Objective: Hypothesizing that aortic outflow velocity profiles contain more valuable information than can be captured by the human eye, features of the complex geometry of Doppler tracings were extracted by a convolutional neural network (CNN) mimicking the human cortex by its architecture.
Methods: After pre-training a CNN (VGG-16) on a large data set (ImageNet data set; 14 million natural images belonging to 1,000 classes), the convolutional part was employed to transform Doppler tracings to 1D arrays. Principal component analysis (PCA) and k-means clustering of those 1D arrays were performed, and association between cluster and 2-year all-cause mortality was hereinafter assessed. Among 366 eligible patients (age: 79.8 ± 6.77 years; 146 [39.9%] women) with pre-procedural echocardiography and right heart catheterization prior to transcatheter aortic valve replacement (TAVR), good quality Doppler tracings from 101 patients were initially analyzed. The remaining 265 patients served for later validation of cluster-related practical evidence.
Results: Among 101 patients from the derivation cohort, estimated 2-year survival was 83.0% (95% CI: 75.1-91.7%). The convolutional part of the pre-trained VGG-16 model in conjunction with PCA and k-means clustering distinguished two shapes of aortic outflow velocity profiles. Kaplan-Meier analysis revealed that mortality in patients from cluster 2 (n = 40 [39.6%]) was significantly increased (HR for 2-year mortality: 3.0 [95% CI: 1.0-8.9]). Apart from reduced cardiac output (4.57 ± 1.42 [95% CI: 4.17-5.04] L/min vs. 5.41 ± 1.17 [95% CI: 5.11-5.70] L/min, p-value: 0.0006) and lower mean aortic valve gradient (22.9 ± 7.37 [95% CI: 20.5-25.2] mmHg vs. 47.7 ± 14.1 [95% CI: 44.4-51.3] mmHg, p-value: 4.6x10-15), patients from cluster 2 were also characterized by signs of pulmonary hypertension (mPAP: 31.9 ± 12.2 [95% CI: 28.5-35.7] mmHg vs. 24.7 ± 10.1 [95% CI: 22.1-27.2] mmHg, p-value: 0.0019), impaired right ventricular function (TAPSE: 18.1 ± 3.82 [95% CI: 17.3-19.2] mm vs. 20.8 ± 3.89 [95% CI: 19.8-21.7] mm, p-value: 0.0014) and right atrial enlargement (right atrial area: 22.0 ± 8.28 [95% CI: 19.5-24.6] cm2 vs. 17.8 ± 5.17 [95% CI: 16.6-18.9] cm2, p-value: 0.0133). Contrarily to the initial expectation, patients from cluster 1 with seemingly less extensive cardiac damage were diagnosed with a more severe obstruction of the aortic valve than patients from cluster 2 (aortic valve area: 0.74 ± 0.21 [95% CI: 0.69-0.80] cm2 vs. 0.90 ± 0.21 [95% CI: 0.84-0.95] cm2, p-value: 0.0001). After training an extreme gradient boosting algorithm on these 101 patients, validation on the remaining 265 patients confirmed that patients assigned to cluster 2 show increased mortality (HR for 2-year mortality: 2.6 [95% CI: 1.4−5.1], p−value: 0.004).
Conclusion: This is a proof-of-principle demonstrating that transfer learning enables sophisticated pattern recognition even in clinical data sets of limited size, thereby alleviating the bottleneck of insufficient training data as often encountered in clinical reality. Importantly, it is the left ventricular compensation capacity in the face of increased afterload, and not so much the actual obstruction of the aortic valve, that determines fate after TAVR.
https://dgk.org/kongress_programme/jt2022/aV1024.html