Investigating the potential for machine learning prediction of patient outcomes: a retrospective study of hospital acquired pressure injuries

Abstract

AbstractBackgroundWhile recent research efforts to reduce pressure ulcers in the clinical context have focused on key retrospective characteristics, little work has focused on creating real-time predictive models to prevent this avoidable hospital-acquired injury. Furthermore, existing machine learning heuristics often fail to surpass traditional statistical models or provide individual-level risk assessments with explanations for each patient. Thus, we sought to compare the predictive performance of five machine learning and traditional statistical modeling techniques to predict the occurrence of Hospital Acquired Pressure Injuries (HAPI).MethodsElectronic Medical Record (EMR) information was collected from 57,227 hospitalizations, containing 241 positive HAPI cases, acquired from Dartmouth Hitchcock Medical Center from April 2011 to December 2016. The five classifiers were trained to predict HAPI incidence and performance was assessed using the C-statistic or Area Under the Receiver Operating Curve (AUC).ResultsLogistic Regression was the best modeling approach (AUC=0.91±0.034). We report discordance between predictors deemed important by the machine learning models compared to traditional statistical model. We provide means to visually assess factors important to every patient’s prediction, regardless of the modeling approach, through Shapley Additive Explanations.ConclusionsMachine learning models will continue to inform decision making processes but should be compared to traditional modeling approaches to ensure proper utilization. Disagreements between important predictors found by traditional and machine learning modeling approaches can potentially confuse clinicians and as such need to be reconciled. Future efforts to analyze time-stamped, prospective medical record data will be enhanced by patient-specific details. These developments represent important steps forward in developing real-time predictive models that can be integrated and readily deployed in electronic medical record systems to reduce unnecessary harm.