Enhance autism spectrum disorder detection using stacking ensemble learning model with explainable AI.

Review Approach involved constructing a multi-layered computational pipeline to process pediatric behavioral records. The team addressed missing entries and categorical variables through standardized data cleaning protocols. They applied Information Gain and Pearson Correlation to isolate the most influential predictive attributes. To balance the distribution of samples, the investigators implemented the Synthetic Minority Oversampling Technique. The core architecture utilized a stacked ensemble strategy with five distinct base learners. A Random Forest meta-classifier integrated these individual outputs to produce final diagnostic predictions. The researchers performed hyperparameter tuning to refine the operational parameters of the entire system. Finally, they employed the Shapley explanation method to quantify the influence of specific features on the generated outcomes.

Key Findings From the Literature indicate that the stacked ensemble model consistently outperformed individual classifiers across all tested datasets. The framework achieved 99% accuracy on the Toddler Saudi dataset. Testing on the Q-CHAT dataset resulted in 98% accuracy. The model reached 99% accuracy when evaluated on both the Nao and fused datasets. These results highlight the superior predictive capability of the combined approach. The authors observed that the meta-classifier effectively synthesized information from the base learners to improve overall reliability. Furthermore, the Shapley explanation method successfully identified the primary behavioral features driving the model predictions. The data suggest that this framework provides a highly effective tool for automated screening applications.

Synthesis and Implications suggest that the integrated computational architecture provides a highly reliable method for identifying early developmental markers. The authors propose that combining multiple base learners significantly improves predictive performance compared to standalone algorithms. Their findings demonstrate that the meta-classifier approach effectively reduces classification errors across diverse datasets. The researchers indicate that incorporating explainable techniques allows clinicians to interpret model-driven decisions with greater transparency. This synthesis highlights the potential for automated systems to support, rather than replace, professional clinical judgment. The authors suggest that their framework offers a scalable solution for screening in resource-limited settings. Their evidence supports the utility of advanced feature selection in refining diagnostic accuracy. The study concludes that such tools represent a promising direction for enhancing pediatric mental health services.