A Multi-Tiered Federated Deep Learning Architecture for Precision Autism Diagnostics and Personalized Intervention: Integrating Multi-Modal Biomarkers with Fairness-Aware Clinical Decision Support

Abstract

Autism Spectrum Disorder (ASD) affects approximately 1 in 36 children, yet traditional diagnostic pathways face critical limitations including prolonged wait times, subjective assessment
variability, and significant disparities in access across demographic groups. While artificial
intelligence has shown promise in autism detection, existing approaches remain constrained
by siloed data modalities, limited generalizability across diverse populations, and a predominant focus on binary classification rather than comprehensive intervention planning. This paper presents AutismCare-Net, a novel three-tiered federated deep learning architecture that
addresses these gaps through integrated multi-modal biomarker fusion, privacy-preserved collaborative learning across 12 clinical sites, and fairness-constrained optimization. The framework incorporates a hierarchical attention-based fusion mechanism combining functional MRI
connectivity patterns, eye-tracking gaze dynamics, acoustic speech prosody features, and structured clinical assessments. We validate our approach on a multi-site cohort of 748 participants
(aged 18–72 months) collected over 26 months, demonstrating significant improvements across
all metrics: diagnostic AUC-ROC of 0.962 (8.7% improvement), F1-score of 0.931 (10.2% improvement), and reduction in demographic parity difference from 0.142 to 0.031. The system
further introduces a novel continuous learning module for longitudinal intervention monitoring
and an explainable recommendation engine for personalized therapy planning. Our findings
establish that integrated, privacy-conscious, and fairness-optimized AI systems can simultaneously advance diagnostic accuracy, clinical utility, and health equity in autism care.