MultimoDal tEnsor fuSion and Completion foR patient-taIlored catheter aBlation of persistEnt atrial fibrillation

Projektbeschreibung

Herzrhythmusstörungen: Computermodelle für eine personalisierte Behandlung

Vorhofflimmern (Atrial fibrillation, AF) ist die häufigste Herzrhythmusstörung und eine der Hauptursachen für Schlaganfälle. Die konventionelle Behandlung besteht in der Ablation, d. h. der Einbringung kleiner Narben in das Herz, um fehlerhafte elektrische Signale zu blockieren. Die Identifizierung der Zielbereiche im Herzen bleibt jedoch subjektiv, und die Anwendung der Ablation hängt vom Urteil des jeweiligen Arztes ab. Mit Unterstützung der Marie-Skłodowska-Curie-Maßnahmen soll im Rahmen des DESCRIBE-Projekts ein rechnergestützter Ansatz entwickelt werden, um automatisch die besten Ablationsziele für jeden Patienten zu ermitteln. Mit der vorgeschlagenen Methode können Herzdaten verknüpft und analysiert werden, auch wenn einige Informationen fehlen. Der DESCRIBE-Ansatz wird die diagnostische Genauigkeit erhöhen und den Behandlungserfolg für die Betroffenen verbessern.

Ziel

Tensor analysis plays a central role in signal processing and machine learning for the representation, analysis, fusion, and classification of data. Responsible for up to 25% of brain strokes, atrial fibrillation (AF) is the most prevalent sustained cardiac arrhythmia and remains the last great frontier of cardiac electrophysiology. Catheter ablation is the most attractive therapeutic option for persistent AF, although the identification of suitable target areas is strongly dependent on practitioners subjectivity. Multi-electrode catheters are increasingly used in ablation as they facilitate the electroanatomical mapping of the atria, but often deliver incomplete data due to lack of contact with the atrial wall. This project aims to improve the personalized characterization and management of AF by proposing novel tensor-based methods for multimodal data fusion in a possibly missing information scenario. New coupled tensor models will be introduced for effectively coupling multimodal information and robust optimization algorithms will be developed for retrieving unknown/unavailable information. It is expected that the optimal exploitation of invasive (intracardiac EGM) and noninvasive (surface ECG) records will allow the automatic identification of the best targets for successful ablation. Encouraging preliminary results have been obtained with the block-term decomposition (BTD) to handle multiple time segments of the ECG for the blind separation of the atrial activity signal. The contribution of EGM into the ECG will be identified by analyzing the common factors obtained by the proposed coupled tensor decompositions. Extensions of coupled BTD to multimodal, possibly missing data will also be proposed. Expected impacts lie in original tensor models and algorithms for data fusion and tensor completion, leading to novel descriptors of AF that can significantly advance the understanding of this prevalent cardiac condition and derive patient-tailored ablation protocols.