Systematic Review on Interpretability in Computer-Aided Alcohol Use Disorder Diagnosis

Numerous studies have investigated the detection of AUD using EEG signals. However, the real-world implementation of diagnostic systems is hindered by users' difficulty in understanding the rationale behind a decision. This study presents a systematic literature review based on PRISMA guidelines, utilizing journal and conference articles from 2019 to 2023 obtained from Scopus. Among 66 collected studies, feature engineering techniques dominated by EEG sub band decomposition processes, followed by time-domain feature extraction, EEG conversion to images, and brain connectivity features. With advancements in computation, comprehensive hybrid features and raw EEG data are emerging as inputs for classifiers. Some studies also use dimensionality reduction techniques such as statistical analysis, optimization methods, feature reduction methods, and manual selection. Moreover, while black box classifier models are widely used, intrinsically interpretable classifiers exhibit no significant differences in accuracy performance. Future work aims to develop AUD detection systems with enhanced interpretability while maintaining robust classification performance, utilizing features informed by experts' domain knowledge to facilitate decision explanation. © 2024 IEEE.