Intelligence Sparse Sensor Network for Automatic Early Evaluation of General Movements in Infants

Abstract General movements (GMs) have been widely used for the early clinical evaluation of infant brain development, allowing immediate evaluation of potential development disorders and timely rehabilitation. The infants’ general movements can be captured digitally, but the lack of quantitative assessment and well‐trained clinical pediatricians presents an obstacle for many years to achieve wider deployment, especially in low‐resource settings. There is a high potential to explore wearable sensors for movement analysis due to outstanding privacy, low cost, and easy‐to‐use features. This work presents a sparse sensor network with soft wireless IMU devices (SWDs) for automatic early evaluation of general movements in infants. The sparse network consisting of only five sensor nodes (SWDs) with robust mechanical properties and excellent biocompatibility continuously and stably captures full‐body motion data. The proof‐of‐the‐concept clinical testing with 23 infants showcases outstanding performance in recognizing neonatal activities, confirming the reliability of the system. Taken together with a tiny machine learning algorithm, the system can automatically identify risky infants based on the GMs, with an accuracy of up to 100% (99.9%). The wearable sparse sensor network with an artificial intelligence‐based algorithm facilitates intelligent evaluation of infant brain development and early diagnosis of development disorders.


Supplementary Note 5. Definition of the Error Band
The mean value (X ̅ ) and the standard deviation (SD) of N samples were calculated as: (2) As a result, the upper and lower error points were obtained as follows: Upper error point = X ̅ + SD (4) The error band defined by these two limits represents the confidence interval of the samples.

Supplementary Note 6. Calculation of the overal degree of overlap
The coverage magnitude (cov) and the total magnitude (tol) at each frequency point were calculated as follows: cov i = min(mag_n i max , mag_r i max ) -max(mag_n i min , mag_r i min ), i=0, 1, 2, …, N-1 (6) where mag_n i and mag_r i represent the power magnitude of normal and risky samples at the ith frequency point, with the superscript max and min representing their maximum and minimum values.The overall degree of overlap (or similarity) was then calculated as follows:

Supplementary Note 7. Definition of the feature's abbreviations
(1) The representative feature "Av.x.P1" in the sensing node represents the mean value of the x-axis angular velocity, with the first term representing the raw data type (Acc.and Av. for the acceleration and angular velocity respectively), the second term representing the dimension of the data (i.e., x., y., z., and sqrt.), and the third term representing the value type (P1 -P29 in Table S3).
(2) The representative entropy feature "Head.Av.x.P35" represents the permutation entropy of the head x-axis angular velocity, with the first term representing the position of the sensor node (Head, Leftwrist, Rightwrist, Leftankle, and Rightankle), the second term representing the raw data type (Acc.and Av.), the third term representing the dimension of the data (x., y., z., and sqrt.), and the fourth term representing the value type (P30 -P36 in Table S3).

Figure S1 .
Figure S1.The overall layout of the circuit board for SWD.

Figure S2 .
Figure S2.Encapsulation process and charging for SWD.(a) Schematic showing the encapsulation processing for SWD based on soft materials.(i) After preparing the metal mold with the bottom encapsulation layer, (ii) casting and curing Ecoflex0030 are followed by (iii) laminating and aligning the fPCB of SWD.(iv) After preparing the metal mold and (v) casting Ecoflex0030, (vi) assembling the two and curing the top encapsulation layer completes the encapsulation process.(b) Optical images showing the easy-to-use charging for SWD based on magnetic pogo pin.

Figure S3 .
Figure S3.Program logic diagram of the LED light in the SWD system.

Figure S4 .
Figure S4.Optical images of the forearm after peeling off the SWD over the course of 1 hour to show no skin irritation.

Figure S5 .
Figure S5.Comparison of high-frequency and low-frequency motion monitoring between SWD sensors and the commercial IMU device.(a) Optical images of the experimental setup with the position of the devices shown in the inset.(b) High-frequency and low-frequency acceleration and angular velocity obtained from SWD sensors and the commercial device.

Figure S6 .
Figure S6.Comparison of motion monitoring and feature calculation between SWD sensors with and without stable/intimate contact.(a) Schematic illustration and (b) optical images of the experimental setup with the position of the devices shown in the inset.(c) Acceleration and angular velocity and (d) relative error of various feature values calculated from the obtained data in (c).

Figure S7 .
Figure S7.Optical image showing the clinical setup for collecting neonatal movement data.

Figure S11 .
Figure S11.Power spectrum of the acceleration and angular velocity obtained from the head.(a) The power spectrum of the x-axis (Acc.x),y-axis (Acc.y), and z-axis acceleration (Acc.z) with error band.(b) The power spectrum of the x-axis (Av.x), y-axis (Av.y), and z-axis angular velocity (Av.z) with error band.The red and blue lines represent the average power magnitude of all "Normal" and "Risk" samples (n = 18).

Figure S12 .
Figure S12.Power spectrum of the acceleration and angular velocity obtained from the right ankle.(a) The power spectrum of the x-axis (Acc.x),y-axis (Acc.y), and z-axis acceleration (Acc.z) with error band.(b) The power spectrum of the x-axis (Av.x), y-axis (Av.y), and z-axis angular velocity (Av.z) with error band.The red and blue lines represent the average power magnitude of all "Normal" and "Risk" samples (n = 18).

Figure S13 .
Figure S13.Power spectrum of the acceleration and angular velocity obtained from the left wrist.(a) The power spectrum of the x-axis (Acc.x),y-axis (Acc.y), and z-axis acceleration (Acc.z) with error band.(b) The power spectrum of the x-axis (Av.x), y-axis (Av.y), and z-axis angular velocity (Av.z) with error band.The red and blue lines represent the average power magnitude of all "Normal" and "Risk" samples (n = 18).

Figure S14 .
Figure S14.Power spectrum of the acceleration and angular velocity obtained from the right wrist.(a) The power spectrum of the x-axis (Acc.x),y-axis (Acc.y), and z-axis acceleration (Acc.z) with error band.(b) The power spectrum of the x-axis (Av.x), y-axis (Av.y), and z-axis angular velocity (Av.z) with error band.The red and blue lines represent the average power magnitude of all "Normal" and "Risk" samples (n = 18).

Figure S15 .
Figure S15.Distribution of features from the left wrist sensing data before (left) and after (right) SMOTE.Red/black: risk/normal samples.

Figure S16 .
Figure S16.Diagram showing the construction of the convolutional neural network (CNN).(a) The data processing flow from subjects' motion data to 2D feature maps.(b) Diagram showing the structure of the convolutional network.

Table S1 .
Comparison between this work and other systems for infant neurodevelopmental assessment in terms of size, quality, system composition, and cost

Table S2 .
Cost analysis of a single SWD