Federated Learning for Privacy-Preserving Autism Research Across Institutions: Enabling Collaborative AI Without Compromising Patient Data Security

Posted: May 11, 2021

• Hammad Khan
  Park University
• Ethan Jones
  University of California Los Angeles
• Sophia Miller
  University of Washington

Abstract

The advancement of artificial intelligence in autism spectrum disorder research
faces significant challenges due to privacy concerns and data governance restrictions
that limit data sharing across institutions. This research presents a comprehen
sive federated learning framework that enables collaborative model development
across multiple healthcare institutions while maintaining patient data privacy and
complying with stringent regulatory requirements. Our approach implements a
sophisticated federated averaging algorithm with differential privacy guarantees,
secure multi-party computation protocols, and adaptive client selection mecha
nisms specifically designed for heterogeneous autism datasets. The framework was
evaluated across six major medical institutions with diverse patient populations,
encompassing data from 4,200 children with autism spectrum disorder and 2,800
typically developing controls. The federated model achieved 92.8% diagnostic accu
racy, comparable to centralized training approaches (93.5%) while providing strong
privacy guarantees with epsilon values as low as 1.2 for differential privacy. The
system demonstrated robust performance across different data distributions and
institutional characteristics, with communication efficiency improvements of 47%
compared to standard federated learning approaches through our adaptive client
selection and model compression techniques. Privacy analysis confirmed that the
framework prevents data reconstruction attacks and membership inference attacks
while maintaining model utility. This research establishes that federated learning
can overcome critical barriers to multi-institutional autism research by enabling
collaborative AI development without sensitive data sharing, potentially accelerat
ing scientific discovery while upholding the highest standards of patient privacy and
data protection. The framework provides a scalable solution for privacy-preserving
medical AI that balances model performance with ethical data handling practices.