Diagnosing post-traumatic stress disorder using electronic medical record data.

Area of Science:

• Medical Informatics
• Computational Psychiatry
• Primary Care Research

Background:

• Post-Traumatic Stress Disorder (PTSD) diagnosis is challenging in primary care settings.
• Electronic Medical Records (EMRs) contain both structured data and unstructured narrative notes.
• Integrating diverse data sources can improve diagnostic accuracy for complex conditions.

Purpose of the Study:

• To develop and evaluate a predictive model for identifying Post-Traumatic Stress Disorder (PTSD) using EMR data.
• To compare the predictive power of structured data versus unstructured narrative notes in EMRs.
• To assess the performance of mixed-data models in PTSD diagnosis.

Main Methods:

• Utilized EMR data from 154,118 patients in the Manitoba Primary Care Research Network (MaPCReN).
• Developed and tested serial and parallel mixed-data models incorporating structured data and unstructured encounter notes.
• Applied feature engineering and evaluated model performance using sensitivity, F-measure, AUC, and PPV on a skewed test dataset.

Main Results:

• The serial mixed-data model achieved the highest sensitivity (0.77), F-measure (0.76), and AUC (0.88).
• The parallel mixed-data model yielded the highest positive predictive value (PPV) of 0.75.
• Unstructured narrative notes demonstrated higher predictive power than structured data alone and significantly enhanced model capabilities.

Conclusions:

• Combining structured EMR data with unstructured narrative notes substantially improves the accuracy of PTSD identification.
• Unstructured clinical notes are crucial for enhancing machine learning models in diagnosing PTSD within primary care.
• These findings support improved quality improvement, research, and disease surveillance for PTSD.