Dynamically Evolving Clinical Practices and Implications for Predicting Medical Decisions

Automatically data-mining clinical practice patterns from electronic health records (EHR) can enable prediction of future practices as a form of clinical decision support (CDS). Our objective is to determine the stability of learned clinical practice patterns over time and what implication this has when using varying longitudinal historical data sources towards predicting future decisions. We trained an association rule engine for clinical orders (e.g., labs, imaging, medications) using structured inpatient data from a tertiary academic hospital. Comparing top order associations per admission diagnosis from training data in 2009 vs. 2012, we find practice variability from unstable diagnoses with rank biased overlap (RBO)<0.35 (e.g., pneumonia) to stable admissions for planned procedures (e.g., chemotherapy, surgery) with comparatively high RBO>0.6. Predicting admission orders for future (2013) patients with associations trained on recent (2012) vs. older (2009) data improved accuracy evaluated by area under the receiver operating characteristic curve (ROC-AUC) 0.89 to 0.92, precision at ten (positive predictive value of the top ten predictions against actual orders) 30% to 37%, and weighted recall (sensitivity) at ten 2.4% to 13%, (P<10(-10)). Training with more longitudinal data (2009-2012) was no better than only using recent (2012) data. Secular trends in practice patterns likely explain why smaller but more recent training data is more accurate at predicting future practices.