Background: Heart failure with preserved ejection fraction (HFpEF) is one of the most frequent cardiac causes of exertional dyspnea. According to the current European Society of Cardiology (ESC) guidelines, echocardiographic and laboratory criteria have to be met to establish the diagnosis of HFpEF(1). An ECG screening method would be desirable, but ECG findings in HFpEF are manifold and there are no known typical and unambiguous ECG features for HFpEF recognizable by human interpretation. This study sought to evaluate whether a deep learning algorithm can identify HFpEF by baseline 12-lead ECGs.

Methods: We included n=1884 patients who presented with exertional dyspnea or equivalent and preserved ejection fraction and clinical suspicion for coronary artery disease. All baseline ECGs were recorded at the index visit. All patients underwent echocardiography, coronary angiography, and left heart catheter. The ECGs were divided in 2-second segments, in order to depict at least one QRS complex per sample, yielding a total of 77.558 samples for analysis. We therefore trained a convolutional neural network (CNN) in order to classify HFpEF and control patients. Data sets were divided in 50% training, 30% internal validation and 20% test sets. The test dataset was withheld and blinded from the network in the beginning to test the accuracy, which was used as external validation cohort.

Results: 720 patients (38%) were identified as HFpEF and 1164 (62%) as controls. HFpEF patients were older (66 years vs. 59 years, p<0.001), more frequently females (46% vs 36%, p<0.001) and had a higher body mass index (31kg/m² vs. 30kg/m², p<0.001). HFpEF patients had higher E/E’ values (11±4 vs. 9±3, p<0.001), left atrial volume indices (29±10ml/m² vs. 25±8ml/m², p=0.005) and higher prevalence of left anterior fascicular hemiblock (n=72 vs. n=66, p=0.03). 115 patients (6%) presented with atrial fibrillation, (n=6 control vs. n=109 HFpEF, p<0.001). Coronary angiography revealed coronary artery disease in 608 patients (52%) of the control group and in 460 patients (64%) of the HFpEF group (p<0.001).

To test the model’s goodness and performance, it was used to  classify the external validation set. The area under the curve of the CNN was 0.92 (95% confidence interval [CI] 0.91–0.94), allowing for a discrimination between HFpEF and controls with a sensitivity of 0.98 (95% CI 0.97–0.99) and a specificity of 0.63 (95% CI 0.59–0.67). This translated into a positive predictive value of 0.66 (95% CI 0.62–0.67) and a negative predictive value of 0.98 (95% CI 0.96–0.99).

Conclusion:

In this study, we report the first deep learning-enabled CNN for the identification of patients with HFpEF according to ESC criteria among patients at risk for HFpEF. By analyzing 12-lead ECGs, the model showed that HFpEF may have electrocardiographic characteristics that can be recognized by artificial intelligence algorithms. Importantly, the excellent predictive value of the CNN to rule out HFpEF was achieved in a cohort of patients with relevant cardiovascular risk factors and high prevalence of conduction abnormalities.

Beyond traditional ECG interpretation, CNN could become a valuable screening tool for the diagnosis of HFpEF using a simple 12-lead ECG. To date, further research is warranted to validate these findings in independent cohorts.
1.Ponikowski P, et al. 2016: ESC Heart Failure Guidelines for the diagnosis and treatment of acute and chronic heart failure. European Heart Journal