Creating new algorithms is a core part of Bodytrak team’s in house software development capability, with the scope to focus our development work on truly differentiating technology. The most recent of these algorithms is potentially one of our most exciting, with enormous potential to advance connected health and safety monitoring.
The Bodytrak Man Down Detection System
Bodytrak has developed a comprehensive man down system using solely kinetic data (no cameras like some of the typical best performing systems) which incorporates a system for detecting slips, trips and falls, as well as a tuneable inactivity monitoring system to detect an extended period of low or no activity. The full algorithm is built around two system modules: a module for detecting different types of falls, and a module for monitoring inactivity.
Both modules can operate in tandem and provide an alert via a remote monitoring platform. With the man down system, both falls and inactivity are being
observed in conjunction with other biometrics. This results in providing a very comprehensive snapshot of the person wearing the Bodytrak device and
far exceeds typical non-biometric fall detection systems. For example, by monitoring physiological parameters such as HR in conjunction with fall detection
we are able to paint a more detailed picture into the overall well-being of the individual following a fall and provide a more useful indication of
the severity of the man down alert to the remote monitoring station for an appropriate response to the situation.
As with all algorithms accuracy and reliability are crucial to success criteria in which existing man down solutions are judged. The Bodytrak algorithm uses machine learning and simulated data (gathered from practice falls) to train a neural network.
Supervised machine learning algorithms feed labelled data into the neural network and through “backward propagation”, gradually improve the classification
performance. For this to be successful the labels or markers which will generate the alert must be highly reliable and consistent. Bodytrak gathered
a considerable amount of data with specific activities being performed to train this network, which included close to 1,000 falls that needed individually
One of the biggest advantages of this system is that it has been designed based on actual fall and inactivity data, and bespoke to the Bodytrak device,
unlike many ‘one-size fits all’ threshold based algorithms which are currently out in the market. . The algorithm in its current state is around 95%
accurate at correctly identifying when a person has fallen and the final posture post fall, and is around 98% accurate for detecting truly inactive
periods. These statistics are in line with those found within the literature, where accelerometer based motion detection systems can claim upwards
of 95% accuracy, although the studies seen tend to focus on detecting falls performed by elderly subjects.
Going forward the benefits of using machine learning will be further enforced, as data is obtained from users in targeted environments. This can be fed
back into the system and the module can be re-trained to continuously improve results.
Inactivity Module - a closer look
It is worth noting the Inactivity Module within the man down system is tuneable; the level of inactivity and the amount of time deemed inactive can be adjusted to meet customer or industry specific requirements. On a standard basis however, the Bodytrak algorithm is optimised to accurately differentiate between classes of:
- Static - little movement - eg. Standing/sitting still
- Inactive - no movement - eg. Unconscious
- Active - any movement - eg. Running
The module's ability to distinguish between 'static' and 'active' is relatively simple, where very little overlap exists. The Inactivity Module’s ability
to identify the subtle differences between 'static' and 'inactive' is its real strength.
Going forward, the team at Bodytrak are very excited for the next phase of our man down and fall detection algorithm development as our hardware evolves
to incorporate more sophisticated sensors. Stayed tuned for our progress in the coming months!