"We designed an ultra‑thin, 128‑element linear‑array ultrasound transducer that can be worn comfortably on the neck," explains Professor Zheng. "It images the right internal jugular vein (IJV) and the common carotid artery (CCA) – two vessels that reliably reflect central venous pressure because the IJV connects directly to the right atrium without valves." The patch probe, with a center frequency of 8.5 MHz, provides excellent spatial resolution and penetration depth. Its dual‑layer acoustic matching and tailored backing ensure broad bandwidth (85% fractional band-width) and high sensitivity, while a solid hydrogel coupling agent and silicone encapsulation maintain stable image quality and skin comfort during prolonged wear – up to 24 hours in feasibility tests.
The wearable device generates continuous cine‑loop videos, but manually analyzing every frame is labor-intensive and impractical. To address this, the team developed the dual‑decoder spatiotemporal attention network (DSTA‑Net), a semi‑supervised segmentation model that requires manual annotation of only about 10% of frames (key frames at maximum and minimum IJV dilation). "The remaining 90% of unlabeled frames are used as training signals through a clever dual‑decoder consistency mechanism," says Dr. Guo. A shared encoder processes all frames, while two decoders – one with temporal attention, one lightweight – enforce cross‑pathway consistency, turning unlabeled data into valuable learning material without fragile teacher‑student updates.
In both internal and external test sets, DSTA‑Net significantly outperformed state‑of‑the‑art fully supervised models (UNet, Swin‑UNet, DeepLabV3+) and semi‑supervised competitors (UniMatch, DWL, AllSpark). For IJV segmentation, it achieved Dice scores of 83.5% (internal) and 75.8% (external) – improvements of ~12% and ~9% over the best supervised baseline. Spearman correlation between DSTA‑Net‑derived vascular indices and expert manual measurements exceeded 0.88 for most parameters, with Bland‑Altman analysis confirming clinically acceptable agreement (percentage error well below the 30% threshold).
The automatically segmented images yield five vascular indices: IJV Max Area, IJV Min Area, CCA Area, IJV Max/CCA Area, and IJV Ratio. These are combined with demographic and physiological variables (age, BMI, blood pressure, heart rate) and fed into a dual‑modality multilayer perceptron (DM‑MLP) – a model designed specifically for tabular clinical data. "Unlike generic architectures such as ResNet or Transformer, our DM‑MLP uses two complementary mixing operations: Attribute‑Mixing to capture cross‑feature dependencies, and Case‑Mixing to refine intra‑feature representations," explains Professor Li. This low‑rank, efficient design outperformed ResNet, DenseNet, and Transformer by 4‑8% in area under the ROC curve (AUC).
In a prospective multi‑center study involving 349 ICU patients (272 from Shanghai Sixth People's Hospital, 77 from Shanghai Tenth People's Hospital), DM‑MLP achieved an AUC of 0.91 (internal test) and 0.87 (external test) for detecting elevated CVP (≥ 8 mmHg), with balanced sensitivity and specificity. Sensitivity analyses using alternative thresholds of 7 and 9 mmHg confirmed the model's robustness. SHAP interpretability analysis revealed that the ultrasound‑derived vascular indices – particularly IJV Max Area and IJV Ratio – were the dominant predictors, far outweighing conventional clinical parameters such as blood pressure or BMI.
"This is not about replacing CVCs in all patients," emphasizes Professor Zheng. "It is about providing a safe, rapid, and repeatable screening tool for patients in whom catheterization is contraindicated or difficult, and for early bedside identification of elevated CVP to guide timely intervention." The system operates at 32 frames per second (30 ms per frame) on a hospital server, enabling near‑real‑time interpretation.
The authors acknowledge limitations: the sample size is modest for AI training, the study focuses on diagnostic performance rather than clinical outcomes, and the semi‑supervised model's internal decision‑making remains relatively opaque. Future work will expand the multi‑center cohort, explore direct prediction of continuous CVP values, integrate interpretability techniques such as Grad‑CAM, and conduct interventional trials to assess whether AI‑guided noninvasive monitoring improves fluid and vasopressor management.
By combining a comfortable, long‑wear ultrasound patch with a semi‑supervised segmentation network and a dedicated clinical predictor, this integrated framework brings automated, noninvasive CVP assessment closer to routine bedside use – a step forward for personalized hemodynamic management in acute and critical care.
Authors of the paper include Liping Dong, Xingxuan Zhang, Meng Li, Yi Li, Jingyi Guo, Shaorong Lu, Zhenyu Peng, Chenxi Zheng, Yufei Hui, Xiaoping Shao, Feiyan Wang, Weikang Jiang, Maoyao Li, Xianshuai Wu, Xu Guo, Yingchuan Li, Yuanyi Zheng, and Liping Zhang.
This work was supported by the Shanghai Municipal Health Commission Clinical Research Program (grant number 20244Y0119) and the Exploratory Clinical Project of the Shanghai Sixth People's Hospital Affiliated with Shanghai Jiao Tong University (grant number yynts202302).
The paper "AI-Enabled Wearable Ultrasound for Noninvasive Central Venous Pressure Monitoring" was published in the journal Cyborg and Bionic Systems on Aug. 8, 2026, at DOI: 10.34133/cbsystems.0653.
