Razi University

This study examines the interpretive challenges in functional magnetic resonance imaging (fMRI), where the blood oxygen level-dependent (BOLD) signal serves as an indirect hemodynamic indicator of neuronal activity and is subject to vascular-neuronal coupling and technical limitations. The aim of this study was to develop and evaluate a reproducible computational framework for estimating hidden neural activity from hemodynamic response function (HRF) modeling and regularized deconvolution.The research methodology involved analyzing public fMRI data from five subjects, comparing preprocessing pipelines (GSR vs. no GSR), and evaluating different predictive models. The results showed that HRF-based deconvolution produced internally consistent hidden estimates with high low-frequency coherence. It is worth noting that nonlinear and machine learning models (e.g., SVR and LSTM) did not outperform the tra  arent linear baseline; The general linear model (GLM) achieved the highest prediction accuracy with a mean root mean square error (RMSE) of 0.1439. An important finding was that “sparse events”—which represent only 15% of the high-amplitude BOLD frames—preserved the functional connectivity structure of the entire time series with a high similarity of 0.89. This paper concludes that although explicit HRF modeling and sparse event analysis enhance the interpretation of the BOLD signal, the estimated signals remain model-derived and should not be considered as direct electrophysiological ground truth.Keywords: functional magnetic resonance imaging; BOLD signal; hemodynamic response function; deconstruction; sparse event connectivity; machine learning