Federated learning for fair autism spectrum disorder screening across age-heterogeneous populations.

Introduction: The detection of Autism Spectrum Disorder (ASD) remains challenging due to the heterogeneity of behavioural manifestations, limited dataset availability, and strict privacy requirements. Conventional centralized machine learning approaches often suffer from overfitting and limited generalizability across different age groups. This study proposes a federated learning (FL) framework to enable collaborative ASD screening across children, adolescents, and adults without sharing sensitive patient data. Methods: A federated learning framework was implemented and benchmarked using multiple FL algorithms, including FedPer, pFedMe, and q-FedAvg. These were compared with traditional centralized machine learning models such as Support Vector Machine (SVM), Random Forest, K-Nearest Neighbors (KNN), and J48. Data preprocessing involved imputation, encoding, scaling, feature selection, and Synthetic Minority Over-sampling Technique (SMOTE) to address missing values, categorical variables, and class imbalance. Model performance, fairness, robustness under non-IID conditions, computational efficiency, and communication costs were evaluated. Results: Customized federated learning approaches achieved superior global accuracy of 97.2% for children, 89.5% for adolescents, and 86.8% for adults. The proposed framework demonstrated improved fairness and robustness in heterogeneous non-IID environments compared to centralized models, while maintaining computational and communication efficiency. Discussion: The findings indicate that personalized federated learning provides a scalable, accurate, and privacy-preserving solution for ASD screening across diverse age groups. By bridging advanced machine learning techniques with ethical clinical practice, the proposed framework supports responsible and effective ASD detection in real-world healthcare settings.