| Sumario: | Introduction: Calcium channel blocker (CCB) poisoning is a critical toxicological emergency that can result in severe complications, particularly cardiovascular effects. This study aimed to evaluate the accuracy of Machine learning (ML) models in predicting the outcomes of CCB poisoning. Methods: This retrospective cross-sectional study analyzed the medical records of patients diagnosed with CCB poisoning at Loghman Hakim Hospital between 2019 and 2024. The accuracy of machine learning (ML) models in predicting the outcomes of CCB poisoning and identifying its predictive factors was evaluated. Various ML models, including XGBoost, CatBoost, Random Forest, and AdaBoost, were trained on clinical and laboratory data. Then, feature selection was performed to identify the most relevant variables. The hold-out set was randomly selected to avoid selection bias. Model performance was assessed using accuracy, precision, recall, F1-score, and macro-averaged area under the receiver operating characteristic (ROC) curve (AUC). Results: 274 CCB poisoning cases with the mean age of 31.99± 17.47 (range: 1.5 to 89) years were evaluated (70.4% female). Feature selection identified 18 key prognostic factors, including body temperature, whole bowel irrigation, need for cardiology consultation, arterial oxygen saturation, Glasgow coma scale (GCS)-eye response, electrocardiography (ECG) findings, serum level of alkaline phosphatase (ALP), pH-venous blood gas (VBG), HCO3-VBG, serum level of lactate dehydrogenase (LDH), blood sugar, pulse rate, fraction of inspired oxygen (FiO2), time elapsed from ingestion to admission, troponin, serum level of alanine aminotransferase (ALT), serum level of creatinine, and serum level of potassium. Among the ML models, XGBoost and CatBoost demonstrated the highest predictive performance, with macro-averaged AUC values of 0.9899 (95%confidence interval (CI): 0.98-0.99) and 0.9983 (95%CI: 0.997-0.999), respectively. These models outperformed traditional statistical approaches, providing enhanced risk stratification for patients with CCB poisoning. Conclusion: This study highlights the potential of ML-based models for predicting outcomes in CCB poisoning, offering a data-driven framework for early risk stratification. The superior performance of XGBoost and CatBoost suggests their clinical applicability. Future research should focus on external validation in multi-center settings and real-time integration into clinical decision-making systems.
|
|---|