Walk into any ICU and listen. The beeping is constant. Monitors, ventilators, infusion pumps, bed alarms — the average intensive care unit generates between 150 and 400 alarms per patient per day, depending on the facility and acuity level. The vast majority of those alarms require no clinical action. Some estimates put the non-actionable rate above 90%.
That volume creates a well-documented problem: alarm fatigue. When nearly every alarm is a false positive, clinical staff stop responding with urgency. The Joint Commission flagged alarm-related patient deaths as a national patient safety concern back in 2013, and the problem has not gone away. A decade of alarm management initiatives has produced marginal improvements, but the underlying architecture — attach sensors, set thresholds, generate alerts when thresholds are crossed — remains the same.
Camera-based vital sign monitoring through remote photoplethysmography offers a different approach. Instead of layering more sensors onto an already overwhelmed patient, rPPG captures heart rate, respiratory rate, heart rate variability, and oxygen saturation from video alone. No wires. No adhesive patches. No alarm thresholds firing every time a patient shifts position and dislodges a lead.
"Up to 99% of alarms in ICU settings are non-actionable, leading to clinician desensitization and delayed response to genuine critical events." — Sendelbach and Funk, AACN Advanced Critical Care (2013)
The alarm fatigue problem in numbers
Alarm fatigue is not a perception problem. It has been measured extensively.
Cvach at Johns Hopkins published one of the definitive analyses in AACN Advanced Critical Care (2012), documenting that a single ICU bed can generate over 350 alarms per day. Most are caused by motion artifact, loose leads, or threshold crossings that resolve on their own within seconds. The clinical staff learns, rationally, that responding to every alarm wastes time. So they filter. They turn down volumes. They widen thresholds. And occasionally, a real alarm gets lost in the noise.
Drew et al. at the University of California, San Francisco, conducted a large-scale alarm audit in 2014, monitoring over 12,000 arrhythmia alarms across 461 patients. The false positive rate exceeded 88%. The study, published in PLOS ONE, found that nurses received so many false arrhythmia alerts that the alerts lost clinical meaning entirely.
The scale of the problem maps predictably to patient harm. A retrospective review by the ECRI Institute identified over 200 alarm-related deaths in FDA databases between 2005 and 2010. The real number is almost certainly higher — alarm-related adverse events are underreported because the causal chain from "ignored alarm" to "patient deterioration" is difficult to document cleanly.
| Alarm metric | Reported range | Source |
|---|---|---|
| Alarms per patient per day | 150-400 | Cvach, AACN Advanced Critical Care (2012) |
| Non-actionable alarm rate | 85-99% | Sendelbach and Funk (2013); Drew et al. (2014) |
| False arrhythmia alarm rate | >88% | Drew et al., PLOS ONE (2014) |
| Alarm-related deaths (FDA, 2005-2010) | >200 reported | ECRI Institute |
| Average nurse response time increase with alarm burden | 2-5x slower | Bonafide et al., BMJ Quality & Safety (2015) |
Bonafide et al. at Children's Hospital of Philadelphia published findings in BMJ Quality & Safety (2015) showing that alarm response times increased measurably as alarm burden rose. The more alarms a nurse had to process, the slower they responded to each one. This is not negligence. It is a predictable human response to information overload.
Why sensors cause part of the problem they are meant to solve
Most ICU alarms originate from contact-based sensors. ECG leads, pulse oximeter clips, blood pressure cuffs, and temperature probes all generate their own alarm streams. Each device has independent thresholds, and those thresholds do not talk to each other. A patient who rolls over might trigger a lead-off alarm from the ECG, a signal quality alarm from the pulse oximeter, and a motion artifact on the blood pressure monitor — three alarms from a single benign event.
The sensor-skin interface is itself a source of complications. Adhesive ECG patches cause skin breakdown in patients with fragile or edematous skin. Pulse oximeter clips can restrict circulation in digits. Patients who are confused or agitated pull at leads and cables, creating a cycle of disconnection, alarm, reconnection, and repeat.
Wearable devices for ICU use were reviewed in a 2025 study published in PMC by researchers examining alternatives to traditional bedside monitors. Their analysis of devices including smartwatches and patch sensors found that while wearables reduced some cable burden, they introduced their own alarm streams and still required skin contact. The review concluded that reducing the total number of contact points on a patient remains a meaningful goal for alarm management. The fewer things attached to a patient, the fewer sources of spurious alerts.
Camera-based monitoring: what it changes
Remote photoplethysmography works by analyzing subtle color changes in skin caused by blood flow with each heartbeat. A standard RGB camera pointed at a patient captures these changes in real time, extracting heart rate, respiratory rate, and derived metrics like HRV without touching the patient.
In the ICU context, this changes three things.
First, it eliminates motion-artifact alarms from dislodged contact sensors. If the monitoring system does not use contact sensors, patients cannot dislodge them. This removes what is arguably the single largest source of non-actionable alarms in intensive care.
Second, it provides continuous trend monitoring rather than threshold-based alerting. Instead of firing an alarm when heart rate crosses 100 bpm, a camera-based system can track the trajectory: heart rate has been gradually increasing from 78 to 95 over the past two hours, respiratory rate is trending upward simultaneously. That trend information is clinically more useful than a binary threshold alert, and it does not generate a blaring alarm every time the patient's heart rate fluctuates above and below a set number.
Third, it reduces the physical burden on the patient. ICU patients are often covered in monitoring hardware — five-lead ECG, pulse oximeter, arterial line, central venous catheter, blood pressure cuff, temperature probe. Removing even some of those contact points improves comfort and reduces skin complications.
Wang and Huang published a 2024 study in IETE Journal of Research investigating the use of existing CCTV infrastructure in hospital ICUs for contactless vital sign extraction. Their approach repurposed surveillance cameras already mounted in patient rooms, adding rPPG processing layers to derive heart rate and respiratory rate from the same video feeds used for patient safety observation. The implementation cost was minimal — the cameras were already there.
A double-center clinical study published in IEEE Journal of Biomedical and Health Informatics tested camera-based multi-parameter vital sign monitoring in ICU environments. The system achieved mean absolute errors of 1.89 bpm for heart rate and 19.04 ms for SDNN (a heart rate variability metric). These figures were obtained from real ICU patients in uncontrolled clinical conditions — not lab settings with cooperative volunteers.
Jorge et al. at the University of Oxford, publishing in npj Digital Medicine (2022), evaluated video-based monitoring of 15 postoperative ICU patients over an average of 16 hours each. Heart rate accuracy was 2.5 bpm MAE, respiratory rate accuracy was 2.4 breaths per minute. The system captured data across various patient positions and ongoing clinical care. After removing periods of privacy or low signal quality, usable monitoring data covered approximately 44-51% of total recorded time — a figure that is likely to improve as algorithms and hardware mature.
What happens when you combine camera data with existing monitors
The most practical near-term application is not replacing ICU monitors but layering camera-derived data alongside them to reduce false alarms.
Consider a common scenario: a patient shifts in bed and partially dislodges an ECG lead. The bedside monitor triggers a lead-off alarm and may also flag an arrhythmia alarm based on the corrupted signal. A camera-based system monitoring the same patient observes a stable heart rate and normal respiratory pattern through the video feed. Cross-referencing the two data streams reveals that the ECG alarm is an artifact. The alert can be suppressed or downgraded automatically.
This cross-validation approach was explored by researchers at JMIR Human Factors (2024), who designed a novel continuous vital sign viewer for ICU clinicians. Their system integrated multiple data streams into a unified display, allowing clinicians to compare trends across monitoring modalities rather than reacting to isolated threshold breaches from individual devices. The interface reduced cognitive load by presenting physiological data as trajectories instead of alarm events.
Park et al. (2025), publishing in JMIR, evaluated a remote early warning system that analyzed continuous vital sign streams from ward patients. Their framework generated rolling deterioration scores from heart rate, respiratory rate, and SpO2 data — the same parameters captured by rPPG. The system identified at-risk patients hours before traditional intermittent monitoring detected problems.
A systematic analysis in JAMIA Open (2022) evaluated 240 combinations of four vital signs for predicting clinical deterioration. Models combining heart rate and respiratory rate outperformed single-parameter models, and the researchers noted that continuous data substantially outperformed intermittent spot checks regardless of which parameters were included. Camera-based rPPG captures both heart rate and respiratory rate continuously, making it a natural fit for these multi-parameter deterioration models.
Skin integrity and infection risk in sensor-heavy environments
ICU patients are already at elevated risk for hospital-acquired infections, pressure injuries, and skin breakdown. Adding contact sensors to compromised skin makes all of these worse.
Medical adhesive-related skin injuries (MARSI) affect an estimated 1.5 million patients annually in the United States, according to data from the National Pressure Injury Advisory Panel. ECG patches, securing tape for IV lines, and adhesive-backed sensors contribute to this number directly. In patients with edema, malnourishment, or immunocompromise — common in ICU populations — even short-duration adhesive contact can cause skin tearing and breakdown.
Central line-associated bloodstream infections (CLABSIs) and catheter-associated complications are among the most closely tracked hospital-acquired conditions. While vital sign monitoring devices are not central lines, the principle applies broadly: every device touching a patient is a potential vector. Reducing contact points is part of the infection prevention toolkit.
Camera-based monitoring eliminates these risks for the vital sign parameters it measures. A camera does not touch the patient's skin. There is no adhesive to remove, no lead wire to clean, no sensor housing to disinfect between patients (though the camera itself still requires standard cleaning protocols).
Technical barriers in the ICU environment
ICU environments are harder for camera-based monitoring than general wards or home settings. Several factors complicate rPPG signal extraction.
Lighting conditions in ICUs vary significantly. Rooms are often dimmed during nighttime hours to support patient sleep cycles, but reduced lighting degrades the rPPG signal. Infrared illumination can solve this — most security cameras already include IR LEDs — but the interaction between IR light and rPPG algorithms designed for visible-spectrum video requires additional validation.
Patient occlusion is frequent. ICU patients wear oxygen masks, have endotracheal tubes, are partially covered by blankets, or have their faces turned away from the camera. Jorge et al. (2022) reported usable data in roughly half of total monitoring time. That gap needs to narrow before camera-based monitoring can serve as a primary or sole monitoring modality.
Skin tone bias remains an unresolved challenge. Dasari et al. at Carnegie Mellon University, publishing in npj Digital Medicine (2021), documented measurable accuracy differences across skin tones in rPPG measurements. ICU populations are diverse, and any monitoring system deployed broadly must perform equitably across all patient demographics.
Debnath and Kim (2025), reviewing 145 articles in BioMedical Engineering OnLine, found that deep learning methods have consistently outperformed traditional signal processing for rPPG accuracy. The improvement trajectory is steep — newer models trained on larger, more diverse datasets show narrowing gaps across skin tones and lighting conditions. But "narrowing" is not "eliminated."
Where ICU monitoring goes from here
The economic pressure on hospitals to address alarm fatigue is real. The Joint Commission's National Patient Safety Goals include alarm management requirements. CMS reimbursement penalties for hospital-acquired conditions create financial incentive to reduce device-related complications. And nursing shortages mean that the cognitive load from alarm processing falls on fewer people.
Camera-based monitoring enters this context not as a replacement for arterial lines or continuous ECG, but as a supplementary layer that handles basic vital sign surveillance without adding to the alarm noise or the sensor burden. Heart rate trending, respiratory rate tracking, and HRV monitoring can move from wired sensors to passive video capture for many ICU patients, particularly those whose acuity does not demand beat-to-beat invasive hemodynamic monitoring.
A clinical trial registered at ClinicalTrials.gov (NCT05202769) is evaluating non-invasive vital sign monitoring using video-based systems in ICU patients, testing rPPG accuracy under varying conditions of lighting, position, and concurrent clinical care. The results will add to the growing evidence base for camera-based ICU applications.
A separate trial (NCT07307521) is evaluating AI-assisted video monitoring for predicting clinical deterioration in ICU patients — combining the continuous data capture of camera-based systems with machine learning for early warning.
Circadify has developed camera-based vital sign technology that captures heart rate, HRV, respiratory rate, and SpO2 from a standard device camera. The parameters this technology measures overlap directly with the vital signs driving ICU alarm burden. As the evidence base for camera-based ICU monitoring grows, the application pathway becomes increasingly clear.
Frequently asked questions
What is ICU alarm fatigue and why does it matter?
Alarm fatigue occurs when ICU staff become desensitized to clinical alarms due to their sheer volume. Studies estimate that 85-99% of ICU alarms are non-actionable, meaning they do not require clinical intervention. This desensitization leads to delayed responses to genuine alerts — the Joint Commission identified alarm fatigue as a top patient safety concern, linking it to preventable patient deaths.
Can camera-based monitoring replace bedside monitors in the ICU?
Not currently. Camera-based rPPG monitoring is positioned as a supplementary layer rather than a replacement for bedside monitors. It adds continuous trend data and can reduce the number of contact sensors attached to a patient, but ICU-grade hemodynamic monitoring — arterial lines, central venous pressure — still requires invasive or contact-based devices.
How accurate is rPPG monitoring for critically ill patients?
Clinical studies report heart rate measurement accuracy within 1.9 to 2.5 bpm mean absolute error compared to contact references. Respiratory rate accuracy falls within 2.4 breaths per minute. These figures come from studies conducted in clinical environments with real patients, though accuracy can decrease with patient movement, low lighting, or skin that is partially obscured.
Does contactless monitoring reduce ICU-acquired infections?
Removing adhesive sensors and contact-based devices from a patient's skin eliminates points of entry for pathogens and reduces skin breakdown. While direct studies linking rPPG specifically to infection reduction are still emerging, the clinical logic is straightforward: fewer devices touching the patient means fewer opportunities for device-associated complications.