Artificial intelligence enhanced ECG to detect reduced LVEF: clinical significance of false positive and false negative AI-ECG classifications

Background Artificial intelligence has been shown to enhance the diagnostic capabilities of the electrocardiogram (AI-ECG), enabling the detection of reduced left ventricular ejection fraction (LVEF). The aim of the present study was to investigate the clinical and prognostic implications of AI-ECG misclassifications, I.e. false positive (FP) and false negative (FN), using a previously validated AI-ECG algorithm to detect reduced LVEF (<50%) at echocardiography. Methods The study population included an independent retrospective cohort of 528 patients (age 68 [56–77] y, 40% women), which consisted of four matched groups according to the concordance between paired AI- ECG predictions and Simpson’s biplane LVEF assessments: true negative (TN), FP,FN, and true positive (TP). Sttudy outcomes included a composite endpoint of cardiovascular death and hospitalization for heart failure, and, development of reduced LVEF (<50%) during follow-up in patients with normal LVEF (FP, TN) at baseline. Results At baseline, LVEF was normal in FP (55%) and TN (56%) patients and impaired in FN (45%) and TP (40%) ones (p<0.001). In contrast, global longitudinal strain (GLS) was preserved only in TN (18%) but progressively worsened across FP (14%), FN (14%), and TP (10%) ones (p<0.001). During median follow-up of 4.5 years (IQR: 2.8–6.5y), the composite endpoint occured in 59 (13%) patients with significantly lower incidence in the TN versus the other three groups (log-rank p<0.001). In multivariable Cox analysis, impaired LVEF or GLS, and higher age emerged as only independent predictors of the composite endpoint (all p < 0.05). During follow-up, a total of 49 out of 131 (37%) FP versus only 12 out of 130 (9%) TN patiens developed overt systolic dysfunction (LVEF < 50%) (p < 0.001). The major causes of LVEF decline were valvular heart- (28.6%) or coronary artery disease (23,2%) and atrial fibrillation (26.8%). In Cox regression analysis, AI-ECG classification was independent predictor of future drop in LVEF (HR 3.78; 95% CI 1.74–8.23; P=0.001) while GLS was not (NS) (Figure 1,2). Conclusions AI-ECG misclassifications are clinically relevant. The FP group has increased risk to develop overt LV dysfunction with adverse outcome, warranting close clinical follow-up. The FN classification is associated with worse outcome driven by LV dysfunction not recognized by AI-ECG, but fortunately, the prevalence of FN in the models is low.