Combining sparseness and smoothness improves classification accuracy and interpretability.

Area of Science:

• Neuroimaging analysis
• Machine learning for neuroscience

Background:

• Sparse logistic regression (SLR) is used for decoding high-dimensional fMRI and MEG data.
• SLR can over-prune features in spatio-temporally correlated signals, leading to overfitting and uninterpretable weights.

Purpose of the Study:

• To introduce smooth sparse logistic regression (SSLR) to address limitations of SLR in high-dimensional neuroimaging data.
• To improve feature selection by encouraging spatio-temporal continuity.

Main Methods:

• Developed a modified sparse logistic regression (SLR) with a spatio-temporal smoothing prior, termed SSLR.
• Applied SSLR to simulated and real MEG data.

Main Results:

• SSLR enhances classification accuracy compared to standard SLR.
• SSLR produces more interpretable weight vectors, reflecting neuroscientific insights.
• SSLR selects continuous spatio-temporal features, unlike the scattered features selected by SLR.

Conclusions:

• SSLR offers a more effective approach for decoding complex neuroimaging data.
• The method improves both the predictive performance and the interpretability of machine learning models in neuroscience.
• SSLR's spatio-temporal smoothing is crucial for analyzing signals with high correlations.