Artificial intelligence assessment of the jugular venous pulse from ultrasound high-resolution B-mode clips: A proof of concept

Objective: The jugular venous pulse (JVP) is a pivotal clinical parameter that currently can only be invasively measured through jugular catheterization and subsequent central venous pressure measurement. The ultrasound B-mode clip of the internal jugular vein cross-sectional area modifications allows to build a JVP curve that significantly correlates with the central venous pressure. However, this process is time-consuming and not suitable for clinical use. The aim of the present study is to verify whether artificial intelligence (AI) allows a rapid and reliable JVP waveform assessment as compared with a human operator. Methods: High-resolution B-mode internal jugular vein clips (558 frames) of a cohort of six human subjects have been blindly analysed in post-processing by three different researchers and a neural network. Agreement was quantified using two complementary measures: Dice similarity coefficient (Dice) and the Hausdorff distance at the 95th percentile (HD95). Furthermore, a noninferiority test was performed comparing the model with the human raters. The null hypothesis (H0) was that the model performs worse than human raters by at least Δ = 0.055 Dice, a difference that is considered clinically negligible. Results: The average processing time per frame was 19.80 ± 5.08 seconds for human operators, compared with 0.03404 seconds ± 0.01806 seconds for the AI model running on a standard consumer-grade laptop. This represents a difference of nearly three orders of magnitude (a difference that could be quantitatively described as 580 times faster). Agreement between human raters was very high, with median Dice 0.959 (95% confidence interval, 0.958-0.960). Agreement between the model and each rater was slightly lower, with a median Dice of approximately 0.907 (95% confidence interval, 0.904-0.909). Human raters had median HD95 values of <5 pixels, reflecting very small boundary differences. The model-vs-rater comparisons showed somewhat higher HD95 values, with medians or approximately 8 to 10 pixels, but still within a clinically acceptable range given the resolution of the images. The Wilcoxon paired test rejected the null hypothesis H0 (P = .004169), showing that the model is not inferior to human raters within this clinically acceptable margin. Conclusions: Our study demonstrates an amazing time efficiency of the entire AI segmentation process, with a precision quite comparable with the human researchers' assessment. Our findings, in perspective, support the clinical introduction of ultrasound AI JVP waveform assessment in a variety of potentially interested medical specialties, including cardiology, critical care, neurosciences, and vascular surgery.