NBS-Predict

NBS-Predict identifies neuroimaging-based biomarkers and predicts behavioral or clinical outcomes by integrating machine learning with graph-theoretical network-based statistics (NBS) within a cross-validation framework applied to rs-fMRI connectivity data.

Key Features:

• Integration of Machine Learning and NBS: Combines machine learning models with network-based statistics (NBS) and graph-theoretical inference to perform statistical inference on brain graphs and control family-wise error rate in mass univariate analyses via cluster-based permutation techniques.
• Cross-Validation Framework: Implements cross-validation to assess and enhance generalizability of biomarker selection and predictive models across datasets.
• Subnetwork Identification from rs-fMRI: Identifies subnetworks and connectivity features from resting-state functional magnetic resonance imaging (rs-fMRI) connectivity matrices for feature selection and localization of effects.
• Performance Evaluation and Benchmarking: Evaluated on simulated datasets with known ground truths and on real rs-fMRI data from the Human Connectome Project 1200-subject release and benchmarked against lasso, elastic net, top 5%, p-value thresholding, and connectome-based predictive modeling (CPM).

Scientific Applications:

• Precision Medicine: Enables individualized inference by providing subject-level biomarker selection that complements traditional group-level statistics for precision medicine applications.
• Schizophrenia Case-Control Classification: Applied to rs-fMRI case-control data for schizophrenia, achieving reported classification accuracy of 90% and identifying subnetworks with reduced connectivity in frontotemporal, visual, motor regions and the subcortex.
• General Intelligence Prediction: Predicts general intelligence scores from rs-fMRI connectivity matrices with a reported prediction correlation of r = 0.2 and identifies a large-scale subnetwork associated with general intelligence.

Methodology:

Integrates machine learning with graph-theoretical network-based statistics (NBS) within a cross-validation framework, uses cluster-based permutation to control family-wise error rate in mass univariate analyses, and is evaluated on simulated ground-truth datasets and Human Connectome Project (1200 subjects) rs-fMRI data with benchmarking against lasso, elastic net, top 5%, p-value thresholding, and CPM.