Machine learning based early mortality prediction in the emergency department.

Area of Science:

• Emergency Medicine
• Artificial Intelligence
• Clinical Informatics

Background:

• Early detection of patient deterioration in emergency departments (ED) is critical for preventing unexpected deaths.
• Managing large volumes of clinical data requires significant physician experience and insight.
• Developing objective tools to aid in risk stratification is essential for emergency care.

Purpose of the Study:

• To evaluate machine learning (ML) models for quantifying the severity of emergency department (ED) patients.
• To identify high-risk patients in the ED using ML-driven mortality prediction.
• To assess the performance of ML models based on routinely available clinical data.

Main Methods:

• A framework utilizing ML and feature engineering was developed for mortality prediction using electronic health records (EHRs).
• Demographics, vital signs, and laboratory tests were extracted from EHRs for 1114 ED patients.
• Nine ML models were trained and validated using 5-fold cross-validation, with hyper-parameters optimized via grid search.

Main Results:

• The LightGBM model, using the entire patient stay record, achieved high performance: 93.6% accuracy, 97.6% AUC, 97.1% recall, and 94.2% precision.
• Prediction performance improved with a more complete time window of available patient data.
• The model demonstrated strong predictive capability without utilizing diagnostic information.

Conclusions:

• The developed ML framework effectively quantifies ED patient criticality and identifies high-risk individuals.
• This approach shows significant potential as a clinical decision support tool for emergency physicians.
• The methodology can be adapted for implementation in other hospitals after validation.