Disease prediction based on functional connectomes using a scalable and spatially-informed support vector machine.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Major psychiatric disorders are linked to widespread neural dysconnectivity.
• Neuroimaging methods are crucial for predicting disorder status.
• High-dimensional functional connectomes present interpretation and computational challenges.

Purpose of the Study:

• To develop a multivariate approach for predicting psychiatric disorders using whole-brain resting state functional connectomes.
• To address the limitations of traditional feature selection methods by incorporating the 6-D structure of functional connectomes.
• To enable interpretable identification of predictive brain regions.

Main Methods:

• Proposed a regularization framework using fused Lasso or GraphNet regularizers to account for the 6-D structure of functional connectomes.
• Developed an efficient optimization algorithm based on augmented Lagrangian and alternating direction method for solving structured sparse support vector machines (SVMs).
• Utilized hinge-loss functions with convex and margin-based loss functions.

Main Results:

• The proposed method successfully identifies spatially contiguous predictive regions within the 6-D connectome space.
• Experiments on simulated data and a schizophrenia dataset demonstrated the effectiveness of the approach.
• The method offers enhanced interpretability for understanding disease processes.

Conclusions:

• The novel regularization framework provides a powerful and interpretable method for analyzing functional connectomes in psychiatric disorders.
• This approach can lead to new insights into the neural underpinnings of various brain diseases.
• The efficient optimization algorithm facilitates the application of this method to large-scale neuroimaging datasets.