Predictive structured-unstructured interactions in EHR models: A case study of suicide prediction.

Area of Science:

• Clinical Informatics
• Computational Psychiatry
• Machine Learning in Healthcare

Background:

• Electronic health records (EHRs) are increasingly used for clinical risk prediction.
• Rising suicide rates necessitate improved prediction and prevention strategies.
• Understanding the predictive value of different EHR data types is crucial.

Purpose of the Study:

• To compare the predictive performance of structured and unstructured EHR data for suicide risk.
• To evaluate machine learning models (Naive Bayes Classifier and Random Forest) using these data types.
• To develop a framework for identifying feature interactions that enhance predictive accuracy.

Main Methods:

• Trained Naive Bayes Classifier (NBC) and Random Forest (RF) models on structured EHR data, unstructured EHR data, and combined data.
• Compared model performance using Area Under the Curve (AUC) and statistical significance.
• Developed and applied a framework to identify significant structured-unstructured feature pair interactions.

Main Results:

• RF models trained on combined data (AUC=0.903) significantly outperformed those trained on structured data alone (AUC=0.887).
• NBC models showed similar performance whether trained on structured data alone (AUC=0.742) or combined data (AUC=0.743).
• Identified feature pairs capturing heterogeneous general concepts improved predictive performance and clinical understanding.

Conclusions:

• Combining structured and unstructured EHR data, particularly with RF models, enhances suicide risk prediction.
• The developed framework effectively identifies informative feature interactions for clinical risk modeling.
• Findings support the integration of diverse EHR data types and advanced modeling techniques for improved patient care.