ASD-Pipeline: An Ensemble Machine Learning Framework Integrating Feature Selection, Behavioural Clustering, and Class Rebalancing for Accurate Autism Spectrum Disorder Prediction

Abstract

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by a variety of behavioral and cognitive patterns. Early and precise detection is critical in enabling timely interventions. Conventional classification models frequently exhibit poor generalization due to irrelevant features, unstructured behavioral data, and severe class imbalance. Despite current advances in machine learning for ASD detection, current models do not integrate adaptive feature selection, behavioral grouping, or imbalanced class handling in a unified, end-to-end pipeline. The lack of incorporation frequently results in suboptimal performance and limited interpretability. This study proposes a new ensemble-based framework called ASD-Pipeline, which integrates flexible feature selection, hybrid clustering, synthetic minority oversampling, and ensemble voting classification to improve the predictive performance for ASD identification. The proposed ASD-Pipeline framework uses a five-stage process to improve the accuracy of autism spectrum disorder prediction. First, the dataset is normalized utilizing Min-Max scaling to guarantee that the feature ranges remain consistent. Next, feature selection is performed utilizing FlexiFeat, an ensemble method integrating filter-based (CfsSubsetEval with BestFirst), wrapper-based (WrapperSubsetEval with GreedyStepwise), and embedded (ReliefF with Ranker) techniques to maintain only the most pertinent feature. The ClusterGroup stage uses K-Means clustering (k=5) and DBSCAN improvement (ε=0.5, minPts=3) within each cluster to create behavioral groups and remove outliers. The ReBalance stage uses Cluster-SMOTE to tackle class imbalance by producing synthetic samples for the minority class and a balanced dataset. Finally, the ASDClassifier stage involves training an ensemble of Logistic Regression, Support Vector Machine, and Gradient Boosting classifiers that are combined using soft voting. Metrics used to assess the model include accuracy, precision, recall, F1-score, and Matthews Correlation Coefficient (MCC). The proposed ASD-Pipeline surpassed existing models, achieving a significantly higher accuracy of 96.18% compared to previous techniques ranging from 76.80% to 90.60%. It also scored 91.51% precision, 91.63% recall, 95.57% F1-score, and 92.51% specificity. These findings emphasize the pipeline's efficacy in enhancing generalization and tackling difficulties such as feature relevance, behavioral grouping, and class imbalance for ASD prediction. The ASD-Pipeline offers a reliable, interpretable, and modular machine learning solution for ASD prediction. Its incorporated method tackles critical challenges in feature relevance, behavioral variability, and data imbalance, rendering it a promising tool for healthcare practitioners and researchers seeking data-driven insights into early ASD detection.