Clinical prediction from structural brain MRI scans: a large-scale empirical study

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Biomedical Data Science

Background:

• Multivariate pattern analysis (MVPA) is crucial in neuroimaging for identifying complex brain associations and building predictive models.
• A lack of standardized benchmark results hinders the comparison and referencing of MVPA methodologies in research.
• This study addresses the need for comprehensive benchmarks in structural neuroimaging prediction.

Purpose of the Study:

• To establish a large-scale, reproducible benchmark for image-based prediction in structural neuroimaging.
• To evaluate the performance of state-of-the-art MVPA algorithms across diverse clinical and demographic variables.
• To identify key factors influencing prediction accuracy in neuroimaging studies.

Main Methods:

• Utilized three classes of advanced MVPA algorithms applied to structural Magnetic Resonance Imaging (MRI) data.
• Analyzed data from over 2,800 subjects across six publicly available datasets.
• Predicted clinical variables including diagnoses (Alzheimer's, schizophrenia, autism, ADHD), age, and cognitive scores.

Main Results:

• Prediction performance is strongly influenced by the biological effect size.
• The selection of image measurement type had a greater impact on prediction accuracy than the choice of MVPA algorithm.
• Cross-validation performance estimates generally correlate well with generalization accuracy, though they tend to be optimistic.

Conclusions:

• Established the largest, most comprehensive, and reproducible benchmark for structural neuroimaging prediction.
• Highlighted the critical role of biological effect size and image measurement selection in predictive modeling.
• Provided insights into the generalizability of cross-validation results and suggested future research directions in neuroimaging prediction.