Background:

The current state of the art in the diagnosis and treatment of cardiovascular diseases has been based on evidence resulting from traditional trials as well as years of clinical experience utilizing patient characteristics, physicians’ experience, and two-dimensional measurements from cardiac imaging. Utilizing technological advances in deep machine learning (ML) we developed an algorithm to assess the individual patient risk and best prosthesis for transcatheter aortic valve replacement (TAVR). Ultimately, personalized treatment recommendations and decisions may be made at the individual patient level. 

Purpose:

The aim of this study was to optimize our ML algorithm using routine clinical imaging and cluster detection to make the algorithm for personalised risk prediction and TAVR-prothesis selection more precise.

Methods:

All patients included in the study were allocated to undergo TAVR. The basic algorithm utilizing 58 baseline features was a random forest classification with high interpretability to predict the outcome. To improve ML model, the whole volume of the pre-procedural cardiac computed tomography (CT) was used. These were histogram normalized, orientation corrected, voxel size adjusted, resized and automatically center cropped. Due to the high computational complexity of and end-end-network dealing with volume data, the 3D-EfficientNet-B0 was used. The random forest and the efficient net were combined as a model cascade. Finally, we performed an adapted clustering using self-organized maps and analysed the performance boosting by a "leave-cluster-out"-approach. The optimised algorithm was compared with the “plain” one. The performance of the estimators was evaluated by a five-fold nested stratified cross-validation. To evaluate the model, receiver operating characteristics and mean area under the curve (AUC) scores were chosen.

Results:

A total of 3324 patient datasets as well as their CT scans were included in this study. The baseline characteristics were consistent with a high cardiovascular risk typical of this collective, and implantation success according to VARC-2 criteria was achieved in 83.8%. Device success was achieved in 83.3%, pacemaker implantation was necessary in 12.2%, and aortic valve insufficiency was observed in 2.5%. The 30-day mortality was 3.4% and one-year mortality 12.7%. The AUC for the 30-day mortality rate was not significantly different (ML_{w/o Cluster}: 0.69 ± 0.3 vs. ML_Cluster: 0.69 ± 0.3; p=n.s.). However, a statistically significant improvement of the risk prediction was observed for 1-year mortality (ML_{w/o Cluster}: 0.68 ± 0.2 vs. ML_Cluster: 0.72 ± 0.02; p<0.05).

Conclusions:

In summary, this study with preliminary data could show promising methods to improve decision making in the context of TAVR evaluation and planning. Methods such as 3D class activation mapping also enable a high degree of interpretability of the classification results. Thus, the integration of ML in the TAVR strategy planning process and prothesis selection offers a valuable tool in providing patient-focused, personalized therapy for routine clinical application.