Data gathering and preprocessing
435 subjects free of cardiac pathology and 89 patients with AS, as confirmed by expert cardiologists, were enrolled in this study, from two clinical centers: University of Leipzig, Leipzig Germany, and Kaplan Medical Center, Rehovot Israel. The free of cardiac pathology dataset includes healthy volunteers, and subjects with some cardiovascular risk factors (smoking, hypertension, diabetic mellitus, and hyperlipidemia) that were enrolled in the hospital with chest pain. The IRB boards in both clinical centers certified the study and all participants signed informed consent. All participants underwent transthoracic echocardiographic scans, using Vivid 7 cardiovascular ultrasound systems (GE Vingmed, Horten, Norway) during which they were at normal sinus rhythm.
For each subject, several clips were obtained in each of the three conventional longitudinal apical views of two, three, and four chambers [26]. The subjects were scanned by different technicians. We choose to zoom in on the apical longitudinal views, from which the EF and GLS indices are derived 44 and are at the center of most clinical studies. Since STE algorithms are sensitive to the image quality [27], a single expert picked for each subject the three clips, corresponding to the three longitudinal views, with the optimal image quality. The expert rated the selected clips with an image quality score of 1 to 5, regarding the ability to detect endocardial boundaries, the visibility of all 6 myocardial segments within the field of view, and the overall image quality. Subjects with clips that received the lowest image quality score (i.e. 1) were excluded from this study.
Strain Analysis
The expert processed the clips using layer-specific 2D STE software termed ‘K-SAD’ [17, 28]. The ‘K-SAD’ operational algorithm flow is similar to the existing commercial software. After bounding the desired heartbeat in the cine loop to be processed and selecting the end-systolic frame, the user places three anatomical points: 2 in the basal endocardium and one in the apical epicardium. The endocardium is then detected automatically and the whole myocardial region of interest (ROI) is defined including the epicardium. From here, the software traces the myocardium back and forth along the clip to generate time strain curves [17]. The expert repeatedly tuned the initial myocardial segmentation until optimal tracking was achieved. Unlike commercial software, the ‘K-SAD’ block matching algorithm is optimized for physiological tissue displacement with minimal smoothing of the tracking results, without applying linear relationship between neighboring layers of the myocardium that are traced independently. The algorithm is less affected by the initial segmentation and enables a thinner segmentation width at the apex. The ‘K-SAD’ results were verified through phantoms, healthy subjects, and certain pathologies, and found to be consistent and even more accurate for commercial vendors [28].
The ‘K-SAD’ software generates TSC for the six myocardial segments in the three layers of the myocardium, corresponding to the 18 segment model [29]. In each segment, the inner and middle tracking chains are averaged to represent the endocardial TSC, the middle and outer tracking chains are averaged to represent the epicardial TSC, and the three tracking chains are averaged to represent the mid-wall TSC [17]. The compact muscle TSC, the middle tracking chain, is not part of the software output and was estimated indirectly algebraically using the endocardial, epicardial, and middle layers TSCs output. The software also provides global TSC across all six segments.
Besides the TSCs, the velocity curves of the basal tracking knots were also measured. The basal tracking knots are the first and last knots of the mid-wall tracking chain, adjacent to the mitral annular plane [28]. These velocity curves closely reflect the mitral annular velocity pattern, routinely assessed by tissue doppler (along with transmitral inflow velocities) to detect abnormal myocardial relaxation [30].
AS grading
The severity of AS was clinically determined according to the acceptable combination score of valve area (using the continuity equation), transvalvular gradient flow, and peak maximal velocity jet. Measurements of wall thickness in the basal segments were also taken among the AS patients [31]. The EF of the LV was calculated, for AS patients and for the free of cardiac pathology subjects, using the modified biplane Simpson’s method for the two and four-chamber apical views [29].
Deformation parameters
In the current dataset, the number of samples (frames) per segmental TSC ranges between 28 and 128, owing to differences in sampling rate and / or heart rate. To adjust the TSC into a fixed length, subsampling or oversampling (interpolation) is needed that may introduce artificial distortions into the non-stationary TSC waveform with the changing cardiac phases. Khamis et al. [32] and Tabassian et al. [33] performed interpolation of the TSC to a length of over 100 samples. Such a procedure raises concerns for overfitting when there are only a few hundreds of subjects to classify. Hence, rather than introducing the classifier the complete segmental TSC, feature selection was implemented by sifting the initial input to a subset of variables containing most of the relevant information [34]. Feature selection also enables to optimize and expedite the system training by excluding redundant information that may interrupt the classifier, and examine the hindered interactions between the selected features [35].
The parameter extraction scheme is illustrated in Fig. 1. For each of the AS patients, chest pain subjects, and healthy volunteers the following STE parameters were extracted from the cine loops of the three views:
-
From the 18-segmental strain curves: peak segmental longitudinal strain (PSLS) amplitude and time to PSLS. The PSLS values were averaged to get the GLS. The mechanical dispersion was defined by the standard deviation of the time to PSLS for the 18-segments.
-
From the 18-segmental strain rate curves: peak strain rate amplitudes at systole (SRs) and at early (SRe) and late diastole (SRa). The values were averaged to get systolic, early, and late diastolic GLSR. Of note, the segmental strain rate values were derived numerically from the corresponding TSC. The SRs was taken as the maximal negative strain rate value during systole and until the aortic valve is closed. The SRe and SRa were taken as the maximal positive strain rate values during the first and last half parts of diastole, respectively.
-
From the 6 basal kernels velocity curves: peak velocity (Vbas) amplitudes at late and early diastole. The values were averaged to get the global Vbas (GVbas) at early and late diastole.
In addition to the parameter extraction scheme, the conventional GLS was taken as the peak global mid-wall TSC.
Data outliers
Potential artefactual strain (rate) curves were identified statistically according to the expected values of the following parameters: PSLS, time to PSLS, SRs, SRe, and SRa. Dalen et al. [36] showed that segmental strain (rate) parameters distributed normally among heterogeneous 1,266 healthy subjects, and we assume this also applies to AS, which is a global myocardial pathology that involves the entire LV [24]. In order to derive the data outliers, and thus to determine which curves are artefactual, the statistical characteristics of the extracted parameters were calculated (see results). Due to sample size and randomness limitations of the sample, the Student's t-distribution can be taken, which has wider tails than a normal distribution. Assuming a normal t-like distribution for the extracted strain (rate) parameters, measurements that exceed the average by more than 1.96 standard deviations (2.5% of the two-tail measurements), were labeled as potential outliers [37].
These outlier cut-off values were tested on the retrieved PSLS and time to PSLS measurements (see Fig. 2). Among the free of cardiac pathology subjects, one can find PSLS outlier values above 30% that were shown to derive from foreshortened probe position and PSLS outlier values below 10% that were shown to derive from faulty ROI segmentation of the epicardium [38, 39]. The same goes for the time to PSLS, where beyond the cut-off values lay outlier measurements that significantly exceed the expected pattern of 1/3 and 2/3 for the systole and diastole phases. Because the extracted parameter values may vary among the myocardial segments [36], we analyzed the 18-segment separately, for the free of cardiac pathology subjects and AS patients. Segmental strain (rate) curves with two or more outlier parameter measurements, were identified as artefactual.
Classification
We employed a soft-margin support vector machine (SVM) classifier, with a Gaussian kernel that was found preferable for classifying TSCs [39, 40]. The classifier was weighted against misclassifying AS patients (i.e. the minority class), with the cost of the inverse ratio of class sizes [41]. The Gaussian scaling factor was optimized through heuristic sequential minimal optimization (SMO) technique [42]. The SVM output score was converted into a posterior probability score as well as into categorical prediction using a two-parametric sigmoid function. The sigmoid function calibrates the SVM output score bias toward the majority class [39, 43].
Validation and Evaluation
For leave-one-out cross-validation, the training set included all but one subject who constituted the test set. The categorical prediction performance was evaluated by computing common statistical indices drawn from confusion matrix [44]: accuracy, specificity (percentage of subjects without cardiac pathology that were correctly identified) and sensitivity (percentage of AS patient that were correctly identified). The posterior probability score performance was evaluated and compared to the EF and GLS indices by computing the area-under-curve (AUC) of the receiver-operating-characteristic (ROC) curve [44].
Counterpart Models
Model performance was compared with the model proposed by Tabassian et al. [33, 45] for classifying echocardiograms of patients with myocardial infarction (MI), and further extender for classifying echocardiograms of patients with heart failure preserved EF (HFpEF). Briefly, the mid-wall segmental strain rate traces, generated by the ‘K-SAD’ algorithm, were taken. The traces were aligned through linear interpolation of the systolic and diastolic phases separately. The partition for training and test sets was conducted similarly to the SVM model. Then, the training and the test sets were both transformed into the PCA domain of the training set. The first five major PCA components, from each of the eighteen segments, were chained into a single vector representing a subject. A k-nearest neighbors (KNN) classifier was implemented, and the number of K neighbors was selected based on the optimal sensitivity result obtained in a range of 1–50. The KNN model was also weighed in favor of the minority AS class [46].
We also grew a binary multi-stage decision tree to classify between the AS patients and the free of cardiac pathology subjects. The hierarchical structure of decision nodes reveals the importance of the different extracted parameters and their (nonlinear) relations. Also, the provided decision thresholds have clinical significance. The decision tree algorithm was weighted in favor of the minority class of the AS patients in the same way as for the SVM classifier, and leave-one-out cross-validation was applied to reduce the error loss. To maintain the simplicity of the model and to avoid overfitting, we limited the number of possible splits to the number of the extracted parameters. For more details on this classifier implementation please see Breiman 64.