Accuracy of artificial intelligence-enabled handheld ultrasound devices to assess left ventricular ejection fraction: A systematic review and meta-analysis
J Winarta, A Gupta, S Zahedmohajerani, R Burgess, C Smith, H Marie North, N Samad, K Witte, S Kamalathasan, J Gierula, M Paton, S StrawAbstract
Introduction
Transthoracic echocardiography (TTE) underpins the diagnosis, phenotyping and subsequent management of heart failure (HF). However, access to high-quality cardiac imaging remains inequitable. Globally, workforce shortages, rising HF prevalence, and increasing service demand contribute to diagnostic delays and variation in care.1-2 Artificial intelligence (AI)-enabled handheld ultrasound devices (HUDs) have emerged as a potential solution, offering the possibility of streamlining or decentralising cardiac evaluation. Despite growing clinical uptake, the diagnostic accuracy of AI-enabled HUDs, particularly when used by novice operators and in real-world clinical settings, has not been comprehensively established.
Purpose
We conducted a systematic review and meta-analysis to determine the diagnostic accuracy of commercially available AI-enabled HUD compared with TTE in determining reduced left ventricular ejection fraction (LVEF).
Methods
The protocol was registered with PROSPERO. We searched PubMed, MEDLINE, Embase, and CENTRAL from inception to 1 February 2026 for studies comparing AI-enabled HUDs with standard echocardiography in adult patients. Two reviewers independently screened and extracted data; studies were included only if 2×2 confusion matrices could be derived for AI-enabled HUD versus TTE. Where this was not possible, but the study appeared potentially relevant, we made requests to corresponding authors. Risk of bias and applicability were assessed using QUADAS-2. All statistical analyses were performed using STATA 19 (StataCorp. 2025. Stata Statistical Software: Release 19. College Station, TX: StataCorp LLC.). We pooled diagnostic accuracy using random-effects and hierarchical summary receiver operating characteristic (HSROC), reported 95% confidence intervals and I², and analysed subgroups by operator experience and LVEF cut-off.
Results
Seventeen studies including 2,879 examinations were included. There was variability across studies in the threshold for reduced LVEF used for diagnostic accuracy analyses. The overall sensitivity was 0.80 (0.74 – 0.87, I2 = 86.34%), specificity 0.93 (0.91 – 0.95, I2 = 82.24%), PPV 0.80 (0.74 – 0.86, I2 = 85.85%), NPV 0.91 (0.87 – 0.95, I2 = 97.02%), and accuracy 0.90 (0.87 – 0.93, I2 = 90.86%). Sensitivity was higher for novices (0.91 [0.85–0.98]) compared to experts (0.70 [0.60–0.81]). Overall, AI correctly classified examinations most commonly when utilising 30% thresholds (accuracy 0.96 [0.95 – 0.98]); it performed worst in 40% thresholds (accuracy 0.85 [0.76 – 0.94], I2 = 96.19%).
Conclusion
AI-enabled HUD demonstrated excellent performance for identifying reduced LVEF, supporting its use as a screening and triage tool. However, substantial inter-study heterogeneity and operator-dependent differences indicate performance is context-specific and should be interpreted with training, scan protocol, and manufacturer information.HSROC CurvePRISMA Flow Chart