AUTISM-FLO: A Federated Learning and Optimization Framework for Privacy-Preserving, Fairness-Constrained, Multi-Modal Autism Spectrum Disorder Diagnostics

Abstract

Autism Spectrum Disorder (ASD) now affects 1 in 36 children in the United States, yet significant disparities persist in diagnostic age, access to care, and clinical outcomes across demographic groups. While artificial intelligence has demonstrated substantial promise in automated
ASD detection, existing systems face five critical barriers to clinical translation: (1) reliance
on homogeneous, single-institution datasets with limited generalizability; (2) absence of formal privacy-preserving mechanisms; (3) systematic algorithmic bias resulting in performance
disparities; (4) opaque decision-making that undermines clinician trust; and (5) computational
requirements exceeding clinical resources. This paper presents AUTISM-FLO, a comprehensive federated learning and optimization framework that simultaneously addresses these barriers through four integrated innovations. First, we introduce a differentially private federated
learning protocol with Renyi accounting that enables collaborative model training across 13 ´
geographically distributed clinical sites (N=748 participants) with formal (ε = 2.0, δ = 10−5
)
privacy guarantees. Second, we develop a hierarchical multi-modal attention network that fuses
fMRI connectivity, eye-tracking gaze dynamics, speech prosody features, and structured clinical assessments through context-dependent weighting, achieving AUC-ROC of 0.963. Third,
we implement a novel fairness-constrained optimization framework with equalized odds projection that reduces demographic predictive disparities by 78.6% while improving subgroup
accuracy for female (+10.4%) and Black (+10.9%) participants. Fourth, we demonstrate edge
deployment viability through knowledge distillation, achieving 96.4% of diagnostic accuracy
with 93% parameter reduction and 19ms inference latency on standard clinical hardware. Extensive validation across 748 participants (458 ASD-positive, 290 ASD-negative) demonstrates
that AUTISM-FLO achieves superior performance across all dimensions: diagnostic accuracy
(AUC-ROC: 0.963, +9.2% improvement), fairness (demographic parity difference: 0.149 →
0.032, 78.6% reduction), privacy (AUC-ROC degradation: 0.963 → 0.957 under ε = 2.0), and
computational efficiency (19ms inference, 5.8MB model size). Our findings establish that privacy preservation, algorithmic fairness, diagnostic accuracy, and clinical deployability are not
competing objectives but can be synergistically optimized within unified architectural frameworks. AUTISM-FLO provides a replicable template for ethical, equitable, and practical AI in
neurodevelopmental disorders and serves as a foundation for global health translation.