Two-step screening for low EF: AI-ECG and AI-guided and interpreted handheld ultrasound

Background/introduction Artificial intelligence (AI) analysis of 12-lead ECG is increasingly being applied to screen for reduced left ventricular ejection fraction (LVEF), yet has limited diagnostic accuracy particularly specificity and positive predictive value. Focused cardiac ultrasound with AI-guidance software (AI-FoCUS) is emerging as a method for inexperienced users to capture basic cardiac images that in turn could be analysed by AI. Whether this could bridge the diagnostic gap of AI-ECG is unknown. Purpose To assess the diagnostic performance of novice-acquired FoCUS scans as an intermediate diagnostic test following an AI-ECG. Methods 1141 prospectively enrolled adults, each undergoing AI-ECG, AI-FoCUS, and a reference comprehensive echocardiogram. AI-FoCUS was performed by one of 9 novice imagers (without any prior clinical or imaging experience) with a handheld ultrasound device guided in real-time by AI. Output of interest was a left ventricular ejection fraction (LVEF) at a threshold of LVEF <40% by an AI-LVEF algorithm. FoCUS acquisition and AI outputs were obtained blinded to clinical data and comparator tests. The reference standard was a clinical transthoracic echocardiography on a GE E95 performed within 15 minutes of the FoCUS. Results Median age 65±15 years; 62% male; BMI range 17-58 kg/m². Prevalence: EF <40% in 4.8%. AI-FoCUS yielded technically feasible images in 99% of cases, with median scan duration of 4 minutes. Diagnostic accuracy for LVEF <40%: AI-ECG: Sensitivity 85.5% (95% CI 73-94%), specificity 89.0% (95% CI 87-91), PPV 28.3% (22-36%), NPV 99.2 (98-100), diagnositic accuracy 88.8% (87-91%). AI-FoCUS: Sensitivity 80.0% (95% CI 67-90), specificity 98.5% (95% CI 98-99), PPV 74.6% (95% CI 62-85), NPV 98.9% (95% CI 98-100), diagnostic accuracy 97.6% (95% CI 97-98). In a 2-step process (AI-ECG first and only if positive then AI-FoCUS): Sensitivity 76.4% (95% CI 63-87), specificity 99.0% (95% CI 98-100), PPV 79.2% (95% CI 66-89), NPV 98.8% (95% CI 98-99), diagnostic accuracy 97.9% (95% CI 97-99). This two-step pathway, rather than FoCUS for all, would reduce the number of patients requiring FoCUS scans to only 14.5%, optimizing resource use. Conclusions AI-guided cardiac ultrasound by truly novice users is highly feasible, rapid, and accurate suggesting that AI-guided handheld ultrasound could bridge the gap between initial screening with the AI-ECG and more resource-intensive diagnostic echocardiography. The combination of AI-ECG and AI-FoCUS may provide a cost-effective, scalable strategy for early and accurate detection of reduced left ventricular ejection fraction.