Privacy-aware Early Alzheimer’s Disease Prediction Using Low-rank Vision Transformers With Machine Unlearning.

 

Background: Early AD diagnosis is critical for timely intervention, but reliable AI models are limited by scarce, highquality MRI datasets. Scanner variability, demographic confounders, and data privacy rules further restrict model training and generalization.

The Challenge: Deep learning models (CNNs, ViTs) show strong potential for AD prediction from neuroimaging, but in data-scarce settings they overfit easily, limiting real-world robustness. There is also a growing need to remove specific patient data from trained models (consent withdrawal, data quality issues) — conventional approaches require full, computationally expensive retraining.

 

Critical Gap: No existing framework combines parameter efficient adaptation for small AD datasets with fast, verifiable patient-data erasure. Conventional unlearning requires full model retraining — computationally expensive and impractical at scale.

Objectives

1. Achieve >95% accuracy on sparse (n<500)
AD neuroimaging via low-rank ViT
2. Cut trainable parameters with LoRA-inspired
adapters, preserving accuracy
3. Suppress scanner/demographic noise via
spectral orthogonal projection
4. Enable fast, certified patient-data erasure
via SVD-based unlearning