2025 Volume 89 Issue 2 Pages 240-250
Background: Implantable cardioverter defibrillators (ICDs) reduce mortality associated with ventricular arrhythmia in high-risk patients with cardiovascular disease. Machine learning (ML) approaches are promising tools in arrhythmia research; however, their application in predicting ventricular arrhythmias in patients with ICDs remains unexplored. We aimed to predict and stratify ventricular arrhythmias requiring ICD therapy using 12-lead electrocardiograms (ECGs) in patients with an ICD.
Methods and Results: This retrospective analysis included 200 adult patients who underwent ICD implantation at a single center. Patient demographics, clinical features, and 12-lead ECG data were collected. Unsupervised learning techniques, including K-means and hierarchical clustering, were used to stratify patients based on 12-lead ECG features. Dimensionality reduction methods were also used to optimize clustering accuracy. The silhouette coefficient was used to determine the optimal method and number of clusters. Of the 200 patients, 59 (29.5%) received appropriate therapy. The mean age of patients was 62.3 years, and 81.0% were male. The mean follow-up period was 2,953 days, with no significant intergroup differences. Hierarchical clustering into 3 clusters proved to be the most accurate (silhouette coefficient=0.585). Kaplan-Meier curves for these 3 clusters revealed significant differences (P=0.026).
Conclusions: We highlight the potential of ML-based clustering using 12-lead ECGs to help in the risk stratification of ventricular arrhythmia. Future research in a larger multicenter setting may provide further insights and refine ICD indications.
Implantable cardioverter defibrillators (ICDs) are highly effective in reducing mortality rates associated with ventricular tachyarrhythmia among high-risk patients with cardiovascular disorders.1,2 Previous studies have attempted to predict ventricular arrhythmia and prevent sudden cardiac death using ICDs.3,4 However, recent findings have revealed adverse long-term prognoses of ICD implantation for primary prevention in patients with non-ischemic cardiomyopathy.5 Although current guidelines rely on left ventricular (LV) ejection fraction (LVEF), New York Heart Association class, and the presence of non-sustained ventricular tachycardia (VT) for ICD implantation indications,6 accurately predicting ventricular arrhythmias remains challenging.
Previous efforts have been aimed at predicting ventricular arrhythmias using 12-lead electrocardiography (ECG),7,8 identifying significant factors, yet facing difficulties in actual prediction. Recent arrhythmia studies use machine learning (ML) approaches and have explored the prediction of future ventricular arrhythmias through ML techniques applied to ventricular signals.9 Furthermore, ML-based models show superior predictive capabilities compared with conventional clinical models. Combining 12-lead ECG data with clinical features in ML-based models significantly enhances the prediction of patient outcomes following ablation for atrial fibrillation.10
In the realm of ICD research, a prior study used ML to forecast electrical storms using remote monitoring data.11 However, to date, no studies have used ML to predict ventricular arrhythmias in patients with implanted ICDs. Our hypothesis was that ML-based models trained on 12-lead ECGs would offer enhanced predictive accuracy in this domain. Therefore, using unsupervised learning, we investigated the prediction and stratification of ventricular arrhythmias requiring ICD therapy using 12-lead ECGs in patients with ICD implantation.
This was a retrospective analysis of 219 consecutive adult patients who underwent ICD/cardiac resynchronization therapy-defibrillator (CRT-D) implantation between January 2001 and August 2023 at a single center. Patients were required to have undergone a 12-lead ECG within 1 week and echocardiography within 3 months before ICD/CRT-D implantation. Patients who underwent ventricular pacing before ICD/CRT-D implantation and those with missing data were excluded from the analyses.
The study protocol adhered to the Declaration of Helsinki and the study was approved by the Yokohama Minami Kyosai Hospital Institutional Review Board, with an opt-out option for patients or proxies. The need for informed consent was waived by the Yokohama Minami Kyosai Hospital Institutional Review Board due to the retrospective nature of the study. The corresponding author had full access to all study data and is responsible for its integrity and analysis.
EndpointThe primary endpoint was the occurrence of ventricular arrhythmias requiring ICD therapy, including antitachycardia pacing (ATP) or shock. This evaluation was conducted by 2 board-certified members of the Japanese Heart Rhythm Society, who assessed the appropriateness of the therapy.
Patient Background, Clinical Features, and Non-Invasive ExaminationsPatient background, extracted from electronic health records, encompassed several key factors, including patient age at the time of ICD/CRT-D implantation, sex, body mass index, and whether the patient had undergone VT ablation before ICD/CRT-D implantation. Documented clinical comorbidities included hypertension, dyslipidemia, diabetes, and chronic kidney disease. Underlying heart diseases were categorized as ischemic cardiomyopathy, dilated cardiomyopathy, hypertrophic cardiomyopathy, valvular heart disease, and channelopathy. The use of antiarrhythmic and cardioprotective drugs was recorded. In cases in which appropriate ICD/CRT-D therapy was observed, the medications taken by the patient at that time were included in the analysis. Alternatively, if no ICD/CRT-D therapy was observed, the medication used at the final follow-up was considered.
To gather 12-lead ECG data, an automated system (ECAPs12c; Nihon-Kohden, Tokyo, Japan) was used. This system measures 12-lead ECG data at the microvolt level, defining the wave amplitude as the absolute distance from the apex of each wave to the baseline. The 12-lead ECG variables were measured using 10-s waveforms. The rate-corrected QT interval (QTc) was calculated by the ECAPs12c system using the modified Framingham formula (QTc = QT + [1,000 − R-R interval] / 7). The middle of the ST level (ST-MID) was defined as the ST level at one-sixteenth of the preceding R-R interval after the J point, and the end of the ST level (ST-END) was defined as the ST level at two-sixteenths of the preceding R-R interval after the J point.12 The 12-lead ECG measurement method is shown in the Supplementary Figure. All echocardiographic examinations were performed using the Vivid system (GE Healthcare, Chicago, IL, USA).
ICD TherapyThe ICDs/CRT-Ds used in this study were manufactured by Biotronik (Berlin, Germany), Medtronic (Minneapolis, MN, USA), Boston Scientific (Marlborough, MA, USA), Sorin (Saluggia, Italy), and Abbott (Abbott Park, IL, USA). Because this was a retrospective study, the selection of single- or dual-chamber devices with or without LV leads and ICD/CRT-D settings was based on the attending physician’s discretion. Parameters, such as the choice between single- and dual-chamber devices, the use of CRT, the number of detection zones, and the detection cycle length, were included in the statistical analysis. In addition, the analysis considered the frequency of ICD therapies in patients who had a history of such treatments.
Statistical AnalysisFisher’s exact test or 1-way analysis of variance (ANOVA) was used to assess the significance of differences in categorical variables. Normally distributed continuous variables are presented as the mean±SD, whereas non-normally distributed variables are presented as the median with interquartile range (IQR). Appropriate statistical tests, such as Student’s t-test, the Mann-Whitney U test, or Kruskal-Wallis tests, were used accordingly. The Bonferroni method was used for pairwise comparisons in ANOVA and Kruskal-Wallis tests. Each cluster was evaluated using Kaplan-Meier curves. The threshold for statistical significance was set at P<0.05. All statistical analyses were performed using EZR version 1.61 (Saitama Medical Center, Jichi Medical University, Saitama, Japan),13 which functions as a graphical user interface for R (R Foundation for Statistical Computing, Vienna, Austria).14
Predictive Model Construction by ML and 12-Lead ECG CharacteristicsThe initial step involved excluding 12-lead ECG parameters with >20% missing data from the analysis. Redundant variables for clustering were reduced by calculating Pearson’s correlation coefficients among the candidate variables. Any pair of variables showing correlation coefficients >0.6 was eliminated from the clustering model (keeping the variable that was most informative and had the least missing values).15 To ensure data completeness, missing values were imputed using the mean value. Dimensionality reduction techniques were used to adjust the parameters, simplifying the data to 1, 2, or 3 dimensions to improve clustering accuracy, as detailed below.
Principal Component Analysis Principal component analysis is a dimensionality reduction technique used to simplify complex datasets by transforming variables into a new set of orthogonal (uncorrelated) variables called principal components. These components capture the maximum variance in the data, allowing for a more compact representation while retaining as much information as possible.16
t-Distributed Stochastic Neighbor Embedding (t-SNE) t-SNE is a non-linear dimensionality reduction technique used to visualize high-dimensional data in lower-dimensional space by measuring similarities between pairs of high-dimensional data points and then optimizing a low-dimensional representation to preserve these similarities as much as possible.17
Uniform Manifold Approximation and Projection (UMAP) UMAP is another non-linear dimensionality reduction technique similar to t-SNE but has computational advantages, aiming to preserve both local and global structures in data by constructing a low-dimensional embedding that optimally represents the manifold structure of the data. UMAP has gained popularity because of its scalability and ability to capture complex patterns in high-dimensional data while being computationally efficient.18
Each method serves as a powerful tool to analyze and visualize high-dimensional datasets, and each has its own strengths and applications.
Python 3.12.1 served as the programming language for ML implementation. K-means clustering and hierarchical clustering were the algorithms applied to the unsupervised learning, as detailed below.
K-means clustering categorizes unlabeled data by grouping them according to their features rather than using predefined categories. Each cluster within this technique is established and characterized by its centroid, which represents the central collection of features. Each data point is then allocated to its nearest centroid, which is typically determined using a chosen distance function. Following the initial assignment of data points, the centroids are recalculated by computing the mean of all data points assigned to each specific cluster. This iterative process of assignment and centroid recalculation continues until certain stopping criteria are met, which may include scenarios in which there are no further changes in cluster assignments, minimization of the sum of distances within clusters, or reaching a specified maximum iteration threshold.19 We applied k-means++ to the algorithm to select the initial values.20
Hierarchical clustering is a cluster analysis method designed to build cluster hierarchies. In agglomerative hierarchical clustering, the algorithm sequentially merges pairs of nodes with the shortest distance until all nodes are aggregated into a single cluster. Ward’s method was used for analysis. The distance between 2 nodes was defined as the weighted Euclidean distance, and the sum of the within-cluster variance was controlled to be minimal at each step.21
The optimal number of clusters was determined based on the elbow method for k-means and silhouette analysis for hierarchical clustering.22,23 Following this, silhouette coefficients were used to evaluate the quality of both clustering methods. Finally, a clustering approach that demonstrated superior performance, as indicated by the silhouette coefficients, was selected for further analysis and implementation.
Kaplan-Meier curves were used to evaluate the risk of developing ventricular arrhythmias within each cluster. Subsequently, the 12-lead ECG characteristics for each cluster were extracted and plotted to scrutinize the relationship between the 12-lead ECG features and the risk of developing ventricular arrhythmias. Graphical representations showing the ECG characteristics of each cluster were generated for a comprehensive analysis.
In all, 219 patients underwent ICD/CRT-D implantation between January 2001 and August 2023 at Yokohama Minami Kyosai Hospital. Seventeen patients were excluded due to ventricular pacing prior to implantation and another 2 were excluded due to missing data, leaving 200 patients for analysis. Of these 200 patients, 59 (29.5%) patients received appropriate therapy.
The mean age of the patients was 62.3 years and 81.0% were male. The mean follow-up period was 2,953 days, with no significant intergroup differences. Statistically significant factors included a history of VT ablation prior to ICD/CRT-D implantation, administration of an angiotensin receptor-neprilysin inhibitor, and sodium-glucose cotransporter 2 inhibitors. Notably, in the context of ICD/CRT-D settings, the group receiving appropriate ICD/CRT-D therapy adhered to more stringent criteria, particularly concerning the maximum detection cycle length and the number of detection zones.
Although this study incorporated ICD/CRT-D implantation for both primary and secondary prevention, the median LVEF was low (38%). There were no echocardiographic parameters that differed significantly between the groups with and without appropriate therapy. A detailed overview of patient characteristics is provided in Table 1.
Baseline Characteristics of the Study Population
| All patients (n=200) |
Appropriate therapy | P value | ||
|---|---|---|---|---|
| Yes (n=59) | No (n=141) | |||
| Follow-up (days) | 2,609.5 [1,257.5~4,533.3] |
2,655.0 [1,583.5~4,564.0] |
2,504.0 [1,098.0~4,321.0] |
0.224 |
| Patient background | ||||
| Male sex | 162 (81.0) | 49 (83.1) | 113 (80.1) | 0.697 |
| Age (years) | 62.3±14.8 | 62.5±15.1 | 62.2±14.7 | 0.903 |
| BMI (kg/m2) | 23.1±5.53 | 22.9±4.54 | 23.2±5.92 | 0.674 |
| Comorbidities | ||||
| Hypertension | 73 (36.5) | 20 (33.8) | 53 (37.6) | 0.748 |
| Diabetes | 75 (37.5) | 24 (40.7) | 51 (36.2) | 0.631 |
| Dyslipidemia | 62 (31.0) | 17 (28.8) | 45 (31.9) | 0.739 |
| CKD | 75 (37.5) | 24 (40.7) | 51 (36.2) | 0.260 |
| Current smoking | 70 (35.0) | 23 (39.0) | 47 (33.3) | 0.516 |
| Cardiac event after ICD implantation | ||||
| Aggravation of HF | 52 (26.0) | 17 (28.8) | 35 (24.8) | 0.598 |
| New appearance of ICM | 24 (12.0) | 8 (13.6) | 16 (11.3) | 0.641 |
| Underlying heart disease | ||||
| ICM | 71 (35.5) | 27 (45.8) | 44 (31.2) | 0.054 |
| DCM | 43 (21.5) | 12 (20.3) | 31 (22.0) | 0.852 |
| HCM | 18 (9.0) | 5 (8.5) | 13 (9.2) | 1.000 |
| Valvular heart disease | 5 (2.5) | 2 (3.4) | 3 (2.1) | 0.633 |
| Brugada syndrome | 26 (13.0) | 6 (10.2) | 20 (14.2) | 0.499 |
| Long QT syndrome | 3 (1.5) | 0 (0.0) | 3 (2.1) | 0.557 |
| ARVC | 1 (0.5) | 1 (1.7) | 0 (0.0) | 0.295 |
| IVF/IVT | 10 (5.0) | 1 (1.7) | 9 (6.4) | 0.286 |
| Others | 23 (11.5) | 5 (8.5) | 18 (12.8) | 0.472 |
| Primary prevention | 123 (61.5) | 34 (59.3) | 89 (63.1) | 0.525 |
| VT ablation before ICD implantation | 28 (14.0) | 14 (23.7) | 14 (9.9) | 0.014 |
| Medication | ||||
| Antiarrhythmic agents | ||||
| Class Ia | 7 (3.5) | 3 (5.1) | 4 (2.8) | 0.423 |
| Class Ib | 18 (9.0) | 6 (10.2) | 12 (8.5) | 0.787 |
| Class Ic | 2 (1.0) | 2 (3.4) | 0 (0.0) | 0.086 |
| Amiodarone | 43 (21.5) | 14 (23.7) | 29 (20.6) | 0.706 |
| Bepricol | 11 (5.5) | 4 (6.8) | 7 (5.0) | 0.735 |
| β-blocker | 137 (68.5) | 36 (61.0) | 101 (66.9) | 0.181 |
| ACEi/ARB | 57 (28.5) | 21 (35.6) | 36 (25.5) | 0.171 |
| ARNI | 31 (15.5) | 4 (6.8) | 27 (19.1) | 0.032 |
| MRA | 56 (28.0) | 14 (23.7) | 42 (29.8) | 0.490 |
| SGLT2i | 28 (14.0) | 3 (5.1) | 25 (17.7) | 0.024 |
| ICD programming settings | ||||
| Dual-chamber device | 124 (64.0) | 40 (67.8) | 84 (59.6) | 0.338 |
| RV pacing rate (%) | 0.0 [0.0~1.0] | 0.0 [0.0~1.0] | 0.0 [0.0~1.0] | 0.900 |
| BiVP rate (%) | 0.0 [0.0~0.1] | 0.0 [0.0~0.2] | 0.0 [0.0~0.1] | 0.672 |
| Subcutaneous ICD | 5 (2.5) | 0 (0.0) | 5 (3.5) | 0.324 |
| CRT-D | 45 (22.5) | 10 (16.9) | 35 (24.8) | 0.268 |
| Detection cycle length (ms) | 330.0 [320.0~362.5] |
360.0 [333.0~400.0] |
320.0 [300.0~360.0] |
<0.001 |
| ICD detection zones | <0.001 | |||
| 1 zone | 105 (52.5) | 18 (29.5) | 87 (62.6) | |
| 2 zones | 75 (37.5) | 31 (52.5) | 44 (30.9) | |
| 3 zones | 20 (10.0) | 10 (16.9) | 10 (7.1) | |
| Echocardiography | ||||
| LA diameter (mm) | 40.7±8.45 | 41.4±8.70 | 40.4±8.36 | 0.467 |
| LVEDd (mm) | 56.8±9.94 | 56.7±8.63 | 56.8±10.5 | 0.976 |
| LVEDs (mm) | 44.4±12.5 | 44.5±10.6 | 44.4±13.3 | 0.945 |
| IVS (mm) | 9.7 [8.4~10.9] | 10.3 [8.4~11.4] | 9.1 [8.4~10.8] | 0.060 |
| LVPWT (mm) | 9.4 [8.1~10.6] | 9.7 [8.6~10.9] | 9.4 [8.1~10.4] | 0.285 |
| LVEF (%) | 38.0 [30.8~57.3] | 37.0 [31.0~52.0] | 38.0 [30.0~58.0] | 0.751 |
Unless indicated otherwise, values are presented as n (%), mean±SD, or median [interquartile ranges]. Categorical variables were compared using Fisher’s exact test, and continuous variables were compared using Student’s t-test for normally distributed data or the Mann-Whitney U test for non-normally distributed data. ACEi, angiotensin-converting enzyme inhibitor; ARB, angiotensin receptor blocker; ARNI, angiotensin receptor-neprilysin inhibitor; ARVC, arrhythmogenic right ventricular cardiomyopathy; BiVP, biventricular pacing; BMI, body mass index; CKD, chronic kidney disease; CRT-D, cardiac resynchronization therapy-defibrillator; DCM, dilated cardiomyopathy; HCM, hypertrophic cardiomyopathy; HF, heart failure; ICD, implantable cardioverter defibrillator; ICM, ischemic cardiomyopathy; IVF, idiopathic ventricular fibrillation; IVS, interventricular septum; IVT, idiopathic ventricular tachycardia; LA, left atrial; LVEDd, left ventricular end-diastolic diameter; LVEDs, left ventricular end-systolic diameter; LVEF, left ventricular ejection fraction; LVPWT, left ventricular posterior wall thickness; MRA, mineralocorticoid receptor antagonists; RV, right ventricular; SGLT2i, sodium-glucose cotransporter 2 inhibitor; VT, ventricular tachycardia.
The 59 ICD-treated patients experienced a median of 2.0 ATP events and 1.0 shocks, with no significant change observed before or after adjustment to the shock reduction setting in 2013. For further details on ICD therapy history, refer to Supplementary Table 1.
Unsupervised LearningWe obtained 260 parameters from the ECAPs12c, 44 of which remained after excluding those with missing data or strong correlations. We applied these parameters to ML and observed that hierarchical clustering after dimensionality reduction to 1 dimension using UMAP yielded the highest silhouette coefficients. The silhouette coefficients for each method are listed in Supplementary Table 2. The number of clusters was then determined by silhouette analysis, resulting in 3 clusters, with 71 patients in Cluster 1, 72 patients in Cluster 2, and 57 patients in Cluster 3. The silhouette analysis and dendrogram for this study are shown in Figure 1.
(A) Silhouette analysis was used to identify the optimal number of clusters for hierarchical clustering. A silhouette curve for clustering after uniform manifold approximation and projection (UMAP) dimensionality reduction is shown, revealing the best value (0.585) when the number of clusters is set to 3. Therefore, these 3 clusters were selected for further analyses. (B) Each Euclidean distance was computed and dendrograms illustrating the proximity classification of each distance are provided. As stated previously, the decision was made to have 3 clusters, resulting in 71 patients in Cluster 1, 72 patients in Cluster 2, and 57 patients in Cluster 3.
Cluster 1 exhibited the highest incidence of ventricular arrhythmias (33.8%) compared with Cluster 2 (29.6%) and Cluster 3 (24.6%). Kaplan-Meier curves demonstrated a significantly higher event rate in Cluster 1 over time. In particular, a significant difference was observed between Cluster 1 and Cluster 3 in the log-rank test (P=0.015). The Kaplan-Meier curves for each of these clusters are shown in Figure 2.
Kaplan-Meier curves were constructed based on the time without implantable cardioverter defibrillator (ICD) therapy in each cluster, revealing significant differences between clusters regarding the presence or absence of ICD therapy. Cluster 3 had the lowest frequency of ventricular arrhythmias requiring ICD therapy, whereas Cluster 1 had the highest frequency. Comparisons between clusters revealed significant differences specifically between Cluster 1 and Cluster 3.
Patient Background of Each Cluster
ECG-based unsupervised learning classified patients into 3 clusters; however, no differences in sex or body mass index were observed among the clusters. Cluster 1, the highest-risk group, had a higher mean age and a significantly higher proportion of cases of ischemic heart disease, whereas Brugada syndrome and other channelopathies were less prevalent in this cluster. Interestingly, Cluster 1 had the highest rate of ICD implantation for primary prevention. Differences were also observed in the ICD type and settings, with Cluster 1 having the highest rate of CRT-D implantation. Cluster 1 was also most commonly associated with treatment settings concerning the detection cycle length. Further details regarding the background of each cluster are presented in Table 2.
Patient Background for Each Cluster
| Cluster 1 (n=71) |
Cluster 2 (n=72) |
Cluster 3 (n=57) |
P value | |
|---|---|---|---|---|
| Patient background | ||||
| Male sex | 57 (80.3) | 61 (84.7) | 44 (77.2) | 0.541 |
| Age (years) | 67.3±11.7 | 64.4±13.6 | 53.3±15.8 | <0.001 |
| BMI (kg/m2) | 22.3±6.0 | 23.9±6.0 | 23.1±4.1 | 0.240 |
| Underlying heart disease | ||||
| ICM | 33 (43.7) | 32 (44.4) | 6 (10.5) | <0.001 |
| DCM | 20 (28.2) | 13 (18.1) | 10 (17.5) | 0.248 |
| HCM | 8 (11.3) | 7 (9.7) | 3 (5.3) | 0.524 |
| Valvular heart disease | 3 (4.2) | 1 (1.4) | 1 (1.8) | 0.535 |
| Brugada syndrome | 0 (0.0) | 4 (5.6) | 22 (38.6) | <0.001 |
| Long QT syndrome | 1 (1.4) | 0 (0.0) | 2 (3.5) | 0.197 |
| ARVC | 0 (0.0) | 1 (1.4) | 0 (0.0) | 1.000 |
| IVF/IVT | 0 (0.0) | 5 (6.9) | 5 (8.8) | 0.030 |
| Others | 6 (8.5) | 9 (12.5) | 8 (14.0) | 0.593 |
| Primary prevention | 52 (70.8) | 43 (59.7) | 28 (49.1) | 0.020 |
| ICD programming settings | ||||
| Dual-chamber device | 56 (78.9) | 53 (73.6) | 15 (26.3) | <0.001 |
| CRT-D | 25 (35.2) | 16 (22.2) | 4 (7.0) | <0.001 |
| Detection cycle length (ms) | 353.0 [320.0~372.5] |
341.5 [320.0~400.0] |
320.0 [300.0~333.0] |
0.001 |
| ICD detection zones | <0.001 | |||
| 1 zone | 27 (38.0) | 36 (50.0) | 42 (73.7) | |
| 2 zones | 35 (49.3) | 26 (36.1) | 14 (24.6) | |
| 3 zones | 9 (12.7) | 10 (13.9) | 1 (1.8) | |
Unless indicated otherwise, values are presented as n (%), mean±SD, or median [interquartile ranges]. Categorical variables were compared using Fisher’s exact test, and continuous variables were compared using Student’s t-test for normally distributed data or the Mann-Whitney U test for non-normally distributed data. Abbreviations as in Table 1.
12-Lead ECG Characteristics of Each Cluster
Of the 44 elements analyzed from the 12-lead ECG, 33 exhibited statistical significance according to either 1-way ANOVA or the Kruskal-Wallis test. Further significant differences among the 3 groups were identified using the Bonferroni method. Significant differences among all 3 clusters were observed for the following variables: ST-END in lead I; T wave amplitude in lead I; ventricular activation time (VAT) in lead I; R wave amplitude in lead II; ST-END in lead II; VAT in lead aVL; R wave amplitude in lead aVF; ST-MID in lead V1; S wave amplitude in lead V2; ST-END in lead V3; T wave amplitude in lead V3; T wave amplitude in lead V5; and ST-END in lead V6. Detailed information on these variables is provided in Table 3. The hazard ratios for ventricular arrhythmias were 1.53 for Cluster 1 and 1.76 for Cluster 2, with Cluster 3 as the reference group (hazard ratio=1.0). An illustration summarizing the odds ratios is presented in Figure 3.
12-Lead Electrocardiogram Characteristics of Each Cluster
| Cluster 1 (n=71) |
Cluster 2 (n=72) |
Cluster 3 (n=57) |
P value | |||
|---|---|---|---|---|---|---|
| Cluster 1 vs. 2 |
Cluster 1 vs. 3 |
Cluster 2 vs. 3 |
||||
| P wave axis (degrees) | 43.1±31.4 | 45.0±43.5 | 52.7±34.3 | 1.000 | 0.500 | 0.780 |
| PQ interval (ms) | 192.0 [171.5~234.5] |
198.0 [182.0~229.0] |
172.0 [154.0~189.5] |
1.000 | 0.004 | <0.001 |
| QRS axis (degrees) | 30.9±55.7 | 16.2±73.3 | 43.9±39.4 | 0.416 | 0.657 | 0.027 |
| QRS duration (ms) | 114.0 [104.0~131.0] |
112.0 [99.5~140.0] |
100.0 [88.0~110.0] |
1.000 | <0.001 | <0.001 |
| T wave axis (degrees) | 150.0 [101~225.5] |
54.5 [11.3~90.0] |
45.0 [20.0~67.0] |
<0.001 | <0.001 | 1.000 |
| QTc modulation (ms) | 454.7±30.4 | 442.4±36.3 | 423.6±41.0 | 0.123 | <0.001 | 0.011 |
| Lead I | ||||||
| P wave amplitude (μV) | 65.0 [46.3~83.8] |
55.0 [42.5~72.5] |
65.0 [47.5~80.0] |
0.580 | 1.000 | 0.660 |
| QRS area (40 ms×μV) | 47.17±377.5 | 177.5±332.8 | 233.3±306.8 | <0.001 | <0.001 | 1.000 |
| R wave amplitude (μV) | 531.9±323.7 | 523.5±287.5 | 599.6±360.2 | 1.000 | 0.720 | 0.560 |
| STJ (μV) | −18.0±27.6 | −14.5±44.5 | 16.5±27.4 | 1.000 | <0.001 | <0.001 |
| ST-END (μV) | −30.0 [−60.0~−10.0] |
10 [−15.0~35.0] |
65.0 [20.0~95.0] |
<0.001 | <0.001 | <0.001 |
| T wave amplitude (μV) | −65.0 [−110.0~−30.0] |
90.0 [20.0~131.3] |
175.0 [102.5~265.0] |
<0.001 | <0.001 | <0.001 |
| VAT (ms) | 58.0 [47.0~69.0] |
46.0 [42.0~51.5] |
40.0 [34.0~57.0] |
<0.001 | <0.001 | <0.001 |
| Lead II | ||||||
| R wave duration (ms) | 72.6±22.1 | 45.9±24.6 | 58.4±18.3 | <0.001 | 0.001 | 0.005 |
| ST-END (μV) | −20.0 [−50.0~12.5] |
32.5 [−11.3~66.3] |
80.0 [35.0~135.0] |
<0.001 | <0.001 | <0.001 |
| T wave amplitude (μV) | 5.6±191.5 | 125.1±148.6 | 248.9±149.3 | <0.001 | <0.001 | <0.001 |
| VAT (ms) | 50.0 [46.0~56.5] |
42.0 [32.0~48.0] |
42.0 [38.0~57.0] |
<0.001 | <0.001 | 1.000 |
| Lead III | ||||||
| QRS area (40 ms×μV) | −52.0 [−523.0~275.5] |
−161.5 [−472.5~219.8] |
150.5 [−11.0~363.0] |
1.000 | 0.014 | <0.001 |
| R wave amplitude (μV) | 265.5 [113.8~517.5] |
190.0 [93.8~416.3] |
437.5 [255.0~862.5] |
0.537 | 0.007 | <0.001 |
| R wave duration (ms) | 46.0 [31.5~60.0] |
30.0 [19.5~44.5] |
46.0 [34.0~56.5] |
0.005 | 1.000 | 0.005 |
| VAT (ms) | 43.2±20.5 | 34.7±21.3 | 39.4±17.0 | 0.045 | 0.873 | 0.586 |
| aVL | ||||||
| QRS area (40 ms×μV) | 216.0 [−19.0~643.0] |
195.5 [−28.8~417.0] |
68.0 [−125.0~193.0] |
0.897 | 0.003 | 0.046 |
| R wave duration (ms) | 62.7±37.9 | 57.9±22.4 | 37.9±20.1 | 0.950 | <0.001 | <0.001 |
| VAT (ms) | 59.5±22.5 | 50.4±11.7 | 36.9±16.4 | 0.008 | <0.001 | <0.001 |
| aVF | ||||||
| R wave amplitude (μV) | 407.5 [253.8~622.5] |
210.0 [117.5~470.0] |
607.5 [353.8~908.8] |
0.002 | 0.007 | <0.001 |
| ST-MID (μV) | 115.0 [55.0~155.0] |
40.0 [−5.0~65.0] |
75.0 [30.0~100.0] |
<0.001 | <0.001 | 0.004 |
| T wave amplitude modulation (μV) | 100.0 [36.3~171.3] |
−56.0 [−145.0~57.5] |
−80.0 [−125.0~25.0] |
<0.001 | <0.001 | 1.000 |
| VAT (ms) | 16.0 [12.0~22.0] |
18.0 [14.0~74.0] |
18.0 [14.0~22.5] |
0.031 | 0.206 | 0.999 |
| V2 | ||||||
| S wave amplitude (μV) | 2,199.0±1,075.1 | 948.2±700.5 | 1,387.5±751.9 | <0.001 | <0.001 | 0.042 |
| S wave duration (ms) | 77.9±23.4 | 51.3±20.6 | 47.9±14.5 | <0.001 | <0.001 | 1.000 |
| V3 | ||||||
| R wave amplitude (μV) | 420.0 [238.8~716.3] |
470.0 [225.0~1,045.0] |
935.0 [483.8~1,567.5] |
0.465 | <0.001 | 0.003 |
| R wave duration (ms) | 28.0 [19.5~44.0] |
39.0 [26.0~53.5] |
40.0 [36.0~46.0] |
0.046 | 0.002 | 1.000 |
| ST-END (μV) | 150.0 [62.5~285.0] |
102.5 [18.8~195.0] |
265.0 [180.0~395.0] |
0.031 | 0.006 | <0.001 |
| T wave amplitude (μV) | 295.0 [170.0~600.0] |
220.0 [91.3~356.3] |
502.5 [323.8~816.3] |
0.032 | 0.003 | <0.001 |
| V4 | ||||||
| QRS area (ms×μV) | 37.0 [−1,146.5~820.0] |
20.5 [−621.0~266.8] |
410.0 [87.0~711.0] |
1.000 | 0.083 | <0.001 |
| VAT (ms) | 39.0 [22.5~50.0] |
36.0 [24.0~44.0] |
36.0 [32.0~40.0] |
1.000 | 1.000 | 1.000 |
| V5 | ||||||
| R wave amplitude (μV) | 1,220.0 [840.0~2,117.5] |
990.0 [681.3~1,665.0] |
1,605.0 [1,150.0~2,035.0] |
0.390 | 0.654 | 0.015 |
| S wave amplitude (μV) | 310.0 [132.5~525.0] |
460.0 [285.0~745.0] |
360.0 [188.8~528.8] |
0.021 | 0.873 | 0.133 |
| T wave amplitude modulation (μV) | −194.4±268.9 | 5.0±195.3 | 234.8±209.9 | <0.001 | <0.001 | <0.001 |
| VAT (ms) | 54.0 [44.0~66.0] |
42.0 [36.0~46.0] |
40.0 [33.0~42.0] |
<0.001 | <0.001 | 0.220 |
| V6 | ||||||
| P wave amplitude (μV) | 55.0 [35.0~75.0] |
52.5 [35.0~68.8] |
55.0 [40.0~66.3] |
1.000 | 1.000 | 1.000 |
| R wave duration (ms) | 74.0 [60.5~91.0] |
48.0 [42.0~58.0] |
48.0 [42.0~58.0] |
<0.001 | <0.001 | 1.000 |
| ST-END (μV) | −88.2±83.6 | 11.0±75.5 | 91.1±104.8 | <0.001 | <0.001 | <0.001 |
| VAT (ms) | 60.0 [50.0~76.0] |
44.0 [37.0~48.0] |
40.0 [36.0~46.0] |
<0.001 | <0.001 | 0.660 |
Unless indicated otherwise, values are reported as the mean±SD or medians [interquartile ranges]. Continuous variables were assessed using 1-way ANOVA for normally distributed data or the Kruskal-Wallis test for non-normally distributed data. The Bonferroni method was used for pairwise comparisons in both the ANOVA and Kruskal-Wallis tests. Bolded items differed significantly among all 3 groups. QTc, rate-corrected QT interval; ST-END, end of the ST level, defined as the ST level at two-sixteenths of the preceding R-R interval after the J point; STJ, ST junction; ST-MID, middle of the ST level, defined as the ST level at one-sixteenth of the preceding R-R interval after the J point; VAT, ventricular activation time.
Hazard ratios of the incidence of ventricular arrhythmias in each cluster. The green circles represent the hazard ratio percentage, and the bars represent 95% confidence intervals (CIs).
Hierarchical clustering after dimensionality reduction using UMAP yielded the highest silhouette coefficient, leading to the classification of patients into 3 clusters. Cluster 1 exhibited the highest incidence of ventricular arrhythmias. Significant differences in 12-lead ECG characteristics were observed, with ST-END in lead I, T wave amplitude in lead I, VAT in lead I, R wave amplitude in lead II, ST-END in lead II, VAT in lead aVL, R wave amplitude in lead aVF, ST-MID in lead V1, S wave amplitude in lead V2, ST-END in lead V3, T wave amplitude in lead V3, T wave amplitude in lead V5, and ST-END in lead V6 showing consistent associations with the risk of ventricular arrhythmia.
Many non-invasive tests have been used to predict ventricular arrhythmias, including T wave alternans,24 signal-averaged ECG,25 and 12-lead ECG. Although some of these examinations are indicated for ICD implantation, the gold standard for ICD implantation for primary prevention still relies on LVEF, the presence of non-sustained VT, and New York Heart Association class. In this context, predicting ICD implantation using a 12-lead ECG alone is challenging. ML has the potential to overcome this problem.
Although a previous study used deep learning to predict ventricular arrhythmia risk, focusing on Brugada syndrome,26 our aim was to use ML to predict and stratify ventricular arrhythmias using the most common test, the 12-lead ECG, in patients with various underlying diseases. As a result, 12-lead ECG was successfully used for risk stratification of ventricular arrhythmias in this study. Previous studies have been conducted to predict ventricular arrhythmia using ECGs; for example, studies examining the influence of T waves and QRS fragmentation leading to sudden cardiac death have been reported.27,28 The present study also suggests that T wave and QRS waveforms, such as T wave amplitude and R wave amplitude, may be involved in risk stratification.
VAT is initially suggestive of LV hypertrophy;29 however, there have been few reports on the association of VAT with ventricular arrhythmias.28 In the present cohort, there was a substantial difference in VAT at induction I and V6 between clusters, which may be attributed to the involvement of LV hypertrophy and arrhythmias or to the high prevalence of ischemic heart disease in Cluster 1. This suggests that myocardial damage due to ischemia contributes to prolonged VAT, potentially leading to ventricular arrhythmia. V2 S wave amplitude is also suggested to be related to LV hypertrophy, which may be involved in ventricular arrhythmias.30 In ML, these factors can be combined to determine risk, likely stratifying more effectively than extracting single factors alone.
This study was conducted in a setting where ICDs were more likely to be used for therapy in Cluster 1, which could have influenced the outcomes. However, this study primarily used unsupervised learning and divided patients into 3 groups based on 12-lead ECG features, independent of ICD therapy. Thus, there is minimal room for treatment bias; the classification was based solely on 12-lead ECG induction. Consequently, it is reasonable to infer that Cluster 1 patients were originally at high risk; thus, the ICD settings were adjusted accordingly, suggesting that the 12-lead ECG in Cluster 1 may predict future ventricular arrhythmias. Therefore, the present study, analyzing the 12-lead ECG data from Cluster 1, holds significant value.
In recent years, risk stratification using clustering, as demonstrated in the present study, has become common in many fields.21,31 To the best of our knowledge, there are no reports indicating that clustering predicts and stratifies future ventricular arrhythmias based on a 12-lead ECG. Although the present study was conducted at a single center and included only Japanese patients, similar studies on a larger scale could provide further insights into ECG trials and offer additional indications for ICDs. For instance, the primary prevention of non-ischemic heart disease, which has shown negative results in some studies,4 could benefit from better patient selection through larger-scale analyses.
This study has several limitations. First, the silhouette coefficients used to assess clustering effectiveness did not exceed 0.6, despite using 3 different dimensionality reduction techniques and 2 unsupervised learning algorithms. Traditionally, the silhouette coefficient threshold was frequently set at 0.5–0.6.32 It is plausible that the single-center nature of this study contributed to this outcome. Second, the inclusion of a variety of diseases presents a limitation. For instance, because of the limited number of cases, non-ischemic heart disease was not clustered in this study, which could affect ICD efficacy for primary prevention. Furthermore, diurnal variations in the ECG observed in channelopathies, such as Brugada syndrome and long QT syndrome,33 may challenge the adequacy of using only a single 12-lead ECG for comprehensive assessment. Lastly, this study aimed to predict ventricular arrhythmias using 12-lead ECG features. However, because ICD therapy is the endpoint, factors such as primary or secondary prevention, ICD settings, and VT ablation can influence the results. Despite these influencing factors, it is challenging to envision using other data to predict ventricular arrhythmias. Therefore, it was difficult to meet all these conditions for analysis in this study.
This study is a small, single-center investigation, and the results are preliminary. To address these limitations, future research in a multicenter setting could include patients from a broader range of backgrounds, which may lead to more refined clustering results.
Our results suggest that clustering by unsupervised learning using a 12-lead ECG can be helpful in the risk stratification of ventricular arrhythmias. In the future, ML may provide further insights into 12-lead ECGs and refine the indications for ICDs.
The authors thank Editage (www.editage.jp) for English language editing.
This research did not receive any specific grants from funding agencies in the public, commercial, or not-for-profit sectors.
None.
This study was approved by Yokohama Minami Kyosai Hospital Institutional Review Board (Reference no. 1_24_3_19).
Deidentified participant data, including patient information and the statistical analysis plan, will be shared up to 36 months after the publication of this paper on a request basis for anyone under approval of Yokohama Minami Kyosai Hospital Institutional Review Board. The data can be used for any kind of analysis and will be shared in Excel format via email.
Please find supplementary file(s);
https://doi.org/10.1253/circj.CJ-24-0269
