AdaCluCSL: an approach for autism spectrum disorder prediction using adaptive clustering smote and cost-sensitive learning

Machine learning struggles to predict autism spectrum disorder (ASD) due to real-world datasets' underlying class imbalance. Adaptive cluster SMOTE and cost-sensitive learning (AdaCluCSL) is a new approach that improves ASD prediction accuracy. Combining adaptive k-means clustering SMOTE with cost-sensitive learning achieves this. Because positive ASD samples are few, conventional classification methods sometimes have poor predictive accuracy. AdaCluCSL uses adaptive clustering to find complex clusters in the underrepresented class and address this disparity. It then creates synthetic samples using SMOTE with cluster-specific weights. Then, the enriched samples are used with the original cases to train a cost-sensitive classifier that allocates additional costs to minority class misclassifications. Experimental tests on benchmark ASD datasets show that AdaCluCSL improves ASD prediction accuracy. Precision and recall rates have improved significantly, outperforming traditional methods. The AdaCluCSL algorithm can advance ASD prediction. It solves uneven data distribution issues, making early diagnosis and intervention more reliable.