A machine learning model using clinical notes to estimate PHQ-9 symptom severity scores in depressed patients

Background: Lack of widespread use of the Patient Health Questionnaire 9-item (PHQ-9) in clinical practice inhibits measurement of treatment follow-up for patients with major depressive disorder (MDD). This study developed, validated and applied a machine learning model to estimate PHQ-9 scores for MDD patients using relevant notes from electronic medical records (EMR). Methods: Information from structured and unstructured sections of prescriber notes from a multi-source real-world mental health database were used to estimate PHQ-9 scores (ePHQ-9). Model performance and agreement were evaluated using binary and categorical PHQ-9 outcomes. The final model strategy was applied to MDD patient encounters without scores to assess the extent of added available PHQ-9 measures. Results: A final model was developed from 48,594 patients and 143,224 clinical encounters with a recorded PHQ-9 score, and then applied to 196,819 MDD patients. Overall model performance was high with an AUC 0.81, PPV 0.71 and NPV 0.76. The addition of ePHQ-9 scores increased the average number of available scores per patient per year by 2.8×. Limitations: The model was developed using prescribing mental health providers' clinical notes, which limits generalizability to other contexts (e.g., primary care). The PHQ-9 is designed to be patient-reported, whereas this model strategy estimates PHQ-9 scores using clinicians' notes, which results in some expected discrepancies. Conclusions: This validated ePHQ-9 model contributes to addressing measurement gaps in depression treatment and research by adding substantially to the number of measures available in real-world data for clinical follow-up.