Bangladesh has roughly 1.17 mental-health professionals per 100,000 people and six child psychiatrists for the entire country — and a new AI framework is trying to fill some of that gap.
Researchers published a feasibility study for ShishuRaksha AI, a decision-support system that fuses four screening signals: standardized questionnaires (SDQ and CPSS), Bengali-language narrative text, House-Tree-Person drawing analysis, and facial affect. The fusion is training-free and clinically weighted, meaning no labeled patient data was needed to build the core model. Because collecting clinical data on abused children raises serious ethical barriers, the team evaluated the system against a synthetic benchmark of 500 cases with four deliberate noise layers. The fused model reached an AUC of 0.874, compared to 0.756 for a questionnaire-only baseline. Explanations are rendered in both Bangla and English, with referral routing to national child-protection services under Bangladesh's Children Act 2013.
The significance here is less about the AUC number and more about the constraint it was built under: no real patient data, one of the world's thinnest mental-health workforces, and no existing Bengali-language screening tool. A system that can offer a clinically weighted, explainable risk flag — without requiring a trained specialist to administer it — has real value in that context, even at the feasibility stage.
The authors are refreshingly direct about what this is not: there is no held-out test set, the text features carry circularity risk, the facial channel was excluded from benchmarking, and an urban-rural subgroup gap remains unresolved. That candor is worth noting in a field where AI health tools sometimes sprint from preprint to press release.