Hybrid CNN-GRU model improves estimation of respiratory and heart rate variations from resting-state BOLD fMRI across the lifespan

A 1D-CNN+GRU deep learning model reduced error in respiratory and heart rate variation estimates from resting-state BOLD fMRI by 7–10% (MAE, MSE, DTW) versus established baselines. Evaluated on three HCP cohorts spanning ages 5–100 years, the model leveraged 630 brain regions an…

• The model processed BOLD signals from 630 cortical, subcortical, white matter, and CSF regions of interest, with six head motion parameters added for respiratory variation estimation.
• Age-specific architectural tuning was applied across the three Human Connectome Project cohorts: Development (5–21 y), Young Adult (22–35 y), and Aging (36–100 y).
• Gains were largest for absolute error and temporal alignment metrics; correlation improvements were more modest (approx. 5–7%).

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.