Sparse Bayesian Non-linear Regression for Multiple Onsets Estimation in Non-invasive Cardiac Electrophysiology

Best paper award FIMH 2017, category: Electrophysiology ; International audience ; In the scope of modelling cardiac electrophysiology (EP) for understanding pathologies and predicting the response to therapies, patient-specific model parameters need to be estimated. Although per-sonalisation from non-invasive data (body surface potential mapping, BSPM) has been investigated on simple cases mostly with a single pacing site, there is a need for a method able to handle more complex situations such as sinus rhythm with several onsets. In the scope of estimating cardiac activation maps, we propose a sparse Bayesian kernel-based regression (relevance vector machine, RVM) from a large patient-specific simulated database. RVM additionally provides a confidence on the result and an automatic selection of relevant features. With the use of specific BSPM descriptors and a reduced space for the myocardial geometry, we detail this framework on a real case of simultaneous biventricular pacing where both onsets were precisely localised. The obtained results (mean distance to the two ground truth pacing leads is 18.4mm) demonstrate the usefulness of this non-linear approach.