Abstract
Background
Stereotactic arrhythmia radioablation (STAR) is a novel therapeutic method for ventricular tachycardia (VT) refractory to catheter ablation (CA) and antiarrhythmic drugs (AADs). Nevertheless, recurrences of VT after STAR remain quite common. A recent study reported that most VT recurrences occurred outside the treated volume [1]. Explainable artificial intelligence (XAI) algorithms could play a key role in assisting clinicians in decision-making on who to treat and how to adjust various treatment parameters.
We have previously shown that an XAI model (LIME) may help identify key features predicting the need for a redo CA following STAR [2]. These features were left ventricular ejection fraction (LVEF), dose of radiation, planned and clinical irradiated target volume (PTV/CTV).
Purpose
Our study is aimed at validating the LIME model with another XAI model, the SHapley Additive exPlanations (SHAP), confirming that the most important features of LIME are the most important ones for clinicians to consider before STAR.
Methods
Twenty patients (16 men and 4 women, age 67±9 y.o.) with refractory VT after failed CA and AADs were treated with STAR. Clinical, paraclinical and interventional parameters were gathered. The cohort was randomly split into a train cohort of 16 patients and a test cohort of 4 patients. This was done five different times for a logistic regression model (ie cross-validation) . The features used to predict the need for redo CA after STAR are shown in Figure 1. Each feature was standardized, ensuring that high cardinality is eliminated as a potential confounder. The XAI algorithm SHAP was used to determine feature importance and how each feature contributed to the final predicted output.
Results
Figure 1 ranks the features using the mean impact on the model output (i.e. redo CA or not). Three out of top five features are validated by SHAP to be most important in determining whether the patient will undergo redo CA or not : (1) PTV, (2) LVEF and (3) radiation dose. In contrast with LIME results for redo CA, two different features are ranked among the top five using SHAP. These are the number of VT ablations before STAR in the last 24 months and the presence of non-ischemic cardiomyopathy (i.e. dilated or hypertrophic cardiomyopathy). A summary of how each value of the top five features influences the final prediction for redo CA is seen with Table 1.
Conclusion
The most important features for the need of redo CA following STAR across both LIME and SHAP are PTV, LVEF and radiation dose. The number of VT ablations prior to STAR seems to play an important role in the SHAP validation for redo CA, while it played a key role for predicting LVEF in the LIME model. The role non-ischemic cardiomyopathy plays in predicting redo CA remains to be further investigated.
SHAP value on a 5-fold cross-validation
Influence of top features on redo CA