Wireless intelligent detection system, a new hope for Parkinson's patients!

By combining radio waves with artificial intelligence, MIT researchers have developed a new detection system: a skeleton-like image of a person moving on the other side of a wall. While it sounds like technology a SWAT team would wish they had before breaking into a door, it's already being used in a surprising way to monitor the movements of Parkinson's patients at home.

“Interest in this technology goes back decades, and there was a major DARPA program trying to use wireless signals to detect people behind walls,” says Dina Katabi, a professor of electrical engineering and computer science at MIT and a senior researcher on the project. “But before this latest work, all those systems could do was show a blob-like image of a person behind a wall.”

Now, the technology is even more accurate: It can depict people in a scene as skeleton-like stick figures, while showing their real-time movements as they do everyday activities, such as walking, sitting, and lying down. The technology focuses on detecting key parts of the human body, including elbows, hips, and feet. "When the person behind the wall starts moving, you see the stick figure created by the system do the same movement," Katabi said. "If the person sits down, you see the stick figure in the system sit down too."

1. Radio signals reflect off the human body and are recognized by the neural network

The RF-Pose system works by transmitting a wireless signal that is 1,000 times less powerful than WiFi, penetrating objects such as walls, reflecting off the human body — which is mostly made of water, which is impenetrable to radio waves — before returning through the wall to the device.

The RF-Pose system uses wireless signals to act as a healthcare system, monitoring a patient's movements from the other side of a wall. (Image credit: Jason Dorfman/MIT CSAIL)

The question now is: How do you interpret the information the system sends back? “That’s where artificial intelligence, and specifically neural network machine learning, comes into play,” Katabi said.

AI researchers train neural networks by adding annotations so that they can infer their own rules from data in order to learn, a process called supervised learning. For example, a neural network that recognizes cats requires people to find a large database of images and label them as "cats" and "not cats." If you want to teach a self-driving car what a traffic light looks like, you need to show images that contain traffic lights and annotate them to tell the AI ​​what a traffic light is.

(Image source: MIT CSAIL)

To solve this problem, the researchers used wireless devices and cameras to collect data. They took thousands of photos of human activities (such as walking, talking, sitting, opening doors, and waiting for elevators). Neural networks are usually used to analyze images, but they can also be used to perform complex tasks, such as translating from one language to another, and even generating new text by imitating the data given.

But in this case, there is a problem: no one can analyze which position is the head and which position is the foot through radio signals. In other words, it is easy to process information through images, but it is not so easy to process information through radio signal data reflected off a person.

(Image source: MIT CSAIL)

The research team's solution is to connect the signal receiver to the camera during training, then label the images created by the camera with information, extract human skeletons (stick figures, used to represent human postures) in different poses, and match them with wireless signals. This combination allows the system to learn the association between wireless signals and stick figures in the scene, helping the neural network to link these activities.

In order for the cameras to actually receive image signals, these training tasks must be completed without walls. Katabi explained: "We use the information tags on the cameras and the radio signals that occur at the same time to train them."

After the training, they were surprised to find that even though the system was only trained on visible people, it could still detect people hidden behind walls. "They can see and depict human figures behind walls, even though they have never seen them during training," Katabi said. Not only that, it can even distinguish people by their gait. With the help of another neural network, the system can record the characteristics of people's walking and identify different people, with an accuracy rate of more than 83% in identifying individuals even through walls.

2. Initial application in the treatment of Parkinson's disease patients

The researchers have begun putting the system into practice in a small study of Parkinson's patients. By installing the devices in the patients' homes, they can monitor the patients' movements in a comfortable environment without the need for cameras. In this sense, it is slightly less invasive than traditional video monitoring. The study involved seven people and lasted for eight weeks.

"The results of the systematic analysis were 'highly correlated' with standard questionnaires used to assess patients," Katabi said. In addition, it revealed more information about the quality of life of Parkinson's patients. Currently, the Michael J. Fox Foundation is funding the team to conduct further research; Katabi said that monitoring patients in this way can help patients avoid "white coat syndrome," that is, patients behave differently when they see a doctor.

The use of this technology also raises corresponding privacy issues, but Katabi explained that it would not be used on anyone without their consent.