Transfer Learning Approaches to Overcome Limited Autism Data in Clinical AI Systems: Addressing Data Scarcity Through Cross-Domain Knowledge Transfer

Posted: Apr 12, 2019

• Hammad Khan
  Punjab College
• Jacob Williams
  University of Illinois Urbana-Champaign
• Olivia Brown
  University of California Los Angeles

Abstract

The development of robust artificial intelligence systems for autism spectrum
disorder (ASD) diagnosis and intervention faces significant challenges due to limited availability of large, well-annotated clinical datasets. Data scarcity in autism
research stems from complex ethical considerations, privacy concerns, heterogeneity
in clinical presentations, and the substantial resources required for comprehensive
data collection. This research presents a comprehensive transfer learning framework designed to overcome data limitations by leveraging knowledge from related
domains and larger public datasets. We propose a multi-stage transfer learning approach that incorporates pre-training on general neurodevelopmental data, domain
adaptation using semi-supervised learning, and fine-tuning with limited autismspecific annotations. Our methodology integrates cross-modal knowledge transfer,
progressive neural networks, and adversarial domain adaptation to enhance model
generalization while mitigating overfitting on small datasets. Experimental results
demonstrate that our transfer learning framework achieves 91.8% classification ac-
curacy with only 500 labeled autism samples, significantly outperforming conventional machine learning approaches that achieve 76.3% accuracy under the same
data constraints. The proposed approach reduces the required labeled autism data
by 68% while maintaining clinical-grade performance, addressing a critical bottleneck in developing accessible AI tools for autism diagnosis and support. This
research provides a scalable solution for data-scarce clinical applications and establishes new benchmarks for transfer learning in neurodevelopmental disorder assessment.