Deep neural network-based clustering of deformation curves reveals novel disease features in PLN pathogenic variant carriers

As described previously [3, 10], we selected PLN p.Arg14del variant carriers from a nationwide registry who underwent transthoracic echocardiography between 2006 and 2019 in the University Medical Center Utrecht, University Medical Center Groningen and Amsterdam University Medical Center. These were both index patients and family members who were identified by genetic screening. While index patients underwent comprehensive genetic testing for cardiomyopathy-related variants, family members underwent targeted testing for the PLN p.Arg14del variant as part of cascade family screening. Index patients with a second pathogenic cardiomyopathy-related variant and subjects with relevant cardiovascular comorbidities such as hypertension were excluded. As defined previously, PLN p.Arg14del variant carriers were classified as pre-symptomatic in case they had no history of ventricular arrhythmias (VA), a premature ventricular complex (PVC) count < 500/24 h and left ventricular ejection fraction (LVEF) > 45% [3].

Controls were derived from the Flemish Study on Environment, Genes and Health Outcomes (FLEMENGHO), which consists of a random population sample from a geographically defined area in Belgium [11]. The prevalence of the PLN p.Arg14del variant in this area was assumed to be negligible, as the prevalence of this founder variant decreases considerably towards the south [9]. We selected participants who underwent transthoracic echocardiography between 2005 and 2009 [12].

For the training and testing of the algorithm, the PLN dataset was split in an 80:20 ratio on the subject level, to make sure no subjects appeared in both datasets (Fig. 1). To maximize the available data for training, multiple echocardiograms per subject were included when available. In the testing dataset, only the first echocardiogram was used, since follow-up echocardiograms may be more affected and may therefore bias the results of diagnostic performance. Every PLN subject in the testing dataset was matched to a control subject using propensity score matching without replacement on age, sex and heart rate, since those parameters may significantly affect the deformation curves. All remaining control subjects were eligible for the training dataset. Propensity score matching was performed during training in a 1:3 PLN to control ratio to account for imbalances in the previously mentioned parameters.

All data used in this study were retrospectively acquired data. The data were anonymized and handled according to the European General Data Protection Regulation. Re-use of the data was permitted by the Medical Ethics Committee. For the PLN variant carriers, the echocardiograms were acquired using a Vivid 7, E9 or E95 machine (GE Healthcare, Horten, Norway) as part of routine clinical care. For the FLEMENGO control group, echocardiograms were acquired using a Vivid 7 machine by a single experienced physician [12]. Longitudinal strain analysis was performed with 2D speckle tracking by two operators using EchoPAC software (version 2.0.3., GE Healthcare) according to the current recommendations [13]. Only apical four-chamber views were used in this analysis and were divided in six segments: apical, mid and basal segments from both the lateral and septal walls. The six regional deformation curves that were computed by the software were included in the deep learning model as raw data and normalized in time to 1 s, with the location of the aortic valve closure at 38% of the RR-interval (mean of the training data). Additional information of the data acquisition and preprocessing can be found in the Supplemental Methods.

We constructed a deep convolutional neural network with exponentially dilated causal convolutions, which is optimized for use on time series such as electrocardiograms (ECGs) [6]. By using dilated causal convolutions, the network can learn complex patterns across large segments of the waveform, while only regarding previous timepoints and therefore taking into account the temporal nature of the signal. The architecture has been described in more detail in the Supplemental Methods and an overview of the architecture can be found in Supplemental Fig. 1. Hyperparameters (i.e., number of feature maps, depth of the network, kernel size, dropout ratio, learning rate and batch size) were optimized in the training dataset using fivefold cross-validation, where complete PLN-control match groups were kept together during the dataset split. The simplest network with the highest F1 score averaged over all five cross-validation folds was selected for testing. The independent test dataset was used once to assess the performance of the final optimized model. To estimate this final performance, all five trained networks (from the five different cross-validation splits) were used as an ensemble model, where the probability for each subject was obtained by taking the mean of these five models’ predicted probability.

To identify the parts of the strain curves that were considered important by the model to classify PLN variant carriers, we used the Integrated Gradients visualization technique [14]. This approach was combined with SmoothGrad-Squared, since recent reports have shown this to be the most robust to produce individual relevance maps [15]. As phenotypical variability was expected for the PLN variant carriers, we performed a time series clustering approach on the relevance maps of the correctly predicted patients (i.e. with a predicted probability over 50%) in the complete dataset. K-means clustering with the Euclidean distance metric was used to divide the normalized relevance maps into four different phenotype clusters. The clusters were visualized by taking the mean and standard deviation in the temporal axis of all patients in that cluster and superimposing the cluster centers of the relevance maps as a heatmap. A Savinsky-Golay filter was applied to each cluster center to smoothen the heat map. The number of clusters was defined empirically by visually assessing whether the clusters showed difference in their morphology, while keeping a minimum of 30 patients per cluster. The mean deformation curve per segment of all the control subjects was used as a reference in the figures. Additional information on the visualization technique can be found in the Supplemental Methods.

To explore the disease course among the different phenotype clusters, follow-up data of PLN variant carriers were derived from an electronic research data platform [10]. The primary outcome variable was sustained VA, which was defined as sudden cardiac arrest/ventricular fibrillation, appropriate implantable cardioverter defibrillator (ICD) intervention or any recorded sustained ventricular tachycardia (> 100 bpm) lasting more than 30 s.

Baseline data are expressed as mean ± standard deviation (SD) or median with interquartile range (IQR), where appropriate. Discriminatory performance of the deep learning model was assessed using the area under the receiver operating curve (AUC) or C-statistic, accuracy, F1-score, specificity and sensitivity. The 95% confidence intervals were derived using 2000 bootstrap samples. For the comparison of clinical characteristics between phenotype clusters, we performed Chi-square, one-way ANOVA or Kruskal–Wallis tests as appropriate. Adjustment for multiple testing was performed using Bonferroni’s correction. The DNN-based classification of the deformation curves into clusters was compared to the manual classification of the deformation curves of this cohort in a previous paper [3]. All statistical analyses were performed using Python version 3.8 (Python Software Foundation).

Overall, 278 PLN variant carriers were included, with a total of 419 echocardiograms (mean of 1.5 echocardiograms per patient). From the FLEMENGO cohort, 621 control subjects were included, with one echocardiogram per subject. Baseline characteristics of the PLN variant carriers and control subjects (stratified by training and test set) can be found in Table 1.

Cross-validated mean C-statistic, accuracy and F1 score on the training dataset were 0.93 ± 0.02, 0.93 ± 0.01, and 0.86 ± 0.05, respectively. The performance of the ensemble model in the independent test set was excellent, with a C-statistic, accuracy, F1 score, sensitivity and specificity at a 50% probability threshold of 0.93 [95% CI 0.87–0.97], 0.90 [95% CI 0.85–0.96], 0.89 [95% CI 0.83–0.95], 0.88 [95% CI 0.79–0.96] and 0.93 [95% CI 0.85–0.98], respectively.

Using the relevance maps generated by the Integrated Gradients visualization technique, we identified four PLN phenotype clusters (clusters A to D, Fig. 2). Cluster O is an additional cluster which represents the PLN variant carriers without any disease features in the deformation curve, who were classified as controls by the DNN (n = 27).

As shown in Fig. 2, the features that were considered most important by the DNN for classification of PLN variant carriers were located in the apical septal segments, mid septal segments and apical lateral segments. Temporally, these features were particularly located at end-systole and in the early diastolic phase. Figure 3 shows representative examples of subjects from each cluster. In clusters A, B and C the relevance maps demonstrated different patterns of delayed relaxation in the apical segments. In cluster A, the deformation curves consisted of one single systolic peak, with a notch during the normal upstroke of the curve after aortic valve closure in the septal apical segment (type A pattern: ‘diastolic notch’, Fig. 3A). Importantly, the deformation curves in this cluster did not show additional shortening after aortic valve closure. By manual classification, 47 of the subjects in cluster A (56%) were previously classified as having normal deformation curves in the apical segments. In cluster B, there were two peaks of shortening in the septal apical segment, of which the first peak occurred before or at aortic valve closure, and the second peak after aortic valve closure (type B pattern: ‘double peak’, Fig. 3B). By manual classification, 25 subjects in cluster B (56%) were previously classified as having normal deformation curves in the apical segments. In cluster C, the deformation curves showed pronounced post-systolic shortening in the apical segments (type C pattern: ‘post-systolic shortening’, Fig. 3C). The relevance maps in this cluster focused specifically on diastolic upstroke of the deformation curve in the apical and mid septal segment, where high diastolic strain rate values were found (Supplemental Table 1). By manual classification, 36 subjects in cluster C (54%) were previously classified as having normal deformation curves in the apical segments. The deformation curves in cluster D were characterized by decreased systolic peak strain values, which was considered most important in the septal and lateral apical segments (type D pattern: ‘reduced peak strain’, Fig. 3D). By manual classification, all subjects in cluster D were previously classified as having abnormal deformation curves in the apical segments.

Additional clinical data of the phenotype clusters can be found in Table 2. Subjects in cluster O, who were recognized as controls by the DNN, were relatively young (36 [IQR 27–46] years), were all identified by family screening and the majority was pre-symptomatic (n = 22, 82%). Conventional echocardiographic measurements in this cluster were without exception within normal limits. Subjects in cluster A were older than subjects in cluster O (40 [IQR 24–53], but still had normal conventional echocardiographic measurements. Subjects in cluster B were older than subjects in cluster A (46 [IQR 34–64] years) with slightly reduced relaxation parameters (E-wave velocity and e’). All other conventional parameters in this cluster were preserved. The subjects in cluster A and B both had more electrical disease expression compared to cluster O with regard to T-wave inversions, low QRS voltages, VA history and PVC burden per 24 h (Table 2). Subjects in cluster C were the youngest of all clusters (36 [IQR 24–44] years), and were characterized by relatively high LV end-diastolic and end-systolic volumes (58.3 [IQR 52.9–65.8] ml/m² and 24.7 [IQR 20.8–27.9] ml/m², respectively) and a particularly high rate of low QRS voltages (38.1%). Finally, cluster D consisted of the oldest subjects (54 [IQR 43–61] years) with the most advanced disease, with high rates of heart failure and VA, and severely impaired conventional echocardiographic LV and RV measurements.

Follow-up data was available for 240/278 PLN variant carriers (86%). During a median follow-up duration of 3.0 years [IQR 1.4–5.2 years], 34 patients (14%) experienced the sustained VA endpoint. These were 4 (5%) from cluster A, 3 (8%) from cluster B, 4 (7%) from cluster C and 23 (46%) from cluster D. None of the subjects that were allocated to cluster O experienced a sustained arrhythmia during follow-up. Figure 4 shows the Kaplan–Meier curves for this endpoint, stratified by the phenotype clusters.

Myocardial deformation curves contain a large amount of information on intrinsic mechanical myocardial properties. However, the interpretation of deformation curves is challenging, which hampers the routine clinical use of these curves. Previously, attempts have been made to classify deformation curves by extracting certain parameters manually and applying disease-specific cut-off values [2]. However, this approach is complicated by the fact that strain values may fall within normal limits during early stages of disease, and peak strain values are influenced by variety of parameters such as pre- and afterload. Instead, assessment of the entire deformation curve to detect disease-specific patterns is probably more appropriate [5]. However, the knowledge of such disease-specific patterns is currently limited. Therefore, we investigated the utility of DNN-based classification combined with an advanced visualization technique to detect and visualize disease features in deformation curves. While current DNN visualization techniques usually only provide insight on the individual subject level, we used a clustering approach to describe different phenotype clusters among the investigated disease population.

This study was specifically designed to explore the unique strain characteristics in patients with a homogeneous genetic background. In a recent observational study, we found that post-systolic shortening in the LV apex is a typical deformation pattern in PLN p.Arg14del variant carriers who are in early stages of disease [3]. In more advanced stages of disease, we observed a global reduction of peak strain. These deformation patterns were now also detected by our DNN-based approach, and are shown in cluster C and D, respectively. The DNN-based approach expanded our knowledge by showing that the reduction of peak strain in cluster D is most pronounced in the apical segments. Post-systolic shortening was also present in this cluster, but the DNN did not consider it to be of additional value on top of the reduced peak strain in the apical segments.

Strikingly, the results of this study also added novel data on specific diastolic myocardial behavior that has not been recognized before. In clusters A and B, we observed specific patterns of delayed relaxation, where overt post-systolic shortening is not (yet) present. It is conceivable that these disease features in early diastole can be explained by the pathophysiological mechanism of the PLN p.Arg14del variant, since it is thought that this variant causes PLN to inhibit calcium reuptake into the sarcoplasmic reticulum, leading to a diastolic calcium overload in the cardiomyocyte [16].

Since the prognostic value for developing VA was similar among clusters A, B and C, one could argue that there is no clinical relevance in distinguishing these clusters. However, it is important to note that by manual classification, a significant part of the deformation curves from clusters A, B and C would be classified as normal (56%, 56% and 54%, respectively). With our approach, we could reclassify the deformation curves from these individuals from normal to abnormal. This highlights the clinical relevance of our approach, considering the difference in development of VA between the normal cluster O and the abnormal clusters A, B and C.

Clusters A to D all contained PLN variant carriers who were correctly classified by the DNN. In addition, we described cluster O which contained variant carriers who were not recognized by the DNN as variant carriers due to the absence of disease features. It is known that the PLN p.Arg14del variant has age-related penetrance with symptoms often beginning around the fifth decade, which implies that there is a pre-phenotypical phase in which disease expression is absent [8]. Therefore, it is conceivable that the subjects in cluster O are the ones who still lack disease expression, and who can therefore not be distinguished from population controls by the DNN. This is supported by the fact that subjects in this cluster were mostly young, pre-symptomatic family members who were identified by family screening. The follow-up data demonstrated that the disease course in this cluster was benign; none of the subjects in cluster O developed a sustained ventricular arrhythmia during 3 years of follow-up.

Since the PLN p.Arg14del variant is characterized by large phenotypical variability, this novel DNN-based approach may be useful for classifying subjects with this particular variant into one of the phenotype clusters. Subjects with this variant who do not exhibit disease features (cluster O) can perhaps undergo low-frequency follow-up, whereas subjects who are classified in the most advanced disease cluster (cluster D) may possibly benefit from more intensive follow-up and appropriate therapeutic intervention, for example ICD implantation. Future studies should elaborate on the prognostic value of these clusters, also considering other clinical variables [17].

Besides using this approach for classification purposes, this approach is very useful to gain insight into characteristic disease features that are concealed in the deformation curves. Ideally, this approach should be applied to other diseases, which will expand our knowledge on disease-specific deformation patterns and potentially improve the interpretation of deformation curves in clinical practice. In this proof-of-concept study we only used the deformation curves from the apical 4-chamber view, but in future studies it would be of interest to include the deformation curves from all apical views, including the atrial and right ventricular deformation curves. Combining the deformation curves with ECG data in a DNN model would also be of interest in future studies.

This study has several limitations to address. First, our control group was derived from a population-based cohort, while the group of variant carriers were scanned and analyzed in other centers. This may have induced center-specific differences between the controls and the variant carriers. Second, the developed DNN was not validated in an external cohort. However, the aim of this study was not to investigate the performance of the algorithm, but to propose a novel way to detect diseases-specific features in deformation curves. Third, the visualization technique that was used in this study only shows the temporal location of important features in the deformation curve, but it does not specify what the feature exactly is. The actual description of the features was performed by visual assessment and should be validated in future studies. Fourth, since detection of regional abnormalities by deformation imaging is limited by inter-vendor variability, it remains unknown whether the results can be generalized to other vendors [18]. Fifth, the strain measurements in our study were performed by different operators. However, inter- and intra-observer agreement of several strain measurements were reported to be good in previous studies by our group [19]. Last, we reported follow-up data to explore prognostic differences among the clusters, but the number of events in this cohort was too low to perform appropriate statistical survival analyses. This remains to be investigated in future studies with longer follow-up intervals.