Increasing the Reliability of Functional Connectivity by Predicting Long-Scan Functional Connectivity based on Short-Scan Functional Connectivity: Model Exploration, Explanation, Validation, and Application.
Authors
Affiliations (3)
Affiliations (3)
- Department of Radiology, The Air Force Hospital From Eastern Theater of PLA, Malu Road, Nanjing, China.
- Department of Radiology, Functional and Molecular Imaging Key Lab of Shaanxi Province, Tangdu Hospital, Fourth Military Medical University, Xi'an, China.
- Department of Health Management Center (HMC), The Air Force Hospital From Eastern Theater of PLA, Malu Road, Nanjing, China. [email protected].
Abstract
Functional connectivity (FC) is a widely used metric in functional magnetic resonance imaging (fMRI) research. However, its reliability has long been a notable concern, especially in studies with small sample sizes. Previous research has demonstrated that FC derived from longer fMRI scans exhibits higher reliability, making the prediction of long-scan FC from existing short-scan FC data a feasible and promising approach. First, we constructed three general linear models (GLMs) to predict long-scan FC from short-scan FC data using the Human Connectome Project (HCP) dataset. Next, we interpreted the models by visualizing their weight distributions. Subsequently, we validated our findings across multiple independent datasets and with different machine learning models. Finally, we applied the models to enhance both the test-retest reliability of FC and the performance of connectome-based predictive modeling (CPM). Our results showed that GLMs based on individual short-scan FC successfully predicted individual long-scan FC values. Moreover, the differences between the three GLMs could be explained by the distinct distribution characteristics of the FC matrices they predicted. Our findings were validated using data from the Consortium for Reliability and Reproducibility (CoRR) project and an in-house local dataset. Additionally, our models outperformed conventional machine learning approaches. Critically, these models effectively improved both the test-retest reliability of FC and the predictive performance of CPM. In conclusion, GLMs built on individual short-scan FC can robustly predict individual long-scan FC values. These models show strong generalizability across different datasets, and can be widely applied to improve the test-retest reliability of FC and the performance of CPM in neuroimaging studies.