Computer modeling to improve measurement of stroke volume with echocardiography

Clinical Snapshot

Key Findings

• P Value: Not reported in abstract

• Effect Size: Mean bias: HEM vs PAC = 3.1 mL; CM vs PAC = 1.75 mL. Diagnostic yield: HEM = 70% (59/84); CM = 92% (78/84) — absolute improvement of 22%

• Primary Outcome: Agreement between echocardiographic SV (HEM and CM methods) and PAC-derived SV by Bland-Altman analysis, and diagnostic yield (proportion of patients with calculable SV)

• Nnt Or Sensitivity: Diagnostic yield improvement: NNT-equivalent = approximately 4.5 patients assessed to gain one additional calculable SV measurement with CM vs HEM. No sensitivity/specificity reported (continuous measurement study). Percentage error for HEM vs PAC and CM vs PAC not explicitly calculated in abstract but LOA suggest clinically wide agreement intervals.

• Confidence Interval: 95% limits of agreement: HEM vs PAC = -28.8 to +34.9 mL (total range 63.7 mL); CM vs PAC = -29.0 to +32.5 mL (total range 61.5 mL). Confidence intervals for mean bias not reported.

Clinical Application

The computer model requires only VTI acquisition (a single Doppler measurement) rather than both LVOTd and VTI, simplifying the workflow and potentially enabling less experienced operators to obtain quantitative SV estimates. However, VTI acquisition itself requires adequate apical five-chamber or three-chamber views and Doppler alignment, which may still be technically challenging in critically ill patients. Full clinical implementation requires integration of the model into echocardiography platforms, which is not yet described as commercially available. Point-of-care echocardiography is increasingly used in Australian ICUs and emergency departments, supported by CICM (College of Intensive Care Medicine) and ACEM (Australasian College for Emergency Medicine) training frameworks. The RACGP does not have specific guidelines for echocardiographic SV measurement in primary care. There is no current TGA-listed or PBS-subsidised machine learning echocardiography tool for LVOTd estimation in Australia. The findings are relevant to Australian intensivists and emergency physicians who perform or interpret point-of-care echocardiography, particularly in resource-limited settings where PAC use is declining and non-invasive haemodynamic assessment is increasingly relied upon. Australian ICU practice increasingly favours less invasive haemodynamic monitoring, making accurate echocardiographic SV measurement clinically important. Adoption would require TGA regulatory approval for any commercial software implementation. Critically ill adult patients in intensive care settings where haemodynamic monitoring is required and echocardiographic SV assessment is clinically indicated. Most directly applicable to patients already receiving invasive haemodynamic monitoring (PAC or equivalent). Potential broader application to perioperative and emergency department settings where point-of-care echocardiography is performed.

Abstract

BACKGROUND: The stroke volume (SV) can be measured by a human expert (HE) using the left ventricular outflow tract diameter (LVOTd) and its velocity time integral (VTI) with echo. If the SV is known, the cardiac output and systemic vascular resistance can be calculated. We have previously described a machine learning computer model (CM) to estimate the LVOTd. The accuracy of LVOTdCM vs. LVOTdHEM in estimation of the SV relative to a pulmonary artery catheter (PAC) is unknown. METHOD: Over 20 months, a prospective observational study was done in patients with a PAC placed for clinical indications. Participants underwent an echocardiogram, and at the same time the SV was measured by PAC. The SV with echo was calculated using the LVOTdHEM and LVOTdCM and the VTI. The bias and agreement between the metrics and PAC were determined with Bland-Altman analysis. RESULTS: Computer modeling improves measurement of stroke volume with echocardiography. Eighty-four patients with a PAC were enrolled. With human expert measurement of the LVOTd and the VTI, the SV could be calculated in 59 (70%), and with the LVOTdCM and the VTI in 78 (92%). Bland Altman analysis comparing PAC and HEM yielded a mean bias of 3.1 with 95% limits of agreement (LOA) -28.8 and 34.9. When comparing PAC and CM, bias was 1.75 and LOA of -29 and 32.5. CONCLUSIONS: A machine learning model of the LVOTd allows accurate calculation of the SV using only the VTI, simplifying the assessment while increasing the yield by 22%. Machine learning tools can lower the bar for obtaining quantitative metrics with echo. Adding objective and repeatable measures will improve the ability of echo to guide therapy in the critically ill.