Cross-region generalizability of early childhood development prediction models across South Asia: Evidence from harmonized MICS surveys

Background Early childhood development (ECD) is crucial for lifelong health and productivity, yet many children in low- and middle-income countries face developmental risks. It remains unclear how well ECD prediction models perform across diverse populations. Methods Using harmonized MICS-6 data from 42,291 children aged 36–59 months across Bangladesh, Nepal, Sindh, Balochistan, and Khyber Pakhtunkhwa, we assessed ECD status using the Early Childhood Development Index (ECDI). Ten machine learning algorithms were compared using Monte Carlo cross-validation, hyperparameter tuning, and SHAP-based interpretation. Cross-region predictive generalizability was evaluated through internal, cross-region external, and leave-one-region-out (LORO) validation. Results The proportion of children developmentally on track ranged from 21.0% (Bangladesh) to 36.3% (Nepal). Literacy–numeracy (66.7–89.5%) and socio-emotional development (58.6–88.7%) were relatively high, but the learning domain remained consistently low (5.4–16.1%). Internal validation yielded moderate performance (AUC: 0.568–0.770), with gradient boosting models performing best. Tuned models achieved AUCs from 0.643 (Bangladesh) to 0.771 (Balochistan). External validation showed reduced and variable performance (AUC: 0.610–0.758), with no model generalizing consistently across regions. Home stimulation and access to learning materials emerged as the strongest predictors across all settings. Conclusion ECD prediction models perform reasonably well within similar populations but show limited cross-context generalizability. The consistent importance of home learning environments highlights low-cost opportunities for intervention, while performance variability underscores the need for locally adapted predictive approaches in resource-constrained settings.