Validity as a gate for interpretability: lessons for developing clinical AI from drawing-based behavioural markers

Open materials for a structured review and methodological framework on AI-based mental health screening from projective drawings (e.g., the House-Tree-Person test). PubMed, Scopus, IEEE Xplore, and arXiv yielded 223 records; after deduplication, 211 were screened, 47 full texts were assessed, and 18 studies were included. Each included study was rated Yes / Partial / No on four criteria: features falsified or drawing ability controlled, validation beyond one dataset or protocol, clinically interpretable output, and clinician evaluation of outputs. The findings motivate a two-pillar framework in which three preregistered falsification tests (residualisation of artistic ability, semantic scrambling, cross-protocol decay) gate entry to a clinician-grounded concept bottleneck model. Files: search strategy with strings as executed; screening workbook with record-level decisions, exclusion reasons, appraisals, and a sheet that recomputes every number reported in the article; PRISMA 2020 flow diagram; preregistration template for applying the framework; supplementary material. Screening and appraisal were performed by a single reviewer.