View all hidden authors and organizations

Edited by Kimani C. Toussaint, Brown University, Providence, Rhode Island, USA, and accepted by Evelyn L. Hu, member of the editorial board, on September 1, 2021 (received for review on March 15, 2021)

Early detection of infant neuromotor disease is essential for timely therapeutic interventions that rely on early neuroplasticity in life. Traditional assessments rely on subjective expert assessments or professional medical facilities, which makes them difficult to expand in remote and/or resource-constrained environments. The results presented here aim to democratize these assessments using a wireless network of miniaturized, skin-integrated sensors that digitize infants’ motor behavior and vital signs in a cost-effective manner. The resulting data generates whole body motion reconstruction in the form of unidentified baby avatars, as well as a series of important cardiopulmonary information. This technical method can quickly and routinely assess babies of any age through an engineering platform that can be used in almost any environment in developed and developing countries.

Early identification of atypical infant motor behavior consistent with underlying neuromotor pathology can speed up the timely registration of therapeutic interventions that utilize inherent neuroplasticity to promote rehabilitation. Traditional neuromotor assessment relies on qualitative assessments conducted by specially trained personnel, most of whom are provided in tertiary medical centers or professional facilities. This method is costly, requires geographical proximity to advanced medical resources, and mainly generates qualitative insights. This article introduces a simple, low-cost alternative in the form of a customized technology for quantitatively capturing the infant’s continuous whole-body movement under free-living conditions in a home or clinical environment, while recording basic vital signs data. The system consists of a wireless network composed of small, flexible inertial sensors, which are placed in key positions of the body and operate in a broadband and time-synchronized manner. These data are used as the basis for reconstructing the 3D movement in the form of avatars, without video recording and related privacy issues, for experts to conduct remote visual evaluation. These quantitative measurements can also be presented in a graphical format and analyzed using machine learning techniques, which has the potential to automate and systematize traditional motion assessment. The clinical implementation in low-risk and high-risk infants of atypical neuromotor development illustrates the application of the system in long-term and subsequent gross motor skill models and quantitative and semi-quantitative assessments of body temperature, heart rate, and respiratory rate. It is measured within 3 months after birth. Engineering is compatible with large-scale deployment, and it is possible to improve the health outcomes of children around the world through early and pragmatic detection methods.

A growing body of literature shows that infants’ motor behavior can reveal a wealth of information about the developing central nervous system. Atypical patterns may indicate underlying neuromotor pathology that reflects or predicts clinical conditions, such as cerebral palsy, autism spectrum disorder, or other mild forms of neurological dysfunction (1⇓ ⇓ –4). Testing in early infancy is essential to initiate clinical interventions that utilize inherent neuroplasticity to promote early recovery and best long-term functional outcomes and quality of life (5⇓ –7). Methods that can be tested before routine clinical identification are of particular interest.

Currently, the diagnosis of infant neuromotor disorders relies on a combination of medical history, clinical examination and imaging methods. Diagnostic screening for atypical development varies with age and risk factors. Among them, children with medical complications before, during, and after birth, or obvious developmental delays that are observed receive enhanced, in-depth clinical testing (8, 9) . Recent international agreements recommend additional evaluations for infants who are at risk early in life, including visual observation of infant general movement (GM) by expert clinicians throughout the preterm and term stages (2, 4, 10). Although the evaluation of GM has value in screening certain neuromotor pathologies, this method is subjective, labor-intensive, and requires specialized training, which may hinder its scalability in remote areas or developing countries sex.

Alternative strategies for high-resolution analysis of infant motor development involve quantitative data from digital imaging systems or commercial wearable sensor packages (11⇓ ⇓ ⇓ –15), usually using EEG (16) or EMG (17). A key disadvantage is that these methods rely on bulky and expensive hardware and are only suitable for tertiary medical centers or specialized facilities. The SI Appendix Table S1 compares systems that support motion analysis, which contains detailed functions and specifications related to equipment size, weight, mechanical structure, and cost. Consumer-friendly imaging platforms, such as the discontinued Microsoft Kinect (18) and related systems, can be used to track body movements, but privacy issues may limit their practical application in community clinics or home environments. This method may also lack the necessary accuracy and resolution to accurately track the natural movements of newborns and young children, whose body structure is different from that of adults, and is usually not substantially reflected in the training data of the pose estimation algorithm (19) . None of these various systems for exercise analysis provide vital signs monitoring capabilities. Therefore, an urgent unmet need is in an affordable, adaptable, high-resolution system to measure and assess infants’ movement and key physiological health indicators across the risk range. Compared with current clinical screening tools, this type of technology can detect and quantify atypical motor development earlier and more accurately. Importantly, the technology that can continuously record and analyze data without the need for trained personnel can be deployed in various environments inside and outside the clinic to better capture the range of natural patterns of sports behavior that may occur in daily life. A scalable way. Improved testing may lead to increased interventions in the early stages of life, spanning a wide range of conditions and situations, thereby facilitating rehabilitation while limiting disability.

To meet these needs, this article introduces an easy-to-use, low-cost, and flexible wireless sensor platform designed for quantitative measurement of infants’ movement patterns and vital signs. The wireless network platform includes a series of miniaturized soft devices, which we call the Infant Adjustment Core Optimization (CORB) sensor. Each sensor operates in a time-coordinated manner to record data from the three-axis digital accelerometer and gyroscope. Flexible CORB sensors are more than three times thinner, five times lighter, and two times smaller than the most advanced commercial motion capture sensors (Xsens MTw Awinda, Xsens Technologies BV) (SI appendix, Table S1), and they also uniquely achieve cardiorespiratory functions And accurate measurement of body temperature. Due to its lightweight structure and miniaturized shape, the CORB sensor can be applied using ready-made medical silicone adhesives (2477P, 3M) or tapes or wraps. When placed in an important position on the body of an infant or child, the CORB sensor can capture rough and subtle movements in the entire space-time scale. The results provide a direct, quantitative assessment of body dynamics and can be used to reconstruct continuous movements in the form of avatars of arms, legs, torso, and head. The high-bandwidth unit placed on the chest simultaneously captures additional data related to standard and non-standard physiological markers of health status (ie, heart rate, heart rate variability, breathing rate and sound, body temperature, and vocal pattern). The following content describes all aspects of the device design and operation, as well as proof-of-concept tests for low-risk and high-risk infants. The full-term correction age for each infant is 1 week, 1 month, and 3 months.

The CORB sensor's thin, miniaturized geometry, soft mechanical characteristics and wireless operation function help to gently bond with the baby's sensitive skin with minimal mass or mechanical load. Continuous, time-synchronized sensing from multiple locations produces quantitative, accurate data about body motion. In a representative use case, these sensors measure three-axis acceleration and angular velocity (Figure 1) from devices installed in strategic locations on the chest, forehead, and limbs to capture the basic characteristics of whole-body motion. The schematic diagram in Figure 1A shows the overall size (32 × 21 × 3 mm) and shape of the device, and each device weighs less than 2.6 g. Thin silicone elastomer forms a waterproof housing (0.3 mm thick, Silbione RTV 4420), while mechanically isolating active and passive components to ensure contact with the skin. The 10 × 1.5 mm hole formed along the side through the functional layer and the encapsulation layer not only enhances the mechanical flexibility of the device when mounted on the skin, but also provides a highly flexible position to start the removal process Peel off. Figure 1B shows 10 such devices installed on a model baby. The layout is designed for whole-body motion assessment according to the GM protocol, spanning the middle of the upper and lower arms, the middle of the thigh and tibia, the chest, and the forehead. Figure 1 CF shows the device bending and twisting in various ways, including those related to peeling (Figure 1 C and E) and bending (Figure S1) at an angle of 45° (SI Appendix, Figure S1) to 90° (Figure S1). . 1 D and F). Finite element analysis confirmed that for bending radii as small as 2.8 mm, the strain in the copper (Cu) traces of the system remained below the yield limit (ε = 0.3%) (Figure 1 E and F and SI appendix, Figure 1) . S1). The miniaturized size of the device and its compliant mechanical properties facilitate installation on small and highly curved parts of the body, with minimal pressure on the skin during application, use, and removal. These characteristics are especially important for the vulnerable patient population considered here.

Images, diagrams and functional flowcharts of small wireless sensors (corb sensors) used to quantify infant gross motor behavior and vital signs. (A) Schematic diagram of an exploded view of the device. Optical image of the device (illustration). (Scale bar, 1 cm.) (B) The measurement configuration used to capture the movement of the whole body, illustrated with a doll. (CF) Image and finite element analysis and calculation of equipment flexibility during various mechanical deformations: peeling (C and E) and 90° angle bending (D and F). (G) Platform function diagram, showing hardware modules, including power management, Bluetooth radio, microcontroller, memory, and six-axis inertial measurement unit for each device. The data collected from each device includes synchronized time stamps to ensure millisecond relative timing accuracy. The user interface on the smartphone or desk controls the device, captures real-time data, and supports the use of local PC to download data and 3D motion reconstruction.

The functional diagram in Figure 1G summarizes the overall architecture and data flow. The system includes three main components: a time-synchronized collection of CORB sensors, a user interface based on a customized application running on a smartphone or tablet, and a set of motion reconstruction implemented on a local personal computer (PC) Algorithm. The main sensor communicates with other sensors to synchronize the local time as the basic information to reconstruct the whole body motion without drift, delay or inconsistency. This synchronization strategy utilizes a custom 2.4-GHz wireless transmission scheme in a star topology, where each node has a 32.768-kHz clock timer to mark data. The master device broadcasts its local time to other sensors every second. Then, each of these sensors synchronizes its local time accordingly. The experimental data shown subsequently shows that all sensors can achieve sub-millisecond time accuracy without any long-term drift. Under the control of the interrogator through Bluetooth Low Energy (BLE), each sensor continuously collects the responses from the three-axis accelerometer and the three-axis gyroscope, and wirelessly transmits the data with these synchronized time stamps to the user interface in real time. The sensor also stores these data in the internal memory module at the same time. Passing the data to the local PC via the BLE protocol allows the processing of motion information generated from each sensor in the network, as well as the static and dynamic directions relative to the gravity vector, thereby allowing the reconstruction of the whole body motion. It can also analyze data from chest sensors to produce heart, lung and vocal sounds, as well as heart and breathing frequency and its periodic changes, as well as estimates of core body temperature.

Figure 2A shows the layout of a flexible printed circuit board (fPCB) platform that supports all the electronic components in the overall layout and is designed to be folded in half to produce the final form of the device. Key components include passive components in 0603 and 1005 (metric code) mm packages, a voltage and current protection integrated circuit (BQ2970, Texas Instruments), a 3.0-V step-down DCDC converter (TPS62740, Texas Instruments), and a 4 -Gb non-volatile NAND flash memory (MT29F4G, Micron), low-power inertial measurement unit (IMU), including accelerometer and gyroscope (BMI160, BOSCH), BLE system on chip (SoC) (nRF52832, Nordic Semiconductor) and a 3.7V lithium polymer battery (7 mAh, 0.9 mm thick; PowerStream), can support more than 5 hours of continuous operation, with real-time data streaming and data storage functions (SI appendix, Figure S2). The dimensions of fPCB before and after folding are 34 × 30 mm and 17 × 30 mm, respectively, located at the positions defined by the yellow dashed line (Figure 2B). A thin top and bottom layer (0.3 mm thick) of silicone elastomer formed using a pair of concave and convex aluminum molds encloses the device, except for two small circular openings (2.0 mm in diameter) to the electrode pads for charging the battery. The packaging process involves placing the folded fPCB between these elastomer layers and filling the remaining space with a low modulus silicone material (Eco-Flex 00-30). This material acts as a soft conformal bumper around rigid electronic components and provides mechanical isolation during operations involving the application and removal of the device from the skin.

Schematic and optical image of sensor configuration and data flow for 3D motion reconstruction. (A) A schematic diagram of a flexible PCB supporting power management and battery protection SoC, 4Gb NAND flash memory, six-axis inertial measurement unit, Bluetooth low energy SoC and lithium polymer battery. (B) Optical image of fPCB with all components installed (left) and folded (right). (Scale bar, 1 cm.) (C) Image of the complete system, including the user interface on multiple time synchronization devices, chargers, and smartphones. (D) Time-synchronized and normalized accelerometer signals obtained from four sensors on the upper left arm (LUA), lower left arm (LLA), upper right arm (RUA) and lower right arm (RLA). The red arrow indicates the time of the jump. The results show that the time synchronization operation is within 10 milliseconds or less. (E) Flow chart of motion reconstruction algorithm. The top frame highlights the steps of deviation correction, signal filtering and time synchronization for each device. The middle frame shows the static attitude extraction from the three-axis accelerometer and the dynamic attitude extraction from the three-axis gyroscope. The bottom frame corresponds to transmitting the result data to ROS for 3D motion reconstruction.

Figure 2C shows a set of sensors and user interface on the smartphone. Figure 2D shows y-axis accelerometer data from four sensors worn on the left arm (upper and lower arm) and right arm (upper and lower arm), which capture the jumping motion performed by the subject. The peak position associated with the jump (red arrow in Figure 2D) indicates that the time difference between the sensors is less than 1 millisecond, which is negligible for the application considered here. When interpreted using an appropriate algorithm, the raw data collected in this way can be directly related to the assessment of dyskinesia without further manipulation. However, in many cases, visual inspection of whole body movement is also useful. The following section focuses on the method of reconstructing this type of movement in the form of an avatar based on sensor data.

The reconstruction algorithm uses the measured values ​​of linear acceleration and angular velocity, and each measured value is along three orthogonal axes, using the Rviz 3D visualization tool to reproduce the points in the three-dimensional (3D) humanoid model based on the Robot Operating System (ROS) platform. The results quantified overall exercise levels and behaviors related to GM assessment and diagnosis of atypical motor development patterns, such as motor development patterns related to cerebral palsy. The block diagram in Figure 2E illustrates the entire process from data collection to motion reconstruction. The process of estimating the attitude and direction of the sensor by analyzing the linear acceleration and angular velocity data is called dead reckoning. Estimating the direction of each sensor at each time step based on the angular velocity measurement value of the time step and the sequential Euler rotation around each axis may introduce systematic errors because the rotation is non-commutative. The simultaneous calculation of rotation minimizes these errors (20), as performed in the software code written in Python 2.7.15 and ROS Melodic, available on the web-based source code management cloud (21). The three main function scripts are used for data conversion, data analysis and calculation. The transformation script captures the position and 3D rotation group (so3 matrix) components, and then broadcasts the result to the Transforms ROS package in a quaternion angle. The parser script imports data from sensor measurements in raw format and parses them into separate containers for triaxial acceleration and triaxial angular velocity. The calculation script calculates the position and orientation of the sensor reference system in the 3D environment, and saves the results in a specific file format for broadcasting to the ROS platform. The software solves the main technical challenges, as described in the simple alternative to dead reckoning.

In dead reckoning, the double integration of the linear acceleration used to produce the 3D position can lead to unacceptable accumulation of small calibration errors, noise sources, and other effects. The additional use of data from magnetometers and/or GPS components can reduce these errors, but their integration with the device platform will greatly increase power consumption and overall size and weight. A simple alternative takes advantage of the constraints associated with the body model and uses a unified robot description format. This solution only relies on the direction determined by the integral of the angular velocity.

All sensors have a certain degree of deviation and noise. Both of these effects can be observed from the measurement results obtained by the sensor in the stationary SI appendix, Figure S3. The deviation is expressed as the distance between the solid line of each measurement axis and the absolute zero of the y​​ axis. Noise corresponds to high-frequency time fluctuations (SI Appendix, Figure S3). Although the deviation can be manually balanced by subtracting each deviation value from the measurement data one by one, the deviation can also be minimized by firmware-level sensor calibration. In this method, compensation is performed every time the sensor is connected wirelessly, as performed in the online firmware. The compensation value stored in the non-volatile memory can be loaded into the register and used to pre-filter the data. The SI appendix, Figure S4 shows the spectral characteristics of a single fixed sensor, where the deviation appears at 0 Hz (SI appendix, Fig. S4A) and a single sensor mounted on the arm of a human subject (SI appendix, Fig. S4B). Consistent with published literature, the relevant frequency range of human motion is between 0 and 20 Hz. The digital low-pass filter (fifth-order Butterworth) integrated in the data analysis script can remove features with frequencies higher than 20 Hz. The frequency response of this filter appears in Figure S5A in the SI Appendix. The figure in Figure S5B of the SI Appendix shows the filtered signal. Relevant filtering schemes can be applied to data from chest sensors to isolate vital signs information and vocal events, as described below.

As mentioned earlier, whole-body motion sensing requires time synchronization of sensor sets installed in strategic locations. The example shown here uses nine sensors-a main sensor in the middle of the body (chest) and a secondary sensor (two for each sensor) in the extremities. Carefully select the wireless transmission power and matching conditions for the antenna of each device to ensure robust and energy-saving operation in synchronous streaming mode, as an alternative to modes involving local storage of data. Figure 3 summarizes the results of sensors with high sampling rate (200 Hz, adjustable to 1,600 Hz) and wide dynamic range (±8 g, adjustable up to ± 16 g) and three-axis gyroscope (angular velocity sensitivity is 262.4 LSB/°/s), high sampling rate (200 Hz, adjustable up to 3,200 Hz) and wide dynamic range (1,000°/s, adjustable to 2,000°/s). These parameters are sufficient to accurately record the movements of children and infants (12).

Representative data collected from child subjects during various arm and leg movements in different angles and directions. (A) Normalized acceleration and gyroscope data during exercise. The subject jumped (approximately 10 seconds), lifted the left and right arms forward 90° (approximately 19 seconds), and lifted 135° (approximately 29 seconds) and 90° (approximately 39 seconds) to the side. (B) Optical and 3D reconstructed images from these motion data. 1: Raise the left and right arms forward by 90°; 2: Raise the left and right arms by 135° to the side; 3: Raise the left and right arms by 90° to the side; 4: Raise the left arm by 30° to the side and the right leg forward by 90° °; 5: Raise the left and right arms to the side by 45° and 120° respectively, and raise the right leg to the side by 80°; and 6: Raise the left arm and right arm to the side by 90°, and the left leg to the side by 30°.

Figure 3 shows an example of a healthy child (7 years old, 118 cm high) with CORB sensors placed on the upper right arm (RUA), lower right arm (RLA), upper left arm (LUA), lower left arm (LLA), and right thigh ( RUL, thigh), right calf (RLL, calf), left thigh (RUL, thigh), left calf (RLL, calf) and chest. In the initial configuration, all sensors are facing forward, and the subject is in a normal standing position (SI Appendix, Figure S6) and placed in the middle of the arm/leg to collect the most accurate motion angle and acceleration for reconstruction. Figure 3A shows representative static and dynamic acceleration and angular velocity data collected during different controlled movements (SI appendix, Figure S7), as described in Figure 3B, with matching optical images and corresponding 3D reconstruction results. In particular, 1) raise the left and right arms forward by 90°, 2) raise the left and right arms to the side by 135°, 3) raise the left and right arms to the side by 90°, 4) raise the left arm to the side by 30° and Right leg is 90° forward, 5) Left arm and right arm are raised 45° and 120° to the side respectively, right leg is 80° to the side, 6) Left and right arm is tilted 90° to one side, left leg Tilt 30° to one side. The detailed information appears in the SI appendix, figure. S8-S10 and movies S1-S4 are used for different movements, including video processing and quantitative comparison between reconstruction results measured from a human arm simulation model (shoulder-upper-lower arm). Measure 100 seconds of controlled upper and lower arm movements (30 repetitive movements, approximately 90° change from the starting position) (SI appendix, Figure S9A) to generate the trajectory angle profile from the reconstruction based on video recording and CORB sensor data, such as The figure is shown in the SI appendix, Figure S9 BD. The results show that for the initial configuration, the difference between the two methods is less than 0.2°, and they increase to a value of 1.3° over the measurement time (SI appendix, Figure S10). Drift (0.013°/s) accumulates during a single session, but resets during the initialization step of ROS.

The synchronized merged video of the results of motion tracking and CORB sensor reconstruction is in movie S1. Movie S2 summarizes the data collected through two sensors on each arm, one on the forearm and the other on the upper arm. The results showed that the reconstruction accurately captured the direction and angle of the subject's arm. In movies S3 and S4, data collected using eight sensors connected to the body are demonstrated. Here, the subject has a sensor on the limbs, both sides of the upper arm, forearm, thigh and tibia while walking. The platform produces a complete 360° 3D movement with quantitative information (SI appendix, Figure S11).

However, the reconstruction method does require some knowledge about the initial body position and the length of the limbs, legs, and core body in order to achieve an accurate and personalized 3D full-body reconstruction model. Indirect sensing (for example, multi-camera) alternatives bypass these requirements, but they often require complex and expensive settings that can only be applied to a single well-defined space and cannot track movement in natural daily activities. Privacy issues represent additional concerns for many parents. As mentioned earlier, due to volume, size, rigid structure, and/or wired connection, direct sensing solutions (eg, inertial sensors, magnetic tracking devices) are not the most suitable for babies (22).

The results presented here focus on the feasibility demonstration of the CORB sensor network on two infants. One is assessed as a low risk (LR) for atypical motor development, and the other is assessed as a high risk (ER) based on various factors. Born at 34 weeks of age, low birth weight, multiple births, and treated in the neonatal intensive care unit for 2 weeks). A total of 10 CORB sensors are placed on both sides of the shoulders, wrists, hips, and ankles to continuously capture the movement of the xiphoid process and forehead in real time (SI Appendix, Figure S12). The baby's exercise includes various postures (ie, prone, supine, supported sitting, supported standing, and horizontal suspension) at about 1 week, 1 month, and 3 months of age. For the 1 week and 1 month cases, only 6 devices were used—wrists, ankles, chest, and forehead—are limited by the small size of infants in these age groups. LR and ER infants standardized three-axis accelerometer and three-axis gyroscope signals at 1 week (SI appendix, Fig. S13), 1 month (SI appendix, Fig. S14), and 3 months (SI appendix, Fig. S15) Appears in the SI appendix, figure. S3-S15.

Fig. Figures 4 and 5 summarize the results of representative LR (subject ID: LR_1) and ER (subject ID: ER_1) infants at 3 months old (23). Visual observation and video analysis (Figure 4 and movie S5) did not show obvious gross motor deficits in ER infants, such as difficulty in lifting or lifting the head in the prone/standing/sitting/horizontal suspension, or in rare or infrequent situations. Shows leg/arm stiffness without exercise. Such behavior at this age may indicate delayed or atypical neuromotor development (23). Both infants captured from the movement reconstructed from sensor data showed changes in body posture. Unlike video analysis, 3D reconstruction can perform quantitative analysis from any perspective, including those that may be hidden in video analysis. For example, video analysis of an ER object in a prone position does not allow to evaluate the symmetry of the movement and direction of the arms and legs. Figure 5 highlights the value of Quantitative Activity Level (QAL), which combines the spectral power of each sensor from 0.1 to 10 Hz (24⇓ ⇓ –27) and integrates the two subjects in equal time To calculate. The details of the calculation method are in the SI appendix, Figure S16. The comparison of this indicator for 3-month-old LR and ER subjects in the prone, supine, supported sitting, supported standing, and horizontal suspension positions is shown in Figure 5A. These initial feasibility data indicate that in most cases examined in these brief events, except for the arms in the supine position, the legs supporting the standing position, and the arms in the horizontal suspension position, the QAL values ​​of LR infants are higher than those of ER infants . The specific values ​​are in Table S2 of the SI Appendix.

Compare the representative gross motor behaviors associated with the LR and ER gross body movements of atypical motor development in 3-month-old infants. The movement behavior of LR and ER infants and the corresponding optical images captured during the 35-minute session and the related 3D movement reconstruction results with different viewpoints. The capture time points of the pictures and reconstruction results are displayed at the bottom of each picture.

Quantitative comparison of long-term and follow-up measurements of representative infants with atypical motor development of LR and ER within 3 months. (A) The QAL comparison of 3-mo-old LR and ER subjects in various postures, such as prone, supine, supported sitting, supported standing, and horizontal suspension, corresponding to the image in Figure 4. (B) 3D scatter plot from 3-mo-old LR and ER three-axis inertial measurement of the subject in the supine position (magenta sphere). Accelerate the left and right arms from the LR (first and second columns of the top row) and ER (first and second columns of the bottom row). The left and right arm angular velocities from the LR (third and fourth columns in the top row) and ER subjects (the first and second columns in the bottom row). XY (red), YZ (blue) and ZX (green) projections on each graph. (C) The overall QAL (left) and each level of the head and chest/arms/legs (right) of the LR and ER infants when the subjects were 1 week, 1 month, and 3 months old.

The original data can be displayed in different graphical forms for further visual analysis. Figure 5B shows an example of a 3D scatter plot of acceleration measured from a 3-month-old LR and ER subject in the supine position (magenta sphere), which is used to quickly assess the symmetry and variability of motor behavior characteristics. The first and second columns and the third and fourth columns show the acceleration and angular velocity distribution of the left and right arms of the LR (top) and ER objects (bottom), respectively. These manifestations clearly indicate that the motor behavior of ER infants is more sporadic and asymmetric than that of LR infants.

The overall QAL value (Figure 5C, left) reveals the difference between LR (subject ID: LR_1) and ER (subject ID: ER_1) subjects at three time points after birth (Figure 5C, right) . Specifically, the QAL values ​​of LR infants were 14%, 66%, and 68% higher than those of ER infants at 1 week, 1 month, and 3 months, respectively. The average value of each sensor (head chest, arms, and legs) provides additional insight, consistent with the qualitative assessment described in Figure 4. For example, at 1 month, the QAL of the head and chest of LR infants is 154% greater than that of ER infants. In a similar way, the QAL of the arms and legs of LR infants is 40% and 64% higher than that of ER infants. Also at 3 months, the QAL of the combined head and chest movement of LR infants was 27% higher in LR than in ER infants. In a similar way, the QALs of the arms and legs were 23% and 124% higher than that of ER babies, respectively. The specific values ​​are in Table S3 in the SI appendix. The various differences described here may correspond to movement variability between subjects, rather than potential neuromotor disorders. Preliminary results of 10 infants (6 LR and 4 ER subjects) showed that there are differences in QAL between LR and ER subjects, especially in terms of arm and leg movements (SI Appendix, Figures S17 and S18 and Table S4). Additional tests and larger-scale data analysis from children with known typical and atypical developmental trajectories will define the normative range of sensor-based exercise indicators throughout early infancy and childhood.

Finally, the sensor can be used not only for evaluation, but also for tracking physical therapies, which involve the controlled placement of infants in different body positions, sometimes referred to as "downtime" in one example. For example, in the SI appendix, Figure S19 shows the body orientation extracted from the chest sensor data of the LR and ER subjects. The results can be used to determine the cumulative time for each posture, including prone and horizontal suspension (body angle> 135°, red shaded area in the SI appendix, Figure S19), supine (body angle <25°), and sitting and standing (25°) <body angle> 135°). These values ​​are not related to the specific therapy here, and are within the expected range for healthy resting newborns and infants (28, 29).

As mentioned earlier, the high-frequency information in the data captured using these same CORB sensors contains vital signs and other physiological information related to the mechanical acoustic characteristics of the internal processes of the body, highlighting the ability to capture heart rate and respiratory rate from the chest sensor. SI appendix, Figure S16 B and C). Figure 6A shows the normalized z-axis accelerometer (black solid), z-axis gyroscope (red solid) signal, and the extracted baseline (blue dot) of the gyroscope signal. For example, the z-axis accelerometer data from the chest sensor of a 1-week-old subject (top in Figure 6A) contains clear information about heart activity as an electrocardiogram (black arrow) with a significant peak caused by the opening of the aortic valve and Mitral valve opening (gray arrow) (30⇓ –32). Using a relatively low sampling frequency setting (200 Hz) for power optimization can prevent the collection of heart sound signals, such as heart murmurs. However, IMUs with high sampling frequency characteristics (3.3 or 6.6 kHz) can capture high frequency spectrum characteristics to measure acoustic signals other than mechanical signals caused by cardiac activity (26, 27). The baseline (blue dashed line) of the gyroscope signal has a clear and periodic response related to exhalation (black dashed line) and inspiration (brown dashed line). Comparing the results of the CORB sensor with the results recorded by the Food and Drug Administration (FDA) approved capnography device (EMMA, Masimo) from three adult volunteers, the respiratory rate is very consistent, as shown in the SI appendix, Figure S20 Shown. The Bland-Altman diagrams of the two methods are in the SI appendix, Figure S21. Compared with the FDA-approved system, measurements made with CORB sensors (three subjects, 1,421 data points) produced an average difference of 0.01 breaths per minute (RPM) and an SD of 0.53 RPM. Due to the size limitation of the capnography device used here, no comparative data was collected from infants. Data collected from subsequent measurements of subjects aged 1 year and 3 months also showed a clear response, but the signal amplitude was different compared with subjects aged 1 week (Figure 6A). Specifically, the z-axis acceleration amplitude of the 1-week-old subjects was 50% higher than that of the 3-month-old subjects. Similarly, compared with the values ​​in the 1-week-old subjects (SI Appendix, Figure S22), the gyro signal fluctuations corresponding to the breathing of the 3-month-old subjects are clearer and stronger, which is related to the temporal and spatial changes. Geometric growth and organ development (33⇓ –35). Figure 6B summarizes the average heart rate and respiratory rate (26, 27) based on these data in the 60-minute recording period in the form of a bar graph. These data are for LR (grey) and ER (red) subjects. Both subjects showed a heart rate of 120 to 160 beats per minute and a breathing rate of 38 to 45 RPM. These same CORB sensors also have a temperature sensing function, which can be used not only to determine the core body temperature estimate of the chest device, but also to determine the surrounding temperature as a measure of overall circulatory health (Figure 6C).

Quantitative long-term and follow-up cardiac and respiratory activity measurements in infants with LR and ER within 3 months. (A) Standardized chest sensor signals for LR infants at 1 week (top), 1 month (middle), and 3 months (bottom). The black line corresponds to the data from the z-axis of the accelerometer (ECG); the red line is the z-axis gyroscope signal; the blue dashed line is the running average of the gyroscope signal (respiratory activity). AO and MO represent the aortic valve and mitral valve openings, respectively. (B) Heart rate (left) and breathing rate (right) are determined based on these data. The themes of LR (grey filled bars) and ER (red filled bars) include error bars. (C) The time-synchronized whole body temperature monitoring of a 3-month-old LR baby collected by the CORB sensor. Temperature change during data collection over 30 minutes (left). Skin temperature determined from each device on the head, chest and limbs, including error bars (right).

The collection of soft and lightweight wireless sensors introduced here has the potential for widespread deployment, a significant improvement over standard clinical screening tools, and it is possible to perform routine automation and quantitative evaluation in a home/community environment for long-term monitoring. The design of these technologies has been optimized for children starting from very young newborns, and features the use of infant safety, low-cost components and high-volume manufacturing technologies. These features, along with inherent ease of use, facilitate access by healthcare professionals and parents. The use of remote computing resources to reproduce 3D motion trajectories provides a behavioral visualization mechanism that can be evaluated by clinicians around the world without privacy issues or geographic restrictions. The output of these sensors can be further linked to established educational resources on sports development (for example, Reference 23) through mobile apps to support early detection of abnormalities in the home and clinical settings. In the future, these basic data can also be used as the basis for automatic scoring of traditional GM assessments and other clinical tests using advanced machine learning technology. At the same time, capturing a wide range of vital signs information can provide additional important information, not only in the context of sports behavior, but also in general health monitoring. A further potential lies in the use of these devices to track physical rehabilitation at home or in the clinic, including strength training and postural stabilization (36, 37). A clinical trial program is currently underway, which will test the effectiveness of the sensor in effectively detecting abnormal movements of newborns and tracking the effects of neuromotor interventions.

Commercial software (AUTODESK EAGLE [Version 9.6.0]) is the basis for generating schematics and layouts for FPCB. Use Segger Embedded Studio to download the customized firmware to the BLE SoC. Place the components, solder them to the FPCB, and then fold the system to complete the manufacture of electronic components. The aluminum mold prepared with a freeformer (Roland MDX 540) defines the shape of the top and bottom layers of silica gel (Silbione-4420, each layer is 300 μm thick). Place the electronics on the bottom layer in the mold, then pour the precursor into the soft silicone elastomer (Eco-Flex 0030, 1:1 ratio), then install the top layer and clamp the system together, ready to be in the oven 95 °C for 20 minutes. After cooling to room temperature, remove the fixture and use a CO2 laser to cut excess material from the periphery to make the final device.

The commercial software ABAQUS (ABAQUS Analysis User's Manual 2010, version 6.10) is used to determine the strain ε distribution in the sensor package and the metal layer under bending deformation. The equivalent bending stiffness of the soft package sensor is about 3.4 Nmm2 (SI appendix, Figure S23), modeled by a tetrahedral element (C3D10), and the thin Cu and PI membranes are modeled by a composite shell element (S4R). The number of elements in the model is approximately 3 × 105, and the smallest element size is one-eighth the width of the narrowest interconnect (100 µm). In all cases, the mesh convergence is guaranteed. For soft packages, the elastic modulus (E) and Poisson's ratio (ν) are EEcoflex = 60 kPa and νEcoflex = 0.5, for copper, ECu = 119 GPa and νCu = 0.34, and for PI, EPI = 2.5 GPa and νPI = 0.34.

These studies were approved by Northwestern University (STU00207900). For the subject, a double-sided medical silicone adhesive (2477P; 3M) was applied to the device and the sensor was fixed to each limb. The child’s legal guardian provided informed written consent to allow the release of research images with blurred faces.

These studies were approved by the institutional review boards of Lurie Children's Hospital (#2018-2098) and Northwestern University (STU00207900). For LR and ER infants, a latex-free self-adhesive elastic bandage (CoFlex; Andover Healthcare) is used to secure the device to the arms and legs. The device on the chest was fixed with hydrogel (Katecho, Inc.). The device on the forehead is held in place using a headband. The legal guardians of these babies provided informed written consent to allow the release of research images with blurred faces.

All research data is included in the article and/or supporting information. Zenodo (https://doi.org/10.5281/zenodo.3690141) provides additional supporting data.

We thank the Ryan Family Foundation for supporting this work. The work described here was implemented as part of the Corbett Ryan-Northwestern-Shirley Ryan AbilityLab-Lurie Early Childhood Infant Detection, Intervention, and Prevention Project.

↵1H.J., SSK and SS have made equal contributions to this work.

Author contributions: H. Jeong, SS Kwak, SS, MKO, AJ, MMD, TS, LSW, SK-J., SX, SWR, RLL and JAR design research; H. Jeong, SS Kwak, SS, JYL, YJL, M​​KO, YP, RA, J.-TK, J.-YY, MI, H. Jang, WO, NS, AT, RAA and AJ conducted research; H. Jeong, JYL, SS Kim and AT contributed New reagents/analysis tools; H. Jeong, SS Kwak, SS, JYL, YJL, M​​KO, RA, J.-TK, YJK, KL and YH analyzed the data; SS Kim completed the user interface application development ; H. Jeong, MKO, RA, YH and JAR wrote this paper.

The author declares no competing interests.

This article is directly contributed by PNAS. KT is a guest editor invited by the editorial board.

View relevant content online, such as comments.

This article contains online support information https://www.pnas.org/lookup/suppl/doi:10.1073/pnas.2104925118/-/DCSupplemental.

This open access article is distributed under Creative Commons Attribution 4.0 (CC BY).

Thank you for your interest in advertising on PNAS.

Note: We only ask you to provide your email address so that the people you recommend the page to know that you want them to see it, and that it is not spam. We do not capture any email addresses.

Feedback privacy/legal

Copyright © 2021 National Academy of Sciences. Online ISSN 1091-6490. PNAS is a partner of CHORUS, COPE, CrossRef, ORCID and Research4Life.