Le Cœur en Sabot: shape associations with adverse events in repaired tetralogy of Fallot

This study retrospectively re-analyzed anonymised images and associated data from a large cohort of adolescents and young adults with rToF recruited from a multicenter study of the German Competence Network for Congenital Heart Defects (study identifier: NCT00266188, title: Non-invasive Imaging and Exercise Tolerance Tests in Post-repair Tetralogy of Fallot—Intervention and Course in Patients Over 8 Years Old). Ethical approval was obtained and all participants gave informed consent. This rTOF dataset constitutes one of the most extensive compiled datasets of this pathology to date. The data was acquired at 14 different sites between 2005 and 2008 on 1.5T and 3T CMR scanners. The primary adverse outcome was defined as any cardiac arrest, ventricular tachycardia, or death until the latest outcome data collection point in 2019 with median follow-up time of ten years (Table 1). Further descriptions of clinical data can be found in [11, 12].

CMR cine exams included multiple short-axis slices covering both RV and LV and one four chamber (4Ch) long-axis slice acquired during breath-hold. End-diastolic (ED) and end-systolic (ES) CMR images were segmented semi-automatically with a subsequent manual correction by a single reader. The LV endocardial, LV epicardial, and RV endocardial contours were defined using cvi42 software (Circle Cardiovascular Imaging, Calgary, Canada) according to the standards published by Society for Cardiovascular Magnetic Resonance [13]. Papillary muscles and trabeculae were included in the chamber volume for this analysis. Tricuspid and mitral valve hinge points and LV apex position were extracted from the 4Ch view. RV epicardial contours were not defined due to the poor reproducibility of manual delineation [14]. Pulmonary regurgitant volume was computed from phase contrast flow imaging across the main pulmonary artery, using flow measurements was performed using dedicated customized software capable of handling data of different vendors [15].

A biventricular model template was automatically customised to the extracted contours using previously validated methods [6, 7]. Briefly, short axis slice positions were first corrected for breath-hold misregistration, and the template was then rigidly aligned and fit to the contours using a diffeomorphic non-rigid registration (see Fig. 1).

Global functional and volumetric parameters were computed from the patient-specific shape model, including ED and ES volumes (LVEDV, RVEDV, LVESV, RVESV), stroke volumes (LV SV, RV SV), LV mass (LVM) and ventricular ejection fraction (LVEF, RVEF).

Besides global volume, regional volumes were also computed using methods similar to Bernardino et al. [16]. The total RV volume was parcellated into three different regions based on RV apex, pulmonary valve, and tricuspid valve centroids position (Fig. 2). The surface vertices were classified as inlet, outlet, and apical regions according to the closest landmark. For each region, partial volume, SV, and LVEF/RVEF were computed.

Patient-specific biventricular geometries were used to build a statistical cardiac atlas characterising the shape variation in patients with ToF late after repair. Details of the procedure are given in [6, 8]. Briefly, a Procrustes algorithm was used to rigidly align the ED shapes to a reference coordinate system defined by the population mean shape. The same transformation was applied to the ES shape, thereby preserving their relative position and the surface deformation from ES to ED.

Principal component analysis (PCA) was performed to build a statistical shape model. Both geometry and motion were encoded by concatenation of the ED and ES shapes for each case. The statistical shape model included RV and LV endocardial and LV epicardial surfaces. However, for clarity, the figures displaying biventricular shape show the RV and LV endocardial surface only.

A PCA component, known as a shape mode (SM), reflects a specific shape variation pattern and encompasses attributes of size and both global and regional shapes. Any individual 3D shape can be characterised by a set of normalised scores (z-scores) specifying each SM’s contribution to the subject-specific geometry. For a robust and meaningful examination, only SMs explaining more than 1% of total variation were selected for further analysis. That resulted in 15 SMs, which in total explained 85% of the variation across the cohort.

Remodelling patterns associated with adverse events were identified using linear discriminant analysis (LDA) combined with an automatic feature selection algorithm according to the area under the curve (AUC) of the receiver operator characteristic, similar to the computational framework proposed by Varela et al. [17]. The shape features used in the LDA included conventional global and regional functional and volumetric measures together with the first 15 atlas-derived SMs. Prior to multivariable LDA, all clinical, demographic, or shape features were standardised and normalised. The ventricular volume and mass were indexed by body surface area (BSA) to mitigate the effects of body habitus.

Three separate preliminary LDAs were performed, each evaluated with stratified four-fold cross-validation (Additional file 1) designed to identify the shape features most associated with adverse outcomes: i) including only conventional measures (total ventricular volumes, EF, left ventricular mass (LVM) index (LVMI), QRS, and blood pressure); ii) regional functional and volumetric features derived from RV volume parcellation, and iii) including only the first 15 SMs. To avoid overfitting, the maximum number of predictors in each step was fixed to three parameters.

Finally, a composite LDA was performed by combining the predictors identified in the three preliminary analyses. The resulting LDA score was used as a new biomarker for risk quantification. The odds ratio (OR) with 95% CIs was computed to quantify the strength of the association with adverse outcomes. The cut-off value was computed by maximising the sensitivity and specificity in the ROC curves analysis.

A surrogate multivariable LDA model was developed to explore the role that PRVI may play in the association between shape and adverse events. This model included 158 cases for which pulmonary regurgitation data were available. The correlations between critical functional and shape predictors and PRVI were studied using univariable linear regression. Then, the shape mode best explaining PRVI was computed using multivariable regression.

Atlas SMs describe both global and regional variation of ventricular shape, but their interpretation remains a challenging task. Calliper measurements have been used to quantify RV regional remodelling response in several pathophysiological conditions in imaging studies [18] and regional shape changes are important in rTOF patients [19]. In this work, multivariable regression was used to determine relationships between SM and calliper-based measures. The resulting associations were used to interpret the effect of the SM patterns.

The calliper measurements were generalised for RV and LV chambers (see Fig. 3), as summarised in Additional file 1: Table S2. Ventricular sphericity (width divided by height) and eccentricity (anterior–posterior width divided by septal-lateral width) were computed for ED and ES using widths at 1/4 (apical), ½ (mid), and 3/4 (basal) of the ventricle height. Longitudinal shortening and apex displacement were computed for both RV and LV. Valve displacement was computed for mitral and tricuspid valves.

The multivariate regression was performed for each calliper-based measure independently using the standardised shape scores as independent variables. The magnitude of the regression coefficients defined the SM’s contribution to the calliper measurement. Finally, the shape scores were interpreted in terms of the effect on calliper distances using the respective regression coefficients to indicate the correlation strength.

Statistical analyses were performed using Python 3.6 using the Scikit-learn package (Python Software Foundation, Wilmington, Delaware, USA). A p value of 0.05 was considered significant.

Of 408 patients enrolled in the main study, 380 patients had CMR examinations, 280 of these had both long axis and short axis slices, and 192 had sufficient 3D information enabling biventricular shape analysis (a long axis view capturing mitral and tricuspid valves and the LV apex) (Additional file 1: Fig. S1). (Additional file 1: Table S1) shows that the characteristics of the 216 excluded cases had similar demographics to the 192 cases included, except for a higher number of pulmonic valve replacement (PVR) after the baseline exam and a higher number of redo surgeries prior to the CMR for the excluded group. Of the 192 patients examined, 16 subjects were identified with primary adverse outcomes (AO) and 176 with no adverse outcomes. Adverse outcomes included 4 deaths, 10 ventricular tachycardia and 2 aborted sudden cardiac death (cardiac arrest requiring defibrillation). Demographic and clinical data are summarised in Table 1. In the studied cohort, the median age at baseline evaluation was 15 years. Among the no adverse outcome group, 26 patients had a PVR in the following two years after baseline evaluation. In the adverse outcome group, one patient had PVR nine years earlier than baseline evaluation, and one patient had a PVR in the following year after baseline evaluation. Six per cent of the entire population had at least one redo surgery. The median number of redo surgical repairs in the adverse outcome group was significantly higher than in the no adverse outcome group.

Functional and volumetric variables are summarised in Table 2. RVEF and LVM was significantly reduced in the adverse outcome group. Partial volume and regional EF for the RV is shown in Table 3. Regional RVEF was significantly reduced in the apical and inlet regions. The total RV volumes were similar between groups; however, the adverse outcome group showed an increased ES inflow and apical regional volume.

In the first LDA model comparing conventional predictors (volumes and EF), the RVEF had the best ability to discriminate the patients with adverse outcomes (AUC = 0.72). The cross-validation feature selection algorithm could not find any combination of additional features improving the model predictive power without the risk of overfitting.

The second LDA model (using predictors derived from RV volume parcellation only) resulted in an equivalent discriminative power (AUC = 0.73) when using apical ejection fraction (RV APEF) as the unique predictor. Both models resulted in a negative correlation between the RVEF and the probability of having an adverse outcome (RVEF coeff = − 0.55, RV APEF coeff = − 0.69).

The third LDA model, which included only PCA SMs, the best discriminative power was obtained (AUC = 0.66) with a linear combination of 2 components: SM6 and SM8. Small but significant correlations were found between SM6 and RVEF and RV APEF (R² = 0.097, p < 0.001 for RVEF and R² = 0.179, p < 0.001 for RV APEF. However no significant correlation was found between SM8 and RV APEF or RVEF ( SM8: R² = 0.004, p = 0.624 and R² = 0.003 p = 0.582 respectively). Additional figures supporting these results are provided in Additional file 1: Fig. S2–S5.

For ease of interpretation, the SMs most associated with adverse events were correlated with calliper measures (Fig. 4). Univariate analysis revealed no significant differences between adverse outcome and no adverse outcome groups. However, SM6 was found to be mostly correlated with LV ES shape and RV ED shape (Fig. 5). A negative SM6 (associated with higher risk of adverse outcomes) resulted in a more elliptical LV ES (decreased eccentricity) and a more annular RV base at ED (decreased mid and basal eccentricity—short axis). SM6 also significantly affected the mitral valve and LV apex position at the ES phase, resulting in a more annular LV at ES (increased sphericity) (Fig. 6). On the other hand, SM8 was mostly correlated with the biventricular mid and basal shape. In the adverse outcome group, the LV sphericity and RV eccentricity increased when compared to the no adverse outcome cohort. Additional material to visualize shape differences between the reference population and patients with adverse events can be found in Additional file 1.

The final multivariable LDA was built using the previously identified predictors: RVEF, RV APEF, and the two shape modes (SM6 and SM8). The optimal model included only RV APEF and SM6 and improved the discriminative performance from AUC = 0.73 to AUC = 0.77. A high LDA score was significantly associated with an increased odds ratio (OR = 2.06; 95% CI = 1.08 to 3.93). The cut-off value maximising sensitivity and specificity was equal to -0.18.

A surrogate model using data from 158 subjects with available pulmonic regurgitation data were computed to investigate the impact on the model discriminative power (Additional file 1: Fig. S6). The LDA analysis showed that the PRVI had lower discriminative power than RV APEF (AUC = 0.60 versus AUC = 0.71) and had no added value when used as a predictor along with RV APEF and SM6 (AUC = 0.73 versus AUC = 0.72).

Further analysis was performed to investigate the PRVI’s relationship with the predictors of adverse events using linear regression (Fig. 6). There was a weak correlation (R² = 0.14, p = 0.000) with the risk score, and no relation was found with either apical or total RVEF.

The shape pattern associated with PRVI was computed using LDA, and it resulted in a linear combination of 3 shape modes: SM7, SM2 and SM4 (R² = 0.50) (Additional file 1: Fig. S7). The biventricular SMs associated with increased PRVI was compared to those identified with AO (Fig. 5). RV dilation was observed in both groups yet with different dilatation patterns. In patients with high PRVI, the RV dilatation was consistent over the RV volume (blue arrows). On the other hand, in the adverse outcome group, the RV dilatation was more prominent in the apical region, with RV basal constriction (red arrows). The combination of LV basal dilatation and RV base constriction led to an elliptical shape of the heart base in patients with adverse outcomes (yellow circle, view 3). Additional files 2 and 3: Videos S1 and S2 show a continuous animation of this remodelling.