Alzheimer's Disease Prediction Using EffNetViTLoRA and BiLSTM with Multimodal Longitudinal MRI Data

TL;DR

This work tackles the challenge of predicting progression from Mild Cognitive Impairment to Alzheimer's disease by leveraging longitudinal MRI and non-imaging biomarkers. It introduces a two-phase, end-to-end architecture that combines EffNetViTLoRA-based feature extraction with a BiLSTM predictor to model four timepoints and forecast status at 48 months. The approach achieves a peak accuracy of 95.05% in distinguishing stable vs progressive MCI, outperforming prior longitudinal prediction studies, and is shown to benefit from integrating imaging with neuropsychological biomarkers and bidirectional temporal modeling. The results highlight the value of fusing local and global image features with temporal dynamics for early AD detection and potential clinical utility, with open-source code to enable replication and further research.

Abstract

Alzheimer's disease (AD) is a prevalent neurodegenerative disorder that progressively impairs memory, decision-making, and overall cognitive function. As AD is irreversible, early prediction is critical for timely intervention and management. Mild Cognitive Impairment (MCI), a transitional stage between cognitively normal (CN) aging and AD, plays a significant role in early AD diagnosis. However, predicting MCI progression remains a significant challenge, as not all individuals with MCI convert to AD. MCI subjects are categorized into stable MCI (sMCI) and progressive MCI (pMCI) based on conversion status. In this study, we propose a generalized, end-to-end deep learning model for AD prediction using MCI cases from the Alzheimer's Disease Neuroimaging Initiative (ADNI). Our hybrid architecture integrates Convolutional Neural Networks and Vision Transformers to capture both local spatial features and global contextual dependencies from Magnetic Resonance Imaging (MRI) scans. To incorporate temporal progression, we further employ Bidirectional Long Short-Term Memory (BiLSTM) networks to process features extracted from four consecutive MRI timepoints along with some other non-image biomarkers, predicting each subject's cognitive status at month 48. Our multimodal model achieved an average progression prediction accuracy of 95.05\% between sMCI and pMCI, outperforming existing studies in AD prediction. This work demonstrates state-of-the-art performance in longitudinal AD prediction and highlights the effectiveness of combining spatial and temporal modeling for the early detection of Alzheimer's disease.

Figures (7)

• Figure 1: Diagram of the proposed prediction method, consisting of two phases. Phase 1 is the EffNetViTLoRA, which serves as a feature extractor. The representative features of the MCI group are obtained from the last hidden layer, and 17 additional non-image biomarkers are appended to each image feature vector to construct the multimodal prediction data. Phase 2 is the prediction phase, which utilizes the features obtained from Phase 1 in combination with the biomarkers to predict Alzheimer’s disease progression based on follow-up data four years after the baseline visit.
• Figure 2: Architecture of the hybrid feature extraction model. The model captures both local and global dependencies in input MRI images by combining EfficientNetV2 with a pretrained ViT. The ViT is fine-tuned using LoRA applied to the Key, Query, and Value weight matrices.
• Figure 3: Internal structure of an LSTM cell.
• Figure 4: Sample axial slices from ADNI for sMCI and pMCI subjects over an 18-month follow-up. Rows (a) and (b) show sMCI cases, while rows (c) and (d) show pMCI cases. Subtle shrinkage in the gray matter and hippocampal regions can be observed across the four time points in the pMCI rows.
• Figure 5: Preprocessing steps of ADNI MRI T1-weighted volumes in our study. The arrows demonstrate the flow of the process.