Stroke volume is a critical metric for assessing cardiac function in critically ill patients, yet obtaining it via echocardiography remains technically challenging. Traditionally, calculating stroke volume requires two specific measurements: the diameter of the left ventricular outflow tract (LVOTd) and the velocity time integral (VTI) of blood flow through that tract. While the VTI can be reliably obtained in over 90% of patients, the LVOTd is frequently unmeasurable in approximately 30% of cases, particularly in intensive care units where mechanical ventilation alters heart position and obscures ultrasound windows. A new study published in Cardiovascular Ultrasound demonstrates that a machine learning model can estimate the LVOTd with accuracy comparable to human experts, significantly increasing the yield of stroke volume measurements.
The research team, led by Sarah Balderston Murthi and colleagues from the University of Maryland, Baltimore, conducted a prospective observational study to evaluate the performance of this computer model. Over a 20-month period from January 2015 to April 2017, the researchers enrolled 84 adult patients who had pulmonary artery catheters (PAC) placed for clinical indications. These patients underwent critical care echocardiography, and their stroke volume was measured simultaneously using the PAC as the reference standard. The study aimed to determine if a machine learning-based estimate of the LVOTd, derived from patient characteristics, could replace the difficult manual measurement without sacrificing accuracy.
The model estimates the LVOTd based on four specific patient characteristics: age, sex, height, and weight. These variables were selected because they correlate with body surface area and anatomical dimensions. In the current study, the computer model’s estimate (LVOTdCM) was compared against the LVOTd measured by a human expert (LVOTdHEM). Both estimates were then used to calculate stroke volume, which was compared to the stroke volume derived from the pulmonary artery catheter data.
The results revealed a significant improvement in the ability to obtain stroke volume measurements. Using the human expert measurement of the LVOTd, stroke volume could only be calculated in 59 out of 84 patients, representing a 70% yield. This limitation was primarily due to the inability to obtain an adequate ultrasound window for measuring the LVOTd in 19 patients. In contrast, when using the computer model’s estimate, stroke volume could be calculated in 78 out of 84 patients, representing a 92% yield. This represents a 22% increase in the successful acquisition of this vital hemodynamic parameter.
When comparing the accuracy of the two methods against the pulmonary artery catheter reference, the computer model performed slightly better than the human expert measurement. Bland-Altman analysis, a statistical method used to assess agreement between two different measurement techniques, showed a mean bias of 3.1 mL for the human expert method, with 95% limits of agreement ranging from -28.8 to 34.9 mL. For the computer model, the mean bias was lower at 1.75 mL, with 95% limits of agreement ranging from -29 to 32.5 mL. The Pearson correlation coefficient was 0.82 for the human expert method and 0.84 for the computer model, indicating a strong correlation between the echocardiographic estimates and the catheter measurements for both approaches.
The study highlights the specific challenges of measuring the LVOTd in the intensive care setting. The LVOTd is a static anatomical length that is best measured from the parasternal long axis view. However, in mechanically ventilated patients, the expansion of the lungs pushes the heart inferiorly and medially, making it difficult to obtain the correct angle for measurement. This technical difficulty is compounded by the fact that the LVOTd is a squared variable in the stroke volume formula, meaning that small errors in diameter measurement can lead to larger errors in the final stroke volume calculation. The computer model eliminates this source of error by relying on stable patient demographics rather than real-time image acquisition.
The participants in the study were a diverse group of critically ill and post-operative patients. Of the 59 patients with complete data, 55% were inpatients in the intensive care unit, including those recovering from liver transplants, coronary bypass grafting, and various valve repairs. The remaining 45% were outpatients undergoing right heart catheterization. The mean age of the participants was 61 years, and 59% were male. Most patients had intact left ventricular function, though some exhibited mild to severe dysfunction.
The authors note that while the computer model improves the yield and slightly reduces bias, it does not eliminate the need for accurate VTI measurement. The VTI must still be obtained via Doppler echocardiography, which requires operator skill and expertise. Therefore, the machine learning tool serves as a complement to, rather than a replacement for, the echocardiographer’s skills. By removing the need to measure the LVOTd directly, the model simplifies the workflow and allows for more consistent and repeatable measurements. This is particularly important in the intensive care unit, where rapid and accurate hemodynamic assessment is crucial for guiding therapy.
The study has several limitations that should be considered. The sample size was relatively small, and the patients were drawn from a convenience sample at a single institution during daytime hours. The data was collected prospectively but analyzed retrospectively, which may introduce potential errors. Additionally, there was a significant time lag between data collection and analysis, partly due to delays related to the pandemic and the difficulty of finding patients with pulmonary artery catheters. The images were obtained by fellows in training under supervision, which may limit the generalizability of the findings to other settings. Despite these limitations, the study provides valuable evidence that machine learning can enhance the utility of echocardiography in critical care.
This research contributes to the growing body of evidence supporting the integration of artificial intelligence into clinical practice. By leveraging machine learning to estimate anatomical parameters, clinicians can overcome technical barriers that have long limited the use of echocardiography for hemodynamic assessment.
Subject of Research: Machine learning application in critical care echocardiography for stroke volume assessment
Article Title: Computer modeling to improve measurement of stroke volume with echocardiography
Article References: Murthi, S. B., Yang, S., Oliveri, P. P., Safadi, S., Fatima, S., & Teeter, W. (2026). Computer modeling to improve measurement of stroke volume with echocardiography. Cardiovascular Ultrasound, 24(1), Article 16. https://doi.org/10.1186/s12947-026-00373-7
Image Credits: AI Generated
DOI: 10.1186/s12947-026-00373-7
Keywords: machine learning, echocardiography, stroke volume, critical care, cardiac output, pulmonary artery catheter, left ventricular outflow tract, clinical ultrasound, hemodynamics, artificial intelligence, Computer, modeling
News Source: Cassandra Pierce. (October 29, 2026). AI Model Boosts Stroke Volume Measurement Accuracy in ICU. Scienmag.
