Privacy-Preserving Autism Detection Using Federated CNN-LSTM Networks: A Spatiotemporal Deep Learning Framework for Decentralized Behavioural Diagnosis

Early detection of autism spectrum disorder (ASD) is essential for effective intervention but remains limited by privacy concerns, fragmented data sources, and constrained access to large-scale, labelled datasets. This paper presents a privacy-preserving diagnostic framework based on federated learning, enabling decentralized institutions to collaboratively train models without sharing sensitive patient data. The proposed architecture integrates Convolutional Neural Networks (CNNs) for spatial feature extraction with Long Short-Term Memory (LSTM) networks to capture temporal behavioural patterns. Training is performed across five heterogeneous ASD datasets in a federated environment, ensuring both model generalizability and data confidentiality. Local preprocessing pipelines standardize input formats, while secure aggregation and differential privacy mechanisms protect model updates. Experimental results demonstrate that the federated CNN-LSTM model achieves 98.9% classification accuracy, outperforming centralized and shallow learning baselines. This work illustrates the feasibility of combining spatiotemporal deep learning with federated protocols for real-world ASD screening and contributes to the growing field of privacy-aware medical AI.