Successful classification of cocaine dependence using brain imaging: a generalizable machine learning approach

Area of Science:

• Neuroscience
• Radiology
• Computer Science

Background:

• Neuroimaging advances understanding addiction neural processes.
• Diagnostic accuracy of these advances in humans remains underexplored.
• Cocaine dependence impacts neural processes, necessitating accurate diagnostic tools.

Purpose of the Study:

• Develop a machine learning framework for classifying cocaine-dependent individuals and healthy controls using brain imaging.
• Identify specific brain regions that differ between cocaine-dependent participants and controls.
• Enhance diagnostic capabilities for substance use disorders through advanced statistical analysis.

Main Methods:

• Utilized Single Photon Emission Computerized Tomography (SPECT) data from 93 cocaine-dependent participants and 69 healthy controls.
• Employed an information theoretic-based feature selection algorithm to reduce voxel count.
• Applied density-based clustering and Support Vectors Machine (SVM) for classification and spatial analysis.

Main Results:

• Successfully classified participants with F-measure accuracies of 0.88 (10-fold cross-validation) and 0.89 (leave-one-out).
• Identified 1,500 voxels in 30 distinct clusters, many relevant to cognitive control and self-referential thought.
• Achieved high sensitivity (0.90) and specificity (0.89) in the leave-one-out approach.

Conclusions:

• The SVM-based approach effectively classified cocaine dependence using SPECT data.
• Identified brain regions align with existing literature on addiction.
• Supports the future use of brain imaging and SVM classifiers for diagnosing substance use disorders and understanding addiction pathology.