Abstract Deep learning (DL) has shown success in predicting Alzheimer’s disease (AD) diagnosis, yet continuous measures such as cognitive assessment remain critical for richer prognosis, trajectory tracking and clinical trial enrichment. Current neurocognitive batteries are time-consuming, and the few DL models predicting cognition require expensive multimodal neuroimaging and longitudinal data. Although magnetic resonance imaging (MRI) is the most clinically accessible modality, on its own it struggles to capture AD heterogeneity in modern DL frameworks. We propose a multitask DL strategy integrating domain knowledge with large pretrained models to predict cognitive scores using only baseline MRI and demographics. By customizing loss functions and leveraging tissue segmentation-tuned latent representations as regularization features, our approach bypasses the need for longitudinal, multimodal or specialized neuroimaging data. This knowledge-informed multitask framework produces accurate diagnosis, segmentation and both current and future cognitive scores from a single baseline scan, with broad implications for early diagnosis, prognosis and clinical trial design.
Ma et al. (2026) studied this question.