Preprocessing Variability in fMRI Predictive Modeling: Effects of Distortion Correction on Functional Connectivity-Based Predictions

Functional connectivity (FC)-based predictive modeling is a widely used approach in resting-state fMRI studies to predict various mental states such as attention, depression, and impulsivity. FC-based predictive modeling often employs a standard procedure: parcellating the brain into regions of interest (ROIs) using a predefined atlas, computing ROI-to-ROI FC, and applying predictive models to estimate behavioral or clinical measures. However, many existing studies focus solely on end-to-end prediction performance and often overlook the influence of preprocessing choices on FC features and downstream predictive model performance. Assessing the preprocessing effect is crucial because it can significantly influence the spatial accuracy of ROI partition, FC measures, and predictive modeling performance, potentially reducing reproducibility. In this study, we investigated the impact of fMRI preprocessing strategies, particularly fieldmap distortion correction, on the resulting FC features and the performance of machine learning models predicting sensation-seeking. We compared two preprocessing pipelines: with distortion correction (DC) and without (NDC). FC matrices were computed from each pipeline and used to train machine learning models to predict sensation-seeking trait. We showed that different preprocessing choices can lead to substantial differences in FC values and model predictions. The prediction model trained on DC data achieved R² of 0.34, while the model trained on NDC data has a lower R² of 0.21. Moreover, we observed notable differences in the key predictive connections between the DC and NDC pipelines, involving the brain regions such as cerebellum, prefrontal cortex, cingulate cortex, and subcortical regions, which also showed the largest voxel shifts following distortion correction. Our findings revealed the important role that preprocessing strategies play in functional connectivity-based modeling and raised the important issue of accounting for preprocessing variability in fMRI predictive modeling.