Machine learning analysis of ECG changes following tocilizumab treatment in out-of-hospital cardiac arrest patients

Abstract Background Inflammatory cytokines, including interleukin-6 (IL-6), are suggested to contribute to cardiac electrophysiological disturbances and affect the risk of arrhythmias. Tocilizumab, an IL-6 receptor antagonist, may affect cardiac function by reducing systemic inflammation. However, traditional electrocardiogram (ECG) interpretation methods may not detect subtle electrophysiological changes linked to cytokine modulation. The FactorECG model 1, a machine learning approach using a variational auto-encoder, generates 32 continuous ECG factors associated with ECG morphology. Trained on a dataset of over a million ECGs, the model offers a transparent method to identify changes and has been demonstrated to predict reduced left ventricular ejection fraction and mortality. It has not previously been used in the context of anti-inflammatory therapies. This is a substudy of a double-blinded, placebo-controlled, single-center, randomized clinical trial 2. The original aim of was to evaluate whether tocilizumab could modulate the systemic inflammatory response in patients with out-of-hospital cardiac arrest. Purpose To assess how tocilizumab impacts cardiac electrophysiology, as reflected by specific ECG factors and their association with the machine learning-predicted probability of reduced ejection fraction. Methods A total of 80 comatose patients who had experienced out-of-hospital cardiac arrest were randomised in a 1:1 ratio to receive tocilizumab or placebo. Twelve-lead ECGs were obtained at baseline, 6 hours and then at 12-hour intervals, and analysed using the FactorECG model to extract 32 ECG factors. ECGs within a window of 12 to 96 hours post-randomization were included, selecting one ECG per patient closest to 24 hours post-intervention. Linear models assessed the effect of the intervention on each ECG factor, and the relationship between significant factors and the machine learning-predicted probability of reduced ejection fraction was evaluated. Results Four ECG factors (Factors 6, 7, 21, and 24) were significantly associated with the intervention group. Factor 6 (Estimate = 1.19, p = 0.0012), related to P-wave size and atrial abnormalities, and Factor 24 (Estimate = 0.012, p = 0.0337), potentially reflecting ventricular function changes, were higher in the tocilizumab group. Linear models predicting the machine learning-derived probability of reduced ejection fraction showed significant effects of Factors 7 (p = 0.024) and 21 (p = 0.0097), indicating an indirect influence of the intervention on reduced ejection fraction probability through specific ECG features. Conclusion Machine learning ECG analysis identified changes in P-wave morphology, ventricular function, and electrophysiological stability following tocilizumab treatment.