Improved Sensitivity For Detection Of Clinical Deterioration When Diagnostic Pathology And Patient Trends Are Included In Machine Learning Models

ObjectivesThis study aimed to develop and validate a machine learning model to predict deterioration using Australian hospital data, paying particular attention to the role of predictors not included in current scoring systems. DesignRetrospective cohort study using electronic health records from a large metropolitan health service. SettingGeneral hospital wards, excluding the Emergency Department, Intensive Care Unit, or Palliative Care. ParticipantsInpatients over the age of 18. Main Outcome MeasuresThe primary outcomes of deterioration were mortality and ICU transfer within 24 hours of a newly available observation. A Gradient Boosted Tree model was estimated using patient demographics, vital signs, pathology results, and linear trends. Resulting feature importance was investigated using Shapley values. The model performance was validated against existing scoring systems, including Between the Flags (BTF) and the Modified / National Early Warning Score (MEWS/NEWS). ResultsA Gradient Boosted Tree was developed from 121,608 patients and tested in 20,605 patients. The model, named aWARE, demonstrated higher discriminative ability (AUROCmortality=0.93, AUROCICU transfer=0.84), and calibration when compared to baseline scores. Overall, the 10 most influential features unique between both outcomes were age, oxygen saturation to inspired oxygen ratio, respiratory rate, white cell count, venous lactate, heart rate to systolic blood pressure ratio, albumin, oxygen saturation, urea and heart rate. Of these, only 3 are included in BTF. ConclusionThe machine learning model proposed in this study identified more deteriorating patients and produced less false positive alerts than Between the Flags. Feature importance highlighted the deficit between strong predictors of deterioration and the parameters used in current scoring systems.