Machine learning-based computerized adaptive testing of PHQ-8 and GAD-7 in psychiatric outpatients.

Abstract

Background: Scalable, accurate and efficient monitoring tools are urgently needed to support psychiatric assessment in routine care. This study aimed to evaluate a machine learning-based computerized adaptive testing (ML-CAT) model for the PHQ-8 and GAD-7 in psychiatric outpatient settings.

Methods: Data were obtained from a longitudinal cohort study of patients with depression or anxiety diagnoses. ML-CAT models were developed using the M5P algorithm. Model performance was evaluated via 10-fold cross-validation. Concurrent validity was assessed with Pearson's r, R², and mean absolute error (MAE). Longitudinal change detection used paired t-tests and Cohen's d. Diagnostic classification agreement (based on ≥10 cutoff between assessments) was examined using percentage agreement and Cohen's κ.

Results: The ML-CAT administered a median of 2 items per instrument (66-71% item reduction) while achieving near-perfect concurrent validity with full-scale scores (r = 0.996-0.997; R² = 0.986-0.991; MAE < 0.62). Both ML-CAT versions reliably detected symptom reductions at 6 months and 1 year. Classification agreement between ML-CAT and full-scale versions was high at 6 months and 1 year for PHQ-8 (κ = 0.85-0.93) and GAD-7 (κ = 0.91-0.96).

Conclusions: The ML-CAT model reproduced PHQ-8 and GAD-7 scores with high accuracy and detected meaningful clinical change across two follow-up waves. The resulting adaptive questionnaire can be presented as a flowchart and administered both online and offline, providing real-world clinical evidence for scalable ML-CAT-based psychiatric assessment.