Differentiating Alcohol and Substance Use Disorders Using Multiclass Machine Learning Models Based on Routine...

Aim: The identification of alcohol and substance use disorders is primarily based on clinical evaluation, and the lack of accessible objective markers remains a challenge. This study aimed to explore whether routine hemogram parameters, analyzed using multiclass machine learning models, could assist in differentiating individuals with alcohol use disorder, substance use disorder, and healthy controls. Method: This retrospective case-control study included 35 patients with alcohol use disorder, 61 patients with substance use disorder, and 132 healthy controls. Routine hematological parameters were obtained from hospital records. Multiclass classification models, including Random Forest, Support Vector Machine (SVM), Artificial Neural Network (ANN), and other conventional machine learning algorithms, were applied. Model performance was evaluated using 10-fold cross-validation with metrics including accuracy, sensitivity, precision, F1-score, and AUC. Results: Significant differences were observed among groups in several hematological parameters, including monocyte count, basophil count, and RDW-CV (p < 0.05). Among the models, Random Forest achieved the highest overall accuracy (81.6%) and AUC (0.93), followed by SVM and ANN models with comparable performance. However, classification performance was not uniform across all classes, and sensitivity was relatively lower for the alcohol use disorder group compared to the control and substance use disorder groups. Conclusions: These findings suggest that machine learning models based on routine hemogram parameters may provide a preliminary, low-cost, and accessible supportive approach for differentiating addiction-related conditions. However, the results should be interpreted as exploratory, given the retrospective single-center design, limited sample size, lack of external validation, and absence of model interpretability analyses. Further studies incorporating larger multicenter datasets, confounding factors, and explainable artificial intelligence methods are required before clinical application can be considered.