Paper Title
PERFORMANCE ANALYSIS OF MACHINE LEARNING AND STATISTICAL MODELS IN PREDICTING OUTCOMES OF CARBON MONOXIDE POISONING: A MULTICENTER RETROSPECTIVE STUDY

Abstract
Background - Carbon monoxide (CO) poisoning is a major public health issue worldwide, with symptoms ranging from mild discomfort to severe conditions like coma or death. Although there are various treatment options available, such as hyperbaric oxygen therapy (HBOT), predicting patient outcomes remains challenging due to the variability in clinical presentations and the lack of standardized guidelines in many countries. Traditional statistical models, like the COGAS and FIRED scores, have been used for outcome prediction, but with the rise of machine learning, there is potential for developing more accurate and reliable predictive tools. Materials and Methods - This project is a multicenter retrospective study using data from the Chang Gung Research Database, focusing on adult patients (age ≥ 17) who were treated for CO poisoning between 2019 and 2022. We will compare the predictive performance of traditional statistical models (FIRED, COGAS, SOFA, PSS) with machine learning models such as Random Forest, Extreme Gradient Boosting (XGBoost), and Multilayer Perceptron (MLP). The data cleaning process will include handling missing values using K-Nearest Neighbors (KNN) imputation, detecting and addressing outliers, and encoding categorical data. Feature selection will be guided by clinical expertise and algorithmic methods. The primary outcome will be the model's ability to predict poor outcomes, defined as death or a Glasgow Coma Scale (GCS) score of less than 13 at discharge. Model performance will be evaluated using metrics such as the Area Under the Curve (AUC), sensitivity, specificity, precision, recall, and F1-score. Result - Preliminary results indicate that the FIRED score may continue to be a strong predictor, with an AUC of 0.93 in external validation datasets. However, machine learning models, particularly XGBoost and Random Forest, also show potential, delivering comparable or slightly lower AUCs (around 0.89). We anticipate further refining these models and exploring whether integrating machine learning techniques with traditional clinical predictors can enhance outcome prediction in CO poisoning cases. This study aims to equip clinicians with more precise tools for predicting patient outcomes, ultimately leading to better-targeted interventions and improved patient care. Conclusion - In this study, the FIRED score remains a strong predictor in external validation, with an AUC of 0.93. Machine learning models, although not fine-tuned or carefully adjusted, showed similar performance to the expert-modified statistical models. Both approaches are clinically sufficient for predicting outcomes in patients with carbon monoxide (CO) poisoning. Keyword - Carbon Monoxide Poisoning, Machine Learning, Predictive Models, Toxicology, Outcome Prediction