Abstract
<title>Abstract</title> <p>Alzheimer's disease progresses heterogeneously across diverse cohorts, yet current predictive models fail to capture this complexity while remaining clinically interpretable. Here we present a multi-dimensional attention framework that simultaneously captures both temporal dynamics and biomarker importance to predict disease progression across three fundamentally different populations: the general late-onset population using the Alzheimer's Disease Prediction Of Longitudinal Evolution (TADPOLE) dataset (N = 1669), cases with Down Syndrome-associated Alzheimer's disease using the Alzheimer's Biomarker Consortium - Down Syndrome (ABC-DS) dataset (N = 396), and cases with autosomal dominant Alzheimer's disease using the Dominantly Inherited Alzheimer Network (DIAN) dataset (N = 425). Trained on each dataset independently, our framework achieved multi-class Area Under the Receiver Operating Characteristic Curve (mAUC) values of 0.793 (TADPOLE), 0.680 (ABC-DS), and 0.902 (DIAN) when predicting individuals' future diagnostic status (cognitively normal/stable, mild cognitive impairment, or Alzheimer's disease) from their longitudinal biomarker history, outperforming conventional approaches. The model generates individual-specific attention maps revealing distinct biomarker importance over time. Transfer learning from TADPOLE—which included neuroimaging data—improved prediction performance on the imaging-free ABC-DS dataset from 0.680 to 0.771, demonstrating that disease mechanisms transcend both etiological boundaries and data modalities. Ultimately, this framework could enable precision medicine approaches for data-limited cohorts across the Alzheimer's disease spectrum.</p>