Radiomics from Cardiovascular MR Cine Images for Identifying Patients with Hypertrophic Cardiomyopathy at High Risk for Heart Failure

Abstract

Purpose

To develop a model integrating radiomics features from cardiac MR cine images with clinical and standard cardiac MRI predictors to identify patients with hypertrophic cardiomyopathy (HCM) at high risk for heart failure (HF).

Materials and Methods

In this retrospective study, 516 patients with HCM (median age, 51 years [IQR: 40–62]; 367 [71.1%] men) who underwent cardiac MRI from January 2015 to June 2021 were divided into training and validation sets (7:3 ratio). Radiomics features were extracted from cardiac cine images, and radiomics scores were calculated based on reproducible features using the least absolute shrinkage and selection operator Cox regression. Radiomics scores and clinical and standard cardiac MRI predictors that were significantly associated with HF events in univariable Cox regression analysis were incorporated into a multivariable analysis to construct a combined prediction model. Model performance was validated using time-dependent area under the receiver operating characteristic curve (AUC), and the optimal cutoff value of the combined model was determined for patient risk stratification.

Results

The radiomics score was the strongest predictor for HF events in both univariable (hazard ratio, 10.37; P < .001) and multivariable (hazard ratio, 10.25; P < .001) analyses. The combined model yielded the highest 1- and 3-year AUCs of 0.81 and 0.80, respectively, in the training set and 0.82 and 0.77 in the validation set. Patients stratified as high risk had more than sixfold increased risk of HF events compared with patients at low risk.

Conclusion

The combined model with radiomics features and clinical and standard cardiac MRI parameters accurately identified patients with HCM at high risk for HF.

Keywords: Cardiomyopathies, Outcomes Analysis, Cardiovascular MRI, Hypertrophic Cardiomyopathy, Radiomics, Heart Failure

Supplemental material is available for this article.

© RSNA, 2024

Summary

The combined predictive model integrating radiomics features with clinical and standard cardiac MRI predictors was able to identify patients with hypertrophic cardiomyopathy at high risk for heart failure events.

Key Points

• ■ In patients with hypertrophic cardiomyopathy (HCM), the radiomics score from cardiac MR cine images was the strongest predictor of heart failure (HF) events, with hazard ratios of 10.37 and 10.25 in univariable and multivariable Cox analyses, respectively (both P < .001).

• ■ A combined predictive model based on radiomics score and clinical and standard cardiac MRI predictors demonstrated high performance in predicting HF events in patients with HCM (the 1- and 3-year areas under the receiver operating characteristic curve were 0.82 and 0.77, respectively, in the validation set).

• ■ Patients at high risk identified by the combined model had more than sixfold increased risk of HF events compared with patients at low risk (hazard ratio, 6.54 in the training set and 6.73 in the validation set).

Introduction

Hypertrophic cardiomyopathy (HCM) stands as the most prevalent monogenic cardiovascular disorder, with a prevalence of at least one in 500 within the general population (1). Among the adverse events in HCM, heart failure (HF) holds particular importance. Limiting symptoms related to HF, such as exertional dyspnea, substantially reduced the quality of life for many patients with HCM. At the same time, with the reduction in sudden deaths due to the use of implantable defibrillators in this disease, HF-related hospitalizations and deaths have become major issues affecting the survival of patients with HCM (1–6).

Primary prevention for patients at risk for HF can be useful to prevent the development of left ventricular dysfunction or new-onset HF (7). However, the risk of HF event occurrence varies among patients with HCM, and it is unnecessary to provide early intervention for patients at low risk. Therefore, accurately identifying patients at high risk has become the primary concern at present. Although several predictors, including quantification of late gadolinium enhancement (LGE) and left ventricular end-systolic volume index, were proven useful in HF risk stratification (3,8), they have achieved only modest success. More powerful predictors as well as predictive models that combine these various predictors are required to enhance the accuracy of HF risk stratification. Radiomics can quantitatively extract a plethora of features hidden within medical images, including texture features. The differences in radiomics features derived from cardiac cine images among patients have been used for prognostic predictions in various diseases (9). In patients with HCM, radiomics features derived from cardiac cine images have also demonstrated substantial value in the phenotyping of scarred myocardium (10,11).

In this study, we aim to investigate the value of radiomics for predicting HF events in patients with HCM and to construct an efficient model combined with radiomics features and clinical and standard cardiac MRI predictors for identifying those at high risk for HF.

Materials and Methods

Study Patients

This study was approved by the institutional ethics committee (2023160X) at Beijing Anzhen Hospital. Written informed consent was waived owing to the study’s retrospective nature. Patients who were diagnosed with HCM and underwent cardiac MRI examination from January 2015 to June 2021 in our hospital were retrospectively reviewed in this study.

Patients were included in the study if they met the following criteria (12): (a) maximal left ventricular end-diastolic wall thickness of 15 mm or more was found at cardiac MRI in the absence of any other causes of hypertrophy in adults, or (b) maximal left ventricular end-diastolic wall thickness of 13–14 mm if a positive genetic test result or HCM family history was present. The exclusion criteria were as follows: (a) patients younger than 14 years old (n = 14), (b) patients with severe coronary heart disease (> 50% coronary stenosis) (n = 25), (c) patients with congenital heart diseases (n = 5), (d) patients with previous septal ablation or myectomy (n = 30), (e) insufficient image quality (n = 49), (f) LGE images that were not available (n = 18); or (g) patients with New York Heart Association functional class III or IV at baseline (n = 65).

Follow-up data were obtained through hospital records, clinic visits, and telephone interviews. We contacted patients who did not visit the hospital regularly and requested that they send us medical records from other hospitals, allowing us to ascertain whether any end points were reached. The end point in our study was HF events, including HF hospitalization and HF death. HF hospitalization was defined as new or worsening signs of HF leading to unplanned hospital admission of more than 24 hours. HF death was defined as death associated with unstable or progressive deterioration of pump function. Blinded to the initial cardiac MRI results, two experienced cardiologists (H.Z. with 6 years of experience and K.H. with 5 years of experience) assessed whether patients experienced HF events.

Cardiac MRI Acquisition

Cardiac MR images were collected using three scanners at 3.0 T (Discovery MR750, General Electric Medical Systems; MAGNETOM Siemens Verio, Siemens Healthcare; Ingenia, Philips Healthcare). Cine images were collected using balanced steady-state free precession sequences. A breath-hold inversion-recovery segmented gradient-echo sequence was used to collect LGE images. Scanning parameters are listed in Table S1.

Cardiac MRI Analysis and Radiomics Feature Extraction

The endocardial and epicardial contours from the base to the apex of the left ventricle in cardiac short cine images were automatically delineated and manually adjusted by two experienced operators (L.Z. and C.Z., with 12 and 10 years of cardiac MRI experience, respectively) independently using cvi42 software (version 5.11.2, Circle Cardiovascular Imaging). Subsequently, left ventricular ejection fraction, left ventricular end-diastolic volume, left ventricular end-systolic volume, and left ventricular mass were derived by the software and were standardized by body surface area.

We defined the region of interest as the left ventricular myocardium from the base to the apex in short cine images at the end-diastolic phase for the extraction of radiomics features. Before feature extraction, the cine images were resampled to an in-plane voxel size of 1 × 1 × 1 mm³, and a bin width of 25 HU was used for gray value discretization (9,10). Subsequently, we used an open-source package (PyRadiomics) for feature extraction through Python software (version 3.7, https://www.python.org). The detailed methods and the distribution of radiomics features are provided in Appendix S1. Radiomics features were standardized by using z scores.

Afterward, myocardial contours were manually delineated on LGE images using cvi42 software (L.Z. and C.Z.). LGE was defined as myocardium 6 SDs above the normal myocardial signal intensity. LGE (positive) was defined as LGE of more than 1% of total left ventricular mass.

A group of 30 patients was randomly selected to investigate the reproducibility of each feature. After 1 month, readers 1 and 2 (L.Z. and C.Z.) performed myocardial segmentations for a second time to evaluate intraobserver reproducibility, while another reader (X.M. with 15 years of cardiac MRI experience) independently conducted a separate segmentation to evaluate interobserver reproducibility. All readers used the same method as described earlier to analyze the images.

Radiomics Feature Selection

First, features with excellent reproducibility (intraclass correlation coefficient > 0.80) in both intraobserver and interobserver analysis were kept for further selection. Then a Spearman correlation matrix for all kept features was calculated, and one random feature was excluded from any feature pair with a Spearman correlation coefficient greater than 0.8 (13). Finally, the least absolute shrinkage and selection operator Cox regression model was used to select HF event–related features with nonzero coefficients from the training set. The radiomics score was generated via a linear combination of selected features weighted by their respective coefficients (14).

Statistical Analysis

Continuous variables are presented as the median (IQR) and were compared using the Mann-Whitney U test. Categorical variables are presented as frequencies or percentages and were compared using the χ² test. Univariable and multivariable Cox proportional hazard models were used to select predictors that are significantly related to HF events. Parameters with a P value < .05 in the univariable analysis were incorporated into the forward stepwise multivariable analysis to construct both the standard model (without radiomics score) and the combined model (with radiomics score). The entering and removing limits were P < .10 and P > .05 in multivariable analysis. To avoid collinearity, only one predictor was included in the multivariable analysis among variables with a variance inflation factor greater than 3 (15,16). We verified the proportional hazards assumption of the models by examining the scaled Schoenfeld residual plots. HF event–free survival probabilities were estimated by using the Kaplan-Meier method. The concordance index (C index) was used to evaluate model discrimination. The time-dependent area under the receiver operating characteristic curves (AUCs) were used to assess prognostic accuracy. A calibration plot using 1000 bootstrap resamples was used to assess model fit. The optimal cutoff value for different models in HF events risk stratification was determined using the X-tile software (version 3.6.1, Rimm Laboratory, Yale School of Medicine) (17). Then the patients were divided into low- and high-risk groups according to the optimal cutoff value of the combined model, and the log-rank test was used to compare the outcomes of different groups. Statistical analysis was performed in SPSS software (version 25.0, IBM), MedCalc software (version 19.3.1, MedCalc Software), and R programming language (version 4.1.2, http://www.r-project.org). The used codes are available at GitHub (https://github.com/codes2345/HCMHF). P < .05 was considered statistically significant.

Results

Study Sample Characteristics

A total of 760 patients with HCM were initially assessed for eligibility, of whom 206 patients were excluded based on the predefined exclusion criteria. Subsequently, 554 patients with HCM were deemed eligible for inclusion in the study. During the follow-up process, 38 of 554 (6.9%) patients were excluded as they could not be contacted. A total of 516 patients with HCM (median age, 51 years [IQR: 40–62]; 367 [71.1%] men and 149 [28.9%] women) were included and were randomly allocated into a training set (n = 361) and a validation set (n = 155) with a ratio of 7:3 (Fig 1). Patient characteristics, cardiac MRI, and echocardiography findings are summarized in Tables 1 and 2.

Figure 1:

Flowchart of patient selection for this study. HCM = hypertrophic cardiomyopathy, LGE = late gadolinium enhancement.

Table 1:

Characteristics of Patients with HCM in the Training Set and Validation Set

Table 2:

Cardiac MRI and Echocardiography Findings in Patients with HCM in the Training Set and Validation Set

The median duration of follow-up was 39.0 months (IQR: 24.0–39.5) for the training set and 41.0 months (IQR: 24.0–52.0) for the validation set. HF events were similar between the training set and validation set (P = .96, log-rank test). A total of 49 of 516 patients (9.5%; 34 patients in the training set and 15 patients in the validation set) met the end point: nine of 516 with HF deaths (1.7%; seven patients in the training set and two patients in the validation set) and 40 of 516 with HF hospitalizations (7.8%; 27 patients in the training set and 13 patients in the validation set).

Radiomics Feature Selection and Radiomics Score Calculation

A total of 1408 features were extracted from end-diastolic cine images; five features were finally identified to calculate the radiomics score after feature selection. The intraobserver reproducibility, interobserver reproducibility, C index, and coefficients for these features are listed in Table S2. The radiomics score is obtained by multiplying these features by their coefficients (which is determined based on the least absolute shrinkage and selection operator Cox regression) and then adding them together (Equation 1). The C indexes of radiomics score were 0.75 (95% CI: 0.66, 0.84) and 0.72 (95% CI: 0.57, 0.87) in the training set and validation set, respectively.

Prediction Models Development and Validation

In univariable Cox regression analysis, a history of nonsustained ventricular tachycardia, a history of hypertension, left ventricular end-diastolic volume index, left ventricular end-systolic volume index, left ventricular ejection fraction, quantification of LGE, and radiomics score were significant predictors of HF events (Table 3). The standard model (without radiomics score) and the combined model (with radiomics score) were built using multivariable Cox regression analysis with the characteristics that were statistically significant in the univariable Cox regression analysis (Tables S3, S4). The standard score (without radiomics score) and the combined score (with radiomics score) were obtained by multiplying the significant variables in the multivariable Cox regression by their hazard ratios and adding them together (Equations 2 and 3, respectively). Calibration curves of the combined model for the training set and validation set are shown in Figure 2A and 2B.

Table 3:

Univariable Cox Regression Analysis in the Training Set

Figure 2:

Graphs of (A, B) calibration curves of the combined model in the (A) training set and (B) validation set and (C, D) time-dependent areas under the receiver operating characteristic curve (AUC) of the combined model, radiomics score (radscore), and standard model in the (C) training set and (D) validation set. Calibration curves show the agreement between model predicted and observed heart failure event–free survival (HFS) outcomes. Time-dependent AUC evaluated the predictive performance of the models at different time points; the standard model exhibited significantly poorer performance compared with the combined model and radscore.

Compared with the standard model (training set, C index: 0.70 [95% CI: 0.60, 0.80], 1-year AUC = 0.78 [95% CI: 0.66, 0.90], 3-year AUC = 0.69 [95% CI: 0.57, 0.82]; validation set, C index: 0.65 [95% CI: 0.47, 0.82], 1-year AUC = 0.76 [95% CI: 0.57, 0.95], 3-year AUC = 0.64 [95% CI: 0.44, 0.83]), the combined model (training set, C index: 0.78 [95% CI: 0.70, 0.85], 1-year AUC = 0.81 [95% CI: 0.70, 0.93], 3-year AUC = 0.80 [95% CI: 0.71, 0.89]; validation set, C index: 0.76 [95% CI: 0.64, 0.89], 1-year AUC = 0.82 [95% CI: 0.69, 0.96], 3-year AUC = 0.77 [95% CI: 0.61, 0.93]) exhibited better predictive performance for HF events in both the training set and validation set (Fig 2C, 2D).

HF Risk Stratification

Due to the superior predictive performance of the combined model, this study primarily relies on this model for risk stratification. Patients were separated into low-risk (combined score ≤ 7.20) and high-risk (combined score > 7.20) groups according to the optimal cutoff value in the training set determined by X-tile. The mean HF event–free survival times for the low- and high-risk groups were 81.5 and 60.6 months, respectively, in the training set (P < .001, log-rank test for two groups) (Fig 3A). Cumulative 1- and 3-year HF event rates in the training set were 2.2% and 5.0%, respectively, for the low-risk group and 16.3% and 31.8% for the high-risk group (Table 4). A similar prognostic impact of this stratification was confirmed in the validation set (P = .001, log-rank for two groups) (Fig 3B). The risk stratification based on radiomics score and standard score is listed in Tables S5 and S6; the cutoff values for radiomics score and standard score were 0.33 and 42.50, respectively. Representative patients of low-risk and high-risk groups are shown in Fig 4A–4D.

Figure 3:

Graphs of Kaplan-Meier curves of patients with hypertrophic cardiomyopathy at different risks of heart failure (HF) events stratified by the combined score in the (A) training set and (B) validation set.

Table 4:

Mean HF Event-free Survival Time and Cumulative HF Event Rates according to Each Risk Group Defined by the Combined Model

Figure 4:

Data and cine images representative of patients in low-risk and high-risk heart failure (HF) groups. (A) Typical patient at high risk (with a high standard score, high radiomics score, and high combined score); HF events occurred 6 months after examination. (B) Typical patient at low risk (with a low standard score, low radiomics score, and low combined score); no HF events occurred during follow-up. (C) Patient classified as high risk based on combined score and radiomics score but with a low standard score; HF events occurred at 28 months after examination. (D) Patient classified as low risk based on combined score and radiomics score but with a high standard score; no HF events occurred during follow-up. Cine images were collected using balanced steady-state free precession sequences, and a breath-hold inversion-recovery segmented gradient-echo sequence was used to collect late gadolinium enhancement (LGE) images. 4ch = four chamber, LVESVi = left ventricular end-systolic volume index, NSVT = nonsustained ventricular tachycardia, 2ch = two chamber.

Discussion

In this study, we demonstrated that radiomics features extracted from cardiac cine images could effectively predict HF events in patients with HCM. Additionally, as the strongest predictor of HF events (hazard ratio: 10.37 in the univariable Cox analyses and 10.25 in the multivariable Cox analyses; both P < .001), incorporating the radiomics score into the predictive model substantially enhances the model’s predictive performance (C index: 0.78 in the training set and 0.76 in the validation set). This result confirms the vital value of radiomics in predicting HF events in patients with HCM. More importantly, the risk stratification achieved by the combined predictive model constructed in this study effectively identified patients at high risk for HF events (hazard ratio: 6.54 in the training set and 6.73 in the validation set; both P < .001), potentially allowing for the stratified management of patients with HCM.

Despite not reaching the predictive performance of radiomics score, clinical and standard cardiac MRI predictors play a crucial role in the combined model. Among all the clinical data, hypertension and a history of nonsustained ventricular tachycardia were identified as significant predictors of HF events in the univariable Cox analysis. This finding is partly consistent with previous studies (3,8,18). Hypertension is a recognized risk factor associated with a relatively high risk of developing HF (7,19). However, the status of a history of nonsustained ventricular tachycardia as a predictor of HF events has been subject to controversy. In the study conducted by Negri et al (18), this variable was not a significant predictor in the univariable Cox regression analysis. Nonetheless, a recent study involving a larger population indicated that the history of nonsustained ventricular tachycardia could indeed serve as a predictive factor for HF events in patients with HCM (3), a finding that aligns with the results of our current study.

LGE in cardiac MRI has been observed to closely correspond to the distribution of myocyte necrosis in the early stages (19–21). Extensive LGE has been identified as a noninvasive indicator of an elevated risk of life-threatening ventricular tachyarrhythmias (22). Additionally, LGE is widely acknowledged as a predictor for HF events in patients with HCM (3,18). In our current study, quantification of LGE emerged as a robust predictor for HF events. The risk of HF events increased by 1.16 times for every 5% increment of LGE in univariable Cox analysis.

As the most robust predictor of HF events in this study, the radiomics score plays an important role within the combined model. Radiomics facilitates the quantification of intratissue heterogeneity in the spatial distribution of voxel intensities, offering a means for outcome prediction (23). In the context of HCM, radiomics has been used to anticipate the presence of scarred myocardium and to differentiate from other conditions that could lead to left ventricular hypertrophy, such as hypertensive heart disease (10,11,24). This study takes a further step by harnessing radiomics for HF events prediction, potentially enhancing clinical management. Moreover, the radiomics score can independently predict HF events without the need for additional information. This characteristic simplifies the prediction process and enables rapid risk stratification after scanning. Furthermore, patients who are contraindicated to contrast agents may benefit from this technique.

The accurate identification of patients at high risk and prone to HF events among patients with HCM is a major contribution of this study. Since the optimal cutoff value was determined in this study, patients with HCM can be stratified into low-risk and high-risk groups. Patients with high risk had more than a sixfold increased risk of HF events compared with patients at low risk. Primary prevention is crucial for patients at high risk because of the high incidence for HF events even after 1 year (16.3% in the training set and 27.3% in the validation set). At the same time, as patients classified as high-risk groups constitute less than 15% of the total population, this has the potential to save medical resources.

This study had several limitations. First, this was a single-center study with a relatively short follow-up time. Although guidelines recommend patients with HCM undergo repeat enhanced cardiac MRI examinations every 3–5 years (12), the median follow-up time of 39 months allows the model in this study to effectively predict HF events between two cardiac MRI examinations, and HF risk can then be reassessed after the next cardiac MRI examination. However, considering that some patients may not adhere to guidelines for repeat cardiac MRI examinations, early identification of patients at high risk for HF is more conducive to early intervention. A multicenter study with a longer follow-up time is needed to assess the clinical utility of the model developed in this study. Second, a retrospective analysis might inevitably cause bias in study sample selection. The models constructed in this study need to be validated with prospective data sets before clinical application. Third, in this study, patients did not undergo repeat cardiac MRI examinations within a short period using different MRI vendors. Therefore, we were unable to assess the reproducibility of radiomics features across different MRI vendors. This issue will be assessed in future prospective studies. Finally, the extraction and analysis process of radiomics features can be quite laborious, which has limited its practical clinical application. To address this, there is a need for the development of deep learning–based fully automated tools that can handle the extraction of radiomics features, their analysis, and calculation of the radiomics score. Such tools would substantially streamline the process and potentially facilitate broader clinical use.

In conclusion, radiomics features extracted from cardiac cine images offer additional value for prediction of HF events in patients with HCM as compared with clinical and standard cardiac predictors alone. A model combined with radiomics score as well as clinical and standard cardiac MRI predictors could aid HF risk stratification in patients with HCM.