ShishuRaksha AI: Training-free screening for Bangladeshi children

ShishuRaksha AI is a training-free multimodal decision-support framework for early screening of abuse-related psychological trauma in Bangladeshi children, presented in a paper submitted 4 Jul 2026 by Salma Hoque Talukdar Koli and Fahima Haque Talukder Jely. The system fuses four screening modalities, renders clinically weighted explanations, and outputs bilingual (Bangla/English) reports with referral routing to national child-protection services under the Children Act 2013.

What is ShishuRaksha AI and how does it work?

ShishuRaksha AI combines validated questionnaires (SDQ and CPSS), Bengali narrative text, House-Tree-Person (HTP) drawing features, and facial affect into a single, training-free fusion design that uses cross-modal attention and a single-modality override rule. The framework is explicitly labeled decision-support, not diagnostic, and every risk score is explained using clinically weighted, perturbation-based additive attribution; outputs are packaged as bilingual reports with referral routing to OCC, DSS and NMHH under the Children Act 2013.

The architecture is clinically informed rather than end-to-end learned: questionnaire scores and modality-specific features are weighted and combined via cross-modal attention into a fused risk score. A single-modality override can trigger when one channel indicates high concern. The authors also generate an explanation for each score using additive attribution grounded in clinical weights.

How was the system evaluated and what were the results?

Evaluation used a noise-aware synthetic benchmark of 500 cases, with 116 labeled positive for abuse-related risk (23.2%), and four deliberate noise layers plus literature-grounded HTP priors. Tree-ensemble surrogate models of the fusion (excluding the facial channel) were tested under 5-fold stratified cross-validation, producing a fused-model AUC of 0.874 with a 95% interval [0.834-0.908]. An SDQ-only baseline reached an AUC of 0.756 [0.705-0.803].

The paper reports ablation, operating-point, subgroup and calibration analyses on the synthetic benchmark. The facial channel was excluded from the surrogate evaluation reported. The authors list key methodological caveats: all data are synthetic, there is no held-out set, text-feature circularity exists, and an urban-rural subgroup gap was observed.

Why does this matter?

Bangladesh has an acute shortage of child mental-health capacity, with an estimated 1.17 mental-health professionals per 100,000 population and only six child psychiatrists nationwide. No Bengali-language, culturally adapted screening tool for abuse-related trauma in children existed before this work. A training-free, clinically weighted system aims to reduce dependence on large labeled clinical datasets while offering explainable, referral-ready outputs that map to national child-protection services. For low-resource settings where collecting ethically valid clinical datasets is not yet feasible, a design that prioritizes explainability and referral routing responds to operational and legal realities captured in the paper.

The reported AUC uplift from 0.756 for an SDQ-only baseline to 0.874 for the fused surrogate suggests that multimodal fusion and clinical weighting materially improved signal in the synthetic benchmark. That improvement matters if it generalizes to real-world data, because higher discrimination could direct scarce clinical time and protection resources more effectively.

Limitations and explicit caveats

The authors state this is a feasibility study and design contribution. Major limitations include synthetic-only data, absence of a held-out set, and text-feature circularity. The facial channel was not included in the surrogate evaluation presented, and an urban-rural subgroup gap was observed. The paper does not claim clinical readiness; the framework is positioned as decision-support and the authors present the limitations openly.

What to watch

External validation on ethically collected clinical or real-world screening data and a held-out test set are the next concrete milestones to assess generalizability. Watch for studies that (1) include the facial channel in evaluation, (2) address the reported urban-rural subgroup gap, and (3) test the bilingual report and referral routing in operational settings connected to OCC, DSS and NMHH.

The paper’s code and data links are listed on the arXiv entry; the submission date is 4 Jul 2026. This work is primarily a systems and evaluation design aimed at low-resource, Bengali-language settings and is explicitly described as a step toward ethically deployable screening, not a finished clinical product.