Cross-modal privacy-preserving synthesis and mixture-of-experts ensemble for robust ASD prediction.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Data Science

Background:

• Autism Spectrum Disorder (ASD) diagnosis is challenged by limited multimodal datasets and data privacy concerns.
• Current diagnostic methods are resource-intensive and rely on unimodal or non-adaptive learning.
• Large-scale, privacy-preserving multimodal data is crucial for advancing ASD research and diagnosis.

Purpose of the Study:

• To introduce AutismSynthGen, a novel privacy-preserving framework for synthesizing multimodal ASD data.
• To enhance Autism Spectrum Disorder prediction accuracy using generated synthetic data.
• To address limitations in data accessibility and privacy in ASD research.

Main Methods:

• A Multimodal Autism Data Synthesis Network (MADSN) using transformer encoders and cross-modal attention within a conditional GAN generates synthetic multimodal data (MRI, EEG, behavioral, severity scores).
• Differential privacy is enforced using DP-SGD (ε ≤ 1.0).
• An Adaptive Multimodal Ensemble Learning (AMEL) module, with heterogeneous experts and a gating network, is trained on real and synthetic data.

Main Results:

• Synthetic data augmentation improved model performance, with validation AUC gains of ≥ 0.04.
• The AMEL module achieved high performance on real data (AUC=0.98, F1=0.99) and near-perfect performance with synthetic data (AUC≈1.00, F1≈1.00).
• Distributional metrics (MMD=0.04, KS=0.03) and BLEU score (0.70) confirmed high fidelity between real and synthetic samples.

Conclusions:

• AutismSynthGen provides a scalable, privacy-compliant method for augmenting multimodal ASD datasets.
• The framework enhances ASD prediction accuracy, offering a valuable tool for researchers and clinicians.
• Future work includes exploring semi-supervised learning, explainable AI, and federated learning for broader, privacy-preserving accessibility.