Data Analytics and the IoTUK Test Beds
This blog forms part of a series that looks at the technical experiences of three IoTUK projects that ran between 2016 and 2018.
Each of the three projects implemented their own IoT technology platform and in this blog we’ll explore how they approached data analytics. You will learn how data from disparate sources is interpreted using machine learning to generate automated alerts.
CityVerve is the UK’s smart cities demonstrator, based in Manchester. CityVerve explored how IoT technologies and the forging of new relationships between the public and private sector can make a city truly smart. The resulting solution was a Platform of Platforms connected by a central portal that exposed data from different elements of the city. The data was open to 3rd party developers allowing them to create new and innovative apps.
Diabetes Digital Coach (DDC) is an NHS IoT Test Bed project and has been developed by consortium of 10 technology and evaluation partners, led by the West of England Academic Health Science Network (AHSN). DDC offers an online service to help people with Type 1 and Type 2 diabetes manage their condition and cut the risk of complications. It is available via computer, smart phone and tablet and brings together a number of digital self-management tools, which provide personalised support. When people sign up for an account with the Diabetes Digital Coach, they provide information about themselves, their health, lifestyle, and how they currently manage their diabetes. The Coach then suggests the most appropriate tools to suit their individual needs. The Coach features five tools which have been carefully selected by both healthcare professionals and people with diabetes; this ‘menu’ covers self-management education, dietitian support, optimising physical activity, insulin and glucose management, and a personal health record.
Technology Integrated Health Management (TIHM) is an NHS IoT test bed aiming to transform support for people with dementia and their carers. The collaboration involves partners from the health, voluntary and technology sectors. Each family in the trial was provided with a home technology pack that was suited to their particular needs. Like the DDC portal, TIHM connects different technology partners but here there is real-time correlation between different data points allowing insights to be gleaned. Any unusual signs are flagged to clinicians who will decide on the appropriate action to take.
TIHM was the only one of the three IoTUK supported projects to apply machine learning techniques on raw disparate healthcare data to generate notifications based on a patient needs. Creating real-time human and/or machine understandable responses was the main analytics challenge in this project.
To provide meaningful and actionable information to clinicians, the raw collected data needed to undergo:
- Pre-processing – The collected data was first validated to comply with the FHIR4TIHM data model and then stored in a Mongo database. Queries were run on the database and algorithms applied to clean, aggregate and filter the raw data.
- Analysis – Techniques including abstraction, pattern detection, machine learning, and adaptive thresholding were employed to extract meaningful information and to provide interpretation for (near) real-time analytical processing. The team were looking for signs of Agitation, Irritability and Aggression (AIA) in patients. By training a number of different machine learning models with real and synthetic data, the team found the accuracy of AIA detection varied significantly from model to model. Full details of the machine learning exercise can be found in this paper. The end result was that raw sources of data could be confidently translated into actionable knowledge to help improve decision making for clinicians. This enhanced the human experience in healthcare and targeted resources towards needs at an earlier stage.
- Interpretation – The output of the data analytics tool was composed of:
- Reminders (for patients) – The reminders were generated if pre-planned measurement were missing and/or incomplete. These messages were generated by the machine learning algorithms. Thus, corresponding companies could notify patients and their carers about re-assessments.
- Flags (for monitoring teams) – Generated flags covered three types of notifications, including ‘Clinical’, ‘Environmental’, and ‘Technical’. Note that each flag contained three main components to provide comprehensive information about:
- the corresponding source of measurement, i.e. user or device
- the pre-set standard thresholds for that measurement and an informative message for the monitoring team
- the severity of flag which is defined in terms of percentage.
- Once a flag had been generated, the monitoring team would take appropriate actions according to its type and severity. The following is a subset of the flags that were raised.
- Technical – Battery level and connectivity status of the implemented sensors
- Clinical – Predictive models were developed using a combination of sources, e.g. environmental and medical sensors could be used to inform the clinical team about the potential Urinary Tract Infection (UTI) by monitoring the body temperature and the number of times a patient used the bathroom.
- Environmental – Detecting the daily routine using data captured from passive environmental sensors and wearable technologies. Unusual behaviour such as falls or wandering could be determined by fall detector and GPS trackers.
- To distinctly display the above information and allow for real-time response mechanisms, a user interface was designed specifically for the TIHM project.
Common patterns across the IoTUK projects
In the DDC project, analytics to enhance user insights were explored with user focus groups and project partners but were not tested within the project
At CityVerve, the option to analyse and act upon the city data was left up to 3rd party developers. There was one exception to this relating to the city’s bike sharing scheme. Analytics were used to accurately determine the number of bikes in circulation.
TIHM showed that it is possible to collect clinical, wearable and environmental data from patients’ homes to learn about their daily patterns. Machine learning algorithms were developed to understand individual patient behaviour but it was important to try a number of techniques to identify the model that performed best. This was the first study of its kind so data was not readily available to train the chosen model, synthetic data was created by the team to overcome this challenge.
The model developed for the project is shown in fig 1:
This decision fusion model is a framework for detecting AIA in people with dementia. The main stages of the decision-making framework are:
- Sense Layer – Collect, pre-process and aggregate data from all the sensors in the trial.
- Decision Layers I & II – Collected data is categorised and analysed to identify anomalies and a decision score is applied.
- Decision fusion – Optimally fuses the decision scores from previous layers so that the final decision is less sensitive to bias and variance.
Full details of the TIHM decision fusion model can be found here. It is important to apply a number of steps when designing a decision-making model and to continuously test it using real data (ideally) or synthesised data.
TIHM used their own servers to test the models but quite often, access to sufficient computation power is a challenge for SMEs. At Digital Catapult, the Machine Intelligence Garage has been created to help startups for whom access to computation power is a barrier to growth.
This was the last of the blogs in this series and we’d love to hear from you especially if you’re about to embark on a large scale IoT solution.
If you want to read the whole series, start here.