AT-Attn: Temporal-Aware Cross-Attention for Longitudinal Multimodal Alzheimer’s Disease Diagnosis

In longitudinal Alzheimer's disease (AD) diagnosis, clinical assessments are often more predictive than structural MRI, which is high-dimensional, noisy, and frequently missing. The authors investigate how to effectively integrate these heterogeneous, irregularly sampled modalities without allowing the weaker MRI signal to distort the more reliable clinical evidence.

The researchers developed AT-Attn, a deep learning framework designed for longitudinal patient-level classification. The architecture includes three key innovations:

• Change-and-Time Encoding: A module that captures both the instantaneous state and the temporal dynamics of clinical and imaging features without requiring explicit, unstable division by irregular time intervals.
• Asymmetric Cross-Attention: A mechanism where MRI features act as queries to extract context from cognitive scales, rather than vice versa, ensuring that the more diagnostic clinical data guides the integration of the imaging data.
• Gated Fusion: A per-visit gating mechanism that balances raw MRI features with cross-attention-enhanced representations, allowing the model to selectively retain direct clinical evidence when MRI data is noisy or missing.

The model was evaluated on 1,520 patients from the ADNI cohort, using a patient-level five-fold cross-validation protocol to ensure robust performance estimation.

AT-Attn outperformed unimodal and naive multimodal fusion baselines, achieving an ROC-AUC of 0.873 and a macro F1 score of 0.721. The results demonstrate that the asymmetric, temporal-aware fusion strategy successfully extracts complementary information from structural MRI, providing a performance boost over models that rely solely on clinical scales or simple concatenation. The study confirms that while MRI is not the primary diagnostic driver in this cohort, it can contribute meaningful, clinically relevant information when integrated via a constrained, temporal-aware architecture.