Clinical trials: how technology is driving digitisation.

The rising cost of clinical trials (CTs), coupled with the commercial risks of failure, mean it’s increasingly important to harness the power of real-world data to inform effective trial designs, better understand outcomes and maximise the likelihood of clinical success.

Chronic stress is a multi-factorial and poorly characterised state which underpins many health problems and impacts the patient’s behaviour (adherence), their physiological response to drugs (metabolism) and their overall wellness. Assessing stress levels unobtrusively to provide clinical insight and influence trial design is imperative for more efficient and successful clinical trials.

So, we decided to develop a solution that would reshape the way we design and conduct clinical trials, using a multi-modal platform to:

  • Capture an individual’s physiological, psychological, behavioural and contextual data through sensors and an app;
  • Create a single numerical estimator of the person’s stress levels via machine learning; and
  • Display the AI driven insights within an clinical trials management system (CTMS) dashboard.

When starting this project we consulted the scientific literature to see what is state-of-the-art in stress research. Although many articles described sensor systems that could be used, none provided insight around which sensor sets constitute the optimal stress correlation system, which is the goal of our project. Finding the best combination depends on the level of insight required by the target audience; decision support information required by clinicians will differ to that of a patient or consumer offering. The latter may require occasional insights and these could be captured using smartphone sensors (e.g. Blood Volume Pulse (BVP), voice and pupil tracking).

Continual and daily average monitoring of physiological parameters (e.g. electrocardiogram (ECG) and electrodermal activity (EDA)) for clinical use requires medical grade sensors that must be able to support the generation of suitable features for training the machine learning algorithm. This in turn needs good noise and interference tolerance and suitable sample rates. Our system uses 500 samples per second where possible in combination with sensors that measure occasionally. When we set out to determine the optimum biometric combination for stress measurement, it soon became apparent that what works in the laboratory does not easily translate in the real world, due to a range of technical and methodological constraints reducing the impact of sensors.

Let’s look at a few different sensor types to illustrate this point, starting with ones that make occasional measurements.

Many researchers agree that stress changes voice pitch and speaking rate, but most of the research done in stress recognition from voice has been carried out in laboratory. It has caused interest primarily because it can be easily measured (people are used to speaking with chat-bots such as Alexa) and it can be implemented in a completely unobtrusive way. If such a test were to be incorporated into an app the voice-based stress analysis can be ineffective both in quiet and noisy spaces, due to the lack of speech recordings and excessive noise respectively. Thought must also be given to where the data is stored and where the processing happens. Edge processing (e.g. on a smartphone) can minimise privacy concerns but has to be implemented with care to minimise battery drainage. In the case of Alexa, the detection of the wake word ‘Alexa’ is implemented on the device itself, with subsequent audio sent to the cloud for analysis thereafter.

BVP is the measurement of the volume of blood that passes over a photoplethysmographic (PPG) sensor with each pulse. This is used to determine Heart Rate Variability (HRV), a popular proxy measurement for stress, and is incorporated into health and wellbeing apps by leveraging the camera and flashlight of smartphones. Data capture involves simply placing a finger over the camera and holding still for a period of time. However, the PPG readings are not very reliable because of a variety of factors, such as:

  • Smartphone models vary, leading to different colour saturation in the captured frames;
  • The user must remain completely still, as finger movement creates motion artifacts which reduces accuracy; and
  • Different fingertip pressure force as well as different features of the tissue change the level of light absorption when it passes through the finger, thereby causing different colour ratios.

Different issues manifest themselves when trying to obtain high quality continual measurements. EDA is known to correlate in a real-time linear fashion with stress, but the placement of the sensor can pose issues. The most sensitive readings occur at the thenar eminence (flesh of palm below thumb) and yet we found it almost impossible to keep a sensor attached for more than an hour without contact being lost due to sweat and hand movements dislodging the sensor. Placing it at the wrist does work but is a compromise as it reduces signal by 50 fold.

Given the multi-modal nature of stress, it is key to develop a system that can fuse information from a variety of sources to increase accuracy of inferred stress. Over the coming weeks we’ll be describing more of the challenges and insights we encountered in defining and designing this system with the aim to demonstrate it at CES between 8-11 January 2019.

Jaquie Finn
Head of Digital Health

With decades of experience in molecular biology, data science, product management, marketing and strategy, Jaquie has significant international business experience, particularly in APAC and the America's.