A preprint reports higher ADNI MRI discrimination and fewer 1.5 T false positives using sparse and unpaired auxiliary data during model training.