DYNAMICALLY EVOLVING CLINICAL PRACTICES AND IMPLICATIONS FOR PREDICTING MEDICAL DECISIONS.

Area of Science:

• Health Informatics
• Clinical Decision Support Systems
• Machine Learning in Healthcare

Background:

• Electronic Health Records (EHR) offer vast data for mining clinical practice patterns.
• Predicting future clinical decisions requires understanding the stability of these patterns over time.
• The optimal source of historical data for predictive models remains an open question.

Purpose of the Study:

• To assess the temporal stability of clinical practice patterns derived from EHR data.
• To evaluate the impact of using different longitudinal historical data sources on predictive accuracy for clinical decision support.

Main Methods:

• An association rule engine was developed using structured inpatient EHR data from a tertiary academic hospital.
• Clinical order patterns (labs, imaging, medications) were compared between 2009 and 2012 training data based on admission diagnosis.
• Predictive model performance was evaluated using future patient data (2013) trained on 2009 vs. 2012 data.

Main Results:

• Clinical practice patterns showed significant variability for unstable diagnoses (Rank Biased Overlap < 0.35) but stability for planned procedures (RBO > 0.6).
• Predicting future orders using recent (2012) training data significantly improved accuracy (ROC-AUC 0.92, Precision@10 37%, Recall@10 13%) compared to older (2009) data.
• Training with more extensive longitudinal data (2009-2012) did not outperform using only recent (2012) data.

Conclusions:

• Clinical practice patterns exhibit temporal instability, particularly for diagnoses, impacting predictive model reliability.
• Recent EHR data is more effective for training predictive models than older or aggregated historical data.
• Secular trends in clinical practice necessitate the use of up-to-date data for accurate clinical decision support.