Comparison of Artificial Intelligence-Driven Echocardiographic Assessment of Diastolic Function With the ASE Guidelines

Figures

• 

  Figure 1. Patient Selection Flowcharts of the 3 Cohorts Included in the Study ARIC = Atherosclerosis Risk In Communities; BioLINCC = Biologic Specimen and Data Repository Information Coordinating Center; HF = heart failure; HTX = heart transplantation; LHC = left heart catheterization; LVAD = left ventricular assist device; MR = mitral regurgitation; MS = mitral stenosis; MV = mitral valve; RHC = right heart catheterization.

• 

  Figure 2. Sankey Diagrams Depicting the Reclassification of Patients Between the Guidelines and the DL Model Given that the DL model was primarily trained for prognostication, the default probability threshold of 0.5 was used to classify patients as high risk or low risk in the ARIC cohort but not in the hemodynamic validation cohorts. Instead, in the first hemodynamic validation cohort, a threshold specifically optimized for identifying elevated LVFP was derived using Youden’s J statistic, yielding a cutoff of 0.99957. To preserve this operating point while avoiding cohort-specific optimization in the second hemodynamic validation cohort, percentile-based threshold transport was applied. Accordingly, the top 32.5% of DL-predicted probabilities were classified as having elevated LVFP, corresponding to a cohort-specific numeric cutoff of 0.98612. ASE = American Society of Echocardiography; DD = diastolic dysfunction; DL = deep learning; EACVI = European Association of Cardiovascular Imaging; LAP = left atrial pressure; LVFP = left ventricular filling pressure; other abbreviations as in Figure 1.

• 

  Figure 3. Prognostic Value of the Guidelines and the DL Model for Predicting the Composite Endpoint C-indices were calculated from univariable Cox regression models containing either the 2016 or 2025 guideline-based classification or the DL-predicted probability of DD. Models in the upper panel were evaluated in the entire ARIC cohort, whereas those in the lower panel were evaluated after excluding indeterminate groups. The 2 guidelines classified different patients as indeterminate; therefore, their performance was evaluated in overlapping but nonidentical subsets after excluding indeterminate cases: the 2016 ASE/EACVI guidelines were evaluated in 4,368/5,450 (80%) patients, whereas the 2025 ASE guidelines were evaluated in 5,326/5,450 (98%) patients. P values for C-index comparison were calculated using the CompareC R package27 in the upper panel, whereas they were calculated using bootstrap resampling in the lower panel. C-indices are reported with 95% CIs. Abbreviations as in Figures 1 and 2.

• 

  Figure 4. Kaplan-Meier Curves Showing the Event-Free Survival of Patient Subgroups in the ARIC Cohort The composite of HF hospitalization or all-cause death was used as the endpoint in this analysis. The P value from the log-rank test for the DL model was <0.001. P values from the pairwise log-rank tests of the guideline-based classifications are reported in Supplemental Tables 9 and 10. Abbreviations as in Figure2.

• 

  Figure 5. Receiver Operating Characteristic Curves Showing the Performance of the Guidelines and the DL Model in Detecting Elevated LVFP (A) ROC analysis was performed first for all patients in the given cohort. (B) Then, indeterminate groups were excluded, and ROC analyses were performed again in the remaining patients. Because the 2016 ASE/EACVI and 2025 ASE guidelines classified different sets of patients as indeterminate, their performance was assessed in overlapping but not identical patient subsets after excluding indeterminate cases: the 2016 guidelines were evaluated in 67/83 (81%) and 91/130 (70%) patients in the first and second hemodynamic validation cohorts, whereas the 2025 guidelines were evaluated in 81/83 (98%) and 118/130 (91%) patients, respectively. To enable the straightforward calculation of diagnostic performance metrics, both guideline-based classifications were dichotomized by merging the “normal diastolic function” and the “DD with normal LAP” groups (as both are supposed to have normal LAP). AUCs are reported with 95% CIs. Abbreviations as in Figure 2.

• 

  Central Illustration. Comparison of Artificial Intelligence-Driven Echocardiographic Assessment of Diastolic Function With the ASE Guidelines In this study, we compared the prognostic and diagnostic performance of the newly released 2025 ASE guidelines, the earlier 2016 ASE/EACVI guidelines, and a rigorously validated DL model for assessing LV diastolic function. Although the 2025 guidelines markedly reduced indeterminate classifications in all 3 cohorts compared with the 2016 guidelines, they demonstrated lower prognostic value and similar diagnostic performance for detecting elevated LVFP in the hemodynamic validation cohorts. The DL model yielded better prognostic discrimination than both guidelines in the ARIC cohort and showed superior performance in detecting elevated LVFP. Nevertheless, because ≥1 guideline-recommended parameters were unavailable in a substantial proportion of patients in the analyzed cohorts, these findings should be regarded as exploratory pending validation in data sets with complete guideline-recommended data. ARIC = Atherosclerosis Risk In Communities; ASE = American Society of Echocardiography; DD = diastolic dysfunction; DL = deep learning; EACVI = European Association…