Abstract Sat304: Predicting In-Hospital Cardiac Arrest Using Serum Electrolytes and Hemoglobin: A Machine Learning Approach

Background: Limited laboratory (lab) values are recently used to predict in-hospital cardiac arrest (IHCA) by machine learning models. The temporal characteristics of lab values in relation to the onset of IHCA is often underreported which is a critical factor in clinical decision-making process for timely prevention. Hypothesis: Machine learning models can identify key laboratory features and their temporal characteristics preceding in-hospital cardiac arrest Methods: We conducted a retrospective, case and control study at a single academic medical center using electronic health records from adult patients between 2014 and 2016. The dataset included 98 cases of adult patients (≥ 18 years old) who experienced IHCA and 131 control patients matched by sex and age. Only the first episode of IHCA was selected. Serum electrolytes including potassium, sodium, chloride, magnesium, hemoglobin, creatinine, and estimated glomerular filtration rate were extracted from electronic health records and analyzed in 4-hour intervals over the 120 hours preceding IHCA. Random Forest (RF), XGBoost, Support Vector Machine (SVM), Multi-Layer Perceptron (MLP), Logistic Regression (LR), K-Nearest Neighbors (KNN), and Naive Bayes. In addition, we designed a soft-voting ensemble model that combined SVM, RF, and LR to integrate both linear and non-linear decision boundaries for improved performance. All models were trained using stratified 5-fold cross-validation and evaluated based on AUROC, sensitivity, accuracy, and F1 score. Results: The clinical characteristics of cases included of 69±15 years old, 66% were male, and 54% admitted to intensive care units. Support Vector Machine (AUROC = 0.856 (0.789-0.924)), and Logistic Regression (AUROC=0.846 (0.783-0.908)) were marginally better than ensemble model 8 hours prior to IHCA. The ensemble model achieved the best AUROC=0.852 (0.779-0.924) after 8 hours to 40 hours prior to IHCA. Performance of ML models declined after 40 hours. Feature analysis showed serum chloride as the most predictive variable from 0 to 8 hours and 12-32 hours, hemoglobin from 8 to 12 hours, serum sodium from 32 to 40 hours. Conclusions: Performance of machine learning models varied across the prediction window, with the highest AUROC values observed closer to the time of IHCA. These findings support the use of dynamic, data-driven models for timely risk identification of IHCA in hospitalized patients.