Assessment of wearable Ultrasound Device Combined with AI for Portable Assessment of Central Venous Pressure Compared with Central Venous Catheterization as the reference Standard in critically ill patients: A Cross-Sectional Study

Read the full article See related articles

Listed in

This article is not in any list yet, why not save it to one of your lists.
Log in to save this article

Abstract

Purpose Currently, central venous catheterization (CVC) remains the reference standard for evaluating central venous pressure (CVP) in critically ill patients, but its invasiveness and associated complications limit its use. This study combined a wearable ultrasound device with artificial intelligence (AI) models to achieve portable noninvasive assessment of CVP in critically ill patients. Methods This study prospectively enrolled critically ill patients who underwent CVC. We divided participants into a CVP≥8 mmHg group and a CVP<8 mmHg group, and recorded the internal jugular vein (IJV) and common carotid artery (CCA) via wearable ultrasound device. A Dual-Decoder Spatiotemporal Attention Network (DSTA-Net) was trained to enable automatic measurement of wearable ultrasound data and was evaluated via Bland-Altman consistency and Spearman correlation analyses. Moreover, it was compared with existing DeepLabV3+, Swin-Unet, TransUNet, and UNet models. The baseline patient parameters were subsequently combined with the wearable ultrasound parameters to construct a dual-mixing multilayer perception (DM-MLP) model to forecast elevated CVP,which was compared with existing VGG, ResNet, and Transformer models. Diagnostic performance was assessed via receiver operating characteristic (ROC) curves. Results Videos from 272 patients (218:54 [80%:20%] in the training and test sets) were used for DSTA-Net training. Both sets were divided into the CVP≥8 mmHg and CVP<8 mmHg groups. In both sets, the CVP and the manually/DSTA-Net-measured IJV Max Area, IJV Min Area, and IJV Max/CCA Area were greater in the CVP≥8mmHg group. Compared with the DeepLabV3+, Swin-Unet, TransUNet, and UNet models, the DSTA-Net model showed higher Dice and IoU values of 82.62±0.18 and 74.97±0.15 for IJV segmentation, and 80.09±0.19 and 71.71±0.17 for CCA segmentation, respectively. Bland-Altman and Spearman correlation analyses confirmed a high degree of consistency and correlation between manually- and DSTA-Net- measured wearable ultrasound parameters. The ROC curve AUCs of the clinical parameters and manually/DSTA-Net measured DM-MLP models were 0.94[0.88,0.94] and 0.88[0.85,0.91], respectively, indicating better predictive performance than the VGG, ResNet, and Transformer models. Conclusion These results suggest that integrating wearable ultrasound devices with DSTA-Net as an image segmentation model and DM-MLP as a clinical prediction model provides a portable, noninvasive method for predicting elevated CVP.

Article activity feed