
• Artificial Intelligence in Healthcare

Background:

• Assessing decision-making capacity in emergency departments (EDs) is crucial for informed consent but often subjective and inconsistently documented.
• Older adults are particularly susceptible to impaired capacity during acute illness, necessitating improved identification methods.
• Existing methods lack scalability and consistent documentation within electronic health records (EHRs).

Purpose of the Study:

• To develop a scalable, EHR-based approach for early identification of older ED patients at risk of decisional vulnerability.
• To utilize the Medical Information Mart for Intensive Care (MIMIC)-IV database for developing and validating the identification model.
• To support conservative, triage-oriented prioritization of older adults with potential decisional impairment.

Main Methods:

• Retrospective cohort study of 51,195 older adult ED patients (≥65 years) using the MIMIC-IV database.
• Development of a weak-supervision expectation-maximization (EM) label model integrating noisy labeling functions from vital signs, acuity, arrival mode, and LLM-classified chief complaints.
• Model performance evaluated on a consensus-derived reference subset using ROC-AUC, PR-AUC, and Brier score, with operating thresholds limited to ≤15% auto-flagging.

Main Results:

• The weak-supervision model achieved an ROC-AUC of ~0.855 and PR-AUC of ~0.837 on the definitive holdout set.
• A discriminative refinement step improved model calibration, reducing the Brier score to ~0.224.
• The model functioned as a conservative alerting strategy, achieving high specificity and positive predictive value under the ≤15% auto-flagging constraint.

Conclusions:

• Demonstrated a methodological proof of concept for using weak supervision to model decisional vulnerability from routine ED EHR data.
• The developed framework supports conservative, triage-oriented prioritization for older adults at risk.
• Prospective validation, external testing, and workflow integration are required prior to clinical implementation.