The system, called MOSAIC, is designed to address a longstanding problem in critical care. Patients in intensive care units can experience sudden and dangerous changes in blood pressure, sometimes within minutes. A pressure that rises too high may increase the risk of stroke, heart attack, and kidney injury, while a pressure that falls too low can reduce blood flow to the brain and other vital organs. Detecting these changes early is often essential, but the current method for doing so continuously requires placing a catheter directly inside an artery.

Arterial lines provide highly detailed, beat-by-beat measurements and are widely regarded as the clinical reference standard for continuous blood-pressure monitoring. Yet inserting an arterial catheter is an invasive procedure. The catheter can cause bleeding, clot formation, infection, or damage to the artery, and it restricts a patient’s movement. The line must also be carefully maintained and monitored by clinical staff. These drawbacks mean that continuous arterial monitoring is generally reserved for patients whose condition is serious enough to justify the risks.

Conventional arm cuffs avoid those risks but introduce a different limitation: they measure blood pressure only intermittently. A cuff inflates periodically, records a pressure reading, and then remains inactive until the next measurement. That approach can miss rapid fluctuations between readings. It also tends to interrupt sleep, limit movement, and provide clinicians with isolated numbers rather than a continuous picture of how pressure changes from one heartbeat to the next.

MOSAIC uses two small sensors to capture the physiological signals needed to estimate blood pressure without entering the bloodstream. One sensor is positioned on the chest to record the heart’s electrical activity, similar to an electrocardiogram. The second is placed on a finger, where it detects changes in blood volume as pulses of blood move through the circulation. This optical or pulse-related signal provides information about the timing and shape of blood flow. Together, the signals describe how electrical activation of the heart relates to the arrival and movement of blood through the body.

The system then sends those measurements to a deep-learning model. Rather than producing only a single systolic and diastolic value, the algorithm reconstructs a blood-pressure waveform, a continuously changing curve that represents pressure over time. The waveform contains information about each heartbeat, including the rise and fall of arterial pressure and the timing of pulse transmission. In principle, this gives clinicians a richer view of cardiovascular changes than a sequence of occasional cuff readings.

The researchers trained and evaluated the system by comparing its reconstructed waveforms with measurements from arterial catheters already being used in patients at Johns Hopkins Hospital. In the initial group of 28 intensive care patients, the signals generated by MOSAIC closely matched the waveforms recorded by the invasive lines. The result does not yet establish that the wearable system can replace arterial catheters in routine care, but it indicates that signals collected from the skin may contain enough information for artificial intelligence to estimate arterial pressure with clinically meaningful detail.

The technical challenge is substantial because blood pressure is influenced by many factors, including heart rhythm, vascular resistance, blood volume, body position, temperature, medications, and movement. The relationship between a pulse detected at the finger and pressure inside a major artery can change from person to person and over time. A deep-learning model can identify complex patterns in the electrical and pulse signals that are difficult to capture with a simple equation, but its performance must remain reliable across different patients, illnesses, skin conditions, sensor placements, and levels of physical activity.

“We reconstruct waveform data in a way that’s meaningful, accurate, reliable and, most importantly, non-invasive,” said senior author Robert Stevens, chief of the Division of Informatics, Integration, and Innovation at Johns Hopkins Medicine. Lead author Carl Harris, a biomedical engineering doctoral student, said the team was seeking an alternative to arterial lines because of their risks of bleeding, clotting, and infection.

The investigators are now testing MOSAIC in a larger cohort of intensive care patients. That next phase will be important because a study of 28 people cannot fully capture the range of physiological conditions encountered in critical care. Larger validation studies will need to determine how accurately the system performs during severe infection, shock, irregular heart rhythms, rapid changes in medication, low blood flow to the extremities, and patient movement. Researchers will also need to establish how the algorithm behaves when sensor signals are noisy or temporarily interrupted.

If the technology continues to perform well, its uses could extend beyond intensive care. A non-invasive continuous monitor could allow clinicians to follow patients in regular hospital wards without transferring them to an intensive care unit simply because they need closer blood-pressure observation. It might also enable earlier recognition of deterioration in patients who are recovering from surgery or serious illness, while allowing them greater freedom to move than an arterial catheter permits.

The researchers also envision a future in which people with hypertension wear similar sensors for extended periods outside the hospital. High blood pressure is among the world’s most common health conditions and a major contributor to cardiovascular disease, stroke, and kidney damage. Current home monitoring usually relies on occasional cuff measurements, which can be useful but may not reveal how pressure changes during work, exercise, sleep, stress, or ordinary daily activity. Continuous data could show patterns that are invisible in isolated readings and might help physicians adjust treatment more precisely.

That possibility resembles the transformation brought by wearable glucose monitors, which allow people with diabetes to observe changes in blood sugar throughout the day rather than relying only on occasional tests. Continuous blood-pressure monitoring could eventually provide a similar window into cardiovascular health, although the comparison should be treated cautiously. Blood-pressure measurement is technically complex, and a device intended for medical decisions must demonstrate accuracy and reliability across a wide population before it can be used broadly.

The system could also help researchers study blood pressure in healthy people. Most detailed blood-pressure data currently come from patients who are already ill or undergoing medical procedures. Less is known about how pressure changes in people as they commute, work, exercise, sleep, and interact with their environment. Long-term wearable monitoring could reveal previously overlooked patterns and improve understanding of how everyday behavior and stress affect cardiovascular risk.

For now, MOSAIC remains an investigational technology rather than a replacement for the arterial line. Its encouraging first test shows that wearable sensors and artificial intelligence can reproduce invasive blood-pressure waveforms with surprising fidelity, but larger studies and rigorous clinical validation are still required. If those studies confirm the initial results, the technology could shift continuous blood-pressure monitoring away from the artery and toward the skin, bringing real-time cardiovascular information to more patients in hospitals and, eventually, at home.

Subject of Research: Non-invasive continuous blood-pressure monitoring using wearable sensors and artificial intelligence.

Article Title: Wearable Sensors and Artificial Intelligence Bring Continuous Blood Pressure Monitoring Out of the ICU

Web References: Johns Hopkins University research announcement; https://www.sciencedirect.com/science/article/abs/pii/S0010482526004257

References: Computers in Biology and Medicine, DOI: 10.1016/j.compbiomed.2026.111861

Keywords

Wearable sensors, artificial intelligence, deep learning, blood pressure, continuous monitoring, arterial line, intensive care, hypertension, biomedical engineering, cardiovascular health

The ICU represents one of the most complex and resource-intensive environments in healthcare, where clinicians must balance an overwhelming array of variables to optimize patient outcomes. Often, the rapid progression or regression of a patient’s condition demands immediate adjustments in treatment strategies. However, current monitoring methods, while comprehensive, are primarily retrospective or reliant on early warning scores that may not fully capture the nonlinear, multifaceted changes occurring in critically ill patients. This new predictive system seeks to fill this crucial gap by integrating diverse physiological signals and treatment data into a unified computational framework that can anticipate patient deterioration or improvement in real time.

At the heart of the system lies state-space modeling, a mathematical technique that describes the evolution of a system’s internal states, which may not be directly observable, through their relationships to measured outputs. In the context of ICU patient monitoring, the internal state corresponds to the underlying physiological status of the patient, while the observed outputs are manifestations such as vital signs, laboratory results, and administered therapies. By applying this model, the researchers were able to create a dynamic representation of patient health that continuously updates as new data is collected, providing a live picture of patient acuity beyond traditional snapshots.

The researchers sourced a rich dataset combining multiple modalities, including cardiovascular measurements, respiratory parameters, biochemical markers, and therapeutic interventions. Through their model, they estimated hidden physiological states that reflect the patient’s intrinsic severity and trajectory. This approach effectively filters out noise and transient fluctuations, isolating clinically meaningful trends that can inform timely interventions. Importantly, the model accounts for the inherent uncertainty in measurements and physiological variability, enabling robust predictions even in the face of incomplete or noisy data typical of chaotic ICU environments.

Key to the practical application of this model is its capability to generate not only estimates of current patient acuity but also forecasts of subsequent therapy requirements. This predictive power allows clinicians to anticipate changes in treatment needs—such as escalation of ventilatory support or initiation of hemodynamic therapies—well before traditional signs manifest. Early identification of impending deterioration could prompt preemptive adjustments, ultimately mitigating complications and improving outcomes. Conversely, the system can suggest when therapy de-escalation is safe, potentially reducing unnecessary interventions and associated risks.

One of the most remarkable aspects of the research is the model’s adaptability to individual patient trajectories. Unlike static scoring systems that apply uniform thresholds, this state-space framework dynamically calibrates itself based on personalized data streams, reflecting the unique physiological response patterns of each patient. This personalized approach addresses a longstanding challenge in critical care: heterogeneity among patients in terms of age, comorbidities, and disease progression. By tailoring predictions to each individual, the system enhances clinical relevance and precision.

Implementing such a system requires integration with electronic health records and bedside monitoring devices, enabling seamless data acquisition and processing. The researchers demonstrated proof-of-concept integration within a high-fidelity clinical environment, highlighting the feasibility of real-time application. They emphasized the importance of human-centered interface design, ensuring that predictions and alerts generated by the model are presented in an actionable, interpretable manner to support rather than overwhelm critical care teams.

Beyond immediate clinical utility, the insights generated through this state-space model have broader implications for healthcare resource management. By identifying patients at high risk of deterioration early, ICUs can better allocate personnel and equipment, potentially reducing length of stay and optimizing bed utilization. The ability to predict therapy trajectories also holds promise for benchmarking and quality improvement initiatives, offering granular feedback on patient responses to various interventions.

This innovative platform marks a significant advance in the field of precision medicine within critical care. By transitioning from episodic assessments to continuous, predictive monitoring, it aligns with emerging paradigms that emphasize preemptive and customized treatment strategies. Though challenges remain—such as ensuring data privacy, addressing model interpretability, and validating generalizability across diverse populations—the foundational demonstration lays a robust pathway for future clinical translation.

In terms of technical sophistication, the researchers employed advanced filtering algorithms such as the Kalman filter and particle filters to update the latent state estimates iteratively. These algorithms excel in handling stochastic processes and measurement noise, which are omnipresent in physiological data. The model was trained and validated on large datasets encompassing thousands of ICU admissions, enabling rigorous evaluation of predictive performance. Performance metrics demonstrated superior accuracy and timeliness in predicting patient acuity changes compared to standard early warning scores or clinician judgment alone.

The model is not merely a black-box predictor. The framework elucidates underlying physiological mechanisms by modeling interactions among clinical variables over time, offering a window into patient-specific pathophysiology. This level of interpretability is crucial for clinical acceptance as it provides rationale behind predictions, fostering trust and enabling expert oversight. By mapping latent states to clinically meaningful constructs, the system serves as an intelligent ally in critical care rather than a cryptic oracle.

Looking ahead, the researchers envision expanding this state-space approach to incorporate additional data sources, such as imaging and genomic information, further enriching the patient model. Coupling with emerging wearable technologies could extend continuous monitoring beyond the ICU, supporting transitions of care and early discharge planning. Moreover, integrating this platform with automated therapeutic delivery systems paves the way for closed-loop critical care, where diagnostic and treatment decisions are increasingly driven by real-time analytics.

The broader implications of this research resonate beyond critical care units. Real-time predictive modeling using state-space frameworks has potential applications in chronic disease management, emergency medicine, and even public health surveillance. As healthcare moves towards data-centric, algorithm-guided practice, this work exemplifies the transformative power of mathematical modeling married with clinical expertise.

Ultimately, this landmark study heralds a new era where the complex, nonlinear dynamics of human physiology are no longer barriers but gateways to smarter, anticipatory medicine. By enabling clinicians to see into the near future of patient trajectories, this technology promises to elevate the standards of ICU care, reduce preventable harms, and improve survival and recovery for the most vulnerable patients. As such systems gain traction, the intensive care landscape may be poised for a renaissance of precision, personalization, and foresight.

Article Title:
Real-time prediction of intensive care unit patient acuity and therapy requirements using state-space modelling

Article References:
Contreras, M., Silva, B., Shickel, B. et al. Real-time prediction of intensive care unit patient acuity and therapy requirements using state-space modelling. Nat Commun 16, 7315 (2025). https://doi.org/10.1038/s41467-025-62121-1

Image Credits: AI Generated