Introduction: Pediatric cardiac arrest occurs in approximately 6,000 patients in the pediatric intensive care unit (PICU) each year in the United States and is often preceded by observable signs of physiological deterioration. Previous attempts to predict in-hospital cardiac arrest (IHCA) have focused on identifying individual risk factors from static patient data or have used only vital signs. Here, we use a combination of demographic data, ECG derived heart rate variability metrics, vital signs derived summary statistics, and medication administration data to predict IHCA within three hours of onset. Methods: We collected 240 Hz ECG waveform data, vital signs time series data, and medications data from 1,145 patients in the PICU at the Johns Hopkins Hospital, including 15 patients from a curated IHCA database. Patient data was split into five-minute windows for feature extraction. 23 heart rate variability (HRV) metrics were extracted from ECG waveforms. 96 summary statistics were calculated for 12 vital signs, such as respiratory rate and blood pressure. Binary features were generated from the medications data to indicate the administration of 46 therapeutic drug classes. The extracted features were split into training and testing data for model development. Six machine learning models to predict IHCA within three hours were evaluated: logistic regression, support vector machine, random forest, XGBoost, LightGBM, and a soft voting ensemble. SHAP (SHapley Additive exPlanations) analysis was conducted to determine feature importance from the models. Results: Of the models evaluated, XGBoost performed the best, with 0.971 auROC, 0.798 auPRC, 99.5% sensitivity, and 69.6% specificity on an independent, held-out test set. SHAP analysis showed that age, respiratory rate, ST-segment of the ECG, and administration of autonomic drugs were among the most important features to predict IHCA. Conclusions: We developed high-performing machine learning models to predict IHCA up to three hours before onset. These models find hidden trends in ECG, vital signs, and medications data that clinicians otherwise may not be able to identify. Our models, therefore, have the potential to assist clinicians in identifying IHCA, enabling them to intervene in a more time-sensitive manner.