Multi-Ethnic Study of Atherosclerosis: Relationship between Left Ventricular Shape at Cardiac MRI and 10-year Outcomes

Abstract

Background

Left ventricular (LV) subclinical remodeling is associated with adverse outcomes and indicates mechanisms of disease development. Standard metrics such as LV mass and volumes may not capture the full range of remodeling.

Purpose

To quantify the relationship between LV three-dimensional shape at MRI and incident cardiovascular events over 10 years.

Materials and Methods

In this retrospective study, 5098 participants from the Multi-Ethnic Study of Atherosclerosis who were free of clinical cardiovascular disease underwent cardiac MRI from 2000 to 2002. LV shape models were automatically generated using a machine learning workflow. Event-specific remodeling signatures were computed using partial least squares regression, and random survival forests were used to determine which features were most associated with incident heart failure (HF), coronary heart disease (CHD), and cardiovascular disease (CVD) events over a 10-year follow-up period. The discrimination improvement of adding LV shape to traditional cardiovascular risk factors, coronary artery calcium scores, and N-terminal pro–brain natriuretic peptide levels was assessed using the index of prediction accuracy and time-dependent area under the receiver operating characteristic curve (AUC). Kaplan-Meier survival curves were used to illustrate the ability of remodeling signatures to predict the end points.

Results

Overall, 4618 participants had sufficient three-dimensional MRI information to generate patient-specific LV models (mean age, 60.6 years ± 9.9 [SD]; 2540 women). Among these participants, 147 had HF, 317 had CHD, and 455 had CVD events. The addition of LV remodeling signatures to traditional cardiovascular risk factors improved the mean AUC for 10-year survival prediction and achieved better performance than LV mass and volumes; HF (AUC, 0.83 ± 0.01 and 0.81 ± 0.01, respectively; P < .05), CHD (AUC, 0.77 ± 0.01 and 0.75 ± 0.01, respectively; P < .05), and CVD (AUC, 0.78 ± 0.0 and 0.76 ± 0.0, respectively; P < .05). Kaplan-Meier analysis demonstrated that participants with high-risk HF remodeling signatures had a 10-year survival rate of 56% compared with 95% for those with low-risk scores.

Conclusion

Left ventricular event-specific remodeling signatures were more predictive of heart failure, coronary heart disease, and cardiovascular disease events over 10 years than standard mass and volume measures and enable an automatic personalized medicine approach to tracking remodeling.

© RSNA, 2022

Online supplemental material is available for this article.

Summary

Cardiac MRI–derived, left ventricular, event-specific remodeling signatures provided quantitative information on subclinical disease and were more predictive of 10-year cardiovascular events than standard mass and volume measures after adjustment for cardiovascular risk factors.

Key Results

• ■ In this retrospective study of 4618 participants from the Multi-Ethnic Study of Atherosclerosis, event-specific remodeling signatures computed from three-dimensional MRI shape analysis improved prediction of 10-year heart failure, coronary heart disease, and cardiovascular disease events.

• ■ Participants with high-risk heart failure scores had a 10-year survival rate of 56% compared with 95% for those with low-risk scores.

• ■ Personalized remodeling signatures can be used to automatically score left ventricular remodeling associated with adverse events with respect to a reference cohort.

Introduction

Left ventricular (LV) mass and volumes have been identified as important metrics of remodeling in patients with myocardial infarction and heart failure (HF) (1,2). LV subclinical remodeling may also occur in asymptomatic individuals prior to the establishment of clinical symptoms in response to exposure to cardiovascular risk factors. In asymptomatic population-based studies, LV end-systolic and end-diastolic volumes and chamber diameters, as well as relative wall thickness and indexed LV mass, have been shown to be predictive of HF (3,4). In the Framingham Heart Study (5), larger LV chamber dimension, lower systolic dimension changes (6), and hypertrophy (7) were associated with future adverse events. In the Multi-Ethnic Study of Atherosclerosis (MESA), LV mass, volumes, and sphericity were associated with both incident coronary heart disease (CHD; low sphericity) and HF (high sphericity) (8,9).

However, current LV mass, volumes, and sphericity measures do not capture all the information available for LV shape. Multidimensional LV shape measures have been shown to have a significant relationship with subclinical disease and risk factors (10–12) and to be more strongly associated with cardiovascular risk factors than traditional mass and volume metrics (10,11,13). In addition, machine learning methods, such as random survival forest analysis, show promise for identification of which factors are most strongly related to cardiovascular outcomes (14). A better knowledge of the preclinical remodeling patterns associated with adverse incident events would aid understanding of the mechanisms of developing disease.

In this study, we aimed to define event-specific “signatures” of preclinical LV remodeling, expressed as a set of event-specific remodeling scores, that are optimally associated with incident HF, CHD, and all cardiovascular disease (CVD) events over a 10-year follow-up period during the MESA. A machine learning pipeline was used to automatically generate patient-specific shape models and determine the event-specific shape signatures. We hypothesized that multidimensional shape signatures would be more strongly associated with incident events than standard LV mass, volume, and ejection fraction measures.

Materials and Methods

Study Design

The MESA is an ongoing prospective, multicenter population-based study in the United States designed to examine disease development from preclinical manifestations to clinical symptoms (15). The study includes 6814 men and women of diverse race and ethnicity with no clinically apparent CVD enrolled from 2000 to 2002. Approximately 28% of the participants are African American, 12% are Asian, predominantly of Chinese descent, 22% are Hispanic, and 38% are White. All participants provided informed consent, and the study was approved by the institutional review boards of all MESA centers. Atlas results using 1991 participants in MESA were previously reported (10). Results of a machine learning atlas generation pipeline study that used 1052 participants in MESA were also recently reported (16). Herein, we report, for the first time, on the relationship between atlas scores and cardiovascular outcomes and propose a framework for using shape atlases in personalized risk prediction and prognostication.

Incident HF, CHD, and CVD events, as defined in the MESA, were used as end points in the current study. Criteria for probable HF included symptomatic HF diagnosed by a physician and treatment, while definite HF also required evidence of one or more other criteria (including pulmonary edema and/or congestion at chest radiography, a dilated ventricle or poor LV function at echocardiography or ventriculography, or evidence of LV diastolic dysfunction). Criteria for CHD included myocardial infarction, resuscitated cardiac arrest, definite and probable angina, and CHD-related death. CVD events included stroke, CHD, atherosclerotic death, stroke death, and CVD-related death. The time to event was defined as the number of days between the baseline examination and event. More information about event definition and adjudication is given in Appendix S1 (online) and reference 8.

MRI Protocol

Cardiac MRI was performed with 1.5-T MRI scanners (Siemens Healthineers and General Electric) at six institutions. Details of the scanners and imaging protocol have been previously described (17). Images were acquired in four-chamber, two-chamber, long-axis, and short-axis sections with a gradient-recalled-echo sequence (typical repetition time, 8–10 msec; echo time, 3–5 msec; flip angle, 20°; field of view, 360–400 mm, pixel size, 1.4–2.5 mm) and a section thickness of 6 mm with 20–30 frames per section (temporal resolution <50 msec).

Event-specific Remodeling Signatures

An automatic end-to-end pipeline, described previously (16), was used to generate three-dimensional segmentation of the images for quantification of LV shape at end diastole and end systole. LV shape models were constructed as described previously (10,13). The processing pipeline is described in Appendix S1 (online) and illustrated in Figure S1 (online). Partial least squares (PLS) decomposition (18) was applied to the three-dimensional patient-specific geometries to extract event-specific remodeling signatures, which represent a set of event-specific remodeling variations, expressed as scores (Appendix S1 [online]). The time to event (in days) was used as the response variable and the LV shape models were used as predictors. PLS latent variable scores (as z scores) were used as the event-specific remodeling signatures. The number of latent variables was set to 30 and associated scores were labeled “RS1, RS2,…RS30” in order of decreasing correlation between shape and time to event.

Relationship of LV Geometry to Events

Three random survival forest models (19) were compared to assess the strength of the relationship between LV shape at baseline and end points over a 10-year follow-up. Model 1 used only demographics and traditional cardiovascular risk factors as predictor variables. These were age, sex, race, body mass index, height, weight, smoking, systolic blood pressure, antihypertensive medication use, diabetes mellitus, fasting glucose level, high-density lipoprotein level, total cholesterol level, resting heart rate, beta-blocker use, and either log (coronary artery calcium score + 1) for CHD and CVD or log (N-terminal pro–brain natriuretic peptide) for HF. Model 2 comprised all model 1 predictors plus the ejection fraction, end-systolic volume, end-diastolic volume, and LV mass. End-systolic and end-diastolic volumes and LV mass were not adjusted for body surface area because height and weight were also included as predictors in the models and the use of ratios in regression analysis can lead to spurious results (20). Model 3 comprised model 1 predictors plus 30 event-specific remodeling signatures derived from the PLS analysis. Each model was trained separately for each end point. For each random survival forest model, an initial random survival forest was built, and variables were ranked based on the mean of the minimal depth of the maximal subtree over the entire forest (Appendix S1 [online]). The lower the depth, the higher the predictive power. The top-ranked variables (those with a depth smaller than the mean depth of all the features) were selected for the final random survival forest analysis, and the model performances were evaluated. The data set was randomly split into 70% training and 30% testing. This process was repeated five times, similar to a fivefold cross-validation, to avoid comparison bias due to training and test data selection.

Statistical Analysis

Analyses were carried out in R (The R Foundation) (21). A P value less than .05 was considered indicative of a statistically significant difference. The discrimination improvement of adding LV shape was assessed using the index of prediction accuracy (IPA) (22) and the time-dependent area under the receiver operating characteristic curve (AUC) for time-to-event outcomes (23).

Time-dependent sensitivity and positive predictive value (PPV) were also extracted from the time-dependent receiver operating characteristic curves. PPV was chosen over specificity due to low event rates. Survival curves were obtained with Kaplan-Meier analysis, and log-rank tests were used to compare survival curves between groups.

Results

Participant Characteristics

Of 5098 participants who underwent a cardiac MRI examination at baseline, 4618 (mean age, 60.6 years ± 9.9 [SD]; 2540 women) were included in this study (Table 1, Fig 1). The median follow-up time was 8.5 years. Of the included 4618 participants, 147 (3%) had HF, 317 (7%) had CHD, and 455 (10%) had CVD events. Note that participants could be associated with multiple outcomes. Participants with events were usually older and more likely to be male. Diabetes, smoking, high cholesterol, and a higher coronary artery calcium score, N-terminal pro–brain natriuretic peptide level, heart rate, and body weight were also associated with events.

Table 1:

Characteristics of the Participants at Baseline according to Incident Event Categories

Figure 1:

Flow diagram of participant inclusion. CMR = cardiac MRI, DICOM = Digital Imaging and Communications in Medicine, MESA = Multi-Ethnic Study of Atherosclerosis, 3D = three-dimensional.

Performance Evaluation

PLS z scores were used as event-specific remodeling signatures. Table 2 shows the average IPA at year 10 from the testing data sets across the fivefold cross-validation for each model and each end point. For all end points, the addition of imaging parameters and proven biomarkers showed improved performance. Inclusion of LV mass and volumes increased prediction performance for each end point (model 2). For all outcomes of interest, model 3 performed best, showing the highest IPA and AUC values. The increase in AUC between model 2 and model 3 was significant for each end point (P < .05). The improvement in IPA and AUC values suggests that LV shape is more strongly related to subsequent outcomes than LV mass and volumes. Improvement in IPA was most pronounced in HF (mean IPA, model 2 [11.9%] vs model 3 [14.6%]), suggesting a stronger relationship between shape and subsequent HF relative to CVD and CHD (mean IPA, model 2 [11.8%] vs model 3 [12.7%] and model 2 [11%] vs model 3 [11.5%], respectively). Model 3 also achieved the highest time-dependent PPV for each end point.

Table 2:

Average Model Performance and Calibration for 10-year Survival across Cross-Validation Folds according to Outcome

Hazard ratios and their 95% CIs from multivariable Cox regression analysis of these parsimonious PLS-based models are shown in Table S1 (online). Proportional hazard assumption was tested using the Schoenfeld residuals test and no violations were found.

Relationship between Selected Predictors and Cardiovascular Events

Table 3 shows the top six variables selected by the final random survival forest model for the prediction of HF, CHD, and CVD events. PLS remodeling signatures could be different among the end points but were among the strongest associations for each event. To interpret each remodeling signature, Figure 2 shows correlation between the top remodeling signatures for each outcome and LV indexed mass and volumes, sphericity, conicity, relative wall thickness, mass-to-volume ratio (LV mass divided by end-diastolic volume), stroke volume (end-diastolic volume minus end-systolic volume), ejection fraction (stroke volume divided by end-diastolic volume), and longitudinal shortening. Calculation of these clinical indexes is shown in Figure S2 (online). Figure 3 shows PLS remodeling signatures for each end point, and the association between LV shape and survival probability is visualized using partial dependence plots that show how each variable (predictor) affected the model’s prediction. Kaplan-Meier survival curves are used to illustrate the abilities of remodeling scores to predict the end points. Partial dependence plots for the top three cardiovascular risk factors (excluding remodeling signatures) for each event group are provided in Figure S3 (online). Animations of the remodeling signatures for each outcome measure are shown in the Movie (online).

Table 3:

Top-ranked Variables according to Outcome

Figure 2:

Correlation matrix heatmap shows the correlation coefficients between the top remodeling signatures for heart failure (HF), coronary heart disease (CHD), and cardiovascular disease (CVD) events. Each ellipse approximates the shape of a bivariate normal distribution with the same correlation. Colors represent the strength and direction of the correlation. Correlation coefficients were multiplied by negative one for better visualization and interpretation because the time to event was used for the regression (ie, a decrease in z score was associated with an increase in probability of event). Only significant values (P < .05) are reported. ED = end diastole, EDVi = indexed end-diastolic volume, ES = end systole, ESVi = indexed end-systolic volume, LVMi = indexed left ventricular mass, RS = remodeling signature.

Figure 3:

The top partial least squares (PLS) remodeling signatures for (A) incident heart failure. Top: Graphics show the first three remodeling signatures (RS1–RS3), whereby the mesh on the left illustrates what a shape in the high-risk category looks like (see bottom row) and the mesh on the right illustrates what a shape in the low-risk category looks like. End diastole is shown as colorless and end systole as colored wireframe. Green indicates endocardium and red indicates epicardium. Middle: Kaplan-Meier survival curves of the two risk subgroups (high risk and low risk) based on patient-specific z scores show the prognostic relevance of each remodeling signature. The optimal cutoff for PLS-derived remodeling signatures separating the two risk groups was determined using classification and regression trees. Participants were free of event at baseline. Bottom: Partial dependence plots show partial values as red dots, Loess curves as black dashed lines, and error bars of plus or minus two standard errors as red dashed lines. The vertical dashed line indicates the threshold determined by using classification and regression trees separating high-risk and low-risk groups for each remodeling signature. st.dev = SD. The top partial least squares (PLS) remodeling signatures for (B) incident coronary heart disease. Top: Graphics show the first three remodeling signatures (RS1–RS3), whereby the mesh on the left illustrates what a shape in the high-risk category looks like (see bottom row) and the mesh on the right illustrates what a shape in the low-risk category looks like. End diastole is shown as colorless and end systole as colored wireframe. Green indicates endocardium and red indicates epicardium. Middle: Kaplan-Meier survival curves of the two risk subgroups (high risk and low risk) based on patient-specific z scores show the prognostic relevance of each remodeling signature. The optimal cutoff for PLS-derived remodeling signatures separating the two risk groups was determined using classification and regression trees. Participants were free of event at baseline. Bottom: Partial dependence plots show partial values as red dots, Loess curves as black dashed lines, and error bars of plus or minus two standard errors as red dashed lines. The vertical dashed line indicates the threshold determined by using classification and regression trees separating high-risk and low-risk groups for each remodeling signature. st.dev = SD. The top partial least squares (PLS) remodeling signatures for (C) incident cardiovascular disease. Top: Graphics show the first three remodeling signatures (RS1–RS3), whereby the mesh on the left illustrates what a shape in the high-risk category looks like (see bottom row) and the mesh on the right illustrates what a shape in the low-risk category looks like. End diastole is shown as colorless and end systole as colored wireframe. Green indicates endocardium and red indicates epicardium. Middle: Kaplan-Meier survival curves of the two risk subgroups (high risk and low risk) based on patient-specific z scores show the prognostic relevance of each remodeling signature. The optimal cutoff for PLS-derived remodeling signatures separating the two risk groups was determined using classification and regression trees. Participants were free of event at baseline. Bottom: Partial dependence plots show partial values as red dots, Loess curves as black dashed lines, and error bars of plus or minus two standard errors as red dashed lines. The vertical dashed line indicates the threshold determined by using classification and regression trees separating high-risk and low-risk groups for each remodeling signature. st.dev = SD.

Movie:

Download video file ^{(6.6MB, mp4)}

Remodeling signature animations (GIF in PPT file).

For incident HF, eight variables were selected by the random survival forest, including age, fasting glucose level, N-terminal pro–brain natriuretic peptide level, systolic blood pressure, and heart rate, as well as three PLS remodeling signatures. The first remodeling signature (hereafter, RS1) had the strongest correlation between shape and outcome and was associated with lower systolic function, shown by a reduced ejection fraction and longitudinal shortening, with increasing risk (Figs 2, 3). The first pattern was also characterized by an increase in LV mass and volumes but no change in relative wall thickness, in accordance with an eccentric hypertrophy at higher risk. The second remodeling signature (hereafter, RS2) was associated with reduced LV volumes and longitudinal shortening, with higher risk. The third remodeling signature (hereafter, RS3) was associated with increased sphericity, mass-to-volume ratio, and relative wall thickness with constant indexed LV mass, with higher risk. This suggests that concentric hypertrophy with a normal ejection fraction and reduced longitudinal systolic function at baseline was also associated with incident HF. The existence of these two types of remodeling associated with HF suggests heterogeneous disease mechanisms. Figure 3 shows that the 10-year survival of participants with high-risk scores was 56% compared with 95% for those with low-risk scores.

For incident CHD, seven variables were selected: age, systolic blood pressure, smoking, high-density lipoprotein level, higher coronary artery calcium score, and two PLS remodeling signatures (RS1 and RS2). The coronary artery calcium score was by far the most important. RS1 was associated with an increase in conicity and a decrease in sphericity at higher risk (Figs 2, 3). This mode was also associated with LV hypertrophy (increase in mass-to-volume ratio and relative wall thickness and decrease in volumes) typical of concentric geometry, as well as a decrease in longitudinal shortening and an increased ejection fraction at higher risk. RS2 was characterized by both a decrease in LV function and a concomitant increase in LV dimension, with a slight or negligible increase in relative wall thickness consistent with eccentric hypertrophy at higher risk.

Incident CVD was associated with 15 variables: higher coronary artery calcium score, age, fasting glucose level, heart rate, high-density lipoprotein level, systolic blood pressure, and nine PLS modes. Age, systolic blood pressure, and higher coronary artery calcium score were top predictors among risk factors for incident CVD. Similar to CHD, the first two most important shape predictors (RS1 and RS2) were associated with concentric remodeling and eccentric hypertrophy, respectively. The third shape predictor (RS3) was characterized by an increase in sphericity and apical conicity at both end-diastolic and end-systolic phases at higher risk.

Figure 4 summarizes each remodeling signature in terms of currently understood remodeling patterns. For incident HF, three patterns were identified as follows: (a) an eccentric hypertrophy with reduced systolic function (RS1), (b) a concentric remodeling with increased sphericity and reduced longitudinal shortening but with preserved ejection fraction (RS2), and (c) increased conicity associated with apical dilation (RS3). The first signatures (RS1) for incident CHD and CVD were both associated with concentric hypertrophy, as reported by Tsao et al (24), with a somewhat increased ejection fraction and reduced longitudinal shortening, likely associated with elevated systolic blood pressure and increased torsion (25). The second signatures (RS2) for both CHD and CVD were associated with decreased longitudinal shortening, which is in agreement with reports by Ng et al and Pavlopoulos et al (26,27).

Figure 4:

Summary of each remodeling signature (RS) in terms of currently understood remodeling patterns. (A) Illustration shows the remodeling signatures associated with heart failure; RS1 was associated with eccentric hypertrophy, RS2 was associated with concentric hypertrophy, and RS3 was associated with apical dilation. (B) Illustration shows the remodeling signatures associated with coronary heart disease; RS1 was associated with concentric hypertrophy and RS2 with eccentric hypertrophy. (C) Illustration shows the remodeling signatures associated with cardiovascular disease; RS1 was associated with concentric hypertrophy, RS2 was associated with eccentric hypertrophy, and RS3 was associated with sphericity and apical dilation. EDVi = indexed end-diastolic volume, EF = ejection fraction, ESVi = indexed end-systolic volume, LS = longitudinal shortening, LVMi = indexed left ventricular mass, M:V = mass-to-volume ratio.

The global remodeling signatures were also associated with local functional changes. Figure S4 (online) illustrates different regional wall thickening patterns captured by the HF RS1 versus CHD RS1. For example, the HF RS1 was associated with reduced inferoseptal wall thickening, whereas the CHD RS1 was associated with increased midventricular wall thickening and somewhat reduced apical radial thickening, consistent with compensated LV remodeling (28,25).

Discussion

Although left ventricular (LV) mass, volumes, and sphericity are known to be associated with adverse events in clinical and subclinical disease, the broader relationship between LV shape and developing disease is still poorly understood. In this study, we identified LV remodeling signatures associated with incident heart failure, coronary heart disease, and cardiovascular disease events over a 10-year follow-up period in an asymptomatic cohort. We then used survival modeling in the context of shape analysis to investigate remodeling signatures indicative of disease mechanisms. Shape signatures offer new analysis tools, which are more powerful than traditional mass and volume measures. This enables a method for capturing fine-grain imaging features that results in additional prognostic power and provides additional knowledge of complex remodeling. In addition, we provided associations of LV geometry with cardiovascular events, and we showed how patient-specific remodeling signature scores related to each end point, incorporating several heterogeneous aspects of remodeling. This confirms findings reported by Tsao et al and Zile et al (24,29), suggesting that measurements of LV geometry carry useful information over and above that conferred by mass, volume, and ejection fraction alone.

Automatic computation of patient-specific remodeling signature scores for each event class enables prospective studies to be conducted to evaluate the effects of treatment on the z scores over time. Shape changes associated with changes in z scores can be visualized and related to factors such as exercise tolerance or other biomarkers. Analogous to polygenic risk scores, which have been shown to be predictive of HF events independently of risk factors (30), event-specific remodeling signature scores enable high-dimensional shape features to be distilled to a few scores, with the added benefit that treatment effects can be automatically evaluated. Patient-specific scores can then be evaluated in terms of disease progression in a precision medicine framework by tracking the amount of remodeling signatures present in each patient.

Unlike many other machine learning algorithms, the computed remodeling signatures can be visualized and interrogated for specific physiologic mechanisms of disease development. The links between these patterns and microstructural and genetic pathways (31,32) can be investigated by analysis of the remodeling signature z scores. Although some information may be captured by mass, volume, and ejection fraction measures, the significantly higher discrimination observed for model 3 versus model 2 indicates that additional prognostic information is present in the remodeling signatures. This improvement was greater than that shown between model 1 and model 2, indicating that shape information is prognostically more powerful than standard mass and volume metrics.

Our results support the finding that both eccentric and concentric hypertrophy are associated with incident HF but with a different magnitude of risk (Fig 3A), in accordance with the study by Velagaleti et al (4). Regarding remodeling signatures, RS1 may be linked to increased wall stress (law of Laplace) whereas RS2 was found to be linked with reduced longitudinal systolic function with a preserved ejection fraction, which is in agreement with previous studies (33,34).

The first limitation of our study is the low rate of events. Incident HF was identified in 3%, CHD was identified in 7%, and CVD was identified in 10% of participants who experienced events, which may lead to model bias toward event-free participants. Therefore, the PPVs reported in Table 2 may seem low because they should range between 0% and 100%. However, as the PPVs are based on prevalence, having a high PPV with such low event rates would require an AUC greater than 0.95 (eg, 95% specificity and 95% sensitivity would give a PPV of 38% for HF). Another limitation of our study is shape bias from the use of gradient-recalled-echo imaging; steady-state free precession imaging is now the current standard for cardiac MRI and shapes differ between the two protocols (35). With further advances in cardiac MRI protocols, transfer learning may be required to adapt the current algorithms. However, previous studies have shown that shape models can be corrected between protocols (35) and atlas analyses are robust to methodology (13). Further validation should be performed on a data set with a higher rate of events to further improve associations. Validation on an independent cohort should also be performed to confirm the generalizability of the remodeling signatures. Further study distinguishing between HF with preserved ejection fraction and reduced ejection fraction should be performed when more participants in each category are available. We have not investigated radiomics features, which have been shown to provide prognostic information (36,37). Finally, future work should include the description of shape changes in the standard American Heart Association segments, which may aid interpretation of the remodeling signatures. Note that regional information was captured by the global remodeling signatures as information from all points were included in the PLS computation (38).

In conclusion, our study has demonstrated that left ventricular shape signatures were independently associated with cardiovascular events over a 10-year follow-up period in a large asymptomatic cohort and were of greater prognostic value than traditional mass and volume measures. These personalized remodeling signatures, automatically calculated from standard medical imaging examinations, provided event-specific signatures and unique mechanistic information about the development of disease in the asymptomatic cohort.