FED-MIND: A Privacy-Preserving Federated Multi-Modal Deep Learning Framework for Equitable Autism Spectrum Disorder Diagnostics and Prognostic Stratification

Abstract

Autism Spectrum Disorder (ASD) affects approximately 1 in 36 children globally, yet profound
disparities persist in diagnostic timing, clinical access, and treatment outcomes across socioeconomic, geographic, and demographic strata. While artificial intelligence has demonstrated
substantial promise in automating ASD detection, existing systems exhibit critical limitations
that have precluded widespread clinical deployment: (1) reliance on homogeneous, singleinstitution datasets with limited generalizability; (2) absence of formal privacy-preserving mech-
anisms that preclude multi-site collaborative learning; (3) systematic algorithmic bias resulting
in significant performance disparities across gender, race, and ethnicity subgroups; (4) exclusive focus on binary classification without prognostic stratification or treatment response
prediction; and (5) opaque decision-making that undermines clinician trust and adoption. This
paper presents FED-MIND (Federated Multi-Modal Integrated Neurodevelopmental Diagnostics), a comprehensive federated deep learning framework that simultaneously addresses these
limitations through five integrated innovations. First, we introduce a hierarchical multi-modal
transformer architecture that fuses four complementary data streams—resting-state fMRI connectivity, eye-tracking gaze dynamics, acoustic speech prosody, and structured clinical assessments—through learned, context-dependent attention weighting. Second, we implement
a differentially private federated learning protocol with Renyi differential privacy account- ´
ing that enables collaborative model training across 14 geographically distributed clinical sites
(N=782 participants) without centralized data aggregation. Third, we develop a novel fairnessconstrained optimization framework that reduces demographic predictive disparities by 79.4%
while maintaining state-of-the-art diagnostic accuracy (AUC-ROC: 0.967, F1-score: 0.941).
Fourth, we introduce a prognostic stratification module that predicts 12-month developmental
trajectories and treatment response profiles with 83.7% accuracy. Fifth, we implement a concept bottleneck explainability layer that provides clinically interpretable rationales aligned with
DSM-5-TR diagnostic criteria. Extensive experiments demonstrate that FED-MIND achieves
superior performance across all dimensions: diagnostic accuracy (AUC-ROC: 0.967, +9.1%
improvement), fairness (demographic parity difference: 0.147 → 0.031, 79.4% reduction),
privacy (AUC-ROC degradation: 0.967 → 0.959 under ε = 2.5, δ = 10−5
), and prognostic accuracy (83.7% for 12-month trajectory prediction). Our findings establish that privacy, equity,
accuracy, and interpretability are not competing objectives but can be synergistically advanced
through principled architectural design. FED-MIND provides a replicable template for ethical,
equitable, and clinically translatable AI in neurodevelopmental disorders and beyond.