W 46: Proof of Concept for the Development of Digital Biomarker using Raw Actigraphy Data from a Wrist Wearable Device
Safira Clinical Research Ltd. Ireland
This poster reports the results of a proof of concept project developed by ICON Plc’s Innovation Team for the development of a digital biomarker for identifying periods of teeth brushing activity from raw 100Hz Actigraphy data.
Actigraphy data was gathered from a subject over the course of ~18hrs. The measurement period contained 7 teeth brushing events. A digital fingerprint for was developed and tested using 4 of these events and finally evaluated using the remaining 3 unseen events.
When considering raw actigraphy data, there is very little value in an individual row of data (the actual accelerometer results for an individual moment in time) – the value comes from considering the pattern or relative change in the data over a period of time, an epoch. Another important consideration is the variation and duration of the event itself. A rolling 1 minute epoch has been selected for this project.
The raw data for each row was summarised into a single figure value, the vector magnitude for the tri-axial accelerometer values, along with the quadrant direction of the vector.
When rolling the raw data up into 1 minute epochs, a number of summary statistics were calculated, along with some descriptive parameters about the pattern of the vector magnitude – such as smoothed peaks per second and average period of peaks.
A digital fingerprint was derived that most closed matched the 4 training events.
A function was defined which calculates the similarity of a 1 minute epoch to the fingerprint and was calibrated on the training events.
Whilst it was possible to obtain a very good fit on the 4 training events used to define the digital fingerprint – it proved difficult to obtain similar levels of accuracy (and particularly specificity) when evaluated on the 3 unseen events.
The limiting factor is a combination of the small number of events to develop a fingerprint from and the degree of variation between all 7 of the events, due to the largely stochastic nature of a teeth brushing event.
A final evaluation test was performed on an additional set of data containing 2 events gathered from a different subject. When evaluated on this dataset, no sufficiently matching patterns were found. Closer investigation showed that the pattern for the new subject, whilst similar, was outside of the similarity bounds defined by the test events.
The raw data collected by actigraphy devices can contain activity patterns not currently captured by existing validated algorithms.
With a sufficient amount of labelled data (i.e. where time periods with the event of interest occurring are marked) it is possible to derive a digital biomarker to measure when and how quantify the number of events occurred.
The quantity of data required to build a robust algorithm will depend on the variance in the pattern between individual events and from subject to subject.
Notwithstanding the limitations in evaluation performance, the project demonstrated that it is possible to derive a new digital biomarker from raw actigraphy data using a small training set.
The following individuals collaborated to develop this abstract:
Louis Smith MSc Data Science Manager, ICON
Marie McCarthy MSc, MBA, Director of Product Innovation, ICON
Michael Philips Phd Director of Product Innovation, ICON
Wilhelm Muehlhausen DVM VP Head of Innovation ICON