Scholarlink

<p>Autism Spectrum Disorder (ASD) diagnosis traditionally relies on centralized data aggregation of sensitive behavioural and physiological information, raising significant privacy and compliance challenges. To address these concerns, a novel Federated Transformer-Based Deep Learning Framework is designed for secure, decentralized, and early autism diagnosis using multimodal behavioural datasets. The system models intricate spatiotemporal connections in behavioural sequences collected from several clinical locations by using the self-attention mechanism of Transformer topologies, all without necessitating the exchange of raw data. Federated Learning (FL) and Transformer encoders are used in the proposed approach to jointly train a global model while protecting edge data privacy. The Autism Brain Imaging Data Exchange (ABIDE) dataset for neuroimaging-derived behavioural characteristics and the SPARK behavioural assessment dataset are two publicly accessible benchmark datasets used to examine this approach. Focused loss functions and federated weighted averaging were used to address the issues of class imbalance and data heterogeneity. The proposed federated Transformer model shows better diagnostic performance compared to baseline CNN models and RNN models in the federated settings, which is supported by experimental results. Evaluation on the ABIDE dataset showed that the model achieved an average accuracy of 99.4%, precision of 99.1%, recall of 98.6%, and an F1-score of 98.3%. Performance on the SPARK dataset resulted in an accuracy of 99.7%, precision of 98.8%, recall of 98.5% and an F1 - score of 98.6%. These results indicate strong generalization and efficient learning of long-range temporal dependencies that occur in behavioural data. In addition, communication efficiency was improved using model compression techniques, which led to a 40% reduction in bandwidth usage without significant reduction in performance. Privacy guarantees were enhanced by adding secure aggregation protocols in the federated training loop. This work is the first to apply Transformer architectures to federated autism diagnosis, and introduces a scalable, privacy-preserving solution for the early and accurate detection of ASD in compliance with data protection regulations.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Federated Transformer-Based Deep Learning Framework for Secure and Early Autism Diagnosis

  • 期刊论文

  • 发表日期：

    2026年2月24日

查看全文
  • R. Maddala Kranthi,
  • R. Venkatesan,
  • C. Amuthadevi,
  • C. Priyadharsini

摘要

Autism Spectrum Disorder (ASD) diagnosis traditionally relies on centralized data aggregation of sensitive behavioural and physiological information, raising significant privacy and compliance challenges. To address these concerns, a novel Federated Transformer-Based Deep Learning Framework is designed for secure, decentralized, and early autism diagnosis using multimodal behavioural datasets. The system models intricate spatiotemporal connections in behavioural sequences collected from several clinical locations by using the self-attention mechanism of Transformer topologies, all without necessitating the exchange of raw data. Federated Learning (FL) and Transformer encoders are used in the proposed approach to jointly train a global model while protecting edge data privacy. The Autism Brain Imaging Data Exchange (ABIDE) dataset for neuroimaging-derived behavioural characteristics and the SPARK behavioural assessment dataset are two publicly accessible benchmark datasets used to examine this approach. Focused loss functions and federated weighted averaging were used to address the issues of class imbalance and data heterogeneity. The proposed federated Transformer model shows better diagnostic performance compared to baseline CNN models and RNN models in the federated settings, which is supported by experimental results. Evaluation on the ABIDE dataset showed that the model achieved an average accuracy of 99.4%, precision of 99.1%, recall of 98.6%, and an F1-score of 98.3%. Performance on the SPARK dataset resulted in an accuracy of 99.7%, precision of 98.8%, recall of 98.5% and an F1 - score of 98.6%. These results indicate strong generalization and efficient learning of long-range temporal dependencies that occur in behavioural data. In addition, communication efficiency was improved using model compression techniques, which led to a 40% reduction in bandwidth usage without significant reduction in performance. Privacy guarantees were enhanced by adding secure aggregation protocols in the federated training loop. This work is the first to apply Transformer architectures to federated autism diagnosis, and introduces a scalable, privacy-preserving solution for the early and accurate detection of ASD in compliance with data protection regulations.