Engineers used artificial intelligence (AI) and cheap finished equipment for converting Wi-Fi signals amplitude into assessment of human heart rate.
The accuracy of this system, called Pulse-Fi, is surprisingly stable in any body position and at a distance, researchers write in a study published on August 5 in research materials. International IEEE Conference 2025 on distributed calculations in intellectual systems and the Internet of Things (DCOSS-IOT).
Many home technologies, such as breast monitors and Smart clockmonitor vital indicators, including heart rate And the frequency of breathing. However, these devices require constant contact with a person and are expensive, which is necessary in contactless technologies.
One of these technologies can use information in Wi-Fi signals, which Radio waves which transmit data between the transmitter and the receiver, for example, between the router and the computer.
“Channel state information” provides an amplitude and signal phase when it passes between these two devices, including when it passes through obstacles, such as moving chests. Since the signals are distorted when passing these barriers, researchers can filter CSI data to catch vital indicators.
Different Now there are examples for determining the heart rhythm Wi-FiBut the cochet and his team claim that a number of restrictions remain. For example, many rely on already non -existent equipment. To eliminate these restrictions, the researchers have developed a new system called Pulse-Fi.
Registration of vital indicators
To collect the data necessary for evaluating Pulse-Fi, the team placed seven people-five men and two women-between two single antennas. ESP32 devices. These microcontrollers radiated Wi-Fi signals, one of them acted as a radiator, and the other as a receiver. The actual heart rate of participants was measured simultaneously using a pulsoximeter attached to the tip of the finger.
Each person participated three times: once at a distance of 3.3 feet (1 meter) from EPS32, and then at a distance of 6.6 feet (2 m) and 9.8 feet (3 m). Each measurement window lasted five minutes.
Then the team developed a machine learning conveyor to assess the heart rate assessment in CSI. The first step was the extraction of information about the amplitude, which belongs to individual heartbeat, and then the removal of random parts of the signal arising due to obstacles in the environment.
Then the engineers added a filter to remove the frequencies of the signal outside the range from 0.8 to 2.17 Hz, which corresponded to 48 to 130 beats per minute (BPM). Then they added a second filter for further smoothing the signal.
Then the team appreciated the heart rate of participants using Recurrent neural network of long-term short-circuent memoryThe form of machine learning, which adds “memory cells” to the processing of successive data, which provides the context necessary to identify dependencies in the data. In this case, these dependencies belong to such elements as the heart rate at rest and BPM resting caused by physical activity.
The team was surprised to find that the heart rate estimates remained accurate at different distances from ESP32 devices. Pulse-Fi underestimates and overestimates the heart rate on average by 0.429 beats per minute at a distance of 1 meter, 0.482 beats per minute at a distance of 2 m and 0.488 beats per minute at a distance of 3 meters.
Then the researchers used previously existing Wi-Fi CSI state data To check how Pulse-Fi copes with various positions of the body and activities. The data was obtained from 118 adult Brazilians who occupied 17 motionless and active positions, including sitting on the spot, walking on the spot and sweeping the floor within 60 seconds. The participants were at a distance of 3.3 feet (1 m) from the transmitter and receiver Wi-Fi, as well as from Raspberry pi 3b+ Used to collect CSI data.
They compared an assessment of the heart rhythm of the neural network with the testimony of smart watches and found that the position of the human body does not affect the Pulse-Fi. A typical error was 0.2 strokes per minute.
Wireless rhythms
This method of early stage is theoretically interesting, said Andreas CarvatA specialist in medical data from the University of Birmingham in the UK, who did not participate in the study.
However, according to him, the key restriction of this study is that the same data was used to train, test, and test the model. The researchers mixed the data each time, but Carvat said that this creates a self -filled prophecy.
“It’s the same as predicting someone’s disease by studying the experience of a person, and then predicting him,” he said to Live Science. “It makes no sense.”
In response to this criticism, the researchers said that although their analysis really included a shuffle, since then they tested the model in real time, where Pulse-Fi studied only on past data, and then was evaluated on a completely new input signal and the environment. This study has not yet been published.
Carvat also explained that smart watches and oximeter used to collect information about the pulse frequency for a compared neural network are not always accurate by 100%, so their data can be biased.
The cochet, bhatia and reverse recognized this restriction of smart watches. However, “a pulsoximeter is usually considered a certified medical device with very high accuracy,” they said.
Currently, the team expands the Pulse-Fi testing to track the heart rate of several people in the room at the same time to see how well the model copes with crowded environments.
The authors said that explicit personal information is not included in the data conveyor, and all heart rate estimates remain in the hardware. Thus, this technology does not cause any problems with data confidentiality. Carwart predicted that at least five to ten years remained before the introduction of this technology.
