The participant characteristics of the 80 non-ischemic training cohort, 20 non-ischemic validation cohort, and the 10 ischemic validation cohort are summarized in Table 1. In brief, the non-ischemic groups were all female cohorts with HIV positive rates of 73.8% (training cohort) and 80% (validation cohort) while the ischemic group was an all male cohort with significantly reduced EF.
ECV-guided LGE analysis in the non-ischemic training cohort
1) ECV criteria development
The averaged ECV value corresponding to the high SI ROI was 34.0 ± 6.6 (%), which was higher than the global ECV value at the mid-LV slice of 27.1 ± 2.8 (%) (P<0.01). In the visual assessment, 50 out of the 80 cases were judged as a scar. The ROC curve presented the AUC of 0.94 (95% CI = 0.90 to 0.99). The single ECV cutoff identified from the ROC curve to differentiate scar from non-scar was 31.5%. Forty-nine cases (61.2%) fit into the ECV≥31.5% scar category while 31 cases (38.8%) fit into the ECV <31.5 non-scar category. Overall, ECV 31.5% cutoff value achieved sensitivity of 90%, specificity of 86.7%, PPV of 91.8%, and NPV of 83.9%. In the nT1 investigation, the averaged nT1 value corresponding to the high SI ROI was 1339 ± 94 (ms). The ROC curve of nT1 against visual scar/non-scar presented the AUC of 0.78 (95% CI = 0.69 to 0.88). The nT1cutoff of 1317 (ms) achieved sensitivity of 68%, specificity of 70%, PPV of 79.1%, and NPV of 56.8%. Based on these results, ECV cutoff was selected to guide the LGE analysis (Figure 4A, B).
2) The optimization of the n-SD threshold in the non-ischemic training cohort
The optimal threshold (n-SD) on average was 7.3 ± 2.9 (SD), which showed a wide range from 3SD minimum to 18SD maximum. The threshold value (n-SD) was not associated with the global LGE scar amount (%) assessed in ECV-guided LGE analysis (β = -0.01, p=0.92) and furthermore, there was no difference between LGE positive (n=61) and negative (n=19) cases. (LGE positive vs. negative = 7.1 ± 2.8 SD vs. 7.7 ± 3.4 SD, P=0.51). Among the 31 cases in which high-SI ROI was judged as non-scar and therefore no scar in the mid-LV slice, 11 of them presented scar in other slices after the propagation of the selected threshold. These 11 cases showed typically a small global scar percentage of 0.46 [0.15 – 1.1] %.
3) Inter-method agreement in the non-ischemic training cohort
LGE scar was detected in all the cases on manual analysis while in 76.3% in ECV-guided LGE analysis (P<0.01). The quantitative global scar amount (%) was significantly larger in manual analysis than that of the ECV-guided LGE analysis (4.5 [3.2 – 6.4] vs. 0.92 [0.1 – 2.1], p<0.01). The inter-method agreement of the global scar (%) between these two methods was fair (CCC= 0.48, P<0.01) with the mean ± 95% LoA by the Bland-Altman plot of -3.2 ± 3.9 (%) (Table 2, Figure 5).
4) Reproducibility analysis in the non-ischemic training cohort
Forty cases were randomly selected for reproducibility analysis. Both inter- and intra-observer reproducibility presented better results in the ECV-guided LGE analysis than the manual analysis at the global level and at the segmental level. The intra-observer reproducibility of global scar (%) by ECV-guided LGE analysis was excellent (CCC=0.94, P<0.01), and was better than that by the manual scar analysis (CCC=0.78, p<0.01). The inter-observer reproducibility of global scar (%) by ECV-guided LGE analysis was excellent (CCC=0.86, p<0.01), and better than that by manual scar analysis (CCC=0.73, p<0.01). Bland-Altman analysis revealed tighter limits of agreement and smaller bias in ECV-guided LGE analysis, for both inter- and intra-observer assessments (Table 3 and Supplement 3). In per-segmental scar (%) analysis, a similar trend of improved inter- and intra-observer reproducibility was observed for the ECV-guided LGE analysis as compared to the manual analysis (Supplement 4).
ECV-guided LGE analysis validation in the non-ischemic validation cohort
The non-ischemic validation cohort presented similar trends of results with the non-ischemic training cohort. The ECV value corresponding to the high SI ROI and the global scar (%) were comparable to those of the non-ischemic trainingcohort [ECV value of the high SI ROI: 32.6 (27.8 – 35.4) vs. 34.0 ± 6.6 (%), P=0.36, the global scar (%) : 0.2 (0 – 1.6) vs. 0.92 (0.1 – 2.1) (%), P=0.14]. The ECV cutoff of 31.5% achieved excellent scar/non-scar differentiation at the high SI ROI in the validation cohort (sensitivity 100%, specificity 81.8%, PPV 81.8%, and NPV 100%) while the nT1 cutoff of 1317ms presented a fair differentiation in the validation cohort (sensitivity 33.3%, specificity 90.9%, PPV 75%, and NPV 62.5%) (Figure 4C, D).The optimal threshold (n-SD) varied from 3SD to 9SD but was lower than that of the non-ischemic training cohort [6 (4 – 7) vs. 7.3 ± 2.9 (SD), P=0.02]. This n-SD threshold was associated with the scar amount (β= -0.53, p=0.02). LGE scar detection rate was higher on manual analysis than on ECV-guided LGE analysis (90% vs. 50%, P < 0.01). The global scar amount (%) was significantly larger on manual analysis than on the ECV-guided LGE analysis (2.5 [1.2 – 3.7] vs. 0.2 [0 – 1.6], P<0.01). The inter-method agreement of the global scar (%) was fair (CCC= 0.59, P<0.01) with the mean ± 95% LoA by the Bland-Altman plot of -1.8 ± 2.5 (%). All of these trends were similar to those observed in the training cohort (Table 2, Figure 5).
ECV-guided LGE analysis validation in the ischemic cohort
The ECV value corresponding to the high SI ROI was 52.2 (49.1 – 54.3) (%) in the ischemic cohort. The optimal threshold (n-SD) was 3.5 (3 – 5) (SD). The inter-method agreement of the global scar (%) by the ECV-guided LGE analysis and the FWHM with manual correction was excellent (25.0 [17.3 – 33.9] vs. 24.6 [18.2 – 27.7], P=0.23, CCC=0.82, P<0.01, the mean ± 95% LoA = 1.8 ± 7.8 (%)) (Table 2, Figure 5). All the intra- and inter-observer reproducibility of the global scar (%) (Table 3 and Supplement 3) were better or comparable in ECV-guided LGE analysis than the conventional analysis. In the per-segmental scar (%) analysis, similar trends were observed in the inter-method agreement and the reproducibility (Supplement 4).