Algorithmic Identification of Potentially High Risk Abdominal Presentations (PHRAPs) to the Emergency Department: A Clinically-Oriented Machine Learning Approach

BackgroundOlder adults presenting to emergency departments (EDs) with abdominal pain have been shown to be at high risk of subsequent morbidity and mortality. Yet, such presentations are poorly studied in national databases. Claims databases do not record the patients symptoms at the time of presentation to the ED, but rather the diagnosis after testing and evaluation, limiting study of care and outcomes for these high risk abdominal presentations. ObjectivesWe sought to develop an algorithm to define a patient population with potentially high risk abdominal presentations (PHRAPs) using only variables commonly available in claims data. Research DesignTrain a machine learning model to predict abdominal pain chief complaints using the National Hospital Ambulatory Medical Care Survey (NHAMCS), a nationally-representative database of abstracted ED medical records. SubjectsAll patients contained in NHAMCS data from 2013-2018. 2013-2017 were used for predictive modeling and 2018 was used as a hold-out test set. MeasuresPositive predictive value and sensitivity of the predictive algorithm against a hold-out test set of NHAMCS patients the algorithm was blinded to during training. Predictions were assessed for agreement with either a chief complaint of abdominal pain (contained in "Reason for Visit 1"), or an expanded definition intended to capture visits which were for abdominal concerns. These included secondary or tertiary complaints of abdominal pain or other abdominal conditions, other abdominal-related chief complaint (e.g. nausea or diarrhea, but not pain), discharge diagnosis of an abdominal condition, or reception of an abdominal CT or ultrasound. ResultsAfter validation on a hold-out data set, a gradient boosting machine (GBM) was the best best-performing machine learning model, but a logistic regression model had similar performance and may be more explainable and useful to future researchers. The GBM predicted a chief complaint of abdominal pain with a positive predictive value of 0.60 (95% CI of 0.56, 0.64) and a sensitivity of 0.29 (95% CI of (0.27, 0.32). Nearly all false positives still exhibited signs of "abdominal concerns" for patients: using the expanded definition of "abdominal concern" the model had a PPV of >0.99 (95% CI of 0.99, 1.00) and sensitivity of 0.12 (95% CI of 0.11, 0.13). ConclusionThe algorithm we report defines a patient population with abdominal concerns for further study of treatment and outcomes to inform the development of clinical pathways.