NBS-Predict: An Easy-to-Use Toolbox for Connectome-Based Machine Learning

Abstract

NBS-Predict is a prediction-based extension of the Network-based Statistic (NBS) approach, which aims to alleviate the curse of dimensionality, lack of interpretability, and problem of generalizability when analyzing brain connectivity. NBS-Predict provides an easy and quick way to identify highly generalizable neuroimaging-based biomarkers by combining machine learning (ML) with NBS in a cross-validation structure. Compared with generic ML algorithms (e.g., support vector machines, elastic net, etc.), the results from NBS-Predict are more straightforward to interpret. Additionally, NBS-Predict does not require any expertise in programming as it comes with a well-organized graphical user interface (GUI) with a good selection of ML algorithms and additional functionalities. The toolbox also provides an interactive viewer to visualize the results. This chapter gives a practical overview of the NBS-Predict’s core concepts with regard to building and evaluating connectome-based predictive models with two real-world examples using publicly available neuroimaging data. We showed that, using resting-state functional connectomes, NBS-Predict: (i) predicted fluid intelligence scores with a prediction performance of r = 0.243; (ii) distinguished subjects’ biological sexes with an average accuracy of 65.9%, as well as identified large-scale brain networks associated with fluid intelligence and biological sex.