Wearables, AI closely track arterial BP in ICU

Clinical takeaway: Sensor-based reconstruction is approaching arterial line fidelity, making wearables a potentially viable option for BP monitoring in a variety of settings beyond consumer use. 

Continuous pressure monitoring in the ICU means accepting a tradeoff, because no better option exists. A catheter threaded into the radial or femoral artery gives a reading with every heartbeat. It also shows the shape of each pulse, which clinicians use to judge perfusion and titrate vasoactive drugs. But it risks bleeding, clots, and infection, and it keeps the patient in bed. 

An automated arm cuff carries none of that risk, but offers no information between readings. A single-center feasibility study tested a two-sensor setup, chest ECG plus a finger sensor, as a potentially less invasive but more informative middle ground. 

Estimates landed within 6.42 mmHg of catheter readings on average, with systolic off by 6.21 mmHg and diastolic by 3.41 mmHg. Overall, the model accounted for roughly 73% of the variation in the catheter signal. Diastolic tracking was the tightest of the three. 

Performance came close to a ceiling model built from clinical-grade bedside ECG and finger signals, which points to the algorithm rather than the sensors as the limiting factor. The model also produced calibrated confidence ranges around each estimate, giving a bedside user some basis for knowing when a number should not be trusted. Commercial cuffless devices generally offer nothing comparable. 

Of 66 patients enrolled across surgical, cardiovascular surgery, and cardiac care units at Johns Hopkins Hospital, 28 contributed analyzable data, yielding 15,489 five-second ECG and photoplethysmography (PPG) recordings. A hybrid deep learning model trained on those segments, but only after they passed recording-quality filters. The authors are explicit that this makes their numbers an upper bound rather than deployment-ready performance. 

The team is validating the sensors and algorithm in a larger cohort of Johns Hopkins ICU patients, and plans to test performance across the full range of real-world unit conditions rather than quality-screened recordings alone. A further question is whether the model detects hemodynamically significant excursions, which matters more clinically than average accuracy. 

Longer term, the researchers describe extending monitoring to general wards and to outpatients with hypertension, comparable to how continuous glucose monitoring changed diabetes care. Nothing in this study supports that step yet. 

"We reconstruct waveform data in a way that's meaningful, accurate, reliable and, most importantly, non-invasive," said senior author Robert Stevens, MD, MBA, chief of the Division of Informatics, Integration, and Innovation at Johns Hopkins Medicine. "It's a possible solution for avoiding the current standard of care for measuring blood pressure, arterial lines, a very invasive procedure with a risk of many complications." 

Source: Harris C, et al. (2026 Aug 15) Comput Biol Med. Non-invasive arterial blood pressure waveform generation in critically ill patients: A sensor-based deep learning approach