Prediction Model for Metabolic Syndrome Comorbidity in Alcohol-Dependent Patients Based on Machine Learning

doi:10.12104/j.issn.1674-4748.2025.23.020

Abstract：Objective: To explore the influencing factors of metabolic syndrome (MS) comorbidity in alcohol-dependent patients and to construct a prediction model for MS comorbidity in alcohol-dependent patients based on machine learning algorithms. Methods: A prospective study was conducted on 276 alcohol-dependent patients treated in the hospital from April 2023 to August 2024. The patients were divided into an MS group (n=75) and a non-MS group (n=201) according to whether they had MS comorbidity. Single-factor analysis was used to identify the influencing factors of MS comorbidity in alcohol-dependent patients. Multivariate logistic regression, classification regression tree (CRT), and back propagation neural network (BPNN) machine learning algorithms were used to build prediction models for MS comorbidity in alcohol-dependent patients using SPSS software. The predictive value of the three models for MS was compared using receiver operating characteristic (ROC) curves. Results: The single-factor analysis showed statistically significant differences between the two groups in age, body mass index (BMI), duration of alcohol dependence, frequency of weekly drinking, alcohol dependence scale (ADS), systolic blood pressure, diastolic blood pressure, fasting blood glucose (FBG), triglycerides (TG), high-density lipoprotein cholesterol (HDL-C), and low-density lipoprotein cholesterol (LDL-C) (P<0.05). Multivariate logistic regression analysis showed that age, BMI, duration of alcohol dependence, frequency of weekly drinking, systolic blood pressure, diastolic blood pressure, and TG levels were independent risk factors for MS (P<0.05). The CRT model showed that BMI, systolic blood pressure, duration of alcohol dependence, LDL-C, and diastolic blood pressure were classification factors for MS. The BPNN model results showed that the importance of variables after standardization was ranked as TG, diastolic blood pressure, duration of alcohol dependence, BMI, LDL-C, and ADS score. The area under the curve (AUC) of the models built by the three machine learning algorithms were all >0.800, with the BPNN model showing the best predictive performance, with an AUC of 0.972, accuracy of 96.01%, sensitivity of 92.00%, and specificity of 97.51%. Conclusion: The prediction model for MS comorbidity in alcohol-dependent patients established through machine learning algorithms demonstrated good predictive ability. Among these models, the BPNN model showed the best diagnostic performance and has the potential to be an effective predictive tool widely applied in clinical practice. Further validation and promotion are needed to test its predictive effect and stability in different environments.