Blood Alcohol Concentration Screening at Emergency Room: Designing a Classification Model Using Machine Learning

Conventional alcohol screening methods in the ER are susceptible to errors, potentially compromising patient safety and treatment efficacy. Thus, the study aimed to develop an improved classification model utilizing blood gas data and machine learning techniques. Patient data collected from the ER formed the dataset, with blood gas data available for each patient. Five distinct machine learning algorithms—K-nearest neighbor, logistic regression, random forest, LightGBM, and support vector machine (SVM)—were employed to construct the classification model. Model performance evaluation encompassed sensitivity, specificity, and mean absolute error metrics. To contend with challenges arising from intricate feature dimensions and skewed data distribution, the study incorporated feature selection techniques like the Filter and Wrapped Methods. Additionally, the SMOTEENN sampling approach was adopted to address data imbalance. Of the evaluated algorithms, the LightGBM model notably excelled, boasting an accuracy rate of 0.908 and an F-score of 0.617. This underscores the efficacy of utilizing blood gas analysis data for BAC prediction. The developed classification model holds potential for enhancing the precision and expediency of alcohol screening procedures within ER settings, enabling more informed medical decisions by healthcare providers.