Sepsis is a life-threatening response of the immune system to infection that can lead to tissue damage, organ failure and the risk of permanent disability or death. It has been reported as the leading cause of in-hospital death and, in the US alone, an estimated 1.5 million sepsis hospital stays cost more than $27 billion to treat.
The future of patient monitoring
Detecting and treating it early can be highly effective in stopping progression. The trouble is that implementing the level of monitoring required to catch sepsis at the earliest stages is tough outside critical care settings. With innovation poised to intervene, I’m going to explore how the emerging non-contact monitoring technologies have the potential to save lives as well as healthcare costs.
Early detection of sepsis is vital
Sepsis doesn’t wait. It can kill in 12 hours and every hour of delayed antibiotics treatment increases mortality by 10%. To help with crucial early detection, a range of early warning scores (EWS) have been developed. They work by measuring a panel of patient attributes, assigning a risk score to the measured value and summing the scores to give an overall assessment of risk. The problem is that they rely heavily on busy front-line care staff identifying and keeping to a suitable frequency of intermittent measurements.
A key drawback here is that deterioration can happen at any time, while continuous patient monitoring should offer the most rapid identification of risk. But conventional monitoring of parameters such as ECG, pulse oximetry and continuous blood pressure uses products that are unacceptably uncomfortable or complex for use outside critical care. EWS also use an unsophisticated summation of expert opinion based risk indices for a general population rather than something tailored to the individual’s risk factors or baseline readings. A data analytics approach is likely to be able to stratify risk much more efficiently and effectively.
Wireless monitoring patches brought to the market to address these issues have had limited commercial success. The price point is too high, and they typically monitor a limited range of physiological signs over a period of just a few days due to trade-offs around power requirements and battery size. Some data supports their superiority over traditional approaches in terms of faster time to respond to a sepsis indication with antibiotic administration, reduced hospital length of stay and lower readmission rates. There is also interest in the use of wearables in this space, but consumer wearables typically favour battery life over medical grade performance.
The advantages of non-contact monitoring
What then, does the ideal solution look like? From my viewpoint it’s a monitoring system that can frequently measure a broad panel of parameters, does not constrain the patient, is sufficiently accurate to detect trends and avoid false alarms and is at a price point that can be deployed widely across lower acuity settings.
Encouragingly, we are seeing the emergence of non-contact monitoring technologies that are full of potential. The advantages include:
- The patient is not restricted
- Intrinsically low per-patient cost when amortised over years of use, therefore scalable
- Avoids the cost of a complex, single-use consumable
- Avoids the requirement of cleaning and potentially re-sterilisation of a wearable
- Does not have the power management issues of patches and wearables
- Is compatible with a business model shift from consumables sales to a service model generating potentially valuable data
If we want success, what do we monitor?
Many EWS systems have the underlying rationale to identify patterns of significant change in core vital organ functions. These changes could indicate the onset of a serious systemic problem and are relatively easy to measure or assess in a general clinical environment without specialist equipment, monitoring or the need to take frequent blood samples.
The National Early Warning Score (NEWS) is a good example of such a scoring system that has been validated as being useful for the early identification of patient deterioration. The monitored parameters, all physiologically associated with early signs of the potential onset of sepsis, are respiratory rate, oxygen saturation, blood pressure, heart rate, temperature and level of consciousness and confusion. Of these, blood pressure and consciousness seem particularly challenging to monitor remotely.
Blood pressure is difficult to measure without direct patient contact. It can, however, be a lagging indicator of a patient being haemodynamically compromised and other correlates may be more useful for identifying early deterioration. Recent research has explored the usefulness of heart rate variability and pulse transit time in monitoring haemodynamics in sepsis patients, with both measurements lending themselves better to non-contact methods.
Levels of consciousness and confusion are measured as part of the NEWS by assessing responses to verbal or physical stimulus. There’s a lot of interpretation involved, and the approach doesn’t lend itself to be adapted to continuous monitoring. There are, however, several other cues that could be used to gain trend data and insights on patient status. Sleep monitoring systems have been developed that use a range of inputs including movement, heart rate, respiratory rate and eye movement to infer not just whether subjects are asleep but also which part of the sleep cycle they are in.
Delirium associated with acute illness is characterised by noticeable physically overactive or underactive states which could be picked up by monitoring. While no single measure is expected to be an adequate monitor of consciousness level, the combination of data from a relatively few sensors together with algorithms is likely to be able to pick out concerning trends in patient status and differentiate these from normal sleep states.
The technology is there, but so are the challenges
Enabling technologies for non-contact monitoring can be broadly split into three categories; sensing, signal processing and location. Patient monitoring requires finding a property that can be measured using a suitable sensor and then processing the signal to give a reading that correlates to the target patient condition. Freeing the patient from the monitoring system places an additional burden of identifying where to gather the signal from.
Let’s start with sensing. Suitable non-contact technologies are probably limited to those based on electromagnetic radiation or sound. The increasing use of cameras in mobile phones and time-of-flight sensing cameras in automobiles is driving performance while reducing cost. Vital signs monitoring systems have been demonstrated using a range of active and passive electromagnetic technologies, from microwaves through to visible light. In addition to spectral response, electromagnetic radiation can be used to measure characteristic movements relating to pulse and breathing.
For example, photoplethysmography (PPG) has been used since the 1930s for contact monitoring. The method uses a light source and a photodetector to monitor the microvascular blood volume changes in the skin caused by the pulsatile nature of the circulatory system. The 1990s saw the development of imaging PPG (IPPG), with camera-based sensors that could be used away from the body. It’s been successful in measuring basic physiology parameters with passive light, capturing systolic and diastolic peaks from the pulse. Several academic research groups have shown the technique works on mobile phones, portable devices and standard webcam to monitor heart rate, heart rate variability and respiration rate.
More can be achieved by introducing advances in the system beyond mobile cameras and passive light. Oxygen saturation has been confirmed with active light. With the addition of a high-speed camera, the pulse transit time (PTT) between the face and the palm is observable, giving further insight into the patient’s haemodynamic status.
Outside the visible spectrum, infrared motion and depth sensing and millimetre wave and microwave doppler measurements have all been demonstrated for application in vital signs monitoring. Sound doesn’t seem to have been explored as thoroughly. We have demonstrated the deployment of very low power algorithms to characterise the sound characteristics of respiration in a patient contacting product, and there has been some interesting work on the use of ultrasound to monitor respiratory airflow in sleep studies. But it’s clear that picking up specific patient sounds in a general ward environment is challenging. Almost certainly it will require fusion with other data sources to show, for example, that a sound such as respiratory noise correlates to a breathing movement.
Highly efficient algorithms are on the way
Imaging techniques bring a heavy signal processing burden. With ‘always on’ monitoring, the large amount of data needs to be rapidly processed very efficiently. In addition, the signal is often hidden in the background and data processing is required to extract important information, identify key trends and spot anomalies. Machine learning techniques underpinned by physiological models are increasingly being used to develop highly efficient algorithms for this purpose, suitable to be deployed on a low power on-device processor.
Non-contact systems bring further significant practical challenges around locating where to measure a signal from. The clinical context fortunately means that patients are usually confined to bed, although they are free to move around within this space. Many sensing technologies need positional tracking of the patient to determine the location for the sensor reading. In addition, reflection off the skin and a busy background are factors that add noise to the collected signals. This means technologies like facial and body tracking, landmark feature recognition – such as forehead for heart rate or centre of the body for breathing – and background subtraction are needed to ensure system viability.
Ideally, a passive system with spatial resolution will be enough for the task, but in some cases actively tracking the patient and directing the sensor may be required to give adequate performance. A further consideration is the sheer processor power needed for broad field high resolution image processing. This can limit the frame rate to a frequency too low to reliably pick out the required signal, so selective filtering to only process the areas of interest may be necessary.
Building a robust solution will mean selecting a set of sensing, signal processing and location technologies that cover the target vital signs while building redundancy into the system.
Translating potential into optimised, data-driven care
Enabling technologies are an encouraging start for a monitoring system. A key concern with such a device is whether the measurement technology would be accurate enough to be clinically useful. In the first instance, it’s likely that such a system would be a trend monitoring adjunct to current EWS practice rather than a replacement. The clinical utility would be faster identification of deterioration between scheduled clinical evaluations and the value proposition based on avoidance of the costs and poorer clinical outcomes associated with the slow detection of sepsis onset.
A system for measurement of multiple parameters will inevitably use a range of sensing modalities. Fusion of data from multiple complementary sensors will allow robustness to be built into the measurement algorithms. Such a system would need to have a high specificity when trading off against sensitivity – it would be preferable to wait and see on marginal cases rather than significantly increase the level of false alarms.
From a data analytics perspective, a mixed model of continuous monitoring and intermittent capture of standard clinical measurements and observations as truth data has interesting possibilities for the development of algorithms. Also, linking monitoring data to clinical events and outcomes will support the development of predictive algorithms from the generated data stream.
While the regulatory implications make ‘live’ learning systems unlikely in the near future, the ability for device manufacturers and hospitals to refine and validate both measurement algorithms, predictive models and care indices offline would be a powerful enabler for data-driven optimisation of care.