Comparing four algorithms in predicting the risk of driving under the influence of alcohol among individuals with alcohol use disorder

Driving under the influence of alcohol (DUIA) is closely associated with alcohol use disorder (AUD) because of significant impacts of alcoholism on brain functions related to response disinhibition and impaired cognitive control. Although previous studies used decision trees (DT) with neuropsychological features to predict the risk of substance use, no investigation utilized machine learning (ML) with neuropsychological outcomes to predict DUIA. Psychological research usually involves small sample sizes in regional settings, whereas ML algorithms are often applied to big data. Thus, it is unclear which ML models are suitable for small sample sizes. Consequently, the objective of this study was to compare the DT with other well-known ML algorithms in predicting AUD-related DUIA, especially with limited samples. Between July 2022 and June 2023, 27 AUD adults (16 DUIA vs. 11 non-DUIA) were recruited from a single tertiary referral center. Fourteen social drinkers served as controls. Based on the two labeled features in response inhibition (RI) task and four in decision-making task, comparisons between the DT, LR, support vector machine (SVM), and k-nearest neighbors (kNN) were conducted. The participants with AUD exhibited excessive alcohol consumption and higher impulsivity compared to controls. Furthermore, the RI was superior to decision-making in differentiating between different groups. The DT algorithm with RI features offered a higher predictability of DUIA compared to the LR, SVM, and kNN. The findings highlighted the significance of RI in detecting the risk of DUIA, while the DT may be an acceptable algorithm in situations involving a limited number of samples.