Improving Machine Learning 30-Day Mortality Prediction by Discounting Surprising Deaths.

Area of Science:

• Emergency Medicine
• Machine Learning
• Health Informatics

Background:

• Machine learning (ML) is explored for predicting end-of-life care needs using mortality as a proxy.
• Deaths unforeseen by emergency physicians may have a weaker association with the emergency department (ED) visit.

Purpose of the Study:

• To develop ML algorithms for predicting unsurprising deaths within 30 days post-ED discharge.
• To assess the performance of ML models in predicting predictable 30-day mortality.

Main Methods:

• Retrospective registry study of ED attendances in Halland, Sweden (2015-2016).
• Deaths within 30 days were classified as surprising or unsurprising by senior emergency medicine specialists.
• Logistic Regression (LR), Random Forest (RF), and Support Vector Machine (SVM) models were developed.

Main Results:

• 76% of 148 total 30-day deaths were deemed unsurprising.
• ML models achieved high predictive performance for unsurprising deaths (ROC-AUCs ranging from 0.944 to 0.950).
• Prediction performance for unsurprising deaths was significantly better than for all mortality (P < .001).

Conclusions:

• A significant majority of 30-day deaths in patients discharged from the ED were not surprising.
• ML models demonstrated significantly improved mortality prediction when focused on unsurprising deaths.
• This approach can enhance the identification of patients for end-of-life care discussions.