Validation of an AI-ECG tool to detect left ventricular systolic dysfunction in community-based patients with suspected heart failure and elevated NT-proBNP

Background Many patients with suspected heart failure (HF) experience long waits for echocardiography, despite risk-stratification using natriuretic peptides (NP). Prioritising echocardiography in those who are likely to have HF with reduced ejection fraction (HFrEF) would be of value in many healthcare settings and to patients, as prognostic therapy could be commenced early. A normal electrocardiogram (ECG), assessed by a cardiologist, has been previously shown to almost exclude left ventricular systolic dysfunction (LVSD). Artificial intelligence (AI) models may also identify patients at high risk of LVSD. Purpose To determine whether AI-based ECG analysis can identify patients with LVSD among symptomatic community-based individuals with suspected HF and an elevated NT-proBNP concentration, referred for echocardiography via a clinical pathway. Methods We evaluated a validated ECG-AI model for detecting LVSD (left ventricular ejection fraction [LVEF] ≤40%) in a multicentre, prospective cohort of patients referred for HF assessment. Of 867 enrolled patients, 783 had an LVEF measured and were included in this analysis. The AI model analysed images of 12-lead ECGs and provided both a categorical risk (“low”, “intermediate”, “high”) and a predicted probability of LVSD. Risk categories were dichotomised into “Unlikely LVSD” (“low” risk) and “Possible LVSD” (“intermediate” + “high” risk). The AI model predictions were compared against an echocardiographic assessment of LVEF. Diagnostic accuracy was evaluated for the categorical output. The continuous probability output from the AI model was also assessed using receiver operating characteristic (ROC) analysis. Results Of 783 patients with an LVEF measured by echocardiography, the AI model generated a risk classification for 763 (97%). The median age was 77 and 51% were female. Patient characteristics, medical history, and baseline medications are shown in Table 1. Of patients categorised as intermediate or high risk of LVSD by the AI model the prevalence of an LVEF ≤40% was 14%. In those categorised as low risk, the prevalence was 1%. For detection of an echocardiographic LVEF ≤40%, the AI model intermediate/high-risk category had a sensitivity of 84 (95%CI: 70-93)%, specificity of 69 (66-73)%, positive predictive value of 14% (10%-19%), and negative predictive value of 99 (97-99)%. The continuous LVSD probability had an area under the ROC curve of 0.87 (95%CI: 0.83–0.93), Figure 1. Conclusion AI-based ECG analysis can identify patients with suspected HF at higher risk of having LVSD, and could be used to streamline clinical pathways to allow prioritisation of echocardiography for these patients to ensure early initiation of life-saving therapy.TableFigure