Developing digital biomarkers from raw accelerometer data


Proof of Concept for the Development of Digital Biomarker using Raw Accelerometer Data

Research based on the use of accelerometer based wrist-worn devices in clinical trials. 

Accelerometers have the ability to capture significant quantities of raw data, potentially containing patterns which, if discoverable, could be used to quantify specific motor movements. The ability to detect such movements has value in identifying periods of tremor or other neuromuscular disorder. We conducted a proof of concept for developing digital biomarkers from raw accelerometer data that we presented as a poster entitled 'Proof of Concept for the Development of Digital Biomarker using Raw Accelerometer Data' from a Wrist Worn Device at the DIA US meeting in 2016.

There is increasing interest in the use on accelerometer based wrist-worn devices in clinical trials. What isn’t well know is that wearables have been around for over 40 years with the first research papers citing the use of accelerometer based wrist-worn devices to capture human activity and sleep patterns appearing in the 1970’s. Accelerometers are electromechanical devices that measure acceleration forces. They have a wide variety of applications and are found in phones, in drones for flight stabilisation, are used by seismologists to detect earthquakes and are key components in the deployment of air bags where they are used to detect car crashes. The most common types used in wearables today are triaxial MEMS accelerometers;  small micro electro-mechanical systems (MEMS), that measure acceleration in 3 dimensions and are very sensitive to human motor movements. These devices are capable of sampling up to 100 hertz- which results in 100 data points been collected every second. Potentially one device worn continuously for 7 days could produce 60million datasets. The management of this level of data was beyond both the memories and the analytics capabilities of these devices. Originally the data generated was compressed and filtered by on-board firmware into epochs of 30seconds[1]*. This pre-processing significantly reduces the amount of data so in the example above, where the device is worn for 7 days, the output is  20 thousand datasets rather than 60 million.  This is still a sizable amount and further analysed by validated algorithms generates discrete sleep and activity endpoints. From a data management perspective, within a clinical trial, the data from accelerometers can be managed in a similar fashion to blood pressure or temperature with 7 days of accelerometer data producing a finite number of clinically relevant endpoints. The use of accelerometer data in this manner is well established, although limited to researchers and clinicians, focusing on the quantification of physical activity and sleep patterns. What is less well understood is the additional value that potentially exists in the raw accelerometer data, which may have potential in the monitoring and assessment of specific motor movements such as tremor. For this experiment we selected tooth-brushing as a surrogate for tremor as it was easy to reproduce the action and the movement was discrete and finite. Context is essential in the management of raw data and we used a diary to timestamp the periods of interest, this allows for the identification of the signal of interest. The training data used to generate the algorithms was different to the data periods that were tested. A set of classical machine models (decision trees, random forest, etc.) were built on the training dataset. A number of algorithms were identified and tested the best one had a positive predicative value of 0.9991.      

This Proof of concept has demonstrated that it is possible to use machine learning techniques to train a classification model from summarized raw accelerometer data to identify periods of specific movement patterns. There are issues regarding scalability but the potential value of accelerometer based devices to provide insights into human physical behaviours is considerable. The advent of Artificial Intelligence and machine learning platforms will transform the ability of these devices to identity new and exploratory endpoints in patients with neuromuscular disease. This approach has potential application in objectively measuring motor movement events in neuromuscular disorders but also in the development of unique personal digital fingerprints. 

Reference:
[1] Other epochs can be selected, earlier models generated data in 1 minute epochs,  30 seconds are most common in sleep studies, with shorter epochs used in the assessment of physical activity.