In the study, through the calibration of professional physicians, we screened out several infants at risk, and then urgently contacted the parents of the infants through the doctors, suggesting that they go to a higher-level hospital for a comprehensive examination. At this moment, we truly felt the importance of what we were doing.

Speaking of a recent study he participated in, Dr. Zhang Senhao and Dr. Bao Benkun from the Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, still feel very meaningful.

Recently, he and his collaborators have jointly created a flexible sparse sensor network system that can achieve ultra-high accuracy in automatic classification in the assessment of infant restlessness.

Currently, the experimental results of this team in collaboration with clinical hospitals show that this technology can quickly and effectively carry out large-scale screening of neonatal cerebral palsy risk.

In a few years, this technology will be able to be promoted to more areas and become a neonatal essential check item similar to "vaccines".Due to the low cost and low resource dependence of this technology, it can be effectively operated even in less developed medical areas such as the central and western regions, which is expected to greatly promote the development of maternal and child health in our country.

Advertisement

"Especially if it can be popularized in areas with low medical standards, and the neurological development and behavioral habits of newborns can be screened, intervened, and rehabilitated as early as possible, it can better reduce the burden on families. Thinking about this, I feel that my efforts are very worthwhile," said Zhang Senhao.

Recently, the relevant paper was published in Advanced Science (IF 15.1) with the title "Intelligence Sparse Sensor Network for Automatic Early Evaluation of General Movements in Infants."

Dr. Bao Benkun and Dr. Zhang Senhao from the Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, are the co-first authors.

Zhang Senhao, Researcher Yang Hongbo and Researcher Cheng Xiankai from the Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, and Professor Cheng Huanyu from Pennsylvania State University, USA, are the co-corresponding authors [1].It is reported that the screening of neonatal cerebral palsy has always been an area of clinical research that draws considerable attention.

However, the methods previously used for clinical assessment mainly relied on experienced professional doctors, who would observe the spontaneous movement patterns of newborns over a period of time to assess whether they are at risk of cerebral palsy, and then confirm whether further imaging examinations are needed.

It is undeniable that this method of assessing neurodevelopment through the spontaneous movements of newborns has a certain convenience and potential for large-scale screening.

However, the shortage of professional doctors limits its further promotion, especially in areas with low-level medical resources.

Zhang Senhao said: "This topic originates from the medical-engineering integration project jointly participated in by our institute's Cheng Xiankai, Yang Hongbo, and the First Affiliated Hospital of Jilin University."This project aims to advance a batch of promising frontline medical-engineering integration and interdisciplinary projects to address the pain points and difficult issues in clinical practice, and this project has been fortunate enough to receive support.

At the beginning, the research team had not found a good solution for wearable methods and data stability.

With the support of the international cooperative partner plan of researcher Yang Hongbo, Zhang Senhao began to participate in the joint doctoral training at the research group of Professor Cheng Huanyu at Pennsylvania State University (PSU, The Pennsylvania State University) in the United States in 2020.

During this period, they discussed together and believed that using a flexible structure and material design methods should be more suitable for this kind of clinical application scenario.

Zhang Senhao said: "Therefore, we started this research with the PSU team. Under the guidance of Professor Cheng Huanyu, we began to design the structure and materials, and optimized the networking method and networking stability."It is understood that previous studies mostly captured motion through the "video method" and then digitized it.

However, this method cannot guarantee privacy, and motion recognition is easily affected by environmental interference. A large amount of background noise can make the data volume extremely large, limiting its development in clinical applications.

Based on this, the research team conceived the idea: Can the digitization of newborn movements be completed by directly collecting motion information?

Considering the fragility of newborn skin, they completed the flexible design of the sensing nodes through the "island-bridge" structure. At the same time, the use of highly biocompatible materials further ensures the safety of newborn skin.

In order not to let the wearing of sensors affect the autonomous movement of newborns, it is necessary to reduce the number of sensing nodes as much as possible.Through means such as algorithm optimization, they only arranged sensors at a total of five locations on the limbs and head of newborns, and constructed a sparse sensing network through low-power Bluetooth.

After the system design was stabilized, they began to initiate clinical verification studies. With the support of the First Affiliated Hospital of Jilin University, Suzhou Children's Hospital, and Quwo County Traditional Chinese Medicine Hospital in Shanxi Province, they successfully obtained a batch of clinical data.

After preliminary data analysis, they found that there were significant differences between children with and without risks in both the time domain and frequency domain. However, there is still no direct threshold that can be used as a dividing standard.

Specifically, they found that relying on traditional classification methods could not achieve high-accuracy screening well, so they began to use machine learning to build classification models.

However, the overall model cannot be too large, otherwise the computational power requirements will increase, making it impossible to deploy in areas with low medical levels.Thus, the goal they pursue is to create a classification model that is as lightweight and compact as possible. At the same time, they do not wish for the algorithm they develop to be a "black box" algorithm.

Therefore, researchers start from the diagnostic criteria of clinical doctors to summarize the extraction of feature values as much as possible.

At the same time, for the obtained feature values, they are sorted according to the degree of relevance to identify those feature values that are more effective for classification and ensure that these feature values can be clinically interpreted.

Only in this way can the dimensionality of the feature values be further reduced, thereby creating a lightweight algorithm model.

During the process of building the model, considering that it is necessary to occupy as few computing resources as possible, Dr. Bao Benkun of the team optimizes from the aspects of feature dimension reduction and algorithm redundancy in order to obtain a small algorithm model with high accuracy.With the assistance of artificial intelligence technology, they successfully developed the required classification model. Moreover, as the volume of data increases, this method of data processing can also be optimized, thereby achieving a more lightweight, faster, and more efficient classification algorithm.

Considering the application and promotion in areas with low-level medical resources, it is essential to develop low-cost hardware systems and easily deployable small-scale automatic recognition algorithms.

Ultimately, with their efforts, the cost of the entire system is below 500 yuan. At the same time, the research team, by reducing the number of feature values and using the logistic regression algorithm, has built the smallest model under the premise of a 99.9% high recognition rate.

It is also reported that this flexible physiological sensing network can not only be used for the detection of newborn restlessness but also achieve the acquisition of more physiological information, thus being used in other clinical aspects.

For example, heart rate and respiratory rate information can be collected through an accelerometer, and at the same time, information such as swallowing ability and activity ability can be collected, thereby playing a role in the ICU intensive care unit.On the other hand, for the rapid screening of neonatal cerebral palsy, the monitoring of fidgety movements is mainly applicable to newborns within 20 weeks.

For older infants who can already crawl, they are also conducting an analysis of crawling behavior and cerebral palsy in order to serve infants of a slightly older age, thereby assessing their neurodevelopmental capabilities.

In terms of application, they are actively contacting some local maternal and child health stations in the Midwest to strive for better promotion of this system.

At the same time, they are also actively trying to see if it is possible to construct a battery-free sensing system.

If wireless power supply can be achieved through radio frequency, the weight of the sensor can be further reduced by removing the battery module, thereby reducing the impact of wearing the sensor on the spontaneous movement of newborns.