Clin Res Cardiol (2022).

Federated Learning of TAVI Outcomes (FLOTO) - A Collaborative Multi-Center Deep Learning Initiative
T. Seidler1, M. Tölle2, F. André3, P. Bannas4, N. Frey3, S. Friedrich5, S. Groß6, A. Hennemuth7, N. Krüger7, A. Leha8, S. Martin9, A. Meyer10, E. Nagel11, S. Orwat12, C. Scherer13, S. Simm14, T. Friede8, S. Engelhardt2, für die Studiengruppe: DZHK-AI/ML
1Herzzentrum, Klinik für Kardiologie und Pneumologie, Universitätsmedizin Göttingen, Göttingen; 2Innere Medizin III: Klinik für Angiologie, Pneumologie und Kardiologie, Universitätsklinikum Heidelberg, Heidelberg; 3Klinik für Innere Med. III, Kardiologie, Angiologie u. Pneumologie, Universitätsklinikum Heidelberg, Heidelberg; 4Klinik und Poliklinik für Diagnostische und Interventionelle Radiologie und Nuklearmedizin, Universitätsklinikum Hamburg-Eppendorf, Hamburg; 5Institut für Mathematik, Universität Augsburg, Augsburg; 6Klinik und Poliklinik für Innere Medizin B, Universitätsmedizin Greifswald, Greifswald; 7Institut für kardiovaskuläre Computer-Assistierte Medizin, Charité - Universitätsmedizin Berlin, Berlin; 8Institut für Medizinische Statistik, Universitätsmedizin Göttingen, Göttingen; 9Institut für Experimentelle und Translationale Kardiovaskuläre Bildgebung, Universitätsklinikum Frankfurt a.M., Frankfurt; 10Klinik für Herz-, Thorax- und Gefäßchirurgie, Deutsches Herzzentrum Berlin, Berlin; 11Kardiovaskuläre Bildgebung, Universitätsklinikum Frankfurt, Frankfurt am Main; 12Klinik für Kardiologie III: Angeborene (EMAH) und Herzklappenfehler, Universitätsklinikum Münster, Münster; 13Kardiologie, LMU Klinikum der Universität München, München; 14Institut für Bioinformatik, Universitätsmedizin Greifswald, Greifswald;
Transcatheter aortic valve implantation (TAVI) relies heavily on the pre-interventional interpretation of computed tomography (CT) data. However, despite sophisticated software to visualize TAVI-relevant anatomy, interpretation is currently limited to "eyeballing" and some simple geometric measurements. Deep learning and, in particular, explainable AI are promising approaches towards more sophisticated data processing and interpretation, but the performance of the algorithms depends on the size of the training datasets. Privacy laws prohibit retrospective pooling of TAVI CT data from different institutions and thus the generation of appropriate training datasets. Federated learning is a new concept that circumvents this dilemma by selectively sharing the algorithm instead of the data between different institutions (Figure 1). We hypothesize that procedure planning and outcome prediction in TAVI can be improved by extracting and processing valuable additional information from TAVI CT images and patient record data using federated deep learning. Therefore, we have established a federated learning infrastructure between nine institutes in Germany. This includes the technical infrastructure for automated decentralized data processing and local model training, as well as connectivity to enable different patterns of iterative routing and training of algorithms across multiple sites. Starting with a limited set of clinical information, including CT image data, patient records, and procedure codes recorded in clinical practice, this platform is designed to incorporate additional sites and data sources for future federated learning projects. To prove the principle, we collaborate to successfully train a deep neural network that helps in intervention planning, prosthesis selection and predicts outcomes of TAVI interventions based on large amounts of multi-site CT image and patient record data. In summary, we have established a collaborative infrastructure between nine institutions to apply deep learning algorithms to each isolated data pool to enable joint training and overall analysis. We demonstrate that applying federated deep learning to otherwise isolated pools of TAVI-related data, provide a privacy- and data security compliant pathway for accessing and integrating multicentric data sources to advance TAVI intervention planning and outcome prediction.