Artificial intelligence-enabled ECG for left ventricular diastolic function and filling pressure

For this study, 274,710 patients who had an ECG and echocardiographic diastolic function assessment within 14 days were identified with no exclusion criteria. Using the recommended algorithm, echocardiography determination of diastolic function was possible in 219,462 patients (80%) but was indeterminate in 55,248 patients (20%). Baseline patient characteristics were similar among training, validation, and testing groups (Supplemental Table 1). There were 20,264 patients with left ventricular ejection fraction <50% (9.2%), 15,548 patients with cardiac amyloidosis (7.1%), 8,161 patients with hypertrophic cardiomyopathy (3.7%), and 5409 patients with moderate to severe aortic stenosis (2.5%). In the test set, estimated diastolic filling pressure by echocardiography was normal in 76,880 patients including 57,301 (58%) with normal diastolic function and 19,579 (19.8%) grade 1 dysfunction, and increased in 21,883 patients including 17,815 (18%) grade 2 and 4068 (4.1%) grade 3 dysfunction. Patients with increased filling pressure determined by echocardiography were older and had more comorbidities (p < 0.001, Table 1). Similarly, patients identified as having increased filling pressure by AI-ECG had more comorbidities (p < 0.001, Supplemental Table 2).

AI-enabled ECG classification performance

In the test set, the AI-enabled ECG for predicting echocardiographically determined increased filling pressure had an area under the curve (AUC) of the receiver operating characteristic (ROC) curve of 0.911 (95% CI: 0.909–0.914) with a sensitivity of 83.2%, specificity of 82.9%, positive predictive value (PPV) of 58%, and negative predictive value (NPV) of 94.5% with the threshold of 0.26, and prevalence of 22.2% (Fig. 1a and Table 2). The AI-enabled ECG’s AUCs for grade ≥1, grade ≥2, and grade 3 were 0.847 (95% CI: 0.844–0.85), 0.911 (95% CI: 0.909–0.914), and 0.943 (95% CI: 0.938–0.948) at thresholds of 0.443, 0.264, 0.058, respectively (Fig. 1b and Table 2). The median output values for the increased filling pressure from the model were significantly higher in the diastolic function grades 2 and 3 by echocardiography compared to normal and grade 1 (Fig. 2). The model showed higher specificity in younger patients and among patients with more comorbidities there was a tendency towards decreased specificity (Supplemental Fig. 1 and Table 3). Echocardiographic diastolic parameters significantly differed between patients identified by the model to have increased and normal filling pressure patients in the testing (Supplemental Fig. 2). Those diastolic parameters were almost identical in patients with normal filling pressure determined by AI-ECG and echocardiography. Those values in patients with increased filling pressure by both AI-ECG and echocardiography were consistent with grade 2–3 diastolic dysfunction and significantly different from values in patients with normal filling pressure. In the indeterminate group, all echocardiographic diastolic parameters except e’ velocity are significantly different between normal and increased filling pressure determined by AI-ECG (Supplemental Fig. 3).

a ROC plot for detecting increased filling pressure. b ROC plots for detecting diastolic function grades using an ordinal scale. ROC receiver operating characteristic, AUC area under the curve.

The distribution was described as a box plot with a kernel density plot. Box plots show median and first and third quartiles with outliers as 1.5 times IQR.

The AI-enabled ECG trained exclusively by ECG Lead I had AUCs of 0.804 (95% CI: 0.801–0.807), 0.875 (95% CI: 0.872–0.878), and 0.915 (95% CI: 0.909–0.921) for grade ≥1, grade ≥2, and grade 3, respectively. The AI-enabled ECG trained by ECG Lead I median beat had AUCs of 0.763 (95% CI: 0.76–0.766), 0.834 (95% CI: 0.83–0.837), and 0.877 (95% CI: 0.87–0.884), respectively.

The AUCs of AI-enabled ECG between before and after the median year of the echocardiography exam, i.e., 2014, were 0.856 and 0.839 for grade ≥1, 0.91 and 0.913 for grade ≥2, and 0.944 and 0.942 for grade 3, respectively (Supplemental Fig. 4).

Survival analysis

Death from any cause was observed in 20,223 (20.5%) of 98,763 patients in the test group and 18,224 (33.0%) of 55,248 in the indeterminate group over a median follow-up of 5.9 years (IQR 2.7, 10.2) and 5.7 years (IQR 2.6, 9.9), respectively. Mortality was significantly higher in patients with increased filling pressure compared to those with normal filling pressure predicted by the AI-enabled ECG after adjusting for age, sex, and comorbidities (hazard ratio (HR) 1.7, 95% CI 1.645–1.757; Fig. 3a). It was similar to the mortality predicted by echocardiographically determined filling pressure (HR 1.65, 95% CI: 1.597–1.705; Fig. 3b). All-cause mortality was also predicted by diastolic function grading determined by ECG with significantly higher mortality in patients with grade 2 and 3 diastolic dysfunction compared to those with normal or grade 1 diastolic dysfunction (HR 1.299, 95% CI 1.279–1.319, Fig. 3c). Diastolic function grading based directly upon echocardiographic parameters had a similar prognostic value (HR 1.298, 95% CI 1.277–1.32, Fig. 3d) even after adjusting for age, sex, and comorbidities. Since some patients had a discordance between AI-ECG and echocardiography determination of diastolic function, the study patients were separated into four groups in each category of diastolic dysfunction: true positive (TP; AI-ECG (+) and echocardiography (+), true negative (TN; AI-ECG (−) and echocardiography (−), false positive (FP; AI-ECG (+) and echocardiography (−), and false negative (FN; AI-ECG (−) and echocardiography (+)). While TP had the worst mortality and TN the best in all 3 diastolic dysfunction groups, FP and FN groups had a similar mortality in grade ≥1 or ≥2, but FP was found to have the same mortality as that of TP which was significantly worse than that of FN (HR 1.402, 95% CI 1.281–1.535) for grade 3 after adjusting for age, sex, and comorbidities (Supplemental Fig. 5).

a Kaplan–Meier curve for a test group of patients according to filling pressure predicted by the AI-enabled ECG. b Kaplan–Meier curve for a test group of patients according to filling pressure by echocardiography. c Kaplan–Meier curve for the test group according to diastolic function grades predicted by the AI-enabled ECG. d Kaplan–Meier curve according to diastolic function grades by echocardiography. e Kaplan–Meier curve for the indeterminate group according to diastolic function grades predicted by the deep learning model. f Kaplan–Meier curve for echocardiographic grade 1 in the testing group according to diastolic function grade normal and grade 1 predicted by the AI-enabled ECG. Number at risk tables are described below Kaplan–Meier curves. Log-rank test is used for the p-value. ECG electrocardiogram.

The risk of death was also greater among patients in the indeterminate group with higher filling pressure predicted from the AI-enabled ECG (HR 1.34, 95% CI 1.298–1.383). Among patients with normal filling pressure by the AI-enabled ECG, grade 1 dysfunction had worse survival than normal grade in both the testing and the indeterminate groups (Fig. 3c, e). Among patients with grade 1 diastolic dysfunction by echocardiography, 54.7% were classified as normal by the AI-enabled ECG, and those who were labeled as normal had lower risk of death than patients who labeled as grade 1 dysfunction by the AI-enabled ECG (Fig. 3f). The AI-enabled ECG successfully discriminated risk of death among specific age groups (≤50 years, 50 < age < 70 years, age ≥ 70 years) even after adjusting for age, sex and comorbidities (Supplemental Fig. 6).