Down, But Not Out

It is always a tragic situation when — despite the many technological advancements that have been made in recent decades — physicians have nothing to offer people suffering from serious medical conditions. But in many ways it is far more tragic when effective treatments are available, yet are not administered in time because the underlying condition was not detected until it was too late.

Perhaps the most avoidable of all serious medical conditions is falls. This is especially true among older adults, where approximately one in four people in this group falls each year. These falls can result in consequences ranging from broken bones to brain injuries and even death. Given the severity and frequency of falls among the elderly, many studies have been conducted to seek out ways to minimize the negative consequences of these events.

Of course all cases cannot reasonably be prevented without placing unacceptable restrictions on the freedoms of these individuals. But it has been noted that when a fall does occur, there is a critical time period of approximately one hour, during which outcomes can be greatly improved if care is provided. Accordingly, if we can, at a minimum, detect a fall the moment that it happens, many of the worst outcomes can be avoided.

Fall detection is most definitely possible these days, with some commercial smartwatches even boasting such features. However, these devices can be on the pricey side, and that prevents them from being widely adopted — especially in the developing world. Engineers Shebin Jacob and Nekhil R put their heads together and came up with a solution that could make fall detection more accessible than it is today. They built an inexpensive, yet very capable, device that can be worn like a wristwatch. When this device detects a fall, it immediately sends a text message to alert first responders or other medical professionals.

The hardware consists only of a Particle Photon 2 Wi-Fi development kit and an ADXL362 accelerometer, with a 400 mAh LiPo battery to provide power. The hardware is housed in a 3D-printed case and attached to a standard watch wristband. A small push button was also included in the build to give users a simple way to interact with the device.

The team’s plan was to use the accelerometer to continually capture motion data from the wearer of the device, then run a machine learning algorithm on the Photon 2’s powerful processor to detect when that data is consistent with the characteristics of a fall.

Building, training, optimizing, and deploying a machine learning algorithm can be pretty challenging, so the team decided to work with the Edge Impulse platform to simplify the entire process. Next, an existing dataset consisting of accelerometer data from people that were either going about their normal daily routines, or falling in a number of different ways, was located and uploaded to Edge Impulse.

That raw data was exactly what was needed to train a classification model to learn the difference between falls and normal activities. A temporal convolutional neural network, in particular, was built and trained as these types of algorithms are especially good at classifying time series data of this sort. The suitability of the model for the task was on full display after the training process completed — an accuracy level of nearly 99 percent had been achieved on the first attempt.

The Photon 2 is supported by Edge Impulse, so the full classification pipeline was packaged up as a downloadable archive for this target, making deployment simple. The team then integrated Twilio into the inference code to enable the device to send an SMS alert the moment that a fall is detected.

This may be a fairly simple device — but that is the point. By keeping costs down and working with highly accessible hardware, this device could conceivably find its way onto the wrists of millions of at-risk individuals. And that could help to reduce the impact of one of the greatest problems facing older adults today.