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

Jul 16, 2026

Upload by Eden

Temporal-aware multimodal Alzheimer’s disease diagnosis support is challenged by irregular clinical visits, noisy or intermittently missing MRI, and the risk that naive fusion lets weak modalities distort stronger clinical evidence. AT-Attn introduces Change-and-Time encoding, time-biased asymmetric cross-attention, and gated fusion to integrate structural MRI with longitudinal cognitive trajectories and static clinical variables. Evaluated on an MRI-retained ADNI cohort of 1,520 patients with patient-level five-fold cross-validation, AT-Attn reaches accuracy 0.719±0.024 and ROC-AUC 0.873±0.013, outperforming unimodal and naive multimodal fusion while remaining competitive with strong tabular baselines.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages9

Reading Time23 min

Introduction

Related Work

Clinical and Biomarker Background

Method Overview

AT-Attn: Temporal-Aware Cross-Attention for Longitudinal Multimodal Alzheimer’s Disease Diagnosis Xinyue Du∗ , Yibo Liu†, Zhenglei Zhou‡, Xuancheng Yao†, Weimin Zhong∗ , Qiuhui Chen∗§ ∗ School of Information Science and Engineering, East China University of Science and Technology, China Email: [xiaoduxiaodu09@gmail.com](xiaoduxiaodu09@gmail.com) †School of Computer Science, Shanghai Jiao Tong University, China ‡Tencent, China § Corresponding author: Qiuhui Chen ([chenqh@ecust.edu.cn](chenqh@ecust.edu.cn)) arXiv :2607 .0709 1v 1 [ cs .CV] 8 Jul 2026 Abstract—In longitudinal Alzheimer’s disease (AD) diagnosis support, clinical and imaging information is often collected at irregular visits. Integrating these multimodal observations may improve diagnostic assessment, but naive fusion can degrade performance when MRI is noisy or intermittently unavailable. We propose AT-Attn, a temporal-aware multimodal framework that combines Change-and-Time encoding, time-biased asymmetric cross-attention, and gated fusion to integrate MRI with longitudinal clinical information. We evaluate AT-Attn on an MRIretained ADNI cohort of 1,520 patients using structural MRI, six cognitive-scale trajectories, and seven static clinical variables under patient-level five-fold cross-validation. The main asymmetric AT-Attn model achieves accuracy 0.719±0.024, macro F1 0.721±0.023, ROC-AUC 0.873±0.013, and PR-AUC 0.783±0.018, outperforming unimodal and naive multimodal fusion baselines while remaining competitive with strong tabular baselines. These results suggest that a temporal-aware and constrained fusion strategy can help structural MRI contribute clinically relevant complementary information for patient-level AD diagnosis support. Index Terms—Alzheimer’s disease, multimodal fusion, longitudinal data, cross-attention, temporal modeling I. INTRODUCTION Alzheimer’s disease (AD) is the leading cause of dementia and is characterized by a progressive transition from cognitively normal (CN) status to mild cognitive impairment (MCI) and eventually dementia [1] . Accurate identification of MCI and current disease stage is clinically important because it can support timely intervention, longitudinal monitoring, and more informed patient management [2], [3] . In both routine clinical workflows and longitudinal cohort studies such as the Alzheimer’s Disease Neuroimaging Initiative (ADNI), patient assessment is inherently multimodal [4] . Structural MRI captures brain atrophy and neurodegenerative patterns [5], cognitive and functional assessments reflect clinical impairment, and demographic or genetic factors such as APOE-ε4 provide complementary risk information [6] . Developing computational methods that can effectively combine these heterogeneous signals therefore remains an important biomedical informatics problem. However, the multimodal setting relevant to practice isnot simply “more modalities are better.” In AD cohorts, cognitive assessments are clinically direct, whereas structural MRI is high-dimensional, noisier, and often unavailable at some visits. MRI may provide complementary information, but naive fusion can also degrade performance by allowing the weaker modality to distort stronger clinical signals. For longitudinal AD diagnosis, this problem is further complicated by irregular follow-up intervals, meaning that cross-modal interaction should account for temporal proximity [7], [8],[9] . These observations motivate not only temporal-aware fusion, but a constrained fusion mechanism that can inject MRI information in a stable and complementary way while preserving direct clinical evidence. To address this problem, we propose AT-Attn (TemporalAware Cross-Attention), a longitudinal multimodal framework for patient-level AD diagnosis support that combines modality-specific encoding, Change-and-Time encoding, timebiased asymmetric cross-attention, and gated fusion. AT-Attn lets MRI representations query cognitive-scale represen

FAQs

What problem does AT-Attn address in longitudinal Alzheimer’s disease diagnosis?

It targets multimodal AD diagnosis when clinical visits are irregular and MRI can be noisy or missing, where naive fusion may degrade performance by distorting stronger clinical signals.

How does AT-Attn model temporal information across modalities?

It uses Change-and-Time encoding and applies time-biased asymmetric cross-attention so MRI representations query cognitive-scale representations with an explicit temporal penalty.

What dataset and evaluation protocol are used to assess AT-Attn?

AT-Attn is evaluated on an MRI-retained ADNI cohort with structural MRI, cognitive trajectories, and static clinical variables using patient-level five-fold cross-validation.

Jul 16, 2026

Upload by Eden