Methods

We previously recruited 120 healthy volunteers (10 men and 10 women each in 6 deciles of age, 20 to 80 years old) in whom we characterized the LV trabeculations. Subsequently, we selected a reference LVNC group (n = 30) as cases in which either echocardiography or CMR diagnosed LVNC by the criteria by Jenni et al and Petersen et al, respectively, and further agreed between 2 independent expert readers after consideration of clinical presentation (abnormal electrocardiogram [68%], arrhythmias [23%], systemic embolization [7%], heart failure [70%], and family history of another cardiomyopathy [20%]). These were empirically divided in 2 subgroups: those with “normal” LV ejection fraction (EF) (>55%) and those with abnormal LV EF (<55%), as determined by CMR. To prospectively validate the new model, we also recruited 73 patients from the regional cardiomyopathy service who were either suspected or confirmed with a diagnosis of cardiomyopathy (dilated, n = 42; noncompacted, n = 31). The study was approved by the Royal Brompton and Harefield NHS Foundation Trust Local Ethics Committee.

All subjects underwent CMR on 1.5-T Siemens Avanto or Sonata scanners (Erlangen, Germany). Three long axes (2, 3-, and 4-chamber views) and a complete short-axis stack (slice thickness/gap = 7/3 mm) of balanced steady state free precession cines were acquired (repetition time/echo time = 3.2/1.6 seconds, flip angle 60°, in-plane pixel size 2.1 × 1.3 mm).

Cine images were analyzed with CMRTools (Cardiovascular Imaging Solutions, London, United Kingdom) for calculation of LV volumes and EF. Manual measurements of the “apparent” end-diastolic (ED) and end-systolic (ES) thicknesses of the T and C layers in 16 segments, excluding the true apex, were made as previously described for each subject ( Supplementary Table 1 ). T/C ratios were computed at ED and ES. For the prospective validation cohort, measurements of T/C myocardium were also performed in long axis, as proposed by Petersen et al.

Statistical analysis was performed in SPSS 19 and SAS v9.3 (IBM and SAS, respectively). Descriptive statistics are expressed as mean ± SEM for normally distributed data and as median (interquartile range) for skewed data. T/C ratios across different groups were compared using the Mann-Whitney test. Significance was reached at p <0.05. Initially, the ability to predict LVNC probability with each individual segment’s T/C ratio was carried out. Separate logistic regression models were fitted with the T/C ratio for each segment as a covariate and LVNC diagnosis as the dependent variable. A forward selection process was used to build a multivariate logistic regression model with the T/C ratios as predictors. Entry into the model was set at p ≤0.05. The process stopped when there were no more significant predictive segments or when the maximum number of predictors was reached in relation to the number of LVNC cases (using the “rule of thumb” that the number should not be more than the number of LVNC cases divided by 10, i.e., 3). By selecting only the statistically significant T/C segment ratios, the forward selection process minimizes collinearity between the segments in the final model. The predictive ability of each of the 2 final multivariate models was assessed using discrimination and calibration methods. Discrimination assesses how good the model is at correctly distinguishing between patients at low probability of LVNC and high probability of LVNC. It is characterized as the area under the receiver operating characteristic curve measured using the C statistic. Calibration examines how the predicted probabilities from the model compare with the actual observed probabilities of LVNC. Calibration was assessed using plots of the predicted probabilities against the observed outcomes in which a straight diagonal line from the bottom left to the top right signifies perfect calibration. Measurements of performance (i.e., specificity, sensitivity, positive predictive value, and negative predictive value) were calculated from each of the 2 final models for a range of predicted cutoffs. The model that performed best (ED or ES) with regards to discrimination and calibration was internally validated and then converted into a simple diagnostic chart.

To test for overfitting of the model (the optimism of the model), we used a bootstrap resampling technique. Three hundred samples were randomly drawn with replacement from the original data set with the same sample size as the original data set (120 volunteers and 30 patients with LVNC). For each bootstrapped sample, the same forward selection process to derive the original model was applied. The C statistic was calculated for each of the 300 bootstrapped models and also for each of the 300 models applied to the original data set. The difference between the 2 C statistics for each sample was found and averaged over the 300 samples. This average difference indicated the optimism of the C statistic in the original model and is an estimate of internal validity.

To facilitate clinical decision making, the final model was converted to a simple diagnostic chart, using those segments found to be predictive of LVNC. Each cell in the chart is color coded as green, yellow, or red equating to increasing categories of LVNC probability (green <15%, yellow 15% to 50%, and red >50% probability of LVNC).

We examined agreement between the Petersen criterion (T/C >2.3 measured from the long axis) and the predicted probabilities from our new model and tested this in 73 prospectively recruited patients.

Results

As seen in Table 1 , the LVNC EF <55% group had significantly greater LV mass and volumes compared with normal volunteers, both before and after indexing for body surface area. The LVNC EF ≥55% group had comparable LV mass and LV ED volume to normal subjects, but the LV ES volume was significantly larger—both before and after indexing—and this resulted in a significantly reduced LV EF compared with normal volunteers.

All T/C ratios are listed in Table 2 . Individual receiver operating characteristic curves were constructed for each segment for the T/C ratio obtained at ES and ED; the analysis was computed to determine the diagnostic accuracy for all patients with LVNC as the positive group versus healthy volunteers. Supplementary Table 2 presents the parameter estimates (95% confidence interval [CI]) and C statistics for separate logistic regression models for each segment.

Stay updated, free articles. Join our Telegram channel

Join