Ejection fraction quantification from ungated chest CT by AI

Figures

• 

  Extended Figure 1.. Discrimination performance for different LVEF thresholds Discriminative power of left ventricular ejection fraction (LVEF) derived from ungated chest CT was assessed using the area under the operating characteristic curves (AUCs) in the pooled external validation. AUCs were presented with 95% confidence intervals (CIs).

• 

  Extended Figure 2.. AI LVEF from ungated CT predicts clinical outcomes in cohort 3 Kaplan-Meier curves of left ventricular ejection fraction (LVEF) < 40% (red) against LVEF ≥ 40% (blue) according to the risk of cardiovascular death (a) and all-cause death (b) in cohort 3 are shown. Unadjusted hazard ratios (HRs) are presented with 95% confidence intervals. P-values from the log-rank test of survival curves are shown above the curves.

• 

  Extended Figure 3.. Feature ranking and influence Features used to predict left ventricular ejection fraction (LVEF) are shown in the order of mean absolute SHAP values. Each dot on each feature row corresponds to an instance that provides the matched explanation, with its color indicating the feature value (blue – low, red – high) and its x position indicating the impact on LVEF prediction (positive – increase, negative – decrease). Larger LV (higher values of volume, represented in red color) showed negative impacts on the prediction, which reduced the derived LVEF. The systolic pressure was another important feature. Lower systolic pressure (represented in blue color) was associated with a lower predicted LVEF. Lower predicted LVEF values were also associated with several other factors, including enlarged LV myocardium, raised heart rate, higher median blood density of the right ventricle (RV) and LV myocardium, lower density of EAT and LV, smaller heart size, higher diastolic pressure, and a higher CAC score. To address multicollinearity, the Modified Index Position (MIP) method iteratively removes the top-performing feature and recalculates SHAP values in a new model with fewer features.…

• 

  Figure 1.. AI LVEF workflow Overview: Cohort 1 included a total of 25,852 patients from 11 geographically separate medical centers (3 countries). Left ventricular ejection fraction (LVEF) was derived from end-diastolic volume (EDV) and end-systolic volume (ESV) in gated cardiac PET and used as the reference for CT-wise artificial intelligence (AI)-enabled estimation that integrated deep-learning CT structures and basic clinical variables. We conducted leave-one-site-out external validation and created 11 models. For each model, one site was reserved for external validation, and the remaining 10 sites were used for model training. This process was repeated 11 times across the entire registry-resulting in 11 independent models and 11 independent external evaluations. We evaluated AI LVEF prediction using both continuous and binary metrics, as well as survival analysis. The model developed using cohort 1 was applied to cohort 2 for comparison with echocardiographic LVEF (n = 128) and to cohort 3 for risk stratification in the National Lung Screening Trial (NLST, n = 24,054).

• 

  Figure 2.. Performance metrics for continuous ejection fraction prediction Continuous left ventricular ejection fraction (LVEF) derived from ungated chest CT was compared with PET LVEF by assessing (a) the bias, (b) the mean absolute error (MAE), (c) the Pearson correlation coefficient, (d) the intraclass correlation coefficient (ICC), and (e) the coefficient of variation in the pooled and site-level external validation with 95% confidence intervals.

• 

  Figure 3.. External classification performance for AI CT ejection fraction < 40% Discriminative power of left ventricular ejection fraction (LVEF) derived from ungated chest CT was assessed by the area under the operating characteristic curves (AUCs) in the pooled (red solid line) and site-level (dotted lines) external validation (a). The prevalences of reduced LVEF according to the reference PET LVEF were presented with corresponding AUCs (b). The positive and negative predictive values (PPV, c and NPV, d) were shown for the pooled and site-level external validation.

• 

  Figure 4.. AI LVEF from ungated CT predicts clinical outcomes in cohort 1 Kaplan-Meier curves of left ventricular ejection fraction (LVEF) < 40% (red) against LVEF ≥ 40% (blue) according to the risk of heart failure-related hospitalization, cardiovascular death, and all-cause death in cohort 1 are shown. Unadjusted hazard ratios (HRs) are presented with 95% confidence intervals. P-values from the log-rank test of survival curves are shown above the curves.

• 

  Figure 5.. Cases for illustration Case (a) shows a male patient in his 60s from site 8 with an increased left ventricular (LV) volume (106 mL/m2) indexed to body surface area. The patient had a reduced left ventricular ejection fraction (LVEF) of 23% based on PET. The CT prediction was 23%. The factors that increased the potential predicted value (red arrows) included heart volume (649 mL/m2), epicardial adipose tissue (EAT) density (−62 Hounsfield Units [HU]). The LVEF prediction was shifted from the base value of 62% to 23%, primarily driven by the LV volume, LV myocardial volume (117 mL/m2), heart rate (88 bpm), systolic pressure (109 mmHg), left lung (LL) emphysema volume (0), right lung (RL) consolidation volume (29 mL/m2), LV median blood density (35 HU), and ascending aorta (AscAorta) max diameter (15 mm/m2) (blue arrows). Case (b) shows a female patient in her 70s from site 1 with an increased LV volume (81 mL/m2). The patient had a reduced LVEF of 32% on PET and 32% on our LVEF CT tool. The predicted value was increased by EAT density (−54 HU), LV density (46 HU), heart volume (576 mL/m2), and no coronary artery calcium (CAC), and was decreased by LV volume, heart rate (99 bpm),…