A team of engineering and health researchers has developed a tool that improves an electronic device’s ability to detect when a human patient is coughing. It has applications in health monitoring. This new tool relies on advanced artificial intelligence (AI) algorithms to help AI better identify uncertainty when faced with unexpected data in real-world situations.
“When an AI is trained to identify cough sounds, this is typically done using ‘clean’ data. Background He doesn’t have a lot of noise or distracting sounds.” Associate Professor of Electrical and Computer Engineering at North Carolina State University. “But the real world is full of background noise and confusing sounds. I thought I was.
“We have developed an algorithm that helps us address this problem by allowing the AI to express uncertainty. , it can also report that it has detected an unfamiliar sound. So the AI is given her third option.
Cough detection technology is of interest for several potential health monitoring applications.
“For example, we might be interested in using wearable health-monitoring devices to detect coughing in asthmatics, which could trigger notifications about increased risk of asthma attacks,” says Lobaton. “There is also interest in using cough detection for things like COVID monitoring.”
However, previous cough detection techniques had high false-positive rates and associated AI reported many unfamiliar sounds as coughs. These false positives severely limited its usefulness.
“In the short term, we will limit reporting of false positives by allowing AI to recognize sounds that it cannot recognize,” says Lobaton. “In the long term, our algorithm should be able to continuously train the AI by telling it whether the unfamiliar sound it is hearing is a cough or an unrelated sound. should allow for more accurate detection over time.”
Additionally, the researchers tested the new algorithm on computational models and found that the modified cough detection AI could operate effectively with far fewer audio samples per second than previous techniques. For example, our previous cough detection tool used about 16,000 sound samples per second, while our new AI tool uses 750 sound samples per second, yielding similar sensitivity with fewer false positives. I’m here.
“Using fewer sound samples is a big advantage for two reasons,” says Lobaton. “First, it means that electronic devices require less computing power, which allows them to be smaller and more energy efficient. Less audio means the technology won’t record intelligible audio, addressing privacy concerns.”
Researchers are now in the process of incorporating the new algorithms into wearable health-monitoring devices that can be used in real-world tests.
Furthermore, the researchers say the approach taken here can be used to address a wide variety of AI applications where AI is likely to encounter unexpected input that it has not been trained to understand.
“We are looking for research partners to help us investigate other health monitoring challenges that this AI modification could help address in a meaningful way,” said Lobaton. increase.
The paper “Robust Cough Detection with Out-of-Distribution Detection” IEEE Journal of Biomedical and Health InformaticsThe lead author of this paper is Yuhan Chen, Ph.D. Student in North Carolina. A co-author of this paper is his Pankaj Attri, a former graduate student at North Carolina State University. Dr. Jeffrey Barajona North Carolina student. He is Alper Bozkurt, professor of electrical and computer engineering at North Carolina State University. Michelle L. Hernandez, Associate and Director of Clinical Research at the Children’s Research Institute at the University of North Carolina at Chapel Hill. and Deleesha Carpenter, Associate Professor of Medicine Outcomes and Policy at the University of North Carolina at Chapel Hill.
This work was supported by the National Science Foundation (NSF) under grant number 1915599 and by NSF-funded Integrated Sensors and Technologies (ASSIST) of North Carolina under grant 1160483. Assisted by advanced self-powered systems. The ASSIST Center’s mission is to create self-powered wearables capable of long-term multimodal sensing without having to replace or recharge batteries.
-Sailor-
Note to editors: Here’s a summary of the research:
“Robust Cough Detection with Out-of-Distribution Detection”
authorBy: Yuhan Chen, Pankaj Attri, Jeffrey Barahona, Alper Bozkurt, Edgar Lobaton, North Carolina State University. Michelle L. Hernandez and Delesha Carpenter (University of North Carolina at Chapel Hill).
Published: April 5 IEEE Journal of Biomedical and Health Informatics
Soil: 10.1109/JBHI.2023.3264783
overview: Coughing is an important defense mechanism of the respiratory system and is also a symptom of lung diseases such as asthma. Acoustic cough detection, collected by a portable recording device, is a convenient way to track potential deterioration in asthma patients. However, the data used to build current cough detection models are often clean and contain a limited set of sound categories, making them suitable for the wide variety of real-world sounds a portable recording device might pick up. performance will be degraded when exposed to the sound of Sounds that are not learned by the model are called Out-of-Distribution (OOD) data. In this work, we propose his two robust cough detection methods combined with an OOD detection module that removes his OOD data without sacrificing the cough detection performance of the original system. These methods include adding a learning confidence parameter and maximizing entropy loss. Our experiments show that 1) the OOD system can produce reliable In-Distribution (ID) and OOD results at sampling rates above 750 Hz. 2) OOD sample detection tends to perform better with larger audio window sizes. 3) The overall accuracy and accuracy of the model increases as the proportion of his OOD samples in the acoustic signal increases. 4) A higher percentage of his OOD data is required to achieve performance gains at lower sampling rates. Incorporating OOD detection techniques significantly improves cough detection performance and provides a valuable solution to real-world acoustic cough detection problems.
