NEURO-FED: A Federated Neuro-Symbolic Framework for Privacy-Preserving, Fairness-Constrained, and Explainable Multi-Modal Autism Spectrum Disorder Diagnostics

Abstract

Autism Spectrum Disorder (ASD) affects approximately 1 in 36 children globally, yet profound disparities persist in diagnostic timing, access to care, and clinical outcomes across demographic, geographic, and socioeconomic strata. While artificial intelligence has demonstrated substantial promise in automated ASD detection, existing systems face five critical
translational barriers: (1) reliance on homogeneous, single-institution datasets with limited
generalizability; (2) absence of formal privacy-preserving mechanisms; (3) systematic algorithmic bias resulting in significant performance disparities; (4) opaque decision-making that
undermines clinician trust and adoption; and (5) computational requirements exceeding the resources of community clinical settings. This paper presents NEURO-FED, a novel federated
neuro-symbolic framework that simultaneously addresses these barriers through five integrated
innovations. First, we introduce a differentially private federated learning protocol with Renyi ´
accounting that enables collaborative model training across 15 geographically distributed clinical sites (N=796 participants) with formal (ε = 1.8, δ = 10−5
) privacy guarantees—the first
demonstration of sub-2.0 DP on multi-modal neurodevelopmental data. Second, we develop
a neuro-symbolic reasoning engine that combines hierarchical multi-modal transformer networks with a differentiable logic programming layer encoding DSM-5-TR diagnostic criteria,
achieving AUC-ROC of 0.968 while providing provably consistent explanations. Third, we
implement a novel multi-objective fairness optimization framework that reduces demographic
predictive disparities by 80.2% while improving subgroup accuracy for female (+11.2%), Black
(+11.8%), and Hispanic (+10.4%) participants with only 0.2% overall accuracy degradation.
Fourth, we demonstrate extreme edge deployment viability through progressive knowledge
distillation, achieving 95.8% of diagnostic accuracy with 96% parameter reduction and 8ms
inference latency on commodity mobile hardware. Fifth, we introduce a continuous learning
protocol that enables sustainable model updating from post-diagnostic outcomes without compromising privacy guarantees. Extensive validation across 796 participants (514 ASD-positive,
282 ASD-negative) demonstrates that NEURO-FED achieves superior performance across all
dimensions: diagnostic accuracy (AUC-ROC: 0.968, +9.6% improvement), fairness (demographic parity difference: 0.152 → 0.030, 80.2% reduction), privacy (AUC-ROC degradation:
0.968 → 0.961 under ε = 1.8), and computational efficiency (8ms inference, 0.38MB model
size on ARM Cortex-A76). Our findings establish that neuro-symbolic integration, differential
privacy, fairness optimization, and extreme edge deployment are synergistic design principles
that can be jointly optimized within unified architectural frameworks. NEURO-FED provides
a comprehensive template for ethical, equitable, and globally deployable AI in neurodevelopmental disorders and establishes a new state-of-the-art for privacy-preserving clinical decision
support.