Residual Gap-Aware Transformers predict Alzheimer's progression 24 months out with 13% less error

A new AI model from Ran Tong and colleagues tackles the challenge of medium-horizon Alzheimer's disease progression prediction by leveraging irregular clinical and biomarker histories. Their approach, called the Residual Gap-Aware Transformer, anchors each prediction at a mild cognitive impairment visit and uses only data observed up to that point to forecast the 24-month change in the Clinical Dementia Rating Sum of Boxes (CDR-SB). The architecture combines a mixed-effects statistical reference (using participant-level random intercepts) with a transformer that learns residuals from pre-anchor histories. To handle sparse and irregular observations, the transformer uses triplet tokenization (time, value, type) and a learned nonnegative time-gap penalty in self-attention, effectively weighting recent and relevant data more heavily.

The model was evaluated on 2,600 labeled anchors from 858 participants in the ADNI database, spanning 7,276 longitudinal rows. Compared to a carefully selected linear mixed-effects baseline, the Residual Gap-Aware Transformer reduced mean squared error by 13.1% and improved prediction-observation correlation by 26.4% across five randomized splits. It also outperformed established methods like GRU-D and STraTS on both metrics. These results demonstrate that combining statistical anchoring with gap-aware residual learning provides a robust structure for forecasting Alzheimer's progression over an 18–30 month window, potentially enabling more personalized treatment planning and clinical trial enrichment.