Predicting acute clinical deterioration with interpretable machine learning to support emergency care decision making

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Decision Support

Background:

• Emergency departments (EDs) face increasing operational pressures.
• Efficiently identifying patients at imminent risk of acute deterioration is crucial.
• Current methods like the National Early Warning Score 2 (NEWS2) have limitations.

Purpose of the Study:

• To systematically compare machine learning algorithms (logistic regression, gradient boosted decision trees, support vector machines) for predicting imminent clinical deterioration.
• To evaluate model performance against NEWS2 using real-world patient data.
• To incorporate interpretable and fairness-aware machine learning techniques.

Main Methods:

• Utilized a dataset of 118,886 unplanned admissions from Salford Royal Hospital, UK.
• Applied machine learning models to cross-sectional patient data from electronic patient records (EPR) at hospital entry.
• Measured clinical deterioration by in-hospital mortality and/or critical care admission within 24 hours.
• Employed Shapely Additive exPlanations (SHAP) for model interpretability and fairness assessment.

Main Results:

• Machine learning models demonstrated superior performance compared to NEWS2, with up to a 0.366 increase in average precision.
• Achieved significant reductions in the daily alert rate (up to a [Formula: see text] decrease).
• Showcased a median 0.599 reduction in differential bias amplification across age and sex demographics.
• SHAP analysis confirmed alignment of model predictions with clinical domain knowledge.

Conclusions:

• Machine learning models offer a promising advancement for predicting patient deterioration in the ED.
• These models have the potential to reduce alert fatigue and identify high-risk patients missed by current scores.
• Further clinical trials are necessary to integrate these data-driven risk modeling tools into practice.