Sensor informatics for wearable healthcare systems

Hospital patients currently have their vital signs (heart rate, breathing rate, blood pressure, oxygen saturation, temperature, and level-of-consciousness) measured by clinicians every 4-8 hours for patients outside the ICU. This leads to delays in detecting when a patient's condition is deteriorating, or missing deterioration entirely, which can lead to patient mortality or unplanned admission to the ICU.
Mortality rates for unplanned admissions to the ICU are extremely high (up to 50%). Accurately predicting patient deterioration, and adjusting treatment to prevent them, promises to improve patient outcomes and reduce NHS expenditure (whereby emergency treatment of deteriorated patients is extremely costly). In order to achieve this aim, patients need to be continuously monitored, and the sensors that do this need to be unobtrusive and wearable, because the majority of hospital patients are ambulatory.
One of the main challenges associated with using wearable sensors in a hospital setting is that of battery life. A device being used on a ward needs to last at least a 72-hour shift (to be usable for a week-end), but constantly collecting and streaming data drains the battery in wearable devices quickly. This project will focus on designing smart algorithms which will enable devices to transmit data only when deemed necessary by using on-board sensor informatics; for example when the patient's vital signs suggest they might be at risk. Predictive algorithms are often computationally expensive (for example, if they use Gaussian processes, as is the current state-of-the-art), but this is not feasible to implement on a wearable device. The device should carry out simpler calculations in order to decide whether to stream data to another platform (such as a nearby smartphone or hospital server). Thus, there is a problem of distributed inference, in which we must determine the optimal balance of computation taking place at the device, on nearby devices (such as bedside tablets / smartphones), and on the hospital cloud. Within this, there are problems concerning the optimal representation of data at all stages in the inference pipeline, from the wearable to the hospital cloud. Introducing a wearable sensor system into hospitals has many other challenges: from patient comfort to the reliability of the devices. The project will be highly collaborative, with insight from hardware engineers, human factors experts, and world-leading clinicians from the Oxford University Hospitals NHS Foundation Trust.
The initial phase of the project will involve exploring machine learning techniques that could be used in the sensors, these will be trained and tested on data which is already available from previous clinical trials. This will inform how the data should be collected, and will identify potential concerns arising from the use of different physiological sensors. Research will be carried out into which devices would be most appropriate to embed the algorithms that this DPhil programme will produce. In order to ensure these devices can best aid clinicians on the ward, specification of the system to be developed will be based on advice from medical experts and reviews from previous trials.
A shortlist of devices will then be rigorously tested in the lab and some will be chosen for a trial. This will allow the devices to be tested in a real hospital setting, and data will be collected for training and improving initial algorithms. Towards the end of the three-year doctoral project, the aim is to have designed a wearable sensor system suitable for integrating into a hospital environment, which can then be quantitatively assessed in a final trial. Funding for this trial exists via the recently-awarded NIHR Biomedical Research Centre.
This project falls within the EPSRC healthcare technology research area, and will contribute towards making the NHS more effective and sustainable, as it adapts to cope with increasing numbers of chro