Arrhythmogenic right ventricular cardiomyopathy/dysplasia (ARVC/D) is a rare hereditary condition presenting at an incidence of 1 in 2500 to 1 in 5000 in the general population, with notable geographical variations in disease prevalence [1]. ARVC/D is characterized by genetic mutations in desmosomal genes [2, 3] and accompanying aberrations in cardiomyocyte cell-cell adhesion, leading to early cardiac regional anatomical abnormalities, typically confined to the right ventricular inflow tract, outflow tract, and apex, which together constitute the “triangle of dysplasia” [4]. Disease progression is in turn dominated by diffuse thinning of the right ventricular wall with cardiomyocyte loss and corresponding fibrofatty replacement of the myocardium [5]. These pathological alterations not only disturb the native electrical conduction system, thereby predisposing affected patients to malignant ventricular arrhythmias and sudden cardiac death (SCD) [6, 7], but also potentially induce left ventricular dysfunction and subsequent dual-chamber cardiac failure [8].

The definitive diagnosis of ARVC/D is challenging owing to the absence of a single set of parameters sufficiently specific to the disease [1]. As such, the current diagnostic criterion seeks to amalgamate a series of clinical, pathological and genetic features most commonly observed in affected patients, amongst which electrocardiographic and echocardiographic parameters are the most prominent [9]. The evident heterogeneity in the phenotypic presentation and complications associated with ARVC/D poses difficulties to optimal management [10]. Current therapies are primarily geared towards the prevention of lethal ventricular arrhythmias, and implantable cardioverter-defibrillator (ICD) placement has hitherto proven to be the only effective strategy in reducing long-term mortality [1]. Such dilemmas in the management of these patients are further compounded by the apparent underreporting of prognostic markers to assist risk stratification in the clinical setting.

The present study aims to investigate the risk of detrimental cardiovascular events in an Asian population of ARVC/D patients, including the incidence of malignant ventricular arrhythmias, new-onset heart failure with reduced ejection fraction (HFrEF), as well as long-term mortality. Moreover, the prognostic importance of several clinical parameters will be examined in an attempt to identify possible markers with predictive value that could improve overall assessment and therapeutic guidance. These markers will be utilized to not only construct scoring systems, but also, for the first time, employ machine learning algorithms designed to enhance outcome prediction.

In this ARVC/D cohort (n = 109), the median age was 61 [46–71] years and 63 (58%) were male. The baseline characteristics are presented in Table 1, with patients stratified according to the development of incident VT/VF. The median ages of the VT/VF (n = 49) and non-VT/VF (n = 60) groups were 65 [45–71] years and 59 [46.5–71.5] years, respectively with similar ages at diagnosis of ARVC/D. Patients who developed incident VT/VF tended to present more often with right ventricular dilatation and systolic dysfunction. This group also demonstrated a significantly longer QRS duration, which took a median value of 108.0 [96.0–129.0] ms, as well as a significantly greater proportion of subjects developing epsilon waves (n = 13; 27%).

Anti-arrhythmic therapy was prescribed in the form of amiodarone (n = 48), sotalol (n = 34) and mexiletine (n = 4), with majority of patients taking these medications after the first episode of VT/VF (amiodarone: n = 28; sotalol: n = 21; mexiletine: n = 4). Some subjects who had received anti-arrhythmic therapy experienced subsequent VT/VF episodes (amiodarone: n = 40; sotalol: n = 27; mexiletine: n = 3), HFrEF (amiodarone: n = 17; sotalol: n = 10; mexiletine: n = 1) or cardiovascular-related mortality (amiodarone: n = 7; sotalol: n = 5; mexiletine: n = 0) post-therapy. A total of 58 patients received implantable cardioverter-defibrillator (ICD) placement either for VT/VF prophylaxis following ARVC/D diagnosis (n = 32) or to prevent VT/VF recurrence after the first episode (n = 26). After ICD implantation, a total of 30 patients experienced at least one episode of VT/VF and six suffered cardiovascular-related mortality.

3.1 Predictors of Adverse Outcomes

In the following ARVC/D cohort, 49 patients and 24 patients developed incident VT/VF and new-onset HFrEF, respectively. A total of 5 patients underwent cardiac transplantation, and 14 patients passed away during follow-up, 10 of which suffered from ARVC/D-related complications, namely SCD and HFrEF, whereas the remaining 4 suffered non-cardiovascular-related deaths. The results of univariate Cox proportional-hazards regression analysis for predicting incident VT/VF, new-onset HFrEF and all-cause mortality are reported in Supplementary Tables 1–3, respectively, with the corresponding Kaplan-Meier survival curves shown in Fig. 1.

Fig. 1.

Kaplan-Meier survival curves for incident VT/VF, new-onset HFrEF and all-cause mortality.

Incident VT/VF (Supplementary Table 1) was associated with prolonged QRS duration, presence of epsilon waves, and syncope, with only the foremost retaining significance after multivariate adjustment (p 0.05). Regarding secondary outcomes, univariate predictors of new-onset HFrEF (Supplementary Table 2) included female gender, prolonged QTc duration, presence of epsilon waves and TWI in any lead except aVR/V1, all of which retained significance in the multivariate model (p 0.05). Likewise, all-cause mortality (Supplementary Table 3) was similarly associated with all univariate predictors of new-onset HFrEF, in addition to prolonged QRS duration (p 0.05). Resultant significant parameters in multivariate Cox regression included female gender, prolonged QRS duration, prolonged QTc duration, and presence of epsilon waves (p 0.05).

3.2 Scoring System for New-Onset VT/VF in ARVC/D

Significant clinical and electrocardiographic parameters in univariate Cox regression (p 0.05) were used to design a scoring system to predict new-onset VT/VF in ARVC/D. Receiver operator characteristics (ROC) curves were used to determine optimal cut-offs for significant continuous variables. The optimal cutoff value for QRS duration was 98.5 ms (AUC: 0.69; sensitivity = 72%; specificity = 67%). After categorization, this parameter retained significance in univariate prediction of incident VT/VF in ARVC/D, and was therefore eligible for inclusion. QRS duration $>$ 98.5 ms, along with presence of syncope and epsilon waves were subsequently used to form the final scoring system (Supplementary Table 4a). Subjects who developed VT/VF presented with a median score that was 0.5 points higher than those who remained free of VT/VF. Cox proportional-hazards analysis revealed that patients with a per unit increase in the score had a 74% higher risk of incident VT/VF (HR: 1.74; 95% CI: 1.30–2.33; p 0.001) (Supplementary Table 4b). Categorization of the VT/VF score using the maximal rank statistics approach (Fig. 2) revealed an optimal cut-off of 1 point. Subsequent Cox proportional-hazards analysis demonstrated that patients with a score $\ge$1 point had an approximate 2-fold increase in risk of new-onset VT/VF (Supplementary Table 4c).

Fig. 2.

Optimal cut-off for the scoring system for incident VT/VF, new-onset HFrEF and all-cause mortality with the maximal rank statistics approach.

3.5 Machine Learning to Predict Incident VT/VF, New-Onset HFrEF and All-Cause Mortality

wRSF models were used to predict primary and secondary outcomes with the identified significant parameters in univariate analysis as input variables (p value 0.05). The optimal tree number used to build each wRSF model was selected by the error rate minimization with iteration approach. The tree number selected to predict VT/VF, new-onset HFrEF and all-cause mortality was 250, 200, and 400 respectively (Fig. 3). The derived importance value ranking of the variables is shown in Table 2. QRS duration $>$98.5 ms was the most predictive variable for incident VT/VF, followed by presence of syncope and presence of epsilon waves. In contrast, prolonged QTc duration demonstrated the strongest predictivity for HFrEF, followed by TWI in any lead except aVR/V1, presence of epsilon waves, and female gender. As it pertains to all-cause mortality, QRS $>$122.5 ms was the most important predictor, followed by presence of epsilon waves, QTc $>$448.5 ms, and female gender. As shown in Table 3, the ability of the wRSF model to predict the primary outcomes was compared with an RSF model and multivariate Cox model, based on evaluation metrics of precision, recall, AUC and C-index. Findings indicate that the wRSF models performed best in the prediction of all three outcomes based on the significant univariate predictors.

Fig. 3.

Optimal tree number of wRSF model and variable importance ranking to predict incident VT/VF, new-onset HFrEF and all-cause mortality.

Table 2.Variable importance ranking generated by wRSF models for primary and secondary outcomes.

Incident VT/VF
Variable	Importance Value
QRS $>$98.5	0.0778
Syncope	0.0054
Epsilon wave	0.0006
New-onset HFrEF
Variable	Importance Value
QTc $>$437.5 ms	0.1180
TWI in any lead except aVR/V1	0.0347
Epsilon wave	0.0147
Female Gender	0.0066
All-cause mortality
Variable	Importance Value
QRS $>$122.5	0.0700
Epsilon wave	0.0377
QTc $>$448.5	0.0291
Female Gender	0.0251

Table 3.Performance comparison between wRSF, RSF, and multivariate Cox model. Italicized values indicate the best model for that particular performance metric.

Incident VT/VF
Model	Precision	Recall	AUC	C-index
wRSF model	0.8352	0.8478	0.8341	0.8202
RSF model	0.8283	0.8372	0.8161	0.8116
Multivariate Cox model	0.7493	0.7793	0.7524	0.7835
New-onset HFrEF
Model	Precision	Recall	AUC	C-index
wRSF model	0.8290	0.8305	0.8363	0.8182
RSF model	0.8053	0.8172	0.8184	0.8021
Multivariate Cox model	0.7493	0.7641	0.7620	0.7770
All-cause mortality
Model	Precision	Recall	AUC	C-index
wRSF model	0.7322	0.7395	0.7549	0.7481
RSF model	0.7071	0.6922	0.6984	0.7012
Multivariate Cox model	0.6734	0.6808	0.6853	0.6744

The present study is novel as it is among the first territory-wide investigations of ARVC/D patients in Hong Kong, allowing for the development of the first risk scores that include both electrocardiographic and clinical parameters for predicting incident VT/VF, new-onset heart failure as well as mortality in ARVC/D. In addition, to our knowledge, this is also the first investigation to apply machine learning algorithms to assess ARVC/D prognosis, in turn demonstrating enhanced risk prediction for outcomes with such algorithms when compared to other analytical models.

The substrate for arrhythmogenesis in ARVC/D is a combination of conduction and repolarization abnormalities associated with structural alterations in the right ventricle [17]. However, recent work has demonstrated that non-classical forms of arrhythmogenic cardiomyopathy, namely left dominant or biventricular forms, are more prone to ventricular arrhythmias than classical ARVC/D [18]. Moreover, the atria, in addition to the ventricles, are also abnormal in ARVC/D [19] with complications such as sinoatrial arrests and atrial fibrillation [20, 21]. Different ECG indices have been identified as risk markers of ventricular arrhythmias [22], amongst which the epsilon wave is the classical pathognomonic feature of ARVC/D [23, 24]. Repolarization criteria, including TWI in inferior leads, a precordial QRS amplitude ratio of $\le$0.48, and QRS fragmentation also constitute valuable variables for predicting adverse outcomes in this disease [25]. Repolarization abnormalities, such as TWI, are important, and electroanatomic mapping areas have shown to be proportional to extent of TWI on 12-lead ECG [26, 27]. Strain imaging by speckle-tracking echocardiography has been used to risk-stratify patients in heart failure [28] and recent work has shown that incorporation of mechanical dispersion can further improve risk prediction [29], as heart failure is typically under-recognized in ARVC/D [30].

Ventricular arrhythmias are a common occurrence amongst patients with so-called inherited cardiac arrhythmias, including ARVC/D, Brugada syndrome, long QT-syndrome and short QT-syndrome, amongst which ARVC/D cohorts have shown to present with the highest rates of frequent ventricular premature complexes, (non)sustained ventricular tachycardias, and malignant ventricular tachyarrhythmias [31]. Several large-scale studies have reported on the clinical characteristics and predictors of ventricular arrhythmias in ARVC/D patients. In 131 definite ARVC/D patients, spontaneous sustained ventricular arrhythmias, cardiac syncope, male gender, proband, and inducibility in electrophysiology study were all significant predictors of incident sustained VT/VF and SCD [32]. The same group further studied the phenotype of ARVC/D patients with late presentation, demonstrating that this subpopulation does not confer a benign prognosis and has a high arrhythmic risk [33]. Another study of 135 patients identified prolonged QRS duration on signal-averaged ECG, non-sustained VT on 24 h-ECG, and the absence of negative T waves in lead aVR on a 12-lead surface ECG as significant predictors of recurrent sustained ventricular arrhythmias and hospitalization due to ventricular arrhythmias in ARVC/D [34]. Moreover, an investigation of 137 patients from France found that low LVEF, positive electrophysiological studies and physical activity $>$6 h/week were shown to be independently associated with the development of ventricular arrhythmias [35]. The findings of the aforementioned studies clearly demonstrate the high incidence of ventricular arrhythmias within ARVC/D cohorts, thereby necessitating the use of prophylactic antiarrhythmics to reduce their occurrence [36], along with radiofrequency catheter ablation for patients who end up developing these rhythm abnormalities [37].

Recently, a systematic review and meta-analysis summarized the current literature, identifying consistently predictive risk factors in patients with definite ARVC across different studies [38]. These were male sex, syncope, TWI in lead V3, right ventricular dysfunction, and prior (non)sustained VT/VF. Our present work extends these findings by demonstrating that longer QRS duration, the presence of epsilon waves and TWI in 2/3 inferior leads were significantly associated with incident VT/VF, albeit only QRS remained a significant predictor after multivariate adjustment. Several parameters retained significance in multivariate prediction of new-onset HFrEF, including longer QTc duration, presence of epsilon waves, TWI in any lead except aVR/V1 and female gender. Likewise, longer QRS duration, presence of epsilon waves, LVEF and age at diagnosis of ARVC/D were all significantly associated with all-cause mortality in multivariate analysis. It was then possible to further enhance risk prediction through the application of wRSF model analysis, which we have recently used to better risk prediction in acquired long QT syndrome [39] as well as Brugada syndrome [40]. The wRSF model was able to improve the risk stratification for incident VT/VF, new-onset HFrEF and all-cause mortality in this ARVC/D cohort.

Furthermore, the clinical heterogeneity typically observed amongst populations with ARVC/D makes the use of scoring algorithms a potentially useful method to amalgamate the different patient parameters for the purposes of risk stratification. Such an approach has been adopted previously in a large cohort of 528 ARVC/D patients to predict the long-term risk of ventricular arrhythmias. The model constructed, which included age, male gender, cardiac syncope in the prior 6 months, prior non-sustained VT, number of PVCs in 24 h, number of leads with TWI and right ventricular ejection fraction, demonstrated an improved ability to estimate risk of ventricular arrhythmias and guide decision-making in ICD implantation for such patients [41]. A meta-analysis identified the following 11 variables as the most important factor for predicting arrhythmic events: (1) male gender, (2) presyncope, (3) left ventricular dysfunction, (4) TWI in inferior leads, (5) proband status, (6) late potentials, (7) syncope, (8) inducibility at electrophysiological study, (9) right ventricular dysfunction, (10) epsilon waves, and (11) premature ventricular contractions greater than 1000/24 h [42]. To our knowledge, such scoring algorithms have not been used to investigate outcomes beyond VT/VF in ARVC/D cohorts. As such, the present study also developed two multi-parametric scores for predicting new-onset HFrEF and all-cause mortality, respectively, both of which demonstrated efficacy in assessing ARVC/D patient prognosis.

This investigation has limitations that should be noted. Firstly, data is primarily based on patients of Chinese ethnicity and therefore lacks the subject variability needed for a comprehensive evaluation of ARVC/D, which itself presents with a heterogeneous phenotype. Secondly, several subjects were prescribed amiodarone and/or sotalol as treatment, both of which have been previously shown to influence certain ECG parameters, for instance by lengthening QTc interval. This could have potentially influenced the reported relationship between QTc duration and new-onset HFrEF as well as all-cause mortality. Finally, the adverse ECG findings, such as prolongation of QRS duration or QTc duration and the presence of epsilon waves, are likely linked to a greater underlying disease severity that in turn leads to malignant ventricular arrhythmias. As such, these parameters possibly only serve as markers, as opposed to outright predictors, of VT/VF, albeit further study is required to confirm this.

The phenotypic variability and adverse prognosis potentially associated with ARVC/D necessitates a multimodality approach for risk stratification that includes both clinical and electrocardiographic parameters. The findings of the current investigation are the first to be demonstrated in an Asian population, thereby extending the generalizability of pre-existing data to that of Asian cohorts. Moreover, the present study is among the first to not only demonstrate the use of scoring systems comprising of both electrocardiographic and clinical parameters in the assessment of long-term outcomes in ARVC/D, but also to employ combinatorial methods involving machine learning algorithms to evaluate prognosis. Such algorithms are able to account for underlying inter-variable interactions, thereby improving overall event and survival prediction. As a result, with further study into their use, machine learning techniques could possibly provide an alternative, more effective avenue to assess patient prognosis in such heterogeneous disease conditions.

Conceptualization—IL and GT; methodology—IL, JZ and QZ; software—IL, JZ and QZ; validation—IL and JZ; formal analysis—IL and JZ; investigation—IL, JZ, SL, KLi, KLe, JH, YL, GL, TL, WW, IW, NM, CM, QZ, GT; resources—GT, TL, QZ; data curation—IL and GT; writing—original draft preparation—IL, JZ, SL, KLi, KLe, JH, YL, GL, TL, WW, IW, NM, CM, QZ, GT; writing—review and editing—IL, JZ, SL, KLi, KLe, JH, YL, GL, TL, WW, IW, NM, CM, QZ, GT; visualization—IL and JZ; supervision—TL, QZ and GT; project administration—QZ and GT; funding acquisition—not applicable. All authors have read and agreed to the published version of the manuscript.

This study was approved by The Joint Chinese University of Hong Kong – New Territories East Cluster Clinical Research Ethics Committee (Approval number: 2019.422). The need for patient consent was waived by the Committee owing to the retrospective and observational nature of the study.

Not applicable.

This research received no external funding.

The authors declare no conflict of interest. Gary Tse, Sharen Lee and Tong Liu are serving as one of the Guest Editors of this journal. We declare that Gary Tse, Sharen Lee and Tong Liu had no involvement in the peer review of this article and has no access to information regarding its peer review. Full responsibility for the editorial process for this article was delegated to Stefan Peters.