Explainable AI for early developmental disability detection: a neuro-fuzzy approach

Background Developmental disabilities affect approximately 1 in 6 children aged 3–17 years. The diagnostic process typically spans 2–3 years and disproportionately affects underserved populations. Traditional machine learning approaches for screening have demonstrated promising accuracy but often lack the transparency required for clinical acceptance. Methods We developed and validated an Adaptive Neuro-Fuzzy Inference System (ANFIS) using a longitudinal observational dataset of 5,000 children aged 1–6 years from 12 early childhood centers (4,311 typically developing; 689 with developmental disabilities, prevalence 13.8%). The ANFIS model implemented Gaussian membership functions and Sugeno fuzzy inference, incorporating age-adjusted ratios and cross-domain interaction features. Model performance was evaluated using standard binary classification metrics and five-fold stratified cross-validation. Results All four assessment domains differed significantly between diagnostic groups ( p < 0.001): Cognitive Scores (60.86 ± 14.51 vs. 51.58 ± 17.09), Behavioral Scores (55.82 ± 9.38 vs. 49.64 ± 11.21), Motor Skills (51.08 ± 7.79 vs. 45.08 ± 8.00), and Social Interaction (51.20 ± 11.66 vs. 42.13 ± 11.19). Family history was present in 85.6% of diagnosed children vs. 45.0% of typically developing children ( p < 0.001). The ANFIS model achieved 96.0% accuracy, 87.5% sensitivity, 97.6% specificity, and AUC = 0.925 on a held-out test set of 100 cases. Five-fold cross-validation yielded a mean accuracy of 89.2% ± 3.4% (95% CI: [85.8, 92.6%]). Cognitive_Social_Ratio was the most influential diagnostic indicator. Conclusion The ANFIS approach demonstrates a clinically relevant balance between diagnostic accuracy and interpretability, positioning it as a viable clinical decision-support tool. Larger external validation studies across diverse populations are required before widespread implementation.