Researchers develop a generation learning model to predict falls

Research published in the journal Research on information systemsShuo Yu and his collaborators of Texas Tech have developed a generated machine learning model to detect instability before a fall occurs. The hope is that the model may work within fall detection devices such as anti-fall airbag vests and medical alert systems to minimize injuries, increase the effectiveness of emergency responses, and reduce medical costs.

“We can treat this as a kind of AI (artificial intelligence),” said Yu, professor of management information systems in the field of information systems and quantitative science at Jerry S. Rawls College of Business. “Detects movement status and predicts whether a fall occurs. It helps to automatically reduce injuries.”

To create the model, Yu and his collaborators worked within two public datasets that monitor nearly 2,000 falls using wearable motion sensor devices. They aggregated the datasets and labeled them as individual data points. They then grouped those points into snippets to determine three hidden states: collapse, impact, and inactivity.

Think of an elevator. People standing in the elevator car are in normal conditions. The button is pressed and the door closes. The sudden upward acceleration of the elevator causes a slight loss of weight. This immediate sensation lies in the ride, in the collapse stage.

That weight loss occurs in a fall, and that's exactly where Yu and his team focused their attention.

“These milliseconds are important,” Yu said. “It takes time for the data to be processed and the airbags to inflate and other protective gear. All of these milliseconds are important when trying to improve this process.”

Rather than following many of the previous studies that relied on simple rules-based models, Yu and his collaborators created a new model that included a hidden Markov model with generative hostile networks (HMM-GAN).

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

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

HMM-GAN works to understand how it looks in the form of a data snippet, even when movements and stages vary considerably from person to person. They also try to predict that someone is likely to fall based on recent exercise patterns.

Over the four experiments, the HMM-Gan model accurately predicted falls, surpassing previous frameworks, making them faster and faster.

For seniors and their families, this new model can increase peace of mind knowing that autumn detection devices can be deployed faster. Researchers point out that hospitals and other facilities where patient falls are common will also benefit from the new model.

Researchers conducted a brief case study to see how catastrophic falls by older adults and how to potentially reduce medical costs in the future. The result was more than $33 million in economic benefits over competing models.

“I'm very happy to see these results,” Yu said. “It's still a proof of concept, but if this work leads to future research in the Faculty of Engineering or related fields and can be transformed into a real product, that's the best.”

Yu also hopes that his work can reduce some of the anxiety surrounding AI.

“I think that's the future of health,” he said. “We already have AI components in our lives like ChatGpt. I think in the future, this kind of device will exist and we can improve our lives in a physical way.”