Central venous pressure reflects the pressure in the right atrium and is widely used as an indicator of venous return, cardiac function, and a patient’s response to fluids or vasoactive medications. The conventional reference method requires placement of a central venous catheter, an invasive procedure that can cause bleeding, infection, thrombosis, pneumothorax, or other complications. Although catheterization remains essential for many critically ill patients, it may be unsuitable or technically difficult in people with coagulopathy, infection, distorted anatomy, or limited venous access. Noninvasive alternatives based on physical examination or intermittent ultrasound can provide valuable information, but they are often operator-dependent and do not deliver uninterrupted monitoring. The new system is designed to bridge that gap by combining a neck-worn imaging device with automated interpretation.

At the center of the platform is an ultra-thin, 128-element linear-array ultrasound transducer developed to sit comfortably over the right side of the neck. The probe operates at a center frequency of 8.5 megahertz, a range that provides a balance between fine spatial resolution and sufficient penetration to visualize the internal jugular vein and common carotid artery. These vessels are important because the internal jugular vein connects directly to the right atrium without intervening valves. Changes in its cross-sectional area can therefore reflect variations in right-sided filling pressure and central venous pressure. The adjacent common carotid artery provides an anatomical reference that helps normalize measurements for differences in probe placement, body habitus, and neck geometry.

The patch’s engineering is intended to preserve image quality during prolonged use. Its acoustic structure includes a dual-layer matching system and a customized backing layer, producing an 85% fractional bandwidth and high sensitivity across a broad range of frequencies. A solid hydrogel coupling material maintains acoustic transmission between the probe and skin without requiring a liquid gel that can dry, leak, or become uncomfortable over time. The electronics and transducer are enclosed in silicone, allowing the device to remain attached while patients move or receive routine care. In feasibility testing, the patch was worn for as long as 24 hours while maintaining stable imaging and acceptable skin comfort, raising the possibility of near-continuous vascular surveillance rather than occasional manual examinations.

Continuous ultrasound, however, generates far more information than a clinician can realistically inspect frame by frame. The device produces cine-loop videos of the jugular vein and carotid artery, and each recording may contain thousands of individual images. To automate this process, the researchers created a semi-supervised artificial-intelligence model called the dual-decoder spatiotemporal attention network, or DSTA-Net. Instead of requiring experts to outline the vessels in every frame, the model uses manual annotations for only about 10% of the images. These key frames correspond to points at which the internal jugular vein is near its maximum or minimum dilation, providing highly informative examples of the vessel’s changing geometry.

DSTA-Net learns from the remaining unlabeled frames through a dual-decoder consistency strategy. A shared encoder first converts each ultrasound image into features that represent vessel boundaries, texture, and surrounding anatomy. Two separate decoders then interpret those features in complementary ways: one incorporates temporal attention to track how structures evolve across consecutive frames, while the other uses a lighter pathway to generate an independent segmentation. The model is trained to make the two pathways agree, allowing unlabeled images to serve as additional learning signals. This approach avoids relying on a continuously updated teacher model, a technique that can be vulnerable to unstable or incorrect predictions in noisy ultrasound data. By exploiting the natural temporal continuity of the cine-loop, the network can follow vessel motion while reducing the annotation burden for medical experts.

Testing indicated that the system could segment the internal jugular vein more accurately than several established deep-learning approaches. On an internal dataset, DSTA-Net achieved a Dice similarity coefficient of 83.5%, while its score on an external dataset was 75.8%. The Dice coefficient measures the overlap between the region identified by an algorithm and the region outlined by an expert, with higher values indicating closer agreement. According to the study, the model improved on the strongest fully supervised baselines, including UNet, Swin-UNet, and DeepLabV3+, by approximately 12 percentage points internally and 9 points externally. It also outperformed semi-supervised systems including UniMatch, DWL, and AllSpark. The model’s derived vascular measurements showed Spearman correlation values above 0.88 for most parameters, and Bland–Altman analyses indicated percentage errors well below the commonly cited 30% threshold for clinical agreement.

The segmented images were converted into five vascular indices: the maximum area of the internal jugular vein, its minimum area, the area of the common carotid artery, the ratio between the maximum jugular and carotid areas, and a jugular-vein area ratio reflecting dynamic changes over time. These measurements were combined with age, body mass index, blood pressure, and heart rate in a second artificial-intelligence system known as a dual-modality multilayer perceptron, or DM-MLP. Unlike image-focused architectures such as convolutional ResNets or vision Transformers, the DM-MLP was designed for structured clinical data. Its Attribute-Mixing operation models relationships among different clinical features, while Case-Mixing refines how each feature is represented across patients. The resulting low-rank architecture uses relatively few parameters while retaining the ability to capture nonlinear interactions, and it outperformed ResNet, DenseNet, and Transformer-based alternatives by roughly 4% to 8% in area under the curve.

The clinical evaluation included 349 intensive-care patients, with 272 enrolled at Shanghai Sixth People’s Hospital and 77 at Shanghai Tenth People’s Hospital. The principal target was elevated CVP, defined as a pressure of at least 8 millimeters of mercury. The model reached an AUC of 0.91 on the internal test set and 0.87 on the external test set, indicating strong discrimination between patients above and below the threshold. Additional analyses using thresholds of 7 and 9 millimeters of mercury produced similarly robust results. SHAP-based interpretability analysis suggested that the ultrasound-derived variables, especially maximum internal jugular vein area and the jugular-to-carotid ratio, contributed more strongly to the predictions than conventional variables such as blood pressure and body mass index. The system processed images at approximately 32 frames per second, or about one frame every 30 milliseconds, on a hospital server, enabling near-real-time analysis.

The researchers stress that the technology is not intended to eliminate central venous catheters in every clinical situation. Instead, it could provide a rapid, repeatable assessment when catheterization is contraindicated, delayed, or unnecessary, and could help identify rising venous pressure before a patient undergoes an invasive procedure. The study remains an early demonstration: its cohort is modest for training modern AI systems, it evaluates diagnostic performance rather than whether the technology improves survival or treatment decisions, and some aspects of the semi-supervised model remain difficult to interpret. Future work will expand the number of participating hospitals, test direct prediction of continuous CVP values rather than categories, add interpretability methods such as Grad-CAM, and assess whether AI-guided monitoring changes fluid and vasopressor management. By uniting wearable ultrasound, temporal image analysis, and clinical prediction, the platform brings automated noninvasive hemodynamic monitoring closer to routine use in acute and critical care.

Subject of Research: Wearable ultrasound and artificial intelligence for noninvasive central venous pressure monitoring in intensive-care patients.

Article Title: AI-Enabled Wearable Ultrasound for Noninvasive Central Venous Pressure Monitoring

News Publication Date: August 8, 2026

Web References: DOI: 10.34133/cbsystems.0653

References: Cyborg and Bionic Systems, “AI-Enabled Wearable Ultrasound for Noninvasive Central Venous Pressure Monitoring.”

Image Credits: Liping Zhang, Department of Emergency Medicine, Shanghai Sixth People’s Hospital, Shanghai Jiao Tong University School of Medicine.

Keywords: wearable ultrasound, artificial intelligence, central venous pressure, internal jugular vein, common carotid artery, intensive care, medical imaging, semi-supervised learning, DSTA-Net, DM-MLP, noninvasive monitoring, hemodynamics

The genesis of this innovative model stems from the critical need to improve nutritional protocols for mechanically ventilated patients, a population characterized by considerable physiological complexity and vulnerability. Previous approaches to predicting nutritional underfeeding have been hindered by static models with limited capacity to adjust predictions as patient conditions evolve. NutriSighT addresses this constraint by incorporating temporal dynamics through transformer-based architecture, a mechanism originally designed for natural language processing but now adapted impeccably for sequential medical data analysis.

At its core, NutriSighT employs multi-headed self-attention mechanisms, enabling the model to weigh multiple clinical variables over time and discern subtle yet pivotal trends indicative of nutritional insufficiency. This capacity for dynamic interpretation allows healthcare providers to anticipate underfeeding episodes before they escalate into critical complications. Furthermore, the model’s transparent decision-making process facilitates clinician trust and acceptance, an essential criterion often lacking in opaque “black-box” AI systems.

A key feature distinguishing NutriSighT from prior work is its dynamic prediction framework embedded within the transformer’s architecture. By continuously integrating incoming patient data — ranging from vital signs, lab results, ventilator settings to nutrition delivery records — the system updates risk assessments in real-time. This ongoing recalibration is crucial in ICU environments where patient conditions fluctuate rapidly, and timely interventions can drastically alter prognoses.

Extensive validation of NutriSighT was undertaken using diverse cohorts of mechanically ventilated patients across multiple medical centers, encompassing thousands of patient-days. This robust dataset allowed the researchers to rigorously evaluate the model’s predictive accuracy, temporal responsiveness, and interpretability. The results demonstrated that NutriSighT outperformed existing models, with superior sensitivity in detecting underfeeding risk and earlier recognition of nutritional inadequacy onset.

Interpreting the predictions generated by NutriSighT is facilitated through attention maps and feature attribution scores which highlight influential clinical parameters modulating the risk of underfeeding. This interpretable framework is transformative for critical care nutrition management because it empowers clinicians to understand the “why” behind each prediction, enabling tailored nutritional adjustments that optimize patient-specific therapy.

The application of this approach is transformative beyond ICUs solely designed for respiratory failure. The generalizability of NutriSighT’s transformer architecture opens avenues for adaptation in other complex, dynamic clinical scenarios where real-time nutritional support is pivotal, such as post-operative care or during severe systemic infections like sepsis. Such versatility exemplifies the potential for AI-driven nutritional optimization across a spectrum of critical care domains.

Moreover, NutriSighT stands at the interface of personalized medicine and machine learning, embodying a shift away from population-level guides toward individualized decision-making. By contextualizing vast arrays of clinical data into actionable predictions, this method aligns with the contemporary healthcare emphasis on precision interventions, promising measurable improvements in patient recovery trajectories and resource utilization.

The study also underscores the importance of multidisciplinary collaboration in developing practical AI tools. The research team combined expertise in critical care, nutritional science, machine learning, and clinical informatics to design an AI model that is both clinically relevant and technically sophisticated. Such integrative efforts ensure that innovative algorithms do not remain theoretical but translate into bedside utility.

Ethical considerations and patient safety lie at the forefront of NutriSighT’s deployment strategies. The transparency and interpretability of AI predictions mitigate risks associated with blind reliance on algorithmic outputs. Clinician oversight combined with model-guided alerts fosters shared decision-making, ultimately enhancing the standard of care delivered to vulnerable patient populations.

In conclusion, NutriSighT represents a monumental stride toward the convergence of AI and critical care nutrition. Its interpretable transformer model introduces a new paradigm of dynamic, real-time prediction of underfeeding, offering hope for reducing one of the persistent challenges faced by intensive care teams worldwide. As AI continues to permeate healthcare, innovations such as NutriSighT exemplify how thoughtfully designed technology can augment clinical judgment, improve outcomes, and pave the way for future breakthroughs in patient-centered care.

The implications of this research reach far into the future of critical care nutrition, potentially influencing nutritional guidelines and protocols internationally. With ongoing validation and integration into electronic health records, NutriSighT could become a foundational tool that continuously guides enteral feeding strategies across diverse ICU settings. This prospective transformation embodies a future where AI seamlessly collaborates with caregivers to provide nutrition that is as dynamic and complex as the patients it seeks to nourish.

As the healthcare community embraces such AI-enabled methodologies, continuous refinement through real-world feedback will be paramount. Future directions include expanding NutriSighT to incorporate multimodal data sources such as imaging and genetic profiles, thereby enriching its predictive capacity and expanding its clinical relevance. This trajectory firmly places Transformer-based models at the cutting edge of AI-driven healthcare innovation.

Significantly, NutriSighT offers a blueprint for developing interpretable machine learning models targeted at other dynamic, longitudinal medical challenges. It serves as a case study demonstrating how transformer architectures and attention mechanisms can be effectively repurposed beyond their original applications to solve pressing healthcare problems with transparency and precision.

The study’s success also highlights the transformative potential of leveraging big data and deep learning to extract actionable insights from the complexity of human physiology. It signifies an epoch where data-driven approaches not only enhance but redefine clinical nutrition, supporting practitioners in delivering timely, evidence-informed interventions.

In sum, NutriSighT sets a new standard in the predictive management of enteral nutrition, showcasing how advanced AI can be harnessed responsibly to augment clinical expertise. It emboldens a vision of care that is simultaneously personalized, proactive, and precise — essential qualities in the evolving landscape of critical care medicine.

Subject of Research: Dynamic Prediction of Underfeeding in Enteral Nutrition for Mechanically Ventilated Patients Using Interpretable Transformer Models

Article Title: NutriSighT: Interpretable Transformer Model for Dynamic Prediction of Underfeeding Enteral Nutrition in Mechanically Ventilated Patients

Article References:
Jangda, M., Patel, J., Vaid, A. et al. NutriSighT: Interpretable Transformer Model for Dynamic Prediction of Underfeeding Enteral Nutrition in Mechanically Ventilated Patients. Nat Commun 16, 11189 (2025). https://doi.org/10.1038/s41467-025-66200-1

Image Credits: AI Generated

DOI: https://doi.org/10.1038/s41467-025-66200-1