Newborn cerebral palsy ultra-early screening method, electronic device, storage medium and system

Abstract

The present application relates to a new-born brain palsy ultra-early screening method, electronic equipment, storage medium and system, the method comprises the following steps: obtaining vibration information and motion data; the vibration information is divided into frequency and the corresponding physiological data is obtained; the emotional state of the new-born in the data collection process is judged through the physiological data; the motion data corresponding to the crying state is screened out, and the motion data corresponding to the natural wakeful state is retained; the motion data corresponding to the natural wakeful state is solved for characteristics, and a characteristic matrix is formed through the obtained characteristics; the characteristic matrix is used as the input of a classification model, and a classification result is obtained.The present application realizes brain palsy risk assessment with the aid of sensor data, reduces the storage and use of data volume, can realize fast result query, ensures that the requirement for computer equipment is greatly reduced in actual use and popularization process; the purity of the entire motion data is improved through the acquisition and analysis of heart rate and respiration, and data analysis is more meaningful.

Description

Technical Field

[0001] This invention relates to the field of early prediction and screening technology for neonatal cerebral palsy, and particularly to methods, electronic devices, storage media, and systems for early screening of neonatal cerebral palsy. Background Technology

[0002] Because the newborn brain has extremely high plasticity during this period, it can recover some or even complete functions through self-guided repair even when damage occurs in early brain development. Simultaneously, through early rehabilitation training, limb motor dysfunction in children with cerebral palsy can be improved to the greatest extent possible. Therefore, ultra-early screening of newborns is crucial. Currently, various methods exist for newborn cerebral palsy assessment and screening, including the Hammersmith Infant Neurological Examinations (HINE), Magnetic Resonance Imaging (MRI), and General Movement Assessments (GMA). At present, there are various forms of General Movement Assessment for early newborn cerebral palsy screening. The most widely used method both domestically and internationally is the video recording method based on General Movements (GMs) quality assessment. This method requires professional physicians to analyze the video and make judgments, and physicians must undergo rigorous training to obtain assessment qualifications. The complex qualification process and the limited availability of professional personnel restrict the application's widespread adoption. This results in few hospitals in China being able to conduct professional whole-body motor assessments. Furthermore, due to the scarcity of medical resources, many city hospitals lack such specialists, hindering the large-scale implementation of this technology. Therefore, a method and system for early neonatal cerebral palsy screening based on whole-body motor assessment is needed to achieve low-cost, intelligent screening, thereby overcoming limitations imposed by medical conditions and enabling newborns in different regions to enjoy equal access to medical resources.

[0003] Existing related patents include Chinese invention patents with application number 202210711384.2, entitled "A Method and System for Classifying Infant Behavioral Characteristics Based on Deep Learning"; Chinese invention patents with application number 202210622793.5, entitled "An Early Screening System, Device and Storage Medium for Cerebral Palsy Based on Infant Dynamic Posture Estimation"; and Chinese invention patents with application number 201710341971.6, entitled "An Intelligent Assessment System for Whole-Body Motion Quality." These patents also analyze the motor behavior characteristics of newborns to determine whether they exhibit restless movements. However, all of these patents acquire newborn motion information data through video capture. The acquired motion information is primarily planar (X-axis and Y-axis) and completely loses depth information. Human movement is spatial, therefore, the raw data for early infant assessment is insufficient. Meanwhile, this method suffers from issues such as image occlusion and inaccurate pose recognition, leading to inaccurate raw data and consequently, errors in the final evaluation results. Furthermore, existing pose acquisition techniques can only capture large movements; minute movements are easily drowned out by inherent fluctuations and noise within the pose points themselves. Additionally, the video acquisition frequency is limited, generally not exceeding 30Hz, preventing the acquisition of effective motion information at higher frequencies and hindering the exploration of the relationship between neonatal cerebral palsy and whole-body movement. In conclusion, video-based early infant motor assessment inevitably suffers from data errors, preventing true analysis in practice and limiting its application to a mere "classification" level. Moreover, deploying such methods often requires large models and networks, making large-scale deployment difficult in practice. Summary of the Invention

[0004] To achieve the above-mentioned objectives and other advantages of the present invention, a first objective of the present invention is to provide a method for ultra-early screening of neonatal cerebral palsy, comprising the following steps:

[0005] Acquire vibration information and motion data; wherein, the vibration information and the motion data are both obtained by high-frequency sampling using sensors;

[0006] The vibration information is subjected to frequency division processing to obtain the corresponding physiological data;

[0007] The physiological data is used to determine the newborn's emotional state during the data collection process; wherein, the emotional state includes crying state and natural awake state;

[0008] Remove motion data corresponding to crying states and retain motion data corresponding to natural, awake states.

[0009] The motion data corresponding to the natural awake state are used to solve for features, and a feature matrix is ​​constructed using the obtained features.

[0010] The feature matrix is ​​used as input to the classification model to obtain the classification result.

[0011] Furthermore, the vibration information is vibration information of the newborn's heart obtained through high-frequency sampling by the physiological data sensor module.

[0012] Furthermore, the motion data includes head motion data, left wrist motion data, right wrist motion data, left ankle motion data, and right ankle motion data.

[0013] Furthermore, the step of performing frequency division processing on the vibration information to obtain the corresponding physiological data specifically involves performing Butterworth bandpass filtering on the vibration information to obtain acceleration data in different frequency ranges; wherein, the acceleration data includes respiratory data and heart rate data.

[0014] Furthermore, determining the newborn's emotional state during the data collection process using the physiological data includes the following steps:

[0015] The acceleration data is subjected to a short-time Fourier transform to obtain the time-frequency information of the newborn's heart rate and respiratory data;

[0016] Calculate the mean frequency intensity and frequency center within a time window of a preset length.

[0017] A support vector machine binary classification model is established based on the mean frequency intensity and frequency center after model training;

[0018] The support vector machine binary classification model is used to determine whether the data within the time window corresponds to a crying state or a naturally awake state.

[0019] Furthermore, the feature extraction of the motion data corresponding to the natural awake state includes the following steps:

[0020] Linear interpolation was performed on the motion data corresponding to the natural waking state.

[0021] The processed motion data is then subjected to low-pass filtering;

[0022] The filtered motion data is segmented using a time sliding window to obtain motion data for the five major joints;

[0023] Solve for the time domain features, frequency domain features, and entropy features of each dimension of all data dimensions for each joint.

[0024] Furthermore, the process of constructing a feature matrix from the obtained features includes the following steps:

[0025] Construct a feature matrix from the time-domain features, frequency-domain features, and entropy features of each joint;

[0026] The feature matrix is ​​sorted according to the order of human joints from top to bottom, and then transformed by matrix transpose to form a feature map.

[0027] Furthermore, the process of constructing a feature matrix from the obtained features also includes the following steps:

[0028] The time-domain features, frequency-domain features, and entropy features are sorted in the order of acceleration first and then angular velocity to form a series of joint feature data.

[0029] Furthermore, the classification model is a convolutional neural network model.

[0030] Furthermore, the classification model is a machine learning model.

[0031] Furthermore, it also includes the following steps:

[0032] Heart sound data and lung sound data are obtained by performing Butterworth high-pass filtering on the vibration information corresponding to the natural awake state.

[0033] Fourier transform is performed on the heart sound data and the lung sound data to obtain frequency information;

[0034] Extract the frequency features of the frequency information;

[0035] The frequency characteristics are compared with the normal heart and lung sound rate information of the newborn to obtain the quality of the heart and lung sounds in the newborn's natural awake state.

[0036] Furthermore, it also includes the following steps:

[0037] The classification results are comprehensively analyzed and evaluated with the quality of heart sounds and lung sounds obtained from newborns in a naturally awake state, and a risk assessment report is given.

[0038] Furthermore, the acquisition of vibration information and motion data specifically involves obtaining physiological data and motion data from the database corresponding to the vibration information and motion data.

[0039] Furthermore, the database is divided according to sensor type, which includes physiological data sensor, head sensor, left wrist sensor, right wrist sensor, left ankle sensor, and right ankle sensor.

[0040] A second objective of the present invention is to provide an electronic device comprising: a memory having program code stored thereon; a processor connected to the memory, and wherein, when the program code is executed by the processor, a method for ultra-early screening of neonatal cerebral palsy is implemented.

[0041] A third objective of this invention is to provide a computer-readable storage medium having program instructions stored thereon, which, when executed, implement a method for ultra-early screening of neonatal cerebral palsy.

[0042] The fourth objective of this invention is to provide a neonatal cerebral palsy ultra-early screening system, implementing the above-mentioned method, comprising a motion data sensor module, a physiological data sensor module, a data display module, and a processor; the motion data sensor module includes a head sensor, a left wrist sensor, a right wrist sensor, a left ankle sensor, and a right ankle sensor; wherein,

[0043] The head sensor is used to collect head movement data at high frequency; the left wrist sensor is used to collect left wrist movement data at high frequency; the right wrist sensor is used to collect right wrist movement data at high frequency; the left ankle sensor is used to collect left ankle movement data at high frequency; and the right ankle sensor is used to collect right ankle movement data at high frequency.

[0044] The physiological data sensor module is used to sample vibration information from the newborn's heart at high frequency;

[0045] The data display module is used to send motion data and vibration information to the processor for real-time display;

[0046] The processor is used to store motion data and vibration information in databases according to sensor type, forming independent databases for each body part; acquire vibration information and motion data; perform frequency division processing on the vibration information to obtain corresponding physiological data; determine the newborn's emotional state during the data acquisition process based on the physiological data; wherein the emotional state includes crying state and naturally awake state; filter out motion data corresponding to crying state and retain motion data corresponding to naturally awake state; perform feature extraction on the motion data corresponding to naturally awake state, construct a feature matrix based on the obtained features; and use the feature matrix as input to a classification model to obtain classification results.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] This invention utilizes sensor data for cerebral palsy risk assessment, significantly reducing data storage and usage. This enables rapid result retrieval and greatly reduces the demands on related computer equipment during practical use and promotion. Simultaneously, by acquiring and analyzing heart rate and respiration, the purity of the overall movement data is improved, ensuring that all movement data originates from the newborn's natural movements rather than disruptive ones, making the data analysis more meaningful. By introducing heart and lung sound data corresponding to the movement data stage, a deeper assessment of the newborn's physiological state in a natural state is achieved, and a comprehensive risk assessment report is provided based on the classification judgments derived from the movement data.

[0049] This invention can be easily and quickly promoted, overcoming the limitations of uneven development in medical conditions, and enabling universal early neurodevelopmental screening for all newborns. This allows for the early detection and screening of children at risk of cerebral palsy, facilitating early intervention and treatment, and maximizing the chances of recovery for these children. For newborns, the earlier rehabilitation training begins, the lower the long-term impact of cerebral palsy will be, and it may even ensure that the child can achieve a normal life later in life.

[0050] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it according to the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Specific embodiments of the present invention are given in detail below with reference to the accompanying drawings. Attached Figure Description

[0051] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0052] Figure 1 This is a schematic diagram of the neonatal cerebral palsy ultra-early screening system of Example 1;

[0053] Figure 2 This is a schematic diagram of the data extraction and analysis module, mathematical model, and classification and evaluation module in Example 1;

[0054] Figure 3 This is a schematic diagram of neonatal movement data from Example 1;

[0055] Figure 4 The motion data processing flow of Example 1 Figure 1 ;

[0056] Figure 5 The motion data processing flow of Example 1 Figure 2 ;

[0057] Figure 6 This is a schematic diagram of the convolutional neural network in Example 1;

[0058] Figure 7 This is a schematic diagram illustrating the classification accuracy of the machine learning model in Example 1;

[0059] Figure 8 This is a flowchart of the neonatal cerebral palsy ultra-early screening method in Example 2;

[0060] Figure 9 This is a flowchart illustrating the process of determining a newborn's emotional state during data collection using physiological data, as described in Example 2.

[0061] Figure 10 This is a flowchart of the feature extraction process for motion data corresponding to a natural waking state, as described in Example 2.

[0062] Figure 11 This is a flowchart of the comprehensive analysis and evaluation process for Example 2;

[0063] Figure 12 This is a schematic diagram of the electronic device in Example 3;

[0064] Figure 13 This is a schematic diagram of the storage medium in Example 4. Detailed Implementation

[0065] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0066] Example 1

[0067] A newborn cerebral palsy ultra-early screening system can assess the quality of restless motor skills in newborns, enabling cerebral palsy screening for newborns with a corrected age of less than five months. Figure 1 As shown, the system includes a motion data sensor module, a physiological data sensor module 2, a data display module 3, and a processor; the motion data sensor module includes a head sensor 1-1, a left wrist sensor 1-2, a right wrist sensor 1-3, a left ankle sensor 1-4, and a right ankle sensor 1-5; among which,

[0068] The head sensor is used to collect head movement data at a high frequency; the left wrist sensor is used to collect left wrist movement data at a high frequency; the right wrist sensor is used to collect right wrist movement data at a high frequency; the left ankle sensor is used to collect left ankle movement data at a high frequency; and the right ankle sensor is used to collect right ankle movement data at a high frequency. In this embodiment, the head sensor 1-1, left wrist sensor 1-2, right wrist sensor 1-3, left ankle sensor 1-4, and right ankle sensor 1-5 use a 100Hz high frequency to collect movement data of the five major parts of the newborn.

[0069] The physiological data sensor module is used to sample vibration information from the newborn's heart at high frequency;

[0070] In this embodiment, the physiological data sensor module 2 uses the same type of sensor as the motion data sensor module and is attached to the newborn's heart. Therefore, the raw data collected by the physiological data sensor module 2 is the same as the raw data collected by the motion data sensor module 1. However, the physiological data sensor module 2 collects vibration information at a high frequency of 2000Hz. Therefore, from a mechatronics perspective, the vibration information it collects contains more physiological data. By collecting vibration information located at the newborn's heart through the physiological data sensor module 2, relevant physiological data (respiratory rate, heart rate, etc.) can be extracted.

[0071] The data display module is used to send motion data and vibration information to the processor for real-time display;

[0072] In this embodiment, the data display module 3 uses Bluetooth communication to acquire motion data from the motion data sensor module 1 and higher frequency vibration information from the physiological data sensor module 2, and displays it on the computer interface in real time.

[0073] The processor is used to store motion data and vibration information in separate databases according to sensor type, forming independent databases for each part, so that they can be retrieved and analyzed later. Figure 3 This demonstrates the motion data of the newborn collected by five sensors in the motion data sensor module. Figure 3 A is a newborn whose doctor assesses them as "at risk". Figure 3 b is a newborn that the doctor assessed as "normal".

[0074] like Figure 2 As shown, the processor includes a data extraction and analysis module 4, a mathematical model 5, and a classification and evaluation module 6. The data extraction and analysis module 4 processes the data stored in various independent databases after passing through the data display module 3. This module has two main processing functions. The first function is to process the data collected by the physiological data sensor module 2. Although this module still obtains acceleration and angular velocity data, because it is ultra-high frequency vibration information, it contains a lot of physiological signal data, such as heart rate, respiratory rate, and higher-frequency heart and lung sounds. Therefore, the corresponding physiological data can be obtained by performing a series of frequency division processes on the data from the physiological data sensor module 2. Specifically, acceleration data in different frequency ranges is obtained by performing Butterworth bandpass filtering on the acquired acceleration data. For example, respiratory data is obtained with a bandpass frequency of 0.08Hz to 0.9Hz, heart rate data is obtained with a bandpass frequency of 1-5Hz, and acceleration data containing heart and lung sounds can be obtained when the bandpass frequency is extended to within 1000Hz.

[0075] By using real-time heart rate and respiratory rate data to differentiate between awake and crying states, the emotional state of newborns during data collection is determined. This emotional state includes both crying and naturally awake states. When a newborn is naturally awake, their breathing and heart rate are stable. However, when a newborn is crying, their heart rate and breathing become rapid, showing a clear distinction. Therefore, respiratory and heart rate monitoring during data collection can serve as a standard for classifying the collection time, dividing the collected data into two main time phases: the naturally awake phase and the crying phase. This allows for the division of newborn movement according to the same time phase, categorizing it into limb movements during the naturally awake phase and those during the crying phase. This effectively purifies the newborn movement data, retaining only the movement data corresponding to the naturally awake state and discarding the movement data corresponding to the crying state. Movement data during the crying state cannot reflect the true movement of newborns, making it unsuitable for newborn cerebral palsy screening and potentially leading to errors in the screening of the true state; therefore, it needs to be deleted.

[0076] Specifically, crying causes a physiological surge in infants, leading to a more active heart than usual, and stronger breathing, resulting in more rhythmic and rapid lung movements. By analyzing the high-frequency acceleration data of newborns acquired throughout the data collection period, a Butterworth low-pass filter was first applied to remove frequencies above 30Hz to eliminate high-frequency noise. Then, a short-time Fourier transform was performed on the acceleration to obtain the frequency components of the acceleration signal over time. As mentioned earlier, the heart rate and breathing of newborns during crying differ from those in the awake state; specifically, the crying period exhibits higher energy in the frequency domain. Based on this characteristic, time-frequency information was acquired from multiple segments of neonatal heart rate and respiratory data. A 5-second time window was used to calculate the mean frequency intensity and frequency center within the window. Based on the two parameters of mean intensity and frequency center, a binary classification model of support vector machine was established through model training to determine whether the neonate was awake within this time window. Because the mean frequency intensity of crying will be higher and the frequency center will shift to the right, the motion data corresponding to the crying state will be deleted to ensure that the obtained motion data consists entirely of motion data from the natural awake state.

[0077] Heart and lung sound data of newborns in a naturally awake state were used as one of the auxiliary diagnostic criteria for newborn cerebral palsy screening. Frequency domain analysis of the obtained heart and lung sound data was performed to obtain the quality of heart and lung sounds in newborns in a naturally awake state, which was used to assist in analyzing the overall health status of newborns.

[0078] Specifically, since complications associated with cerebral palsy can lead to potential muscle control problems, which in turn can affect lung and heart development, acceleration data at a sampling frequency of 1000Hz was acquired. A Butterworth high-pass filter was used to filter out frequencies below 50Hz to remove acceleration changes caused by neonatal movement. Then, a Fourier transform was performed on the preprocessed data to obtain frequency information. Basic frequency characteristics were then extracted from this information. The infant's frequency information was compared with the heart and lung sounds of a normal infant to make a basic assessment, thereby determining whether there are deeper impacts on the newborn's lungs and heart.

[0079] The second processing function of the data extraction and analysis module 4 is to process data from five independent databases corresponding to the motion data sensor module. First, based on the natural awakening stage and crying stage obtained after the first processing function, the data in the five independent databases are purified to ensure that the data analyzed later are all motion data in the natural awakening stage. Then, feature extraction is performed on the motion data to obtain time domain features, frequency domain features, and entropy features, and a single-joint data feature model is established.

[0080] like Figure 4 As shown, each joint has six motion time-series data points, one square root acceleration data point, and one square root angular velocity data point, totaling eight dimensions of motion data. First, the data undergoes linear interpolation, followed by low-pass filtering to retain motion data within 10Hz. The motion data is then segmented, including a 60-second sliding window and a 10-second motion overlap. After the sliding window, a standard one-minute motion data point for the five major joints is obtained. Time-domain, frequency-domain, and entropy features are calculated for each of the six data dimensions for each joint, resulting in 36 feature data points. Therefore, each joint of each newborn forms a 36×8 feature model.

[0081] Mathematical Model 5 consists of six parts, representing data from six different parts of the newborn's body. The five outermost parts of Mathematical Model 5 represent feature data from five motion data sensor modules. This feature data has been processed to remove data from the crying stage, retaining only the data from the natural awakening stage, thus greatly improving the purity of the data. Figure 5 and Figure 6 The specific structure of mathematical model 5 is shown. Figure 5 This demonstrates the data graph construction process. After initial data purification and time-windowing (60s window length, 10s overlap), a multi-dimensional data matrix is ​​constructed. The time-frequency domain and entropy characteristics of the data under the corresponding window length are then calculated. Finally, the data sources are divided into head, left and right wrists, and left and right ankle data graphs, as shown in the five outer circles of mathematical model 5. The establishment of this data graph ensures the smoothness of the subsequent overall model construction. Figure 6This demonstrates the specific model layout in Data Model 5, using the Keras framework to design a two-layer convolutional neural classification network, specifically in the following order: First convolutional layer: 128-channel 2×2 convolution with ReLU activation; Second convolutional layer: 64-channel 3×3 vector; Fully connected layer: 100 neurons with ReLU activation; Last output layer: two neurons with SOFTMAX activation, used to output the classification result. The feature matrix is ​​used as input to the classification model to obtain the classification result.

[0082] like Figure 5 As shown, after time windowing, the neonatal motion data yields multiple one-minute new sample data points. These sample data belong to the newborn's own motion data, thus enabling the construction of a multi-dimensional data matrix. For example, after time windowing, a newborn's motion data has 7 new sample data points. Each sample data point contains five joint data points, and each joint motion data point contains 8 dimensions of data. Each dimension has 1800 data points (60 seconds × 30Hz), ultimately constructing a 7 × 5 × 1800 × 8 data matrix.

[0083] Feature extraction is performed on each new one-minute sample data point. For each joint within the new one-minute sample data point, time-domain, frequency-domain, and entropy features are calculated, and an 8×36 feature matrix is ​​constructed. For example... Figure 5 As shown, there are a total of 5 8×36 feature matrices based on the joint type. The five feature matrices are sorted according to the vertical order of the human joints, namely head, left wrist, right wrist, left ankle, and right ankle, and then transformed by matrix transpose to finally form a 40×36 feature map.

[0084] In one embodiment, there is another mode for mathematical model 5, namely, constructing a machine learning model. Instead of building a three-dimensional feature matrix, the features are solved directly using time series data. Thus, each joint obtains a feature sequence, and each newborn has a two-dimensional feature matrix composed of 5 feature sequences. The classification model is constructed using machine learning models such as the nearest neighbor algorithm (KNN), support vector machine (SVM), decision tree (DT), and random forest. Figure 7 This indicates the classification accuracy of the eleven machine learning models established after the above data processing steps. The classification accuracy obtained using the above systematic method generally reaches over 75%, with the accuracy of logistic regression (LR) approaching 100%.

[0085] Based on the aforementioned six-dimensional data (three-axis acceleration and three-axis angular velocity), the motion time series data of each joint are obtained by solving the time domain, frequency domain, and entropy features of each dimension. The feature data are then sorted in the order of acceleration first and then angular velocity to form a complete series of joint feature data.

[0086] The classification and assessment module 6 provides a comprehensive risk assessment report based on the results of mathematical model 5 and the analysis results of heart sounds and lung sounds in physiological data sensor module 2.

[0087] Specifically, the motion data is processed by mathematical model 5 to obtain three classification results: high risk, low risk, and normal. These results are used to indicate the degree of risk of cerebral palsy in children, to judge the quality of newborn movement, and to provide a basic judgment for newborn cerebral palsy screening. Based on this, the heart sound and lung sound data in physiological data sensor module 2 are compared with the frequency information of normal newborns after the above analysis process to determine whether the newborn's development has been affected at a deeper level.

[0088] Example 2

[0089] A method for ultra-early screening of neonatal cerebral palsy can assess the quality of restless motor movements in newborns, enabling cerebral palsy screening for newborns with a corrected age of less than five months. Figure 8 As shown, the method includes the following steps:

[0090] S1. Acquire vibration information and motion data; where both vibration information and motion data are obtained using high-frequency sampling by sensors; the vibration information is vibration information of the newborn's heart area obtained through high-frequency sampling by the physiological data sensor module. The motion data includes head motion data sampled at high frequency by the head sensor, left wrist motion data sampled at high frequency by the left wrist sensor, right wrist motion data sampled at high frequency by the right wrist sensor, left ankle motion data sampled at high frequency by the left ankle sensor, and right ankle motion data sampled at high frequency by the right ankle sensor.

[0091] In this embodiment, head sensor 1-1, left wrist sensor 1-2, right wrist sensor 1-3, left ankle sensor 1-4, and right ankle sensor 1-5 use a high frequency of 100Hz to collect motion data of the five major parts of the newborn.

[0092] Physiological data sensor module 2 uses the same type of sensor as the motion data sensor module and is attached to the newborn's heart area. Therefore, the raw data collected by physiological data sensor module 2 is the same as that collected by motion data sensor module 1. However, physiological data sensor module 2 collects vibration information at a high frequency of 2000Hz. Therefore, from a mechatronic perspective, the vibration information it collects contains more physiological data. By collecting vibration information located at the newborn's heart area through physiological data sensor module 2, relevant physiological data (respiratory rate, heart rate, etc.) can be extracted.

[0093] The acquisition of vibration and motion data specifically involves retrieving physiological and motion data from corresponding databases. These databases are categorized by sensor type; that is, motion data and vibration information are stored separately for each sensor type, forming independent databases for each body part, facilitating later retrieval and cross-analysis. Sensor types include physiological data sensors, head sensors, left wrist sensors, right wrist sensors, left ankle sensors, and right ankle sensors.

[0094] S2. The vibration information is processed by frequency division to obtain the corresponding physiological data; specifically, acceleration data in different frequency ranges is obtained by Butterworth bandpass filtering of the vibration information; among which, the acceleration data includes respiratory data and heart rate data.

[0095] Although the physiological data sensor module still acquires acceleration and angular velocity data, due to the ultra-high frequency vibration information, it contains a wealth of physiological signals, such as heart rate, respiratory rate, and even higher-frequency heart and lung sounds. Therefore, the corresponding physiological data can be obtained by performing a series of frequency division processes on the data from physiological data sensor module 2. Specifically, acceleration data within different frequency ranges is obtained by performing Butterworth bandpass filtering on the acquired acceleration data. For example, respiratory data is acquired at bandpass frequencies of 0.08Hz to 0.9Hz, heart rate data is acquired at bandpass frequencies of 1-5Hz, and acceleration data containing heart and lung sounds can be obtained when the bandpass frequency is extended to below 1000Hz.

[0096] When a newborn is moving in a naturally awake state, their breathing and heart rate are stable. However, when a newborn is moving while crying, their heart rate and breathing become rapid, showing a clear distinction between the two. Therefore, by monitoring breathing and heart rate during data collection, we can use this data to classify the data into two main time phases: the naturally awake phase and the crying phase. This allows us to categorize newborn movement according to the same time phase, dividing it into limb movements during the naturally awake phase and limb movements during the crying phase. This effectively purifies the newborn movement data, retaining only the movement data corresponding to the naturally awake state and discarding the movement data corresponding to the crying state. Movement data during the crying state cannot reflect the true movement of newborns; it is not only unsuitable for newborn cerebral palsy screening but also leads to errors in screening based on the true state, therefore it needs to be deleted.

[0097] S3. Determine the newborn's emotional state during the data collection process using physiological data; the emotional state includes crying and naturally awake states.

[0098] Crying causes a physiological surge in infants, leading to a more active heart than usual, and more forceful breathing, resulting in more rhythmic and rapid lung movements. Figure 9 As shown, determining a newborn's emotional state during the data collection process using physiological data includes the following steps:

[0099] S31. As mentioned earlier, the heart rate and respiration of newborns during crying are different from those in the awake state. Specifically, the crying period has higher energy in the frequency domain. Based on this characteristic, a short-time Fourier transform is performed on the acceleration data to obtain the time-frequency information of the newborn's heart rate and respiration data.

[0100] S32. Calculate the mean frequency intensity and frequency center within a time window of a preset length (e.g., 5 seconds).

[0101] S33. Based on the mean frequency intensity and frequency center, a support vector machine binary classification model is established through model training;

[0102] S34. Because the average frequency intensity of crying increases and the frequency center shifts to the right, a support vector machine binary classification model is used to determine whether the data within the time window corresponds to the crying state or the natural awake state. Motion data corresponding to the crying state is deleted to ensure that the obtained motion data consists entirely of motion data from the natural awake state.

[0103] S4. Filter out the motion data corresponding to the crying state and retain the motion data corresponding to the natural awake state.

[0104] S5. Solve the feature extraction of the motion data corresponding to the natural awake state, and construct a feature matrix using the obtained features;

[0105] like Figure 4 As shown, each joint has six motion timing data points, one square root acceleration data point, and one square root angular velocity data point, for a total of eight dimensions of motion data. Figure 10 As shown, feature extraction of motion data corresponding to a natural waking state includes the following steps:

[0106] S51. Perform linear interpolation on the motion data corresponding to the natural waking state;

[0107] S52. Perform low-pass filtering on the processed motion data to retain motion data within 10Hz;

[0108] S53. The filtered motion data is segmented by a time sliding window, including a 60-second time sliding window and a 10-second motion overlap. After the sliding window processing, there will be a standard one-minute motion data of the five major joints.

[0109] S54. Solve for the time domain features, frequency domain features, and entropy features of each dimension of all data dimensions for each joint. There are a total of 36 feature data. Thus, each joint of each newborn forms a 36×8 feature model.

[0110] In one embodiment, constructing a feature matrix from the obtained features includes the following steps:

[0111] Construct a feature matrix from the time-domain features, frequency-domain features, and entropy features of each joint;

[0112] The feature matrix is ​​sorted according to the order of human joints from top to bottom, and then transformed by matrix transpose to form a feature map.

[0113] Mathematical Model 5 consists of six parts, representing data from six different parts of the newborn's body. The five outermost parts of Mathematical Model 5 represent feature data from five motion data sensor modules. This feature data has been processed to remove data from the crying stage, retaining only the data from the natural awakening stage, thus greatly improving the purity of the data. Figure 5 and Figure 6 The specific structure of mathematical model 5 is shown. Figure 5 This demonstrates the data graph construction process. After initial data purification and time-windowing (60s window length, 10s overlap), a multi-dimensional data matrix is ​​constructed. The time-frequency domain and entropy characteristics of the data under the corresponding window length are then calculated. Finally, the data sources are divided into head, left and right wrists, and left and right ankle data graphs, as shown in the five outer circles of mathematical model 5. The establishment of this data graph ensures the smoothness of the subsequent overall model construction. Figure 6 This demonstrates the specific model layout in Data Model 5, using the Keras framework to design a two-layer convolutional neural classification network, specifically in the following order: First convolutional layer: 128-channel 2×2 convolution with ReLU activation; Second convolutional layer: 64-channel 3×3 vector; Fully connected layer: 100 neurons with ReLU activation; Last output layer: two neurons with SOFTMAX activation, used to output the classification result. The feature matrix is ​​used as input to the classification model to obtain the classification result.

[0114] In another embodiment, constructing a feature matrix from the obtained features further includes the following steps:

[0115] The time-domain features, frequency-domain features, and entropy features are sorted in the order of acceleration first, then angular velocity, to form a series of joint feature data. Each newborn has a two-dimensional feature matrix consisting of 5 columns of feature sequences, and a classification model is constructed using machine learning models such as the nearest neighbor algorithm (KNN), support vector machine (SVM), decision tree (DT), and random forest. Figure 7This indicates the classification accuracy of the eleven machine learning models established after the above data processing steps. The classification accuracy obtained using the above systematic method generally reaches over 75%, with the accuracy of logistic regression (LR) approaching 100%.

[0116] S6. Use the feature matrix as input to the classification model to obtain the classification result.

[0117] Because complications associated with cerebral palsy can lead to potential muscle control problems, which can affect lung and heart development, heart and lung sound data obtained from newborns in a naturally awake state are used as one of the auxiliary diagnostic criteria for newborn cerebral palsy screening. Frequency domain analysis of the obtained heart and lung sound data is performed to obtain the quality of heart and lung sounds in newborns in a naturally awake state, which is used to assist in analyzing the overall health status of newborns. Figure 11 As shown, it also includes step S7:

[0118] S71. By performing Butterworth high-pass filtering on the vibration information corresponding to the natural awake state to remove frequency information below 50Hz and eliminate acceleration changes caused by neonatal movement, heart sound data and lung sound data are obtained.

[0119] S72. Perform Fourier transform on the heart sound data and lung sound data to obtain frequency information;

[0120] S73. Extract frequency features from frequency information;

[0121] S74. Compare the frequency characteristics with the normal heart and lung sound frequency information of the newborn to obtain the quality of the heart and lung sounds in the newborn's natural awake state, thereby determining whether there is a deeper impact on the newborn's lungs and heart.

[0122] S75. The classification results are comprehensively analyzed and evaluated with the quality of heart sounds and lung sounds obtained from the newborn in a naturally awake state, and a risk assessment report is given. Specifically, the motor data is processed by mathematical model 5 to obtain three-category results: high risk, low risk, and normal, which are used to represent the degree of risk of cerebral palsy in children, to judge the quality of the newborn's motor function, and to provide a basic judgment for newborn cerebral palsy screening. On this basis, the heart sound and lung sound data in physiological data sensor module 2 are compared with the frequency information of normal newborns after the above analysis process to determine whether the newborn's development has been affected at a deeper level.

[0123] Example 3

[0124] An electronic device 200, such as Figure 12As shown, the method includes, but is not limited to: a memory 201 storing program code; and a processor 202 connected to the memory, which, when executed by the processor, implements a method for early screening of neonatal cerebral palsy. For a detailed description of the method, please refer to the corresponding description in the above method embodiments, which will not be repeated here.

[0125] Example 4

[0126] A computer-readable storage medium, such as Figure 13 As shown, it stores program instructions, which, when executed, implement a method for ultra-early screening of neonatal cerebral palsy. For a detailed description of the method, please refer to the corresponding description in the above method embodiments, which will not be repeated here.

[0127] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0128] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0129] The above are merely embodiments of this specification and are not intended to limit the scope of one or more embodiments of this specification. Various modifications and variations can be made to one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of one or more embodiments of this specification.

Claims

1. A neonatal cerebral palsy ultra-early screening system, characterized in that: It includes a motion data sensor module, a physiological data sensor module, a data display module, and a processor; the motion data sensor module includes a head sensor, a left wrist sensor, a right wrist sensor, a left ankle sensor, and a right ankle sensor; wherein, The head sensor is used to collect head movement data at high frequency; the left wrist sensor is used to collect left wrist movement data at high frequency; the right wrist sensor is used to collect right wrist movement data at high frequency; the left ankle sensor is used to collect left ankle movement data at high frequency; and the right ankle sensor is used to collect right ankle movement data at high frequency. The physiological data sensor module is used to sample vibration information of the newborn's heart at a higher frequency than that of the motion data sensor module. The data display module is used to send motion data and vibration information to the processor for real-time display; The processor is used to store motion data and vibration information in separate databases according to sensor type, forming independent databases for each body part; acquire vibration information and motion data; perform frequency division processing on the vibration information to obtain corresponding physiological data; determine the emotional state of the newborn during the data acquisition process based on the physiological data; wherein, the emotional state includes crying state and naturally awake state; filter out motion data corresponding to crying state and retain motion data corresponding to naturally awake state; perform feature extraction on the motion data corresponding to naturally awake state, construct a feature matrix based on the obtained features; and use the feature matrix as input to a classification model to obtain the classification result. The step of performing frequency division processing on the vibration information to obtain corresponding physiological data specifically involves performing Butterworth bandpass filtering on the vibration information to obtain acceleration data within different frequency ranges; wherein, the acceleration data includes respiratory data and heart rate data; The process of determining the newborn's emotional state during the data collection process using the physiological data includes the following steps: The acceleration data is subjected to a short-time Fourier transform to obtain the time-frequency information of the newborn's heart rate and respiratory data; Calculate the mean frequency intensity and frequency center within a time window of a preset length. A support vector machine binary classification model is established based on the mean frequency intensity and frequency center after model training; The support vector machine binary classification model is used to determine whether the data within the time window corresponds to a crying state or a naturally awake state. The process of feature extraction from motion data corresponding to a natural waking state includes the following steps: Linear interpolation was performed on the motion data corresponding to the natural waking state. The processed motion data is then subjected to low-pass filtering; The filtered motion data is segmented using a time sliding window to obtain motion data for the five major joints; Solve for the time domain features, frequency domain features, and entropy features of each dimension of all data dimensions for each joint; The process of constructing a feature matrix from the obtained features includes the following steps: Construct a feature matrix from the time-domain features, frequency-domain features, and entropy features of each joint; The feature matrix is ​​sorted according to the order of human joints from top to bottom, and then transformed by matrix transpose to form a feature map.

2. The neonatal cerebral palsy ultra-early screening system as described in claim 1, characterized in that: The vibration information is obtained by high-frequency sampling of the newborn's heart area through a physiological data sensor module.

3. The neonatal cerebral palsy ultra-early screening system as described in claim 2, characterized in that: The motion data includes head motion data, left wrist motion data, right wrist motion data, left ankle motion data, and right ankle motion data.

4. The neonatal cerebral palsy ultra-early screening system as described in claim 1, characterized in that: The process of constructing a feature matrix from the obtained features also includes the following steps: The time-domain features, frequency-domain features, and entropy features are sorted in the order of acceleration first and then angular velocity to form a series of joint feature data.

5. The neonatal cerebral palsy ultra-early screening system as described in claim 1, characterized in that: The classification model is a convolutional neural network model.

6. The neonatal cerebral palsy ultra-early screening system as described in claim 4, characterized in that: The classification model is a machine learning model.

7. The neonatal cerebral palsy ultra-early screening system as described in claim 1, characterized in that: It also includes the following steps: Heart sound data and lung sound data are obtained by performing Butterworth high-pass filtering on the vibration information corresponding to the natural awake state. Fourier transform is performed on the heart sound data and the lung sound data to obtain frequency information; Extract the frequency features of the frequency information; The frequency characteristics are compared with the normal heart and lung sound rate information of the newborn to obtain the quality of the heart and lung sounds in the newborn's natural awake state.

8. The neonatal cerebral palsy ultra-early screening system as described in claim 7, characterized in that: It also includes the following steps: The classification results are comprehensively analyzed and evaluated with the quality of heart sounds and lung sounds obtained from newborns in a naturally awake state, and a risk assessment report is given.

9. The neonatal cerebral palsy ultra-early screening system as described in claim 3, characterized in that: The acquisition of vibration information and motion data specifically involves obtaining physiological data and motion data from the database corresponding to the vibration information and motion data.

10. The neonatal cerebral palsy ultra-early screening system as described in claim 9, characterized in that: The database is divided according to sensor type, which includes physiological data sensor, head sensor, left wrist sensor, right wrist sensor, left ankle sensor, and right ankle sensor.

Images