| Popis: |
BackgroundPredicting mortality and morbidity amongst hospitalized patients has long been a struggle for inpatient Internal Medicine physicians. To prevent physician burnout, hospital organizations are turning to shift work for the care of hospitalized patients. Such shift work frequently leads to a handoff of patients, with many physicians often being the sole provider in the hospital for close to one hundred patients.ObjectivesWe propose developing an artificial intelligence model that helps predict which patients will be most at risk of increased mortality. This model would assist providers and frontline staff in focusing their efforts on improving patient outcomes.Materials and MethodsRecords of patients who were transferred from non-intensive care units to intensive care units were queried from the Veteran Affairs Corporate Data Warehouse (CDW). Two thousand four hundred twenty-five records were identified. The patient outcome was designated a dependent variable, with bad outcome defined as the patient dying within 30 days of admission and good outcome as the patient being alive within 30 days of admission. Using twenty-two independent variables, we trained sixteen machine learning models, of which six best-performing ones were fine-tuned and evaluated on the testing dataset. Finally, we repeated this process with twenty independent variables, omitting the Length of Stay and Days to Intensive Care Unit Transfer variables which are unknown at the time of admission.ResultsThe best results were obtained with the LightGBM model with both datasets, one that included Length of Stay and Days to Intensive Care Unit Transfer variables and the other without these two variables. The former achieved Receiver Operating Characteristics Curve - Area Under the Curve (ROC-AUC) of 0.89, an accuracy of 0.72, a sensitivity of 0.97, and a specificity of 0.68, while the latter achieved a ROC-AUC of 0.86, an accuracy of 0.71, sensitivity of 0.94 and specificity of 0.67 respectively.ConclusionsOur predictive mortality model may offer providers a means for optimizing the utilization of resources when managing a large caseload, especially with shift changes. |
