Predicting ward transfer mortality with machine learning.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Predictive Analytics

Background:

• Internal medicine physicians face challenges in identifying patients at risk of increased mortality.
• Accurate risk stratification is crucial for optimizing patient care and resource allocation.

Purpose of the Study:

• To develop and evaluate artificial intelligence (AI) models for predicting patient mortality risk.
• To identify key clinical variables for mortality prediction in patients transferred from non-ICU to ICU settings.

Main Methods:

• Extracted data from 2,425 patient records from the Veteran Affairs Corporate Data Warehouse (CDW).
• Created two datasets: one with 22 variables, another with 20 variables (excluding admission-unknown factors).
• Trained and evaluated 16 machine learning models, focusing on the LightGBM algorithm.

Main Results:

• The LightGBM model demonstrated high performance on both datasets (ROC-AUC up to 0.89).
• The model using 20 clinically relevant variables achieved a ROC-AUC of 0.86 and accuracy of 0.71.
• Key predictors included lab values (lymphocyte, hemoglobin) and transfer time variables.

Conclusions:

• A clinically relevant AI model can effectively predict patient mortality risk.
• This tool can assist providers in optimizing resource utilization and managing patient caseloads.
• The model's insights are particularly valuable during critical care transitions and shift changes.