Autoregressive ConvLSTM Framework for fMRI Time Series Forecasting in Alzheimer’s Disease

Ahmed Alshembari, Anima Kujur, Zahra Monfared

Keywords: LSTM, fMRI, Time Series, Alzheimer Disease, ConvLSTM, Deep Learning

TL;DR: We propose a ConvLSTM-based autoregressive framework with a custom loss that predicts future fMRI brain states in Alzheimer’s disease more accurately and robustly than LSTM, or CNN-LSTM baselines.

Abstract: Resting-state functional magnetic resonance imaging (fMRI) provides a noninvasive window into brain dynamics and has emerged as a powerful tool for studying neurodegenerative disorders. We develop an autoregressive deep learning framework that employs convolutional long short-term memory (ConvLSTM) units to forecast future brain states in resting-state fMRI sequences from patients with Alzheimer's disease (AD). Unlike traditional linear autoregressive models or hybrid CNN-LSTM approaches, which often ignore spatial structure or flatten brain images, our method integrates convolution directly into the LSTM gates. This design reduces the number of parameters and maintains spatial coherence, preserving the intrinsic 2D structure of brain images while capturing temporal dependencies. To enhance prediction quality, we introduce a custom loss function that jointly optimizes mean squared error and structural similarity index. Experiments on the ADNI fMRI dataset demonstrate that our model generates high-fidelity brain state predictions and achieves substantial performance gains over pure LSTM, and CNN-LSTM baselines. Cross-validation further confirms the robustness of our approach across subjects, which highlights its potential for early biomarker discovery and disease progression monitoring in AD.

Submission Number: 54