The UK Biobank is a large, population-based prospective cohort study comprising more than 500,000 participants aged 49 to 69 years, enrolled from 22 assessment centers in the United Kingdom between 2006 and 2010. Details of the UK Biobank protocols are accessible online [21]. Participants diagnosed with stroke at baseline (n=7780, 1.5%), lacking follow-up information (n=20,251, 4%), or without CMR data (n=437,713, 87.2%) were excluded, resulting in a final cohort of 36,467 (7.3%) participants.

We included the following CMR measures in the study: LV stroke volume (LVSV), LV myocardial mass (LVMM), LV end-diastolic volume (LVEDV), LV end-systolic volume (LVESV), LV ejection fraction (LVEF), LV global longitudinal strain (LVGLS), left atrial (LA) maximum volume (LAMV), and LA ejection fraction (LAEF) (Figure 1).

Stroke diagnosis was ascertained through linkage to self-reported medical conditions, hospital inpatient data, and death register records based on International Classification of Diseases codes, with the earliest recorded date of outcome provided in Table S1 in Multimedia Appendix 1. The follow-up time was defined as the interval from the date of recruitment to the occurrence of stroke, all-cause mortality, loss to follow-up, or September 30, 2023, whichever came first. Details of covariates were shown in Table S2 in Multimedia Appendix 1. This study was reported according to the transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) guidelines (Checklist 1).

Inspired by the algorithm introduced by Bellfield et al [20], we used a Gaussian mixture model to train the GTM model and visualize the distribution of data in the latent space. GTM uses a soft assignment method to estimate the probability of each participant being assigned to a cluster. The high-dimensional data were then visualized in a 2D latent space. After preprocessing the dataset, we used the ugtm Python (Python Software Foundation) package for GTM modeling and applied 10-fold cross-validation with the expectation-maximization algorithm to optimize the parameters. After GTM modeling, the generated latent space illustrated the distribution of participants in each cluster, with darker colors indicating higher densities. The modeling parameters (cardiac function variables) demonstrated the impact of cardiac function on each latent grid node in the GTM model. We aggregated the latent grid nodes by calculating the Euclidean distances between their modeling parameters and applied agglomerative hierarchical clustering using the Ward minimum variance method to derive phenotypes [22]. The characteristics of the phenotypes were compared to evaluate the distribution of risk factors for cardiovascular aging.

We used Cox proportional hazards models to evaluate the association between cardiac function phenotypes and long-term stroke risk, reporting results as hazard ratios (HRs) with 95% CIs. In multivariable analyses, model 1 was an unadjusted model; model 2 was adjusted for age at recruitment, sex, Townsend deprivation index at recruitment, smoking status, alcohol intake frequency, systolic blood pressure, diastolic blood pressure, and BMI; and model 3 was additionally adjusted for atrial fibrillation, type 2 diabetes, coronary heart disease, heart failure, antihypertensive drugs, and statins. In sensitivity analyses, we explored the association of cardiac function phenotypes with the risk of different stroke types. Additionally, we performed a competing risk analysis for stroke risk using the Fine and Gray method, accounting for mortality as the competing event.

Following the categorization of participants into different phenotypes through the GTM approach, we deployed several supervised ML models to predict cardiac function phenotypes. Hyperparameters for each model are detailed in Table S3 in Multimedia Appendix 1. Hyperparameter optimization was conducted using grid search and Bayesian optimization. Discriminatory ability was evaluated using the receiver operating characteristic curve. Calibration performance was assessed using the calibration curve. We selected the best-performing method as the final algorithm for developing a phenotype prediction model. We constructed a website using the Flask framework and the DeepSeek-R1 API (application programming interface). The predicted probabilities and individualized metrics were sent to the DeepSeek-R1 model to generate a detailed analysis of health status and provide professional recommendations. Detailed descriptions of the methods are provided in Multimedia Appendix 1.

The ethical framework of the UK Biobank was approved by the North West Multi-centre Research Ethics Committee (11/NW/0382). All participants provided informed consent through electronic signature at enrollment.

Normally distributed continuous data are presented as mean (SD), and group comparisons were performed using 2-tailed t tests. Categorical data were presented as frequency (percentage), and group comparisons were conducted using chi-square tests. To account for multiple comparisons, P values were adjusted using the false discovery rate method. Statistical tests were conducted with R software (version 4.2.1; R Foundation for Statistical Computing), and a 2-sided P value <.05 was considered statistically significant.

The study design is shown in Figure 2. Of the 36,467 participants included in the study (Figure 2), the mean age was 54.9 (SD 7.5) years, with 17,442 (47.8%) male participants. In total, 14,373 (39.5%) of participants were current or former smokers, 8180 (22.4%) had daily or almost daily alcohol intake, 1047 (2.9%) had type 2 diabetes, 345 (0.9%) had atrial fibrillation, and 48 (0.1%) had heart failure at baseline. During a mean follow-up time of 14.7 (SD 1.1) years, 500 (1.4%) and 664 (1.8%) participants developed stroke and died, respectively (Table 1).

Figure 3 illustrates the latent space generated by the GTM model. The number of participants in clusters is represented using the “viridis” color scheme and visualized as circles, where a deeper blue hue and larger circle diameter indicate a higher number of participants. LVSV, LVMM, LVEDV, LVESV, LVEF, LVGLS, LAMV, and LAEF are aligned point-to-point with their latent space in a light teal blue scheme. In the dendrogram plot, 2 cardiac function phenotypes are identified, clearly separating the dataset into 2 major branches at a relatively large linkage distance. Figure S1 in Multimedia Appendix 1 illustrates that stroke diagnoses among participants were mainly distributed in the area of phenotype 1.

To further characterize the phenotype patterns, we compared the distributions of investigatory variables and visualized them on the latent space with the light orange color scheme (Figures S2-S7 in Multimedia Appendix 1). Significant differences were observed between the 2 cardiac function phenotypes across demographic characteristics, cardiovascular risk factors, and cardiac function parameters. Participants in phenotype 1 exhibited larger cardiac chamber volumes and greater myocardial mass, with higher LVEF and LAEF but lower LVSV, LVMM, LVEDV, LVESV, LVGLS, and LAMV compared with phenotype 2 (all P<.001). In addition, phenotype 1 demonstrated a higher burden of cardiovascular risk factors and comorbidities, including a greater prevalence of major adverse cardiac events, atrial fibrillation, heart failure, coronary heart disease, and type 2 diabetes (P<.001). Participants in phenotype 1 also had higher blood pressure, arterial stiffness index, and more adverse metabolic profiles. Taken together, these structural, functional, and clinical differences suggest that phenotype 1 represents a cardiovascular aging–related phenotype, characterized by cardiac remodeling, increased cardiovascular risk burden, and elevated incidence of stroke and mortality (Table 1 and Table S4 in Multimedia Appendix 1).

Multivariable modeling identified that cardiovascular aging phenotypes were associated with the long-term risk of stroke and mortality. Compared with phenotype 1, phenotype 2 was significantly associated with a decreased risk of stroke in the unadjusted model (HR 0.579, 95% CI 0.484-0.691; P<.001); model 2 (HR 0.671, 95% CI 0.541-0.833; P<.001); and model 3 (HR 0.695, 95% CI 0.559-0.864; P=.001; Table 2).

Kaplan-Meier survival curves revealed that the phenotype 1 group had a higher risk of stroke (P<.001; Figure 4).

Regarding stroke subtypes, cardiac function phenotypes were associated with cerebral infarction and other nontraumatic intracranial hemorrhage (all P<.05; Table S5 in Multimedia Appendix 1). When mortality was considered as a competing risk, phenotype 2 remained significantly associated with stroke risk (HR 0.578, 95% CI 0.484-0.691; P<.001; Figure S8 in Multimedia Appendix 1).

Basal metabolic rate, sex, testosterone, standing height, weight, forced vital capacity, waist circumference, creatinine, forced expiratory volume in 1 second, and urate were included in the ML models (Table S6 in Multimedia Appendix 1). By integrating the accuracy in the training and validation sets, we selected the random forest model as the optimal model based on its excellent accuracy (area under the curve 0.914, 95% CI 0.911-0.918 for training and 0.867, 95% CI 0.858-0.876 for validation; and accuracy 0.837, 95% CI 0.833-0.841 for training and 0.805, 95% CI 0.804-0.806 for validation) without overfitting the training set (Figure 5). Calibration plots exhibited that the random forest model had good performance in the calibration quality (Brier score 0.111, 95% CI 0.109-0.113 for training and 0.132, 95% CI 0.127-0.137 for validation; Table 3 and Tables S7 and S8 in Multimedia Appendix 1). We drew a feature importance plot to better visualize the contributions of the predictors to the predictive power of the random forest model (Figure S9 in Multimedia Appendix 1).

After entering the participant information, the local model calculated the corresponding phenotype, and the DeepSeek-R1 model provided an analysis of the participant’s health status along with recommendations for potential clinical application (Figure S10 in Multimedia Appendix 1).

In this study, we identified distinct cardiac function phenotypes related to cardiovascular aging using the GTM model based on CMR parameters. These phenotypes were significantly associated with long-term stroke risk and could be reliably predicted by supervised ML models.

CMR is important for evaluating cardiovascular pathologies and for assessing the etiology and prognosis of patients with stroke. A previous study demonstrated that cardiovascular aging is associated with reduced LV volumes and increased LV concentricity on CMR, with notable sex-specific differences in systolic cardiac function [9]. CMR also exhibited high sensitivity in detecting intraventricular thrombi and thrombi within the LA appendage [23]. Additionally, CMR imaging could affect the management of acute stroke by detecting aortic plaques, cardiac structural abnormalities, and intracardiac thrombi in patients with stroke [24,25]. However, CMR metrics represented multidimensional data with complex interrelations, potentially involving collinearity and interaction effects. Using unsupervised ML may improve the ability to reveal the inherent structure of the data and achieve more accurate classification.

GTM incorporated probability distribution functions to refine the self-organizing map algorithms and mitigated their inherent limitations, such as convergence issues, insufficient neighborhood preservation, and the absence of a distinct objective function. Moreover, GTM was capable of visually representing phenotypes through latent spaces [17]. Sattarov et al [26] applied GTM to explore the latent space of Simplified Molecular Input Line Entry System–based autoencoders for generating targeted molecular libraries and visualized the autoencoder latent space on a 2D topographic map. These findings indicated a promising application of GTM in structure-based affinity assessments. However, unsupervised ML faces challenges related to limited interpretability and scalability, making the derived phenotypes and their clinical relevance difficult to translate into real-world medical practice. Therefore, we innovatively applied supervised ML to relearn the phenotype labels obtained from GTM, constructing a random forest model that accurately predicted the 2 distinct phenotypes. Given the high cost and inconvenience of CMR scans for patients with stroke [27], using clinical variables allowed for a reduction in model application requirements while preserving accuracy, making the classification model more suitable for clinical use.

The phenotypes derived from CMR metrics may reflect distinct patterns of cardiac structural and functional remodeling associated with cardiovascular aging. Key variables contributing to phenotype differentiation, including LVMM and concentric geometry, have been linked to vascular aging markers, supporting the connection between cardiovascular aging and cardiac structural changes [28]. LA enlargement and functional impairment are also recognized markers of cardiovascular aging and are related to chronic diastolic dysfunction and atrial fibrillation [29]. In addition, accelerated cardiovascular aging driven by cumulative risk factors such as hypertension, obesity, diabetes, and coronary artery disease may further promote cardiac remodeling, consistent with the differing comorbidity profiles observed between the phenotypes in our study [30]. Notably, the cardiovascular aging–related phenotype in our study was characterized by LV and LA abnormalities indicative of volume or pressure overload and myocardial dysfunction. These alterations have been associated with increased risks of cardioembolic stroke and subclinical cerebral infarction [31,32]. Therefore, the CMR-derived phenotypes may capture different stages of cardiovascular aging and help explain the heterogeneity of cardiac remodeling and its potential contribution to long-term stroke risk.

In our study, supervised ML models identified several clinical parameters as predictors for cardiac function phenotypes. Notably, basal metabolic rate is the energy required to maintain essential functions in a healthy state. Prior research indicated that basal metabolic rate was independently linked to heart failure, atrial fibrillation, and flutter and acted as a risk factor for cardiovascular mortality [33,34]. Furthermore, a pooled analysis of 8 cohorts showed that obstructive and restrictive physiology were associated with incident heart failure with reduced or preserved ejection fraction [35]. Moreover, a distinctive J-shaped curve was noted between BMI and waist circumference and the risk of heart failure [36]. Additionally, renal impairment in chronic heart failure was typically linked to decreased cardiac output and subsequent renal hypoperfusion, and targeting volume overload may improve outcomes in cardiac dysfunction [37]. These findings suggested that the ML model could reflect systemic metabolic, cardiopulmonary, and renal alterations that may explain the clinical characteristics of the identified cardiac function phenotypes.

To the best of our knowledge, this was the first study to use GTM for clustering CMR parameters and to identify cardiovascular aging phenotypes associated with long-term stroke risk. However, there were several limitations to our study. First, the CMR parameters extracted from the UK Biobank were collected according to the cohort’s standardized protocols. Because the CMR examination protocol may differ in other studies, this should be considered when applying our model to other cohorts. Second, while trained on the UK Biobank’s large sample, the model’s applicability to other cohorts may be limited by environmental, genetic, and cultural variations. Third, given the predominantly White and healthy population in the UK Biobank, selection and health biases may be inevitable. The generalizability of our findings to ancestrally diverse individuals will assist in determining whether more appropriate and ethnically relevant decisions are needed.

In conclusion, we used GTM to identify cardiac function phenotypes based on CMR parameters. The identified phenotypes were significantly associated with stroke risk, with the high-risk phenotype characterized by impaired cardiac function and features related to cardiovascular aging. Furthermore, we relearned phenotype labels using supervised ML models, enabling phenotype prediction from clinical variables. These models may facilitate the identification of individuals at higher risk and support the development of preventive and therapeutic strategies. Further studies are warranted to explore the potential mechanisms linking cardiovascular aging, cardiac phenotypic heterogeneity, and stroke risk.