Utilization of Data Mining for Generalizable, All-Admission Prediction of Inpatient Mortality

The all-condition prediction of patient mortality at the time of hospital admission has significant clinical value and broader implications for patient care and clinical decision support capabilities. In this study we have applied machine learning models to predict inpatient mortality, that is whether a patient will die during the hospital stay, as predicted from a time near to admission. We have utilized an Agency for Healthcare Research and Quality-provided large dataset of hospital discharges, to develop and evaluate a number of machine learning models. We report on the performance of the best performing of these models, with the best performing model having an AUC score of 0.802. We also evaluate the generalizability of the models via evaluating these on a separate large dataset corresponding to a different time period. We describe the results and provide an analysis and discussion of their significance.