Abstract
In order to address a long standing challenge for internal medicine physicians we developed artificial intelligence (AI) models to identify patients at risk of increased mortality. After querying 2,425 records of patients transferred from non-intensive care units to intensive care units from the Veteran Affairs Corporate Data Warehouse (CDW), we created two datasets. The former used 22 independent variables that included "Length of Hospital Stay" and "Days to Intensive Care Transfer," and the latter lacked these two variables. Since these two variables are unknown at the time of admission, the second set is more clinically relevant. We trained 16 machine learning models using both datasets. The best-performing models were fine-tuned and evaluated. The LightGBM model achieved the best results for both datasets. The model trained with 22 variables achieved a Receiver Operating Characteristics Curve-Area Under the Curve (ROC-AUC) of 0.89 and an accuracy of 0.72, with a sensitivity of 0.97 and a specificity of 0.68. The model trained with 20 variables achieved a ROC-AUC of 0.86 and an accuracy of 0.71, with a sensitivity of 0.94 and a specificity of 0.67. The top features for the former model included "Total length of Stay," "Admit to ICU Transfer Days," and "Lymphocyte Next Lab Value." For the latter model, the top features included "Lymphocyte First Lab Value," "Hemoglobin First Lab Value," and "Hemoglobin Next Lab Value." Our clinically relevant predictive mortality model can assist providers in optimizing resource utilization when managing large caseloads, particularly during shift changes.