Predicting the Higher Energy Need for Effective Defibrillation Using Machine Learning Based on an Animal Model

Background: Early defibrillation improves outcomes in cardiac arrest, but the optimal defibrillation strategy and energy requirements remain debated. This study investigated whether arterial blood gas (ABG) parameters could predict optimal defibrillation energy requirements for achieving the highest first-shock success rates in an animal model. Our study focused on clinical scenarios where ABG measurements are readily available, such as ventricular tachycardia and ventricular fibrillation storms requiring multiple shock deliveries. Materials and Methods: In the experimental setting, ventricular fibrillation was induced by 50 Hz direct current (DC), and the defibrillation threshold (DFT) was determined using a stepwise defibrillation protocol. ABG parameters were measured before each defibrillation attempt, recording partial arterial pressure of carbon dioxide (PaCO₂) and oxygen (PaO₂), pH, hematocrit (Hct), sodium (Na⁺), potassium (K⁺), and bicarbonate (HCO₃^-) levels. The relationships between ABG parameters and the DFT were analyzed for 15 subjects using classical data analysis techniques and machine learning (ML) algorithms. Multiple ML models were trained and tested to predict the higher energy needed for successful defibrillation based on the ABG parameters. Results: Statistically significant differences were found in Hct and Na⁺ levels between the two DFT categories, above 130 Joules (J) and below 40 J (p < 0.01). The DFT negatively correlated with PaO₂ and positively correlated with Hct and Na⁺. However, other ABG parameters did not show significant correlations with DFT. Using ML, we predicted cases requiring higher defibrillation E. Our best-performing model, the Extra Trees Classifier, achieved 83% overall accuracy, with 100% and 67% precision rates for higher and lower DFT categories, respectively. We validated the model using bootstrap resampling and 10-fold cross-validation, confirming consistent performance. We identified Hct, PaCO₂, and PaO₂ as significant contributors to model prediction based on the feature importance value. Conclusions: Modern data analysis techniques applied to ABG parameters may guide personalized defibrillation energy selection, particularly in controlled clinical environments such as catheterization laboratories and intensive care units where ABG measurements are readily available.