Synthetic hematocrit derived from the longitudinal relaxation of blood can lead to clinically significant errors in measurement of extracellular volume fraction in pediatric and young adult patients

This single-center, retrospective, observational study was approved by the Vanderbilt University Institutional Review Board and completed between January 2013 and February 2017. Consent was obtained for all participants. Patients who underwent a CMR using T1 mapping with modified Look-Locker inversion (MOLLI) recovery sequence and who had a Hct drawn within 3 months of the date of CMR were eligible and all Hct values were determined in a central laboratory. Pediatric, young adult, and adult congenital patients were included. None of the patients in this study had general anesthesia and none of the patients received intravenous fluids aside from the small volume administered as part of the contrast injection. Pertinent clinical data, including underlying cardiac diagnosis, were collected.

CMR was performed on a 1.5 Tesla Siemens Avanto (Siemens Healthcare, Erlangen, Germany) with an 8 channel cardiac coil. Functional imaging was performed as previously described using balanced steady state free precession (bSSFP) imaging [19]. Breath-held MOLLI sequences were performed prior to and 15 min after contrast administration at the level of the base, mid-LV, and apex in the short axis plane [20, 21]. MOLLI sequences were motion-corrected, ECG-triggered images obtained in diastole with typical imaging parameters: non-selective inversion with a 35° flip angle, single shot SSFP imaging, initial inversion time of 120 ms with 80 ms increments, field of view 340 × 272 mm², matrix size 256 × 144, slice thickness 8 mm, voxel size 1.3 × 1.9 × 8.0 mm³, TR/TE 2.6 ms/1.1 ms, parallel imaging factor of two. The matrix size was decreased to 192 × 128 for heart rates >90 (approximate voxel size 1.8 × 2.1 × 8 mm³). Intravenous Gd-DTPA contrast (gadopentate dimeglumine or gadobutrol, Bayer Healthcare Pharmaceuticals, Wayne, NJ, USA) was administered through a peripheral intravenous line (PIV) at a dose of 0.2 mmol/kg. The pre-contrast MOLLI acquired 5 images after the first inversion with a 3 beat pause followed by 3 images after the second inversion, or 5(3)3 [22]. Early in the protocol, this was modified to the equivalent of a 3 s pause, or 5(3 s)3, to reduce bias from higher heart rates (3 beat pause for heart rate of 60, 4 beat pause for heart rate of 80, and so on). The post-contrast protocol was acquired at a 4(1)3(1)2. Motion correction as described by Xue, et al. was performed and a T1 map was generated on the scanner [23]. One of the investigators was present for all scans and reviewed the MOLLI sequences with the technologist for adequacy at the time of the scan. Any image felt to be inadequate due to poor breath holds or poor motion correction was repeated at the time of the scan.

One reader (FR) traced all contours in this study. Regions of interest (ROIs) were manually drawn on native and post-contrast T1 images in blood pool, septum, and free wall of mid-LV. ROIs of the myocardium were carefully contoured so as to only include the mid-myocardium, avoiding contamination with blood pool or epicardial fat. Given the significant variability in underlying cardiovascular disease that could include patients with myocardial ischemia, areas of LGE were not included in the ROIs. Synthetic Hct was calculated using the published model [18]: [Hct_MOLLI = (866.0 · [1/T1_blood]) - 0.1232]. A locally-derived linear regression model was also created (see Results for details). ECVs were calculated as described above at the mid-septum and mid-LV free wall using both measured and synthetic Hct values. Finally, to evaluate whether the synthetic Hct provides added value over simply assigning the same Hct to all patients (equivalent to the partition coefficient), the ECV was recalculated in all patients after assigning Hct values of both 40% and 45% (reported as static Hct). The Hct of 45% provided ECV values closest to the measured ECV, so those results are reported below.

ECV results were compared to our laboratories normal cut-off of 28.5%. This cut-off was conservatively set at 3 standard deviations above the mean ECV derived from a previously published cohort of healthy young adult patients [8]. For comparison, analysis was repeated for a normal cut-off of 27.0%, representing 2 standard deviations above the mean ECV.

Continuous data are presented as a mean and standard deviation. A locally derived synthetic ECV was created from the longitudinal relaxivity of blood (R1, or 1/T1). This model was created using linear regression modeling with R1 as the predictor variable and the measured hematocrit as the outcome. All three models (published, local, and static) were compared with measured values using multiple statistical methods. Correlations between synthetic and measured Hct and synthetic and measured ECVs were evaluated using linear regression modeling. Agreement between individual sets of synthetic and measured variables were represented graphically using Bland-Altman plots [24]. This allowed for assessment of intrinsic bias in each model. Intraclass correlation coefficients were used to determine how strongly synthetic and measured Hct and ECVs correlated overall and within subgroups. It was assumed that repeated CMRs from the same subject were independent. Analyses were performed with SPSS (v 24, International Business Machines, Inc. Chicago, Illinois, USA). Study data were collected and managed using REDCap (Research Electronic Data Capture) electronic data capture tools hosted at Vanderbilt [25].

A total of 163 CMRs from 114 children and young adults were included in the study. The characteristics for each participant at each CMR are shown in Table 1. Demographics are for individual patient characteristics at each CMR. The average age at the time of CMR was 16.4 ± 6.4 years. All but 17 patients had Hct drawn on the same day as the CMR. There were a disproportionate number of patients with muscular dystrophy in our cohort as this is the primary population for which our center reports ECV. This also led to a male predominance in our study population.

Using the equation for calculation of synthetic Hct published by Treibel and colleagues [18] for T1 MOLLI sequences, a mean synthetic Hct of 44.0% ± 3.7% was calculated (Table 2). The mean measured Hct for the cohort was 41.8% ± 3.4% and there was poor correlation between synthetic and measured Hcts by linear regression modeling (r² = 0.16, p < 0.001, Fig. 1a). Bland-Altman analysis demonstrated a − 2.2% bias (Fig. 1b) in the calculated synthetic Hct. The difference between measured and synthetic Hct ranged from −16.6% to 6.5%. Despite this poor correlation for the Hct, there was a strong correlation between measured and synthetic ECVs at the mid-free wall of the left ventricle (r² = 0.80, p < 0.001, Fig. 1c) and the mid-septum (r² = 0.72, p < 0.001, Additional file 1: Figure S1A). Bland-Altman analysis demonstrated a small but significant bias in the synthetic mid-free wall ECV (Figs. 1d) and mid-septum ECV of 1.2% (Additional file 1: Figure S1B). This bias was expected given the bias in the synthetic Hct and its relationship with ECV. There was significant variation in the difference between measured and synthetic ECV, ranging from −4.5% to 10.2% for the mid-septum and from −4.7% to 9.4% for the mid-free wall.

In order to eliminate the bias seen with the published model, a local linear regression equation was determined from the blood pool native T1 and the measured Hct. The following locally derived linear regression model was created: [HctMOLLI = (315.1 · [1/T1_blood]) – 0.213]. (Fig. 2). While the calculated local model synthetic Hct had a similar fit with the measured Hct (r² = 0.16, p < 0.001, Fig. 3a), the mean synthetic Hct was equivalent to the measured Hct at 41.8% ± 1.4% (p < 0.001, Table 2). The difference between measured and synthetic Hct ranged from −8.5% to 12.2%. Bland-Altman analysis of the local model demonstrated elimination of the bias seen in the published model, although there was a tendency for there to be greater discrepancy between measured Hct and synthetic Hct at the extremes (Fig. 3b). There was improved correlation by linear regression for both synthetic ECV at the mid-free wall (r² = 0.87, p < 0.001, Fig. 2c) and the mid-septum (r² = 0.83, p < 0.001, Additional file 2: Figure S2A). Minimal bias was observed on Bland-Altman analysis for either the mid-free wall ECV (Fig. 3d) or the mid-septum ECV (Additional file 2: Figure S2B). There was substantial variation in difference between measured and synthetic ECV, ranging from −8.4% to 4.3% for the mid-septum and from −12.6% to 15.8% for the mid-free wall.

Previous studies have used the partition coefficient, which excludes the (1-Hct) term in the ECV equation, to characterize myocardial extracellular matrix expansion [26, 27]. For this model, ECV calculations were performed using a static Hct value of 45% for all ECV calculations, which is statistically equivalent to using the partition coefficient for linear regression analyses. Similar regression fit was seen compared to the other synthetic ECV models at the mid-free wall (r² = 0.85, p < 0.001, Additional file 3: Figure S3A) and the mid-septum (r² = 0.80, p < 0.001). Bland-Altman analysis demonstrated a modestly larger 1.6% bias (Additional file 3: Figure S3B) in the calculated synthetic ECV at the mid-free wall. There was substantial variation in the difference between measured and synthetic ECV, ranging from −8.0% to 6.0% for the mid-septum and from −8.0% to 6.3% for the mid-free wall.

ICCs were calculated for Hct and the ECV at mid-septum and mid-free wall using the published model, our local model, and the static Hct equal to 45% as an estimation of the partition coefficient (Table 3). The published model showed modest correlation for Hct overall as well as in control, muscular dystrophy, and congenital subpopulations. The ICCs for mid-septum and mid-free wall ECV with this model were substantially better. For the local model, the ICC for Hct was worse than the published model, however the ICCs for ECV for both mid-septum and mid-free wall were slightly better than the published model. Use of the partition coefficient demonstrated similar results with strong ICCs.

Using our lab’s cutoff for abnormal ECV of 28.5%, we evaluated each of the models for accuracy compared to the measured ECV (Tables 4 and Additional file 5: Table S4). All three of the models demonstrated significant miscategorization of patients, with total fraction miscategorized ranging from 37% for the published model to 23% for the local model. While the local model had fewer errors in categorizing patients (27 and 37% fewer errors compared to the partition coefficient and published models, respectively), a larger proportion of these were false positives and the absolute number of clinically significant miscategorizations was still high. The rates of miscategorization were similar for patients with or without same day Hct (Tables 4 and Additional file 4: Table S4). Repeat analysis using 27.0% (2 standard deviations above the mean) as an abnormal ECV cut-off demonstrated a similar distribution of miscategorizations (Table 4). There were slightly fewer unique miscategorizations due to the fact that some of the borderline abnormal cases shifted solidly into the abnormal category with the decreased threshold.