Researcher develops generative learning model to predict falls

In a study published in the journal Information Systems Research, Texas Tech University’s Shuo Yu and his collaborators developed a generative machine learning model to detect instability before a fall occurs. The hope is that the model could work within fall detection devices, such as anti-fall airbag vests or medical alert systems, to minimize injuries, increase emergency response effectiveness and lower medical costs.

“You can treat this as a kind of AI (artificial intelligence),” said Yu, Wetherbe Professor of Management Information Systems in the Area of Information Systems and Quantitative Sciences at the Jerry S. Rawls College of Business. “It detects your moving status and predicts if there’s going to be a fall. It can help mitigate injuries automatically.”

To create the model, Yu and his collaborators worked within two publicly available datasets that used wearable motion-sensor devices to monitor nearly 2,000 falls. They combed through the datasets and labeled individual data points. They then grouped those points into snippets and determined three hidden states of a fall: collapse, impact and inactivity.

Think of an elevator. A person standing in an elevator car is in a normal state. The button is pressed and the doors shut. With the sudden upward acceleration of the elevator, there’s a slight loss of weight. This immediate feeling, milliseconds into the ride, is the collapse phase.

That loss of weight happens in falls, and it’s exactly where Yu and his team focused their attention.

“Those milliseconds are what matter,” Yu said. “You need time for the data to process and to inflate the airbags or activate other protective equipment. All those milliseconds matter when you’re trying to improve this process.”

Rather than follow much of the past research that relied on simple rule-based models, Yu and his collaborators created a new model which includes a hidden Markov model with generative adversarial network (HMM-GAN).

HMM is a statistical model for understanding sequences over time and consists of two types of variables: observations and hidden states. In this instance, motion data was used to mark the observations and hidden states.

GAN is a machine learning model consisting of two parts: a generator that tries to create realistic fake data and a discriminator that tries to tell the difference between real and fake data.

Combined, HMM-GAN works to understand what a fall looks like in the form of data snippets, even if the movements and phases vary quite a bit from person to person. It also tries to predict when someone is likely to fall based on recent movement patterns.

Across four experiments, the HMM-GAN model accurately predicted falls and did so faster, outperforming previous frameworks.

For senior citizens and their families, this new model could provide increased peace of mind, knowing that fall detection devices could be deployed faster. The researchers note that hospitals or other facilities where patient falls are common would also benefit from this new model.

The researchers ran a simple case study to see how their model could potentially reduce catastrophic falls by senior citizens and any subsequent medical costs. The result was more than $33 million of economic benefits over competing models.

“I feel very happy seeing these results,” Yu said. “It’s still a proof-of-concept, but if this work can lead to future research in engineering departments or related fields and can be turned into actual products, that would be the best.”

Yu also hopes his work can lessen some of the anxieties surrounding AI.

“I think that’s the future of health,” he said. “We already have AI components in our lives like ChatGPT. I believe, in the future, this kind of device can come into existence and improve lives in a physical manner.”