Confidence-Based Training for Clinical Data Uncertainty in Image-Based Prediction of Cardiac Ablation Targets

Ventricular radio-frequency ablation (RFA) can have a critical impact on preventing sudden cardiac arrest but is challenging due to a highly complex arrhythmogenic substrate. This work aims to identify local image characteristics capable of predicting the presence of local abnormal ventricular activities (LAVA). This can allow, pre-operatively and non-invasively, to improve and accelerate the procedure. To achieve this, intensity and texture-based local image features are computed and random forests are used for classification. However using machine-learning approaches on such complex multimodal data can prove difficult due to the inherent errors in the training set. In this manuscript we present a detailed analysis of these error sources due in particular to catheter motion and the data fusion process. We derived a principled analysis of confidence impact on classification. Moreover, we demonstrate how formal integration of these uncertainties in the training process improves the algorithm’s performance, opening up possibilities for non-invasive image-based prediction of RFA targets.