Central Venous Waveform Patterns in the Fontan Circulation Independently Contribute to the Prediction of Composite Survival

Abstract

It is well appreciated that the Fontan circulation perturbs central venous hemodynamics, with elevated pressure being the clearest change associated with Fontan comorbidities, such as Fontan-associated liver disease (FALD) and protein-losing enteropathy (PLE). Our group has better quantity of these venous perturbations through single- and multi-location analyses of flow waveforms obtained from magnetic resonance imaging of Fontan patients. Here, we determine if such analyses, which yield principal components (PC) that describe flow features, are associated with Fontan survival. Patients with a Fontan circulation (N = 140) that underwent free-breathing and mechanically ventilated cardiac MRI were included in this study. Standard volumetric and functional hemodynamics, as well as flow analysis principal components, were subjected to univariate and bivariate Cox regression analyses to determine composite clinical outcome, including plastic bronchitis, PLE, and referral and receipt of transplant. Unsurprisingly, ventricular function measures of ejection fraction (EF; HR = 0.88, p < 0.0001), indexed end-systolic volume (ESVi; HR 1.02, p < 0.0001), and indexed end-diastolic volume (EDVi; HR = 1.02, p = 0.0007) were found as specific predictors of clinical events, with specificities uniformly > 0.75. Additionally a feature of IVC flow (PC2) indicating increased flow in systole was found as a highly sensitive predictor (HR = 0.851, p = 0.027, sensitivity 0.93). In bivariate prediction, combinations of ventricular function (EF, ESVi, EDVi) with this IVC flow feature yielded best overall prediction of composite outcome. This suggests that central venous waveform analysis relays additional information about Fontan patient survival and that coupling sensitive and specific measures in bivariate analysis is a useful approach for obtaining superior prediction of survival.

Introduction

Patients with single-ventricle heart disease, a rare and severe congenital heart defect, are increasingly reaching adulthood as a result of improvement in medical and surgical care during the staged palliative surgeries culminating in the Fontan circulation [1, 2]. The circulation is defined by passive, non-pulsatile pulmonary flow in the total cavopulmonary connection (TCPC) or where the pulmonary arteries anastomose with the Fontan conduit [1]. Congestion in the TCPC leads to a range of Fontan-associated comorbidities following the completion of this surgical palliation, including protein-losing enteropathy (PLE), plastic bronchitis (PB), and Fontan-associated liver disease (FALD) [2]. In addition, there is significant risk for circulatory failure that remains clinically diverse and can present as congestive heart failure, reduced ventricular function, lymphatic leak, or a combination of comorbidities [2, 3]. The causes of PLE, PB, FALD, and other organ dysfunction are poorly understood in this population, and the current standard of care for end-organ surveillance has not substantially improved overall patient outcomes [2]. The presence of end-organ dysfunction increases the risk of patient mortality and additional comorbidities [4, 5], warranting investigation into alternative diagnostic measures. The overall goal of this work is to develop a new prognostic model for patients with SVD through interrogation of TCPC hemodynamics.

Pulsatility loss in the vasculature leads to decreased hemodynamic shear stress on the endothelium, which is necessary in regulation of vascular tone, structure, and remodeling [6]. The lack of a sub-pulmonary ventricle in the Fontan circulation results in non-pulsatile TCPC flow, which ultimately alters the pulmonary artery endothelial response [7, 8]. Decreased endothelial nitric oxide synthase, an enzyme required for production of the protective molecule nitric oxide [9], was found in a Fontan model that displayed slower vasodilation response time and an increased likelihood of developing pulmonary hypertension compared to biventricular counterparts [8]. Nitric oxide also plays an important role in lowering pulmonary vascular resistance (PVR) and a study has determined basal PVR is higher in Fontan patients of a worse cardiac functional class [7]. Additional studies have also associated non-pulsatile flow in the TCPC with greater shear blood viscosity, which correlates to decreased pulmonary blood flow and alters flow patterns in the Fontan circulation [10, 11]. These studies suggest that venous hemodynamics are altered in this circulation, a now widely accepted concept in the field [1, 2], although very few groups, aside from our own [12, 13], have attempted quantification of this notably contributing factor to Fontan failure.

A number of studies [14–19] suggest that ventricular size and function are associated with increased likelihood of Fontan complications and death, but prognostication of this population has not markedly improved since dissemination of these findings [2, 20, 21]. Additionally, our group has established that novel assessment of waveform patterns in patients following the Fontan operation are related to parameters of pulmonary, renal, and ventricular function [12, 13]. However, no research groups have investigated patient outcomes by comprehensive analyses of the central systemic and venous vasculatures within the single-ventricle circulation and thus is the novel component of this research study. We hypothesized that a multifactorial approach, with inclusion of quantitative venous flow, may provide greater predictive value than just ventricular function or venous measurements alone.

Methods

Single-ventricle patients that received a cardiac MRI (cMRI) at Children’s Hospital Colorado (CHCO) between July 2011 and August 2022 were included in this outcomes study, permitted by the Colorado Multiple Institutional Review Board (COMIRB) as a portion of the Fontan at Altitude Registry for Outcomes (FAROUT) protocol at our institution. A waiver of consent was obtained by the patient or legal guardian in accordance with required regulations. Hemodynamic measurements gathered from cMRI included aortic and TCPC flow over time, collateral flow burden, valve regurgitation, and differential pulmonary blood flow for 140 total patients. Study time was defined as the number of days from cMRI date to the date of the first composite event or the number of days from cMRI to time to follow-up preceding December 1, 2022. A composite outcome was defined as the development of PB (n = 1) or PLE (n = 2), referred to transplant (RTT, n = 9), received a transplant (n = 4), or death (n = 0) from the time of cMRI to time to follow-up. The demographics and the aforementioned measures of ventricular size & function, cardiac output, and collateral flow obtained from cMRI are displayed in Table 1. Patients were excluded from further analyses if two superior vena cavae or an obstructed outflow tract was present.

Table 1.

Mean or median demographic and hemodynamic measures with corresponding standard deviation or interquartile range for the Fontan cohort studied

Demographics and hemodynamics	Mean or median	SD or IQR
Study time (days)	779	460 –1297
Age at Fontan surgery (yrs)	3	2—4
Age at cMRI (yrs)	14	9—16
Ao valve regurgitation (%)	1.52	2.66
BSA (m2)	1.35	0.42
EF (%)	48	8
EDVi	94.4	30.4
ESVi	50.5	23.8
CI (L/min*m2)	3.52	1.37
LPA flow (%)	45	11.5
Aortic collateral flow (%)	18	10—30.5
PA collateral flow (%)	18	11.5—29.5

cMRI Acquisition

Phase velocity maps and tissue intensity sequences of the AAo, SVC, IVC, and LPA were obtained using a PC-MRI, ECG-gated sequence as previously described [12, 22], by utilizing a 1.5 or 3.0 Magnetom Avanto (Siemens Medical Solutions, Erlangen, Germany) or Ingenia (Philips Medical System, Best, the Netherlands) Tesla magnet and a phased-array body surface coil. A free-breathing PC-MRI sequence was used under the following conditions: time to repetition, 14–28 ms/25–40 cardiac phases; time to echo, 2.2–3.5 ms; matrix, 160 × 256; flip angle, 25 degrees; 100% k space sampling; cross-sectional pixel resolution, 0.82 × 0.82 mm2 and 1.56 × 1.56 mm2; and slice thickness, 5 mm. Heart rate-dependent, PC-MRI acquisition varied 2–3 min for each vessel. Aliasing was accommodated for using the following velocity-encoding values: AAo, 150–200 cm/second; SVC and IVC, 75–100 cm/second; and LPA, 50–100 cm/second. The AAo, SVC, and IVC images were acquired in axial cine stack and the LPA in vertical long axis, all orthogonal to flow.

Flow Profile Analysis

Flow profile characteristics were assessed as previously described and graphically illustrated [12, 13]. Briefly, flow waveforms were acquired by precise parallel segmentation of 2D phase contrast image series (25, 30, or 40 images/series) in Circle CVi42 (Calgary, Canada). Flow data were captured for each patient at the SVC, IVC, LPA, and AAo and imported into MATLAB (Natick, MA). Each waveform was normalized by dividing each flow point by patient body surface area to minimize size effect on the raw data. Data were interpolated using cubic spline interpolation to 40 points, guaranteeing size-matched array lengths for further analyses; as a result, 2 waveform sets were upsampled from 25 points and 91 waveform sets were upsampled from 30 points. A comprehensive error analysis examined changes in frequency content in these upsampled waveforms and concluded that, on average, such upsampling induces negligible impacts (all < 0.5%, scaled by spectral power) on the resulting waveforms. A single point represents a measurement at a time point in the cardiac cycle.

Single vessel data matrices for each location were created consisting of [rows, number of patient samples] × [columns, 40 flow data points]. Coupled vessel data matrices were created as described previously by our group [12], where the AAo flow data lay in columns 1–40 and the venous flow data (SVC, IVC, or LPA) are positioned in columns 41–80, resulting in matrices of [rows, number of patient samples] × [columns, 80 flow data points]. A graphical depiction of these matrices and how this analysis proceeds is additionally provided there [12]. Seven data matrices were created overall, with the size of each data matrix differing because, in addition to the inherent differences in the single and coupled matrices, flow data were not available or unusable (due to artifact from stent placement, aliasing, lack of patient compliance, or any other limiting factor). The size of each data matrix varied and is as follows: single-vessel matrices, SVC (124 × 40), IVC (132 × 40), LPA (125 × 40), and AAo (132 × 40) and coupled data matrices, AAo SVC (124 × 80), AAo IVC (132 × 80), and AAo LPA (132 × 80). All matrices were normalized along rows using a mean and standard deviation from all data or the vessel-specific columns, where AAo is normalized using columns 1–40 and the venous (SVC, IVC, or LPA) using columns 41–80.

Principal component analysis (PCA) is a data decomposition technique that returns a matrix of uncorrelated coefficients (PCs) that describe the variability of the data analyzed by identifying axes of greatest variance [23]. PC1 represents the most variable component in the original dataset, while the subsequent PCs represent components of decreasing variance. The first two principal components from each matrix were investigated based on our previous findings [12, 13]. The mechanistic impact of each PC on the original flow data was interpreted visually by plotting the site-specific, normalized average flow and the average flow plus and minus the PC vector of interest. Roughly, the first 20 points approximate systole, while the last 20 estimate diastole. Our group has previously applied this machine learning technique to flow data in the Fontan circuit as a novel vascular stiffness analysis with promising results [12].

Statistics

Statistical analyses were performed in GraphPad Prism and began by determining the univariate Cox hazard proportion (HR) or ratio, for each PC and cMRI hemodynamic measures. Single and coupled PCs, in addition to cMRI measures, that had statistically significant hazard ratios and the greatest concordance indices (c-statistic) were used to create receiver operating characteristic (ROC) curves. The c-statistics represents the concordance probability or the probability that a patient’s risk decreases given longer survival time, although in highly censored data the parameter is known to overestimate survival and will be considered in combination with other measures [25]. The ROC curves were used to calculate the area under the curve (AUC) and Youden’s index, defined as $\left[\mathrm{s}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{i}\mathrm{t}\mathrm{y}\right(\mathrm{x})–\mathrm{s}\mathrm{p}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{f}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{t}\mathrm{y}(\mathrm{x}\left)\right]-1$. The maximum Youden’s index and corresponding cut-off value, sensitivity, and specificity were recorded and defined grouping for further Kaplan–Meier survival analyses. The Mantel–cox log-rank test was used to determine if a significant difference existed between cut-off-defined Kaplan–Meier curves. Univariate parameters with the greatest c-statistic were used; up to two parameters at a time based on the total composite outcomes, for multivariate Cox hazard regression analysis. Akaike’s information criteria (AIC) was used to compare univariate and multivariate predictive models. AIC is an estimation method that explains deviance from a best fit predictive model and thus a lower AIC has greater predictive value and a difference of 2 between models is considered significant [24].

Results

140 SVD patients were included in this study, and the most common type of SVD is HLHS (n = 44), followed by tricuspid atresia (TA, n = 31), double outlet right ventricle (DORV, n = 9), double inlet left ventricle (DILV, n = 7), or a combination of the previously mentioned diagnoses with additional cardiac abnormalities (n = 50). The median age of patients at the time of Fontan completion was 3 years (IQR 2–4), and the majority of patients received an extracardiac Fontan (n = 114), followed by the lateral tunnel Fontan (n = 18) or a unique anatomic reconstruction, labeled as other (n = 8). The median age at the time of the cMRI acquisition was 14 years (IQR 9–16) (Table 1). Body surface area (BSA, 1.35 ± 0.42 m2) and study time (time to event or time to follow-up, 779 days, IQR 460–1297) were also noted.

Hemodynamic measures gathered from cMRI included ejection fraction (EF, 48 ± 8%), aortic valve regurgitation (1.52 ± 2.66%), end-diastolic volume index (EDVi, 94.4 ± 30.4), end-systolic volume index (ESVi, 50.5 ± 23.8), cardiac index (CI, 3.52 ± 1.37 L/min*m²), differential LPA flow (45 ± 11.5%), aortopulmonary flow as a percent of the net aortic flow defined by ${\mathrm{Q}}_{\text{aorta}}-\left({\mathrm{Q}}_{\mathrm{S}\mathrm{V}\mathrm{C}}+{\mathrm{Q}}_{\mathrm{I}\mathrm{V}\mathrm{C}}\right)$ (AAo col, 18; 10–30.5%), and aortopulmonary flow as a percent of the net pulmonary flow defined by $\left({\mathrm{Q}}_{\mathrm{R}\mathrm{P}\mathrm{V}}-{\mathrm{Q}}_{\mathrm{R}\mathrm{P}\mathrm{A}}\right)+\left({\mathrm{Q}}_{\mathrm{L}\mathrm{P}\mathrm{V}}\right.-{\mathrm{Q}}_{\mathrm{L}\mathrm{P}\mathrm{A}})$ (PA col, 18; 11.5–29.5%), as seen in Table 1.

Principal Component Analysis and Statistics

PCA was performed on each matrix, both single and coupled, and the percent of variance in the original data is described by the PC % explained in Table 2. The PC that accounted for the most variance in an original dataset is LPA PC1 and in all cases, the first two PCs accounted for more than 60% of the original waveform data variance. Cox hazard ratios were calculated for PCs and cMRI hemodynamic measures as univariate predictors of composite outcomes. Table 2 displays the hazard ratios, the corresponding 95% CIs, the associated p-values, c-statistic, and AIC. A parameter was considered statistically significant if the CI did not cross 1, which represents no hazard, or the null hypothesis and AIC was used to compare predictive models. AIC is displayed in italics if the model was missing > 15 samples, or patients’ measurements, and therefore are not comparable due to differences in data size and were not considered for multivariate analysis.

Table 2.

The % of each PC explained from both single-site & coupled PCA are displayed, in addition Cox hazard ratios for all PCs and cMRI hemodynamic measures, their p-values, and the AIC and c-statistics for model comparison and risk classification

	Variable	% PC Explained	HR Estimate	95% CI	p value	c-statistic	AIC
Hemodynamics	Age at Fontan	-	1.64	1.02 to 2.58	0.0356	0.684	99.5
	EF	-	0.880	0.830 to 0.935	< 0.0001	0.734	130
	EDVi	-	1.02	1.01 to 1.03	0.0007	0.673	139
	ESVi	-	1.02	1.01 to 1.03	< 0.0001	0.709	135
	CI	-	1.24	0.858 to 1.67	0.194	0.641	136
	LPA flow %	-	0.979	0.941 to 1.02	0.325	0.594	124
	Ao col %	-	1.01	0.981 to 1.03	0.417	0.650	85.9
	PA col %	-	1.04	0.999 to 1.08	0.0520	0.672	61.2
	Ao regurg fraction %	-	1.02	0.708 to 1.22	0.904	0.531	78.1
Single	AAo PC1	54.6	0.811	0.654 to 1.02	0.0654	0.666	135
	AAo PC2	31.5	1.35	1.01 to 1.84	0.0458	0.599	134
	SVC PC1	53.3	0.926	0.803 to 1.04	0.251	0.562	125
	SVC PC2	29.7	0.715	0.476 to 1.05	0.139	0.605	124
	IVC PC1	51.1	1.06	0.947 to 1.17	0.320	0.569	137
	IVC PC2	24.7	0.844	0.735 to 0.992	0.0265	0.691	134
	LPA PC1	56.4	1.05	0.949 to 1.18	0.343	0.592	133
	LPA PC2	27.4	1.00	0.908 to 1.25	0.990	0.392	134
Coupled	AAo SVC PC1	44.5	0.939	0.817 to 1.05	0.334	0.542	125
	AAo SVC PC2	23.7	0.673	0.448 to 1.03	0.0792	0.627	123
	AAo IVC PC1	42.1	1.07	0.959 to 1.18	0.228	0.582	137
	AAo IVC PC2	23.2	0.851	0.743 to 0.989	0.027	0.690	134
	AAo LPA PC1	45.9	1.06	0.951 to 1.18	0.330	0.592	133
	AAo LPA PC2	23.1	0.978	0.898 to 1.16	0.719	0.645	134

AIC in italics represents > 15 missing samples. Significant p-values and the greatest c-indices are bold

The lowest AIC, excluding those with incomparable sizes, and therefore the greatest univariate outcomes predictor is EF (Table 2, AIC = 130, c = 0.734). Age at the time of Fontan surgery (HR 1.64, p = 0.0356, c = 0.684), EF (HR 0.880, p < 0.0001), EDVi (HR 1.02,p = 0.0007, c = 0.673), ESVi (HR 1.02, p < 0.0001, c = 0.709), single AAo PC2 (HR 1.35, p = 0.0458, c = 0.599), single IVC PC2 (HR 0.851, p = 0.0265, c = 0.691), and coupled AAo IVC PC2 (HR 0.851, p = 0.027, c = 0.690) were all significant, with ventricular volume measures tending to be more specific, and age and IVC measures being more sensitive (Fig. 2 table). In a sub-cohort (due to missing samples), the greatest HR was found from age at the time of the Fontan operation (HR = 1.64), which suggest patients that are older at the time of the Fontan operation are at a higher risk, and the size effect of all parameters can be seen in Fig. 1.

Fig. 2.

ROC curves generated for parameters predicting composite outcomes in the Fontan population. Curves shown had the greatest c-statistics and the calculated AUC, p value, the optimum sensitivity and specificity, Youden’s index, and the cut-off value were established for further survival analyses and interpretation. Red-dashed line represents the null hypothesis

Fig. 1.

Forest plot displaying the hazard ratio of each hemodynamic measure gathered from cMRI and of the PCs from waveform analysis. Bars represent 95% confidence intervals

ROC curves were generated for parameters that returned the greatest Harrell’s concordance [26] or c-statistic, which determines a parameter’s ability to classify a patient at risk of a composite outcome (Fig. 2). The ventricular function measures that EF (p = 0.0056, AUC = 0.713) and ESVi (p = 0.0241, AUC = 0.674) were significant specific classifiers, with EDVi (p = 0.0925, AUC = 0.630) approaching significance and showing greatest specificity (0.973). In comparison, AAo IVC PC2 (p = 0.023, AUC 0.681) and IVC PC2 (p = 0.029, AUC = 0.674) were significant sensitive classifiers, with IVC PC2 having greatest overall sensitivity (0.933). Age at the time of the Fontan procedure had the second greatest AUC at 0.705 (p = 0.022) and the second highest sensitivity of 0.917.

Prior to Kaplan–Meier analyses, the proportionality assumption was confirmed using Schoenfeld residuals vs time plots (Supplementary Fig. 1), and no trends were observed over time. Univariate Kaplan–Meier curves were generated for each parameter that displayed a significant ROC curve and Youden’s index (Fig. 2) was used to define groups for univariate Kaplan–Meier analysis. Kaplan–Meier curves can be seen in Fig. 3 in addition to the p value from the Mantel Cox log-rank test, which determines if curves are significantly different. Waveform patterns with significantly different Kaplan–Meier curves were graphed for physiologic interpretation as the average flow, and the average flow plus or minus 2(PC of interest) in Fig. 4. The coupled AAo IVC PC2 waveform showed significant differences in the IVC waveform primarily and both the coupled and single IVC waveforms described aberrations in systolic flow rate and acceleration (Fig. 4). Increased IVC flow in systole, accompanied by early systolic acceleration are less likely to experience a composite outcome compared to patients that have late systolic acceleration in the IVC (Figs. 3, 4).

Fig. 3.

Kaplan–Meier curves for univariate predictors of composite outcomes and log-rank Mantel Cox p-value. Groups were created using the Youden’s index-defined cut-off values

Fig. 4.

The average AAo and IVC flow and the coupled waveform aberrations described by coupled AAo IVC PC2. The red-dashed line represents y = 0

Parameters that were strong univariate predictors, as determined by the cox hazard c-statistic, were considered input variables for a multivariate Cox regression hazard models. The results can be seen in Table 3 and display promising AICs for models, including the covariates EF, ESVi, IVC PC2, and IVC AAo PC2, or models A-D. The most predictive and significant multivariate Cox HR model includes single IVC PC2 and EF (p values, 0.0970 and < 0.0001) with an AIC of 120 (Table 3).

Table 3.

Results from multivariate Cox hazard regression with corresponding statistical values

	Covariates	HR	CI	p-value	AIC
A	IVC AAO PC2	0.871	0.738 to 1.04	0.115	120
	EF	0.880	0.827 to 0.937	< 0.0001
B	IVC PC2	0.857	0.724 to 1.04	0.0970	120
	EF	0.879	0.828 to 0.936	< 0.0001
C	IVC AAO PC2	0.857	0.740 to 1.00	0.0495	124
	ESVi	1.02	1.01 to 1.034	< 0.0001
D	IVC PC2	0.847	0.729 to 1.01	0.0448	124
	ESVi	1.02	1.01 to 1.03	< 0.0001
E	IVC AAO PC2	0.864	0.751 to 1.01	0.0543	126
	EDVi	1.02	1.01 to 1.03	0.0001
F	IVC PC2	0.852	0.737 to 1.008	0.0458	126
	EDVi	1.02	1.009 to 1.030	0.0001
G	EF	0.891	0.801 to 0.986	0.0290	132
	ESVi	1.00	0.982 to 1.02	0.772

Discussion

In this study, we explored univariate and multivariate associations between composite outcomes in a relatively large, singe-center Fontan cohort, evaluating standard and ML-derived parameters. Unsurprisingly, EF, as a reliable and well-studied measure of systolic function, was the most significant univariate predictor of hard outcomes. Additionally, our novel waveform measures that describe variations in IVC flow, alone and coupled with the AAo, were significant univariate predictors and approached or remained significant in multivariate models.

EF is an established clinical measure used to assess heart health in the Fontan population, although a subset of Fontan patients may experience preserved EF with circulatory failure [27, 28]. Our findings reinforce that while EF has significant association (AUC = 0.713, p = 0.0056, AIC = 130) with composite outcomes, it is more specific (0.793) than sensitive (0.625). Thus, while EF provides certainty that patients with declining ventricular function are at higher risk, it of course cannot identify adverse events with preserved EF, which supports our previous suggestions of including peripheral measures of Fontan function in a predictive model [12]. Thus, we explore the significance of single and coupled distal waveform patterns in prognosis of Fontan patients.

Univariate Cox HR analyses of the novel waveform measures revealed IVC PC2 as the strongest waveform-based univariate predictor (HR = 0.844, p = 0.0265) and the most sensitive (0.933) of all parameters. PC2 describes changes in systolic IVC flow, and our study found that patients with early acceleration and increased flow in the IVC during systole are less likely to experience a composite outcome. IVC PC2 accounted for nearly 25% of the variance in the IVC waveform data, suggesting a meaningful discrepancy in systolic IVC flow in our single-center cohort (N = 140). Our findings support our previous work, although in a smaller cohort (n = 95), that suggested that the waveform pattern which significantly correlated to patient health was AAo/IVC PC2 and that it described aberrations in systolic IVC flow [12]. Additionally, the waveform patterns in Fig. 4 (and the PC eigenvector, not shown) demonstrate that most of the variance in this parameter resides in venous flow, as inclusion of the AAo waveform in coupled PCA had little to no effect on IVC flow patterns. This is preliminary evidence that ventricular and venous function are in fact independent of one another and that TCPC single-site hemodynamics focused on the central venous circulation should be the primary focus of this work moving forward. As the most sensitive parameter, we also postulate that IVC PC2 could be most useful to rule out the occurrence of an adverse event in the Fontan population.

The multivariate models explored included IVC PC2, AAo IVC PC2, EF, EDVi, and ESVi. Our most predictive models, with an AIC of 120, included measures of central venous function and EF. The results indicate stronger predictive value than the univariate models of both AAo IVC PC2 (AIC = 134), IVC PC2 (AIC = 134), and EF (AIC = 130). Our findings provide initial support our primary hypothesis that inclusion of quantified venous measures may have a strong influence on prognosis and that, when considered in conjunction with the gold standard for systolic function, are stronger predictors of end-organ dysfunction than EF alone. Further, given that the best combination of sensitivity and specificity typically yields the best univariate classifier, we speculated, then found, that bivariate combinations of a variable with high sensitivity (such as IVC PC2) with another with high specificity (such as EF) yielded best predictive value. Undoubtedly, more work is required in larger multicenter cohorts to fully understand and confirm the role of central venous hemodynamics in the failed Fontan circulation, but our work provides a promising foundation for further research.

Several limitations exist within this study. First, our cohort size (N = 140) and number of events (16) are fairly small, indicating we do not have high power to detect differences in our parameters; however, our significant results do suggest large effect sizes for the differences found. We are continuously collecting more cMRI data to address such issues and plan to explore these new parameters in a multicenter study. Additionally, parameters included in this study were limited, and inclusion of pulmonary, blood, and additional cardiac measures from echocardiogram and catheterization should be explored to determine which parameters contribute to global Fontan decline. The exclusion requirements for this study also may have limited the scope of our study, as patients with LPA stents, image captures of a double outflow tract, anatomic abnormalities, such as double SVCs, combined ventricle, or a patient that already received a heart transplant at the time of cMRI. This study may also have greater clinical implications given further sub-group analyses and risk stratification, although this is an area of intended work. The statistical approach of PCA also comes with inherent drawbacks, including linear data reduction, which does not consider non-linear reduction techniques. As data reduction suggests, compressing the original data for usability is also accompanied by a loss of, ideally insignificant, original data. Further work must be done to understand how influential clinical parameters, like heart rate and ventricular dominance, contribute to the PC measures defined in this study.

To conclude, this study determined the role of cMRI and novel venous hemodynamic measures, both single and coupled, in composite outcome development in the Fontan population. EF was found to be the most significant predictor of outcomes, although EF was significantly more predictive (as determined by AIC) in a multivariate model combined with measures of central venous function. This is the first study, to the best of our knowledge that attempts to incorporate venous stiffness measures in a prognostic model of Fontan failure. Further work is necessary to verify the model described in this work, although this study provides a promising avenue of further research in Fontan prognostication.

Abbreviations

SVD

Single-ventricle disease

PLE

Protein-losing enteropathy

PB

Plastic bronchitis

FALD

Fontan-associated liver disease

PC-MRI

Phase contrast magnetic resonance imaging

VVCR

Ventricular vascular coupling ration

VO₂

Rate of oxygen consumption

AAo

Ascending aorta

SVC

Superior vena cava

IVC

Inferior vena cava

LPA

Left pulmonary artery

cMRI

Cardiac MRI

TCPC

Total cavopulmonary connection

PCA

Principal component analysis

PCs

Principal components

EF

Ejection fraction

EDVi

End-diastolic volume index

ESVi

End-systolic volume index

CI

Cardiac index

BNP

B-type natriuretic peptide

GGT

Gamma-glutamyl transferase

AST

Aspartate aminotransferase

SaO₂

Arterial oxygen saturation

FEV1

Forced expiratory volume in one second

mSVCP

Mean SVC pressure

mPAP

Mean pulmonary artery pressure

HLHS

Hypoplastic left heart syndrome

TA

Tricuspid atresia

DORV

Double outlet right ventricle

DILV

Double inlet left ventricle

HRHS

Hypoplastic right heart syndrome

TCPC

Total cavopulmonary circuit

FVC

Forced vital capacity

RV

Residual volume

TLC

Total lung capacity

Alk phos

Alkaline phosphatase

BUN

Blood urea nitrogen