Revolutionizing Patient Monitoring: The Non-Invasive CVP Solution
Innovative Design of the Wearable Ultrasound Patch for CVP Monitoring
Professor Zheng highlights the creation of an exceptionally thin, 128-element linear-array ultrasound transducer engineered for comfortable wear around the neck. This device is designed to image the right internal jugular vein (IJV) and the common carotid artery (CCA). These two blood vessels are critical because the IJV connects directly to the right atrium without valves, making them reliable indicators of central venous pressure.
Advanced Acoustic Properties and Stability of the Patch Probe
The patch probe boasts a center frequency of 8.5 MHz, ensuring superior spatial resolution and adequate penetration depth. Its design incorporates a dual-layer acoustic matching system and a custom backing, which together provide a wide bandwidth (85% fractional bandwidth) and high sensitivity. A solid hydrogel coupling agent and silicone encapsulation further enhance stable image quality and patient comfort, allowing for continuous wear for up to 24 hours during preliminary trials.
Automating Image Analysis with the Dual-Decoder Spatiotemporal Attention Network (DSTA-Net)
To overcome the challenge of labor-intensive manual analysis of continuous cine-loop videos generated by the wearable device, the research team developed the dual-decoder spatiotemporal attention network (DSTA-Net). This semi-supervised segmentation model minimizes the need for manual annotation, requiring it for only about 10% of frames, specifically key frames showing maximum and minimum IJV dilation. Dr. Guo explains that the remaining 90% of unlabelled frames are utilized as training signals through an innovative dual-decoder consistency mechanism. A shared encoder processes all frames, while two distinct decoders—one with temporal attention and another lightweight version—enforce cross-pathway consistency, transforming unlabelled data into valuable learning material without relying on potentially unstable teacher-student updates.
Superior Performance in Vascular Segmentation and Clinical Agreement
In both internal and external validation studies, DSTA-Net significantly outperformed cutting-edge fully supervised models, including UNet, Swin-UNet, and DeepLabV3+, as well as other semi-supervised approaches like UniMatch, DWL, and AllSpark. For IJV segmentation, it achieved Dice scores of 83.5% (internal) and 75.8% (external), representing improvements of approximately 12% and 9% over the leading supervised baseline. The Spearman correlation between DSTA-Net's vascular indices and expert manual measurements exceeded 0.88 for most parameters, and Bland-Altman analysis confirmed clinically acceptable agreement, with percentage errors well below the 30% threshold.
Integration of Vascular Indices and Clinical Data for CVP Prediction with DM-MLP
The automatically segmented images yield five crucial vascular indices: IJV Max Area, IJV Min Area, CCA Area, IJV Max/CCA Area, and IJV Ratio. These indices are then combined with demographic and physiological data, such as age, BMI, blood pressure, and heart rate, and fed into a dual-modality multilayer perceptron (DM-MLP). This model is specifically tailored for tabular clinical data. Professor Li elucidates that unlike generic architectures like ResNet or Transformer, their DM-MLP employs two complementary mixing operations: Attribute-Mixing, which captures dependencies across different features, and Case-Mixing, which refines intra-feature representations. This efficient, low-rank design demonstrated a 4-8% improvement in the area under the ROC curve (AUC) compared to ResNet, DenseNet, and Transformer.
Clinical Validation and Interpretability of the AI-Powered CVP Monitoring System
A prospective multi-center study involving 349 ICU patients (272 from Shanghai Sixth People's Hospital and 77 from Shanghai Tenth People's Hospital) demonstrated the effectiveness of the DM-MLP. The model achieved an AUC of 0.91 (internal test) and 0.87 (external test) for detecting elevated CVP (defined as ≥ 8 mmHg), exhibiting balanced sensitivity and specificity. Robustness was confirmed through sensitivity analyses using alternative thresholds of 7 and 9 mmHg. SHAP interpretability analysis revealed that the ultrasound-derived vascular indices, particularly IJV Max Area and IJV Ratio, were the most influential predictors, significantly outweighing conventional clinical parameters like blood pressure or BMI.
Future Directions and Clinical Significance of Non-Invasive CVP Monitoring
Professor Zheng emphasizes that this technology is not intended to replace central venous catheters (CVCs) entirely but rather to provide a safe, rapid, and repeatable screening tool. It is particularly valuable for patients where catheterization is contraindicated or challenging, and for early bedside identification of elevated CVP to guide timely interventions. The system operates efficiently at 32 frames per second (30 ms per frame) on a hospital server, allowing for near real-time interpretation. The authors acknowledge that future work will involve expanding the multi-center cohort, exploring direct prediction of continuous CVP values, integrating interpretability techniques like Grad-CAM, and conducting interventional trials to assess the impact of AI-guided non-invasive monitoring on fluid and vasopressor management.