Explainable AI Screens Child Trauma in Low-Resource Settings

In Bangladesh, a severe shortage of mental health professionals, particularly child psychiatrists, highlights the urgent need for early screening tools for abuse-related psychological trauma in children. This research introduces ShishuRaksha AI, a decision-support framework designed for low-resource settings. It employs a training-free, multimodal approach, integrating data from validated questionnaires (SDQ, CPSS), Bengali narrative text, House-Tree-Person (HTP) drawing features, and facial affect. The system fuses these modalities using clinical weighting and cross-modal attention, with an override rule for single modalities. Crucially, every risk score is explained through clinically weighted, perturbation-based additive attribution, generating bilingual (Bangla/English) reports that include referral routing to national child-protection services. Since ethical collection of clinical data on abused children is currently impossible, the framework was evaluated on a noise-aware synthetic benchmark, demonstrating an AUC of 0.874, significantly outperforming an SDQ-only baseline.