Genetic algorithms for feature selection when classifying severe chronic disorders of consciousness

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Diagnosing and predicting outcomes for patients with severe chronic disorders of consciousness (DoC) remains a significant clinical challenge, with a high incidence of misdiagnosis.
• Novel diagnostic tools are crucial for improving accuracy and consequently, patient prognosis.
• Functional Magnetic Resonance Imaging (fMRI), particularly resting-state scans, has emerged as a promising non-invasive technique for assessing brain function in DoC patients.

Purpose of the Study:

• To identify specific brain regions of interest (ROIs) that are most effective in distinguishing between different patient groups (e.g., vegetative state, minimally conscious state) and healthy controls.
• To evaluate the utility of resting-state fMRI data in conjunction with advanced computational methods for improved DoC diagnosis.
• To optimize diagnostic performance by focusing on the most discriminative ROIs identified through feature selection.

Main Methods:

• Preprocessing of resting-state fMRI data using a standard pipeline.
• Extraction of correlation matrices from 132 predefined regions of interest (ROIs).
• Application of a genetic algorithm and a support vector machine (SVM) for feature selection to identify the most relevant ROIs for classification.

Main Results:

• Feature selection successfully identified a subset of ROIs that significantly contribute to differentiating between patient groups and healthy individuals.
• Classification models trained using only the most frequently selected ROIs demonstrated substantially improved performance compared to models using all ROIs.
• The study highlights the potential of specific brain network features derived from resting-state fMRI for accurate DoC assessment.

Conclusions:

• Resting-state fMRI, when analyzed with advanced feature selection techniques like genetic algorithms and SVM, offers a powerful tool for improving the diagnosis of severe chronic disorders of consciousness.
• Identifying and utilizing the most informative brain regions can enhance diagnostic accuracy and potentially refine prognostic predictions.
• This approach holds promise for clinical application, aiding in more precise patient stratification and management.