Longitudinal Alzheimer's Disease Classification from 2.5D MRI and Clinical Data via Adaptive Gated Fusion

Reliable discrimination between cognitively stable and impaired individuals remains a clinically important challenge in Alzheimer's disease (AD) assessment. Existing multimodal deep learning approaches have predominantly relied on computationally intensive 3D convolutional networks, while relatively few studies jointly model longitudinal imaging and non-imaging data or explicitly account for patient-specific modality contributions. In this work, a multimodal longitudinal framework is proposed that combines 2.5D MRI feature extraction with bidirectional LSTM-based longitudinal sequence encoding. A residual gated fusion mechanism adaptively weights imaging and clinical features for each patient, enabling joint learning of longitudinal and multimodal representations. Structural MRI and clinical data from 204 patients across three longitudinal visits from the Alzheimer's Disease Neuroimaging Initiative (ADNI) were evaluated using patient-level 5-fold cross-validation to prevent information leakage across visits. The proposed framework achieved a mean AUC of 0.870 ± 0.035, balanced accuracy of 0.802, sensitivity of 0.756, and specificity of 0.847, consistently outperforming unimodal baselines and alternative architectural variants. Ablation studies demonstrated that longitudinal modeling, the 2.5D representation, and residual gated fusion each contributed independently to the overall performance. Quantitative Grad-CAM analysis further revealed cross-visit spatial consistency and significantly greater medial activation in cognitively impaired individuals than in stable individuals, indicating that the model focuses on neuroanatomically plausible regions associated with neurodegenerative change. Our work underscores the potential of multimodal longitudinal deep learning to improve AD assessment while leveraging routinely collected imaging and clinical data.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-18
DOI
https://doi.org/10.21203/rs.3.rs-11605408/v1
Primary Topic
Dementia and Cognitive Impairment Research
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Longitudinal Alzheimer's Disease Classification from 2.5D MRI and Clinical Data via Adaptive Gated Fusion

Sowmya Kamath S., Malathi S. Bhat, Santosh Rai P V, Sharvari Hemant Mohokar
4 links
Research Square
Dementia and Cognitive Impairment Research
preprint

Longitudinal Alzheimer's Disease Classification from 2.5D MRI and Clinical Data via Adaptive Gated Fusion

Sowmya Kamath S., Malathi S. Bhat, Santosh Rai P V, Sharvari Hemant Mohokar
preprint en

Abstract

Reliable discrimination between cognitively stable and impaired individuals remains a clinically important challenge in Alzheimer's disease (AD) assessment. Existing multimodal deep learning approaches have predominantly relied on computationally intensive 3D convolutional networks, while relatively few studies jointly model longitudinal imaging and non-imaging data or explicitly account for patient-specific modality contributions. In this work, a multimodal longitudinal framework is proposed that combines 2.5D MRI feature extraction with bidirectional LSTM-based longitudinal sequence encoding. A residual gated fusion mechanism adaptively weights imaging and clinical features for each patient, enabling joint learning of longitudinal and multimodal representations. Structural MRI and clinical data from 204 patients across three longitudinal visits from the Alzheimer's Disease Neuroimaging Initiative (ADNI) were evaluated using patient-level 5-fold cross-validation to prevent information leakage across visits. The proposed framework achieved a mean AUC of 0.870 ± 0.035, balanced accuracy of 0.802, sensitivity of 0.756, and specificity of 0.847, consistently outperforming unimodal baselines and alternative architectural variants. Ablation studies demonstrated that longitudinal modeling, the 2.5D representation, and residual gated fusion each contributed independently to the overall performance. Quantitative Grad-CAM analysis further revealed cross-visit spatial consistency and significantly greater medial activation in cognitively impaired individuals than in stable individuals, indicating that the model focuses on neuroanatomically plausible regions associated with neurodegenerative change. Our work underscores the potential of multimodal longitudinal deep learning to improve AD assessment while leveraging routinely collected imaging and clinical data.

Research Square
National Institute of Technology Karnataka (IN), Manipal Academy of Higher Education (IN), Kasturba Medical College, Mangalore (IN)
Good health and well-being
Dementia and Cognitive Impairment Research
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon