About 550,000 Americans receive maintenance hemodialysis, most of them three times a week for four hours per session. That adds up to roughly 86 million dialysis sessions annually in the United States alone. During each session, blood is drawn from the patient, filtered through a dialyzer, and returned. The process removes excess fluid and waste, but it also stresses the cardiovascular system in ways that are difficult to predict. Blood pressure can drop suddenly. Heart rate can spike. A patient who was stable 10 minutes ago might become symptomatic without warning.
The standard monitoring protocol in most dialysis centers involves a blood pressure cuff measurement every 15 to 30 minutes. Between those checks, nobody is watching the patient's hemodynamics. That gap is where complications happen, and researchers are now testing whether continuous and contactless monitoring technologies could close it.
"Patients with intradialytic hypotension had earlier and more profound reductions in systolic and diastolic blood pressure during treatment. Nearly all advanced vitals differed between groups." — Ilan et al., Biomedicines (2024)
The intradialytic hypotension problem
Intradialytic hypotension is the most common serious complication during hemodialysis. Depending on how you define it, IDH occurs in 15% to 50% of all sessions. The Kidney Disease Outcomes Quality Initiative (KDOQI) and European Best Practice Guidelines define IDH as a systolic blood pressure drop of at least 20 mmHg or a mean arterial pressure reduction of 10 mmHg, combined with clinical symptoms requiring nursing intervention.
The consequences go well beyond the immediate session. Repeated IDH episodes are associated with myocardial stunning, a form of cardiac injury where regions of the heart temporarily lose function due to inadequate blood flow during dialysis. Researchers at the University of Nottingham, including work by McIntyre and colleagues, have documented that this cardiac injury accumulates over time, contributing to the high cardiovascular mortality rate among dialysis patients. Brain and gut ischemia can also occur during severe IDH events.
The underlying mechanism is straightforward in principle but hard to predict in practice. During ultrafiltration, fluid is removed from the blood. If the removal rate exceeds the rate at which fluid refills the vascular space from surrounding tissue, blood volume drops and pressure follows. But individual patients respond differently based on cardiac function, vascular tone, hydration status, and dozens of other variables that change session to session.
How dialysis patients are monitored today vs. what's emerging
| Monitoring approach | What it measures | Frequency | Contact required | IDH detection capability | Current status |
|---|---|---|---|---|---|
| Standard cuff BP (current standard) | Systolic/diastolic BP | Every 15-30 min | Yes, arm cuff | Detects IDH after it occurs | Universal in dialysis centers |
| Crit-Line blood volume monitor | Hematocrit changes (fluid proxy) | Continuous | Inline sensor | Indirect, shows fluid trends | Available (Fresenius) |
| Wearable PPG device (e.g., Biobeat) | HR, cuffless BP, SV, CO, SVR | Continuous (every 5s) | Wrist/chest patch | Early hemodynamic changes | Research phase (Ilan et al., 2024) |
| ML prediction from dialysis machine data | IDH risk score | Real-time | None (uses existing data) | 15-75 min advance warning | Research phase (Yun et al., 2024) |
| rPPG camera-based monitoring | HR, RR, SpO2 estimate | Continuous | None | Heart rate and respiratory changes | Early research, not yet dialysis-specific |
| Implantable hemodynamic monitors | Pulmonary artery pressure | Continuous | Implanted device | Direct hemodynamic measurement | Limited use, high cost |
The gap between current practice and what researchers are testing is wide. A blood pressure reading every 15 minutes gives you 16 data points across a four-hour session. A wearable device recording every 5 seconds gives you roughly 2,880.
Wearable continuous monitoring during dialysis
The most direct evidence for continuous monitoring in dialysis comes from Ilan et al., published in Biomedicines in 2024. Their team used a PPG-based wearable device (Biobeat BB-613WP) to continuously monitor hemodialysis patients, recording heart rate, cuffless systolic and diastolic blood pressure, stroke volume, cardiac output, and systemic vascular resistance at 5-second intervals.
Across 98 dialysis sessions, IDH occurred in 22 sessions (22.5%). The wearable blood pressure readings correlated well with cuff-based measurements (r > 0.62, p < 0.001). More importantly, the continuous data revealed that patients who eventually developed IDH showed earlier and more pronounced drops in blood pressure than the periodic cuff measurements caught.
This finding matters because it suggests the hemodynamic deterioration leading to IDH doesn't happen suddenly. It builds over time, and periodic spot-checks miss the trajectory. Continuous monitoring could give clinicians enough lead time to reduce the ultrafiltration rate, administer fluid, or take other preventive measures before the patient becomes symptomatic.
Machine learning prediction of hemodynamic instability
A parallel line of research skips new hardware entirely and instead applies machine learning to data the dialysis machines already collect. A study published in Nephrology Dialysis Transplantation by researchers including work at Taichung Veterans General Hospital developed ML models to predict IDH 15 to 75 minutes before onset using standard dialysis session parameters.
Yun et al., highlighted in Kidney News (2025), demonstrated that explainable deep learning models could provide real-time predictions of both intradialytic hypotension and hypertension. Their approach used time-series data from dialysis machines, including blood flow rates, ultrafiltration volumes, and venous pressures, to flag patients heading toward hemodynamic instability.
A systematic survey published in the International Journal of Advanced Computer Science and Applications (2024) reviewed multiple ML approaches and identified LightGBM as particularly effective for IDH prediction. The review noted that gradient-boosted models consistently outperformed logistic regression and basic neural networks on this task, likely because they handle the mixed numerical and categorical features in dialysis data well.
The practical appeal of ML-based prediction is that it requires no new equipment. Dialysis machines already record the relevant data streams. The prediction model runs on top of existing infrastructure.
Where contactless camera monitoring fits
Camera-based rPPG has not yet been specifically validated in dialysis settings. But the clinical rationale for applying it there is strong. Dialysis patients sit in recliners for 3-4 hours, relatively still, facing forward. This is close to the ideal scenario for rPPG signal acquisition, which struggles most with patient movement and unpredictable lighting.
A camera mounted above a dialysis station could continuously track heart rate and respiratory rate without any device touching the patient. In a setting where patients already have a fistula needle in one arm and often a blood pressure cuff on the other, removing the need for additional contact sensors has obvious practical value.
The Oxford University Engineering Science department, through work by Clifton and colleagues on continuous vital sign monitoring in renal patients, laid early groundwork showing that physiological instability during hemodialysis follows detectable patterns. Their research focused on automated alerting systems that could identify deterioration from continuous vital sign streams, which is conceptually aligned with what rPPG could provide.
The 550,000-patient scale problem
The United States has approximately 7,800 dialysis centers. Most run three shifts per day, six days a week. A typical center has 15-30 treatment stations. The staffing model relies on a ratio where one nurse oversees multiple patients simultaneously, monitoring them visually and through periodic vital sign checks.
This creates a surveillance problem that technology could address. One nurse cannot watch eight patients' hemodynamics simultaneously. But a camera system generating continuous heart rate and respiratory rate data, processed through algorithms that flag concerning trends, could function as an early warning layer.
The economics of dialysis also matter here. Medicare spent approximately $37 billion on dialysis-related care in recent years. IDH events lead to extended treatment times, emergency interventions, and in some cases hospitalization. If continuous monitoring prevented even a fraction of severe IDH episodes, the cost savings could justify the investment.
Infection control considerations
Dialysis centers already face serious infection control challenges. Patients with compromised immune systems sit in close proximity. Equipment is cleaned between patients but shared across shifts. Any monitoring approach that reduces the number of physical touchpoints, devices applied to skin, or pieces of equipment transferred between patients has infection control benefits.
Contactless camera monitoring eliminates the device entirely from the patient's body. There's nothing to clean between patients, nothing to apply or remove. For a population already vulnerable to bloodstream infections and other healthcare-associated infections, that matters.
What still needs to happen
Camera-based vital sign monitoring in dialysis faces the same validation challenges as in other clinical settings, plus a few specific to this population. Dialysis patients are often elderly, frequently have darker skin tones (given the disproportionate burden of kidney disease on Black and Hispanic populations), and may have anemia or fluid shifts that alter the rPPG signal.
Skin tone bias in rPPG algorithms remains an active concern. Nowara et al. at Rice University documented reduced accuracy on darker skin tones across multiple rPPG algorithms. For a technology intended to monitor a population where racial health disparities are already severe, addressing this bias isn't optional.
Blood pressure estimation from rPPG is also less mature than heart rate measurement. Since IDH is defined by blood pressure thresholds, a camera system that only tracks heart rate and respiratory rate provides useful but incomplete information. Combining camera-based heart rate monitoring with ML models running on existing dialysis machine data could offer a more complete picture than either approach alone.
The most likely near-term path is a layered system: existing dialysis machine sensors providing fluid removal and pressure data, ML models processing that data for IDH prediction, and camera-based monitoring adding continuous heart rate and respiratory rate without requiring additional contact devices. No single technology solves the problem. The combination might.
Circadify has developed camera-based vital sign measurement technology and is working to bring it to various clinical environments. Dialysis monitoring represents a setting where the combination of long session duration, limited patient movement, and clear clinical need for continuous surveillance aligns well with rPPG's current capabilities.
Frequently asked questions
Why is continuous monitoring important during hemodialysis?
Intradialytic hypotension occurs in 15-50% of hemodialysis sessions and is linked to increased morbidity and mortality. Standard monitoring checks blood pressure every 15-30 minutes, which can miss rapid drops between measurements. Continuous monitoring could catch hemodynamic changes earlier and give clinicians time to intervene.
Can cameras measure vital signs during dialysis treatment?
rPPG technology can estimate heart rate and respiratory rate from facial video. Blood pressure estimation from video remains an active research area with limited clinical validation. Camera-based systems could supplement existing dialysis machine sensors rather than replace them.
What is intradialytic hypotension and why is it dangerous?
Intradialytic hypotension is a drop of at least 20 mmHg in systolic blood pressure during dialysis, often accompanied by symptoms like dizziness, nausea, or cramping. Repeated IDH episodes are associated with cardiac injury, brain ischemia, and increased mortality risk over time.
How accurate are machine learning models at predicting intradialytic hypotension?
Recent studies show machine learning models can predict IDH 15-75 minutes before onset. Yun et al. demonstrated that deep learning models achieve useful predictive accuracy using real-time dialysis machine data, though clinical deployment is still in early stages.
Related articles
- What is rPPG Technology? — A complete overview of remote photoplethysmography and how it measures vital signs from video.
- Contactless Heart Rate Monitoring — Detailed analysis of camera-based heart rate measurement accuracy and applications.
- Contactless Blood Pressure Measurement — How rPPG-based blood pressure estimation works and current research validation.