Participant-Wise Validation of Transcript-Only Models for Adolescent Depression and Anxiety Triage in MMPsy

Winata Putri (Department of Pediatrics, CMHC Research Center, Palembang, Indonesia)
Mischa Chantal Adella (Department of Forensic Medicine and Medicolegal, Phlox Institute, Palembang, Indonesia)
Agnes Mariska (Department of Internal Medicine, Jayawijaya Private Clinic, Jayapura, Indonesia)
Husin Sastranagara (Department of Neurology, CMHC Research Center, Palembang, Indonesia)

Scientia Psychiatrica
Vol. 7 No. 4 (2026): Scientia Psychiatrica

Background. Digital interviews may offer scalable signals for identifying adolescents needing further assessment, but language models can appear accurate because of participant leakage, response length, or tuning-data evaluation. Objective. To develop and internally validate transcript-only probabilistic models of Patient Health Questionnaire-9 (PHQ-9) and Generalized Anxiety Disorder-7 (GAD-7) screening positivity in the Multi-Modal Psychological Assessment (MMPsy) corpus and quantify incremental performance beyond response length. Methods. This secondary prediction-modeling study used 1,275 participants with Mandarin transcripts and complete PHQ-9 and GAD-7 scores. Transcripts were represented by term frequency-inverse document frequency character 2–4-grams and modeled with elastic-net logistic regression. Five-fold outer cross-validation estimated performance in unseen participants; four-fold inner cross-validation selected penalization, fitted Platt calibration, and set fold-specific thresholds targeting 0.80 sensitivity. Comparators were response length and training-fold prevalence. Intervals used 2,000 participant bootstraps. Results. Screening positivity occurred in 258 (20.2%) PHQ-9 and 136 (10.7%) GAD-7 participants. Transcript areas under the receiver operating characteristic curve were 0.791 (95% confidence interval 0.761–0.820) for PHQ-9 and 0.822 (0.784–0.857) for GAD-7; precision-recall areas were 0.453 and 0.356. Calibration intercepts were 0.02 and −0.02; slopes were 1.10 for both. At sensitivity-directed thresholds, sensitivity was 81.4% and 80.1%, positive predictive value 36.1% and 23.5%. Improvements over response length were small and uncertain. Decision-curve patterns favored transcript content more consistently for GAD-7 than PHQ-9. Conclusion. Participant-wise transcript models identified screening positivity with useful internal ranking and high negative predictive values, but incremental value and clinical utility were outcome-dependent. External, prospective, subgroup, and workflow validation are required before operational use.