Modeling physician variability to prioritize relevant medical record information

Area of Science:

• Machine Learning in Healthcare
• Clinical Informatics
• Data Science in Medicine

Background:

• Electronic Medical Record (EMR) systems aim for efficient patient information retrieval.
• Physician information-seeking behavior varies, impacting EMR usability.
• Machine learning can predict physician information needs in clinical contexts.

Purpose of the Study:

• To develop and compare hierarchical and non-hierarchical machine learning models for identifying relevant patient information in EMRs.
• To explicitly account for inter-physician variability in information-seeking behavior.
• To enhance the efficiency of patient data retrieval in intensive care unit (ICU) settings.

Main Methods:

• Critical care physicians identified relevant data items from ICU patient cases for morning rounds.
• Hierarchical logistic regression (HLR) and standard logistic regression (LR) models were derived.
• EMR data served as predictors to assess model performance in identifying relevant information.

Main Results:

• HLR models demonstrated a statistically significant higher area under the receiver operating characteristic curve (0.81) compared to LR models (0.75).
• HLR models achieved significantly lower expected calibration error (0.07) than LR models (0.16).
• Physician variability in data selection was observed and explicitly modeled by HLR.

Conclusions:

• Hierarchical models significantly outperform standard logistic regression in both discrimination and calibration for EMR information retrieval.
• Explicitly modeling physician-related variability enhances the performance of predictive models.
• Hierarchical models offer a superior approach for identifying relevant patient information in EMRs, especially in diverse clinical settings like ICUs.