Reconstructing Heart Rate Variation from Bold-FMRI Using 1D-CNN and GRU Networks with Dynamic Functional Connectivity-Based ROIs

This paper introduces a machine learning model to estimate Heart Rate Variation (HRV) directly from blood oxygen level-dependent (BOLD) functional magnetic resonance imaging (fMRI) signals, addressing the limitations of traditional ex-ternal cardiac monitoring devices. Our proposed method leverages a hybrid architecture combining one-dimensional Convolutional Neural Networks (1D-CNN) and Gated Recur-rent Units (GRU) to process BOLD signals from 628 regions of interest (ROIs) spanning cortical, subcortical, and white matter regions. Cortical and subcortical ROIs were derived from the dynamic functional connectivity-based Brain Functional Atlas (D-BFA), while white matter ROIs were obtained from the Johns Hopkins University Diffusion Tensor Imaging (JHU DTI)-based atlas. The integration of a dynamic functional atlas and white matter ROIs, working in concert wtih the machine learning network, improved performance metrics-Mean Absolute Error (MAE), Mean Squared Er-ror (MSE), correlation coefficient (r), and Dynamic Time Warping (DTW)- by 7%, compared to using a static atlas, excluding white matter, or using CNN or GRU alone. The approach demonstrates promise for HRV estimation directly from fMRI, particularly when external monitoring is unreli-able or unavailable.