Clin Res Cardiol (2022). https://doi.org/10.1007/s00392-022-02002-5
Artificial intelligence enabled prediction of treatment success in patients undergoing transcatheter tricuspid valve repair using transoesophageal echocardiographic imaging
M. Unterhuber1, K.-P. Kresoja1, K.-P. Rommel1, C. Besler1, T. Noack2, M. Orban3, V. Fortmeier4, J. Hausleiter3, V. Rudolph5, H. Thiele1, P. Lurz1
1Klinik für Innere Medizin/Kardiologie, Herzzentrum Leipzig - Universität Leipzig, Leipzig; 2Universitätsklinik für Herzchirurgie, Herzzentrum Leipzig - Universität Leipzig, Leipzig; 3Medizinische Klinik und Poliklinik I, LMU Klinikum der Universität München, München; 4Klinik für Allgemeiner und Interventionelle Kardiologie/Angiologie, Herz- und Diabeteszentrum NRW, Bad Oeynhausen; 5Allgemeine und Interventionelle Kardiologie/Angiologie, Herz- und Diabeteszentrum NRW, Bad Oeynhausen;

Background:

Transcatheter tricuspid valve repair (TTVR) has been shown to be a safe and effective treatment option for patients with severe tricuspid regurgitation (TR) deemed at prohibitive surgical risk. With the rise of TTVR, effective patient selection is crucial to improve technical and patient directed outcomes. Therefore, we sought to create and evaluate a prediction model based only on preinterventional transoesophageal echocardiography (TOE) to predict treatment success.

 

Methods:

Patients with severe symptomatic TR treated with TTVR between 2016 and 2020 from three high-volume centers in Germany (Munich, Bad Oeynhausen and Heart Center Leipzig) were included in this study. Treatment success was defined as TR grade ≤2 at first follow-up after TTVR. Preinterventional transgastric short-axis view B-mode cines of minimum 22 frames length were standardized and cropped. For every frame, movement vectors with starting and ending coordinates in relation to the timeframe were extracted. This information was fed to a long-short-term-memory neural network (LSTM-NN) to predict treatment success. Seventy percent (449 patients) of the study population served as derivation group, whereas 193 (30%) were split into internal (120 patients, 19%) and external validation cohorts. Finally, the model was tested on the remaining 73 patients, (11%). This dataset was withheld and blinded from the network from the beginning to test the accuracy on a “never-seen” dataset and served as external validation cohort.

 

Results:

Overall, 642 patients were included in the present study. The mean age was 78±6 years, 52% were women. Overall, treatment success was achieved in 553 patients (86%). The LSTM-NN showed a suitable performance with an area under the curve of 0.77 (95% CI 0.64–0.89) in predicting treatment success on the test set with a specificity of 91% and a sensitivity of 63%, a positive predictive value of 98% and a negative predictive value of 30%.

 

Conclusion:

This study shows that treatment success of TTVR can be predicted by an artificial intelligence algorithm using preinterventional TOE cines with a high positive predictive value, allowing to assess the feasibility and success probability before performing a procedure and paving the way for optimized patient selection even in centers with less experience in TTVR. These findings warrant further research and validation in larger cohorts.
https://dgk.org/kongress_programme/jt2022/aV1284.html