Prospective and External Validation of Machine Learning Models for Short- and Long-Term Mortality in Acutely Admitted...

Area of Science:

• Biomedical Informatics
• Clinical Prediction Models
• Machine Learning in Healthcare

Background:

• Emergency department (ED) mortality prediction faces challenges in balancing model simplicity and performance.
• Developing effective prognostic models for ED patients is crucial for timely intervention.
• Utilizing minimal data from a single blood sample offers a practical approach for prediction.

Purpose of the Study:

• To develop simple yet effective machine learning models for predicting short- and long-term mortality in ED patients.
• To evaluate the predictive power of routine clinical biochemistry from a single blood sample upon admission.
• To assess model performance across various mortality time points (10, 30, 90, 365 days).

Main Methods:

• Analysis of three cohorts (1 retrospective, 2 prospective) from Danish university hospitals (2013-2022).
• Development of prediction models using Light Gradient Boosting Machines based on routine blood biochemistry at ED admission.
• Evaluation using metrics including Area Under the Receiver Operating Characteristic Curve (AUC), sensitivity, specificity, and Matthews correlation coefficient (MCC).

Main Results:

• Analysis included 43,648 unique patients and 65,484 admissions.
• Machine learning models demonstrated high accuracy with AUC values ranging from 0.87 to 0.93.
• Excellent predictive performance was observed across different short- and long-term mortality intervals.

Conclusions:

• Routine clinical biochemistry from a single blood sample at ED admission is a powerful predictor of mortality.
• Simple, data-efficient models can achieve high accuracy in predicting both short- and long-term mortality.
• This approach offers a feasible method for risk stratification in emergency care settings.