MSSFN: Multi-stimulus stereo spatiotemporal fusion network with pattern disentanglement for Alzheimer’s disease diagnosis

Eye movements have been established as biomarkers for Alzheimer’s disease (AD). Current methods use diverse cognitive stimuli to elicit diagnostic responses, but face two challenges: the lack of 3D spatiotemporal modeling for comprehensive fusion and the entanglement of cognitive patterns within and across stimuli. To overcome these challenges, we propose the multi-stimulus stereo spatiotemporal fusion network (MSSFN) for AD diagnosis. The MSSFN contains two core modules: (1) stereo spatiotemporal embedding fuses binocular spatial fixation distributions, temporal scanpath dynamics, and stereo priors to establish unified 3D spatiotemporal eye movement modeling; (2) local-sensitive contrastive disentanglement implicitly extracts and decouples intrinsic cognitive patterns via a local-sensitive attention and a cognitive contrastive constraint to separate diagnostic cognitive patterns across stimulus paradigms. In addition, demographic-guided cognitive aggregation incorporates participant information as auxiliary covariates in the final decision. On a demographically matched, clinic-recruited cohort of 100 clinically diagnosed AD patients and 100 healthy controls (age: 72.12  ±  6.18 years), participants completed three stimulus paradigms: saccade, visual search, and visual attention. Under subject-wise 5-fold cross-validation, MSSFN demonstrates superior performance, achieving 0.87 accuracy, 0.87 AUC, 0.88 sensitivity, and 0.86 specificity, surpassing the strongest baseline MCGRL (0.81 accuracy) by 7.4%. Ablation studies further verify the contributions of the proposed modules. These results suggest that MSSFN enables efficient eye movement based AD screening as a complement to existing AD biomarkers, supporting future large-scale deployment.