Discrimination of alcoholics from non-alcoholics using supervised learning on resting EEG

Abstract

Alcoholism is a widespread problem that can have serious medical consequences. Alcoholism screening tests are used to identify patients who are at risk for complications from alcohol abuse, but accurate diagnosis of alcohol dependence must be done by structured clinical interview. Scalp Electroencephalography (EEG) is a noisy, non-stationary signal produced by an aggregate of brain activity from neurons close to the scalp. Previous research has identified a relationship between information extracted from resting scalp EEG and alcoholism, but it has not been established if this relationship is strong enough to have meaningful diagnostic utility. In this thesis, we investigate the efficacy of using supervised machine learning on resting scalp EEG data to build models that can match clinical diagnoses of alcohol dependence. We extract features from four minute eyes-closed resting scalp EEG recordings, and use these features to train discriminative models for identifying alcohol dependence. We found that we can achieve an average AUROC of .65 in males, and .63 in females. These results suggest that a diagnostic tool could use scalp EEG data to diagnose alcohol dependence with better than random performance. However, further investigation is required to evaluate the generalizability of our results.