Combining Multiple Resting-State fMRI Features during Classification: Optimized Frameworks and Their Application to...

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Machine learning models are increasingly used to predict disease states from resting-state functional magnetic resonance imaging (fMRI) data.
• Previous studies often utilized limited feature types (<4), leaving the optimal feature selection for classification uncertain.
• Resting-state fMRI data offers diverse feature types from local and network measures that may contain complementary information.

Purpose of the Study:

• To investigate the effectiveness of combining multiple resting-state fMRI features for improved disease classification.
• To compare different feature combination frameworks (feature, kernel, and classifier combination) using support vector machines.
• To identify the most informative resting-state features for predicting disease states.

Main Methods:

• Calculated multiple resting-state fMRI features from local and network measures.
• Employed an optimized grid-search approach for feature selection based on statistical tests.
• Tested three optimized frameworks: feature combination, kernel combination, and classifier combination, utilizing support vector machines.
• Validated the approach on a cohort of 100 smokers and 100 non-smokers using 10-fold cross-validation.

Main Results:

• Feature combination and classifier combination frameworks achieved 75.5% accuracy in predicting nicotine addiction.
• The kernel combination framework achieved 73.0% accuracy.
• All combination frameworks demonstrated improved classification performance compared to single-feature methods (best single-feature accuracy: 70.5%).

Conclusions:

• Combining multiple resting-state fMRI features significantly enhances classification performance for predicting disease states.
• The study highlights the discriminative power of resting-state fMRI data and the efficacy of multimodal feature integration.
• This approach offers a promising strategy for improving diagnostic accuracy in neurological and neuropsychiatric disorders.