[Paper] Early Warning Index for Patient Deteriorations in Hospitals

Source: arXiv - 2512.14683v1

Overview

The paper presents Early Warning Index (EWI), a multimodal machine‑learning system that continuously evaluates a patient’s risk of serious deterioration—ICU transfer, emergency response team activation, or death—by pulling together structured and unstructured data from a hospital’s electronic health record (EHR). By embedding clinicians in the loop and using SHAP‑based explanations, the authors turn a “black‑box” model into a practical triage dashboard that is already being used at a large U.S. hospital.

Key Contributions

• Multimodal risk model that fuses bedside vitals, lab results, medication orders, scheduled surgeries, and operational metrics (e.g., ward census) into a single early‑warning score.
• Human‑in‑the‑loop design: clinicians set alert thresholds and co‑interpret model outputs, ensuring the system aligns with real‑world workflow.
• Explainability via SHAP: each patient’s risk score is broken down into contributing factors, making the predictions transparent to physicians and administrators.
• Scalable feature extraction from both structured tables and free‑text clinical notes, eliminating the need for manual data wrangling.
• Live deployment in a hospital dashboard that stratifies patients into three risk tiers, currently supporting proactive care management.
• Strong predictive performance on a retrospective cohort of 18,633 patients (C‑statistic ≈ 0.80) while maintaining interpretability.

Methodology

1. Data Ingestion – The pipeline pulls real‑time feeds from the EHR (vitals, labs, medication orders) and operational systems (surgery schedules, bed occupancy). Unstructured notes are processed with a lightweight NLP tokenizer to extract key concepts (e.g., “shortness of breath”).
2. Feature Engineering – Time‑window aggregations (e.g., rolling averages of heart rate), categorical encodings (e.g., surgery type), and interaction terms (e.g., “high census × post‑op status”) are automatically generated.
3. Model Architecture – A gradient‑boosted decision tree (XGBoost) is trained to predict a binary label representing any of the three adverse events within the next 24 hours. The model is calibrated to output a probability that serves as the EWI score.
4. Human‑in‑the‑Loop Calibration – Clinicians review a validation set, adjust the decision threshold that maps probabilities to the three risk tiers (low, medium, high), and provide feedback on false positives/negatives.
5. Explainability Layer – SHAP values are computed for each prediction, highlighting the top clinical and operational drivers (e.g., rising lactate, upcoming surgery, high ward census). These explanations are displayed directly on the dashboard.
6. Evaluation – The model is assessed on a hold‑out test set using AUROC (C‑statistic), precision‑recall, and calibration plots. A prospective pilot measures alert adoption and time saved for physicians.

Results & Findings

The model reliably distinguishes patients who later required ICU care or experienced a rapid response event, while the SHAP explanations align with clinicians’ intuition (e.g., postoperative status, abnormal labs). The dashboard’s three‑tier stratification helped care teams focus early interventions on the “high” group, reducing unnecessary alerts and improving trust.

Practical Implications

• Proactive Triage – Hospitals can embed EWI into existing patient‑monitoring dashboards to automatically surface high‑risk patients, freeing clinicians from manual chart reviews.
• Resource Allocation – By surfacing operational drivers (e.g., high ward census), administrators can adjust staffing or postpone elective surgeries to mitigate systemic risk.
• Improved Patient Flow – Early identification of deterioration reduces unexpected ICU admissions, smoothing bed turnover and potentially lowering readmission rates.
• Explainable AI Adoption – The SHAP‑based UI demonstrates a viable path for integrating interpretable ML into regulated healthcare settings, addressing compliance and clinician trust concerns.
• Scalable to Other Institutions – The multimodal pipeline is built on standard HL7/FHIR feeds and generic NLP, making it portable to hospitals with different EHR vendors.

Limitations & Future Work

• Single‑site validation – Results are based on data from one large U.S. hospital; external validation across diverse health systems is needed to confirm generalizability.
• Temporal drift – Model performance may degrade as clinical protocols or patient populations change; continuous monitoring and periodic retraining are required.
• Unstructured data depth – The current NLP component extracts only high‑level concepts; richer language models could capture subtler clinical nuances.
• Alert fatigue risk – Although tiered alerts reduce noise, the optimal threshold balance may differ across units; future work will explore adaptive, unit‑specific thresholds.
• Integration with downstream actions – The study stops at risk stratification; linking EWI alerts to automated care pathways (e.g., order sets, rapid response team activation) is a logical next step.

Authors

• Dimitris Bertsimas
• Yu Ma
• Kimberly Villalobos Carballo
• Gagan Singh
• Michal Laskowski
• Jeff Mather
• Dan Kombert
• Howard Haronian

Paper Information

• arXiv ID: 2512.14683v1
• Categories: cs.LG
• Published: December 16, 2025
• PDF: Download PDF