Predicting LVEF improvement after stereotactic arrhythmia radioablation using machine learning
A Mircea, M Grehn, A Luca, J Solana-Munoz, C Teres, L Schiappacasse, A Zaman, F Hohendanner, M Kruska, D Duncker, R Tilz, E Lian, J Boda-Heggemann, O Blanck, E PruvotAbstract
Background
Stereotactic arrhythmia radioablation (STAR) is used to control refractory ventricular tachycardias (VT). Animal studies showed that left ventricular ejection fraction (LVEF) improves after STAR. Our previous work identified key features for predicting the need for redo catheter ablation (CA) after STAR using LIME and SHAP as post-hoc explainable artificial intelligence (AI) methods. These features were planned target volume (PTV), LVEF prior to STAR, administered radiation dose, presence of hypertrophic or dilated cardiomyopathy (HCM/DCM) and number of VT ablations before STAR.
Objective
To determine whether a machine learning (ML) model can predict the improvement in LVEF following STAR.
Mtehods
38 patients (67 years old median age, 14 ischemic cardiomyopathy, 6 females/32 males) treated by STAR for refractory VT from the German and Swiss registries were included. LVEF change after STAR was defined as a binary outcome, 1 standing for improvement and 0 standing for no improvement or reduction. A logistic regression model was trained on 30 patients and validated on 8. A point biserial test was used to assess whether the input features (PTV, administered radiation dose, number of VT ablations before STAR and hypertrophic/dilated cardiomyopathy) correlate with LVEF improvement.
Results
Overall LVEF improved after STAR as shown in Figure 1 by the cumulative positive LVEF changes across patients. The logistic regression model achieved 87.5% accuracy in predicting LVEF improvement following STAR. PTV showed an inverse correlation with LVEF improvement (r = –0.32, p = 0.04). Similarly, the number of prior VT ablations before STAR showed an inverse correlation (r = –0.30, p = 0.06). Baseline LVEF, radiation dose, and HCM/DCM showed no significant correlation.
Conclusion
ML models can predict LVEF improvement after STAR in patients with VT. Higher PTV and a greater number of VT ablations prior to STAR are associated with a lower likelihood of LVEF improvement.