Postoperative vertebral fracture patient risk grading system based on neural network

Abstract

The invention discloses a vertebral fracture postoperative patient risk grading system based on a neural network, and relates to the technical field of artificial intelligence and medical imagines.The vertebral fracture postoperative patient risk grading system comprises the steps that image data, physiological index data and motion sensing data of a patient are collected; calculating a three-dimensional displacement track of the cone based on the image data, and extracting a sudden change time point in a displacement direction and a displacement amplitude to obtain an instability feature set; performing time alignment on the motion sensing data and the instability characteristic set, and analyzing a muscle activation signal through a neural network to obtain a compensation ratio sequence; counting the abrupt change frequency in the displacement direction in the instability characteristic set and calculating a risk transfer intensity value in combination with the compensation ratio sequence growth rate; and determining the postoperative complication risk level of the patient according to the numerical value interval of the risk transmission intensity value and the abnormal fluctuation condition of the physiological indexes. The objective quantitative evaluation of the postoperative complication risk is realized.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and medical imaging technology, specifically to a risk grading system for postoperative vertebral fracture patients based on neural networks. Background Technology

[0002] Vertebral fractures are a common bone disease, and the assessment of postoperative complication risk has a significant impact on patient rehabilitation and prognosis. Currently, clinical assessment of postoperative complication risk mainly relies on imaging examinations and physician experience, lacking objective and quantitative assessment standards.

[0003] Existing risk assessment methods primarily rely on static imaging data analysis, which fails to effectively reflect the dynamic stability of the vertebral body during changes in patient position. Furthermore, they do not adequately consider the impact of muscle compensation mechanisms on vertebral stability, thus limiting the accuracy and reliability of risk assessment results.

[0004] Traditional assessment methods typically analyze imaging data, physiological indicators, and motion parameters separately, failing to establish correlations between various data and making it difficult to comprehensively reflect the patient's risk status for postoperative complications.

[0005] With the development of medical imaging technology and wearable sensing devices, it has become possible to collect multi-dimensional patient data. How to effectively integrate multi-source heterogeneous data, establish an objective risk assessment model, and achieve accurate classification of postoperative complication risks remains a key focus and challenge of current research. Summary of the Invention

[0006] The purpose of this invention is to provide a risk grading system for postoperative patients with vertebral fractures based on neural networks, aiming to at least solve one of the technical problems existing in the prior art.

[0007] The technical solution of this invention is: a risk grading system for postoperative vertebral fracture patients based on neural networks, characterized in that the system comprises: The data acquisition module is used to collect imaging data, physiological index data, and motion sensor data of patients after vertebral fracture surgery; The displacement trajectory analysis module is used to calculate the three-dimensional displacement trajectory of the vertebrae during the change of body position based on the vertebral spatial coordinates of continuous time points in the image data, extract the time points and displacement amplitudes of the abrupt change in displacement direction in the three-dimensional displacement trajectory, and obtain the set of instability features. The neural network processing module is used to time-align the body position change time sequence in motion sensing data with the time points in the instability feature set, obtain the muscle activation signal corresponding to the time point, and calculate the ratio of the peak intensity to the baseline intensity of the muscle activation signal through the neural network to obtain the compensation ratio sequence. The risk intensity calculation module is used to statistically analyze the frequency of sudden changes in displacement direction in the set of instability characteristics, and calculate the risk transmission intensity value by combining the numerical growth rate of the compensation ratio sequence. The risk grading module is used to determine the level of postoperative complication risk for patients based on the numerical range of the risk transmission intensity value and the abnormal fluctuations in physiological indicator data, and output the grading results.

[0008] The operation of the displacement trajectory analysis module includes the following steps: Collect vertebral body image data at continuous time points and extract the grayscale matrix from the vertebral body image data; Region segmentation is performed based on the gray value matrix to obtain the vertebral body boundary point set, and the pedicle position coordinates and vertebral endplate position coordinates are extracted from the vertebral body boundary point set. Calculate the coordinates of the centroid of the vertebral body based on the coordinates of the pedicle and the coordinates of the endplate of the vertebral body. Construct a local coordinate system with the coordinates of the centroid of the vertebral body as the origin and the direction of the pedicle coordinates as the principal axis. Project the pedicle and vertebral endplate position coordinates onto the local coordinate system and calculate the displacement direction and displacement amplitude between adjacent time points. A three-dimensional displacement trajectory is constructed based on the displacement direction and displacement amplitude. The moment point of sudden change in displacement direction is identified by calculating the gradient value of the three-dimensional displacement trajectory. The moment point is combined with the corresponding displacement amplitude to obtain the set of instability features.

[0009] The operation of the neural network processing module includes the following steps: Muscle activation signals and body position change time sequences are collected from motion sensing data, and the muscle activation signals are filtered to obtain signal waveforms. Extract the time points of body position changes from the body position change time sequence, and divide the signal waveform into multiple signal segments according to the time points of body position changes; The time points of body position changes are time-aligned with the time points in the set of instability characteristics, and the signal segments corresponding to the time points are selected based on the time alignment results. The signal segment corresponding to the time point is subjected to time-frequency transformation to generate a time-frequency feature matrix, and the time-frequency feature matrix is ​​input into a neural network to extract the spectral energy distribution; The signal intensity at rest is identified as the baseline intensity based on the spectral energy distribution, and the signal intensity at contraction is identified as the peak intensity. The ratio of the peak intensity to the baseline intensity is calculated to obtain a compensation ratio sequence.

[0010] The time points of body position changes are time-aligned with the time points in the instability feature set. Based on the time alignment results, signal segments corresponding to the time points are selected, including: The raw signals from motion sensing data are collected, and the time points of body position change are constructed into a body position change time series. The raw signals are segmented according to the body position change time series to obtain the initial signal segments. Calculate the time difference between adjacent time points in the body position change time series and generate a time difference sequence; The temporal distance between the time difference sequence and the time point sequence in the instability feature set is calculated based on the dynamic programming algorithm, a temporal distance matrix is ​​constructed, and a time alignment matching path is extracted from the temporal distance matrix. A time-to-time correspondence table is generated based on the time alignment matching path, and time alignment position information is extracted from the correspondence table. The initial signal segment is decomposed into a multi-scale decomposition matrix using wavelet transform. The coefficient values ​​at the corresponding time points are located and extracted from the decomposition coefficient matrix based on the time alignment information. The coefficient values ​​are reconstructed using the inverse wavelet transform to obtain the signal segment corresponding to the time point. The signal segment is then subjected to piecewise smoothing to obtain the final signal segment.

[0011] The process of inputting the time-frequency feature matrix into a neural network to extract the spectral energy distribution includes: A neural network structure is constructed, which includes a feature extraction module and a fully connected module. The feature extraction module consists of convolutional layers and pooling layers, and the fully connected module consists of multiple layers of connection units. The time-frequency feature matrix is ​​divided into training data matrices according to time windows, and the training data matrices are normalized by maximum and minimum values ​​to obtain normalized feature matrices. The normalized feature matrix is ​​input into the feature extraction module, and frequency domain feature vectors are obtained through convolution and pooling operations. The frequency domain feature vector is input into the fully connected module, and the energy mapping vector is obtained by mapping through the activation function. Perform inverse normalization on the energy mapping vector and construct a spectral energy curve by arranging them in chronological order. Extract the peak and trough position coordinates from the spectral energy curve and generate a spectral energy distribution based on the peak and trough position coordinates.

[0012] The operation of the risk intensity calculation module includes the following steps: Extract the moment points of sudden displacement direction changes from the set of instability features, arrange the moment points of sudden displacement direction changes in chronological order, and calculate the time interval between adjacent moment points of sudden displacement direction changes. The sampling interval is set based on the time interval, and the number of times the displacement direction change occurs within the sampling interval is counted to obtain the frequency of displacement direction change. The compensation ratio sequence is truncated according to the sampling interval, and the numerical change of the compensation ratio within the sampling interval is calculated to obtain the numerical growth rate of the compensation ratio sequence. The risk transmission intensity value is obtained by weighting the frequency of abrupt changes in displacement direction with the numerical growth rate of the compensation ratio sequence.

[0013] The operation of the risk classification module includes the following steps: Collect historical patient risk transmission intensity values ​​and sort them to obtain the numerical range of historical risk transmission intensity values; Based on the numerical range, a dividing benchmark is set, and the numerical range is divided into multiple risk level ranges using the dividing benchmark, and a correspondence between the risk level ranges and the risk levels is established. The fluctuation range of blood pressure and heart rate values ​​in physiological indicator data is statistically analyzed, the fluctuation range is divided into multiple fluctuation level intervals, and the correspondence between fluctuation level intervals and risk correction values ​​is established. The initial risk level is obtained by positioning the risk transmission intensity value of the patient to be evaluated into the risk level range, and the risk correction value is obtained by positioning the physiological index data of the patient to be evaluated into the fluctuation level range. The initial risk level is combined with the risk correction value to calculate and determine the level of postoperative complication risk for the patient, and the classification result is output.

[0014] This invention provides a method for risk stratification of patients after vertebral fracture surgery based on neural networks, characterized in that the method includes: Collect imaging data, physiological data, and motion sensor data of patients after vertebral fracture surgery; Based on the vertebral spatial coordinates at continuous time points in the image data, the three-dimensional displacement trajectory of the vertebral body during the change of body position is calculated, and the time points and displacement amplitudes of the abrupt change in displacement direction in the three-dimensional displacement trajectory are extracted to obtain the set of instability features. The time sequence of body position changes in motion sensing data is time-aligned with the time points in the set of instability features to obtain the muscle activation signals corresponding to the time points. The ratio of the peak intensity to the baseline intensity of the muscle activation signals is calculated by a neural network to obtain the compensation ratio sequence. The frequency of abrupt changes in displacement direction in the statistical instability feature set is used to calculate the risk transmission intensity value by combining the numerical growth rate of the compensation ratio sequence. Based on the numerical range of the risk transmission intensity value and the abnormal fluctuations in physiological indicator data, the level of postoperative complication risk for patients is determined and the classification result is output.

[0015] One technical solution provided in this embodiment of the invention is an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the aforementioned systems.

[0016] One technical solution provided in this embodiment of the invention is a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the steps in any of the aforementioned systems.

[0017] This invention analyzes muscle activation signals based on neural networks, which can accurately assess a patient's compensatory capacity and the intensity of risk transmission; it improves the accuracy of risk assessment by using time-series alignment technology to achieve collaborative analysis of multi-source data; it makes the risk assessment results more comprehensive and reliable by using a hierarchical mapping mechanism combined with abnormal fluctuations in physiological indicators; it automatically completes the risk level determination, reducing the workload of doctors and improving assessment efficiency; and it establishes objective and quantitative assessment standards, providing an important basis for the formulation and adjustment of postoperative rehabilitation plans. Attached Figure Description

[0018] Figure 1 A schematic diagram of a risk grading system for postoperative vertebral fracture patients based on a neural network, provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the comparative analysis of muscle compensation indices under different body position change events according to an embodiment of the present invention. Figure 3 A flowchart of a risk grading method for postoperative vertebral fracture patients based on neural networks, provided for an embodiment of the present invention. Detailed Implementation

[0019] like Figure 1 As shown, Figure 1 A schematic diagram of a neural network-based risk grading system for postoperative vertebral fracture patients provided in this embodiment of the invention, the system comprising: The data acquisition module is used to collect imaging data, physiological index data, and motion sensor data of patients after vertebral fracture surgery; The displacement trajectory analysis module is used to calculate the three-dimensional displacement trajectory of the vertebrae during the change of body position based on the vertebral spatial coordinates of continuous time points in the image data, extract the time points and displacement amplitudes of the abrupt change in displacement direction in the three-dimensional displacement trajectory, and obtain the set of instability features. The neural network processing module is used to time-align the body position change time sequence in motion sensing data with the time points in the instability feature set, obtain the muscle activation signal corresponding to the time point, and calculate the ratio of the peak intensity to the baseline intensity of the muscle activation signal through the neural network to obtain the compensation ratio sequence. The risk intensity calculation module is used to statistically analyze the frequency of sudden changes in displacement direction in the set of instability characteristics, and calculate the risk transmission intensity value by combining the numerical growth rate of the compensation ratio sequence. The risk grading module is used to determine the level of postoperative complication risk for patients based on the numerical range of the risk transmission intensity value and the abnormal fluctuations in physiological indicator data, and output the grading results.

[0020] Based on the vertebral spatial coordinates at continuous time points in the image data, the three-dimensional displacement trajectory of the vertebral body during the postural change process is calculated. The time points and displacement amplitudes of abrupt changes in displacement direction in the three-dimensional displacement trajectory are extracted to obtain the set of instability features, including: Collect vertebral body image data at continuous time points and extract the grayscale matrix from the vertebral body image data; Region segmentation is performed based on the gray value matrix to obtain the vertebral body boundary point set, and the pedicle position coordinates and vertebral endplate position coordinates are extracted from the vertebral body boundary point set. Calculate the coordinates of the centroid of the vertebral body based on the coordinates of the pedicle and the coordinates of the endplate of the vertebral body. Construct a local coordinate system with the coordinates of the centroid of the vertebral body as the origin and the direction of the pedicle coordinates as the principal axis. Project the pedicle and vertebral endplate position coordinates onto the local coordinate system and calculate the displacement direction and displacement amplitude between adjacent time points. A three-dimensional displacement trajectory is constructed based on the displacement direction and displacement amplitude. The moment point of sudden change in displacement direction is identified by calculating the gradient value of the three-dimensional displacement trajectory. The moment point is combined with the corresponding displacement amplitude to obtain the set of instability features.

[0021] Imaging data of patients undergoing vertebral fracture surgery under different post-operative positions was acquired. Continuous scanning of the patient was performed using medical imaging equipment to obtain vertebral body imaging data at consecutive time points. The acquired image data can be CT or MRI image sequences, each containing two-dimensional cross-sectional information of the vertebral body. After importing all image sequences into an image processing system, a grayscale matrix was extracted from each image. Each element in the grayscale matrix represents the pixel intensity at the corresponding location in the image, with a value ranging from 0 to 255, where 0 represents black, 255 represents white, and intermediate values ​​represent different degrees of gray.

[0022] Vertebral body region segmentation is performed based on the extracted grayscale value matrix. Region segmentation employs a combination of adaptive thresholding and edge detection to identify the boundaries between the vertebral body and surrounding soft tissues. For each pixel in the grayscale value matrix, the difference in grayscale value between it and surrounding pixels is calculated. When the difference exceeds a preset threshold, the pixel is marked as a boundary point. This method yields the boundary point set of the vertebral body. The coordinates of the pedicles and the vertebral endplates are extracted from the boundary point set. The pedicle coordinates are determined by identifying the two structures extending outward from the posterior part of the vertebral body, while the vertebral endplate coordinates are determined by identifying the boundaries of the upper and lower planes of the vertebral body. The pedicle positions can be marked as two coordinate points, corresponding to the center positions of the left and right pedicles, respectively; the vertebral endplate positions can be marked as the coordinates of the center point of the upper and lower endplates.

[0023] The coordinates of the vertebral body centroid are calculated based on the extracted coordinates of the pedicle and endplate positions. A weighted average is then taken between the coordinates of the left and right pedicles and the center points of the superior and inferior endplates. Assuming the left pedicle coordinates are (x1, y1, z1), the right pedicle coordinates are (x2, y2, z2), the center point of the superior endplate is (x3, y3, z3), and the center point of the inferior endplate is (x4, y4, z4), the vertebral body centroid coordinates (x0, y0, z0) can be obtained by taking the arithmetic mean of the coordinate components of these four points. The calculated vertebral body centroid coordinates are used as the origin of the local coordinate system. The direction from the midpoint of the line connecting the two pedicles to the midpoint of the anterior edge of the vertebral body is taken as the positive x-axis, and the vector perpendicular to the x-axis and along the vertebral height direction is taken as the positive y-axis. The positive z-axis is determined using the right-hand rule. This establishes a local coordinate system with the vertebral body itself as the reference.

[0024] The coordinates of the pedicle and vertebral endplate positions are projected onto a local coordinate system. The projection process involves two steps: translation and rotation. Translation is achieved by subtracting the vertebral centroid coordinates from the coordinates of the points in the global coordinate system; the translated coordinates are then transformed to the local coordinate system using a rotation matrix. This method eliminates the influence of changes in the patient's overall position, focusing instead on the displacement changes of the vertebral body itself. In the local coordinate system, the coordinate differences between adjacent time points for the pedicle and vertebral endplate positions are calculated to obtain the displacement vector. This displacement vector can be decomposed into three components: anterior-posterior displacement, superior-inferior displacement, and lateral displacement. The direction of displacement is determined by the direction angle of the displacement vector, while the magnitude of displacement is the modulus of the displacement vector.

[0025] Based on the calculated displacement direction and amplitude, a three-dimensional displacement trajectory of the vertebral body is constructed. This trajectory is a spatial curve representing the vertebral body position at continuous time points, reflecting the motion pattern of the vertebral body during changes in body position. By calculating the gradient value of this curve, the time points where the displacement direction changes abruptly are identified. For each point on the trajectory, the rate of change of the directional angle between it and its adjacent points is calculated. When the rate of change of the directional angle exceeds a preset threshold, the point is marked as a displacement direction change point. For example, a patient who underwent vertebral fracture surgery was monitored continuously for 72 hours, and vertebral body imaging data was collected every 15 minutes, totaling 288 time points. After processing using the above method, significant displacement direction changes were found at time points 37 and 163, with corresponding displacement amplitudes of 3.2 mm and 4.5 mm, respectively. Combining these time points with the corresponding displacement amplitudes yields the vertebral instability characteristic set for this patient.

[0026] The acquired set of vertebral instability features is input into a pre-trained neural network model for risk classification. This neural network model employs a multilayer perceptron structure: the input layer receives data from the instability feature set, the hidden layers extract and transform the features, and the output layer provides the patient's risk level. During model training, patient data with known risk levels is used, and the network parameters are optimized through backpropagation to achieve high risk prediction accuracy. Based on the risk level output by the model, healthcare professionals can develop corresponding nursing plans and treatment strategies for patients.

[0027] This invention achieves an objective and quantitative assessment of the risks of patients after vertebral fracture surgery by accurately calculating and analyzing the three-dimensional displacement trajectory of the vertebrae during changes in body position. Through comprehensive analysis of abrupt changes in displacement direction and displacement amplitude, potential vertebral instability risks can be detected early, providing medical staff with a scientific basis for developing targeted interventions, effectively reducing the incidence of postoperative complications; alleviating the workload of medical staff; and improving the efficiency of medical resource utilization.

[0028] The temporal sequence of body position changes in motion sensing data is time-aligned with time points in the instability feature set to obtain the muscle activation signals corresponding to those time points. The ratio of the peak intensity to the baseline intensity of the muscle activation signals is calculated using a neural network to obtain the compensation ratio sequence, which includes: Muscle activation signals and body position change time sequences are collected from motion sensing data, and the muscle activation signals are filtered to obtain signal waveforms. Extract the time points of body position changes from the body position change time sequence, and divide the signal waveform into multiple signal segments according to the time points of body position changes; The time points of body position changes are time-aligned with the time points in the set of instability characteristics, and the signal segments corresponding to the time points are selected based on the time alignment results. The signal segment corresponding to the time point is subjected to time-frequency transformation to generate a time-frequency feature matrix, and the time-frequency feature matrix is ​​input into a neural network to extract the spectral energy distribution; The signal intensity at rest is identified as the baseline intensity based on the spectral energy distribution, and the signal intensity at contraction is identified as the peak intensity. The ratio of the peak intensity to the baseline intensity is calculated to obtain a compensation ratio sequence.

[0029] When performing risk stratification for patients after vertebral fracture surgery, it is necessary to collect muscle activation signals and temporal data of body position changes from the patient's motion sensing data. Muscle activation signals are acquired using a surface electromyography (EMG) device. Temporal data of body position changes are acquired using an accelerometer fixed to the patient's torso to record the acceleration information of changes in body position in real time. The acquired muscle activation signals undergo filtering processing, including using a low-pass filter to remove high-frequency noise above 50Hz, a high-pass filter to remove low-frequency baseline drift below 0.5Hz, and a notch filter to eliminate power frequency interference. The resulting signal waveform is smoother and effectively eliminates interference from external environmental factors.

[0030] Postural change time points were extracted from the time-series data. The extraction method involved calculating the synthetic acceleration from the accelerometer data; when the amplitude of the synthetic acceleration change exceeded a preset threshold, it was marked as a postural change time point. The threshold was set to three times the standard deviation of the synthetic acceleration under static conditions. Based on the extracted postural change time points, the filtered muscle activation signal waveform was divided into multiple signal segments. Each signal segment corresponds to the muscle activation status before and after a postural change. The signal segments were divided using a 30-second time window, meaning that the signal from 30 seconds before to 30 seconds after each postural change time point was considered a complete signal segment.

[0031] The time points of body position changes are time-aligned with the time points in the instability feature set. This time alignment process considers potential time differences between devices by finding the closest body position change time point to the instability time point, with a time difference threshold set to within 5 seconds. For example, for the first time point 13:45:28 in the instability feature set, the closest time point in the body position change sequence is found to be 13:45:30, with a time difference of 2 seconds, meeting the threshold requirement. Therefore, these two time points are determined to correspond to the same body position change event. The signal segment corresponding to the instability time point is then selected based on the time alignment result.

[0032] The time-frequency transformation employs the Short-Time Fourier Transform (SFT) method, with a window length of 256 sampling points and an overlap rate of 50%. The SFT converts the time-domain signal into a time-frequency domain representation, forming a time-frequency feature matrix. Each element in the matrix represents the signal energy at a specific time point and frequency. The rows of the time-frequency feature matrix represent the frequency dimension, the columns represent the time dimension, and the matrix value represents the energy magnitude at the corresponding time-frequency point. The generated time-frequency feature matrix is ​​then input into a pre-trained neural network to extract the spectral energy distribution. This neural network uses a convolutional neural network structure, comprising three convolutional layers and two fully connected layers. The convolutional layers extract local feature patterns from the time-frequency feature matrix, while the fully connected layers synthesize these features to identify the spectral energy distribution. The network training utilizes a large amount of labeled muscle activation signal data, and the network parameters are optimized using a backpropagation algorithm.

[0033] The signal intensity at rest is identified as the baseline intensity based on the spectral energy distribution of the neural network output. At rest, muscles are typically dominated by low-frequency energy with a relatively flat spectral energy distribution. The baseline intensity is determined by calculating the average value of the low-frequency band (5-15Hz). For a signal segment acquired from a patient, the calculated baseline intensity is 0.023mV. The signal intensity during muscle contraction is then identified as the peak intensity. Muscle contraction typically involves enhanced mid-to-high frequency energy with a distinct peak in the spectral energy distribution. The peak intensity is determined by calculating the maximum value of the mid-to-high frequency band (20-150Hz). For a signal segment from the same patient, the calculated peak intensity is 0.187mV. The ratio of peak intensity to baseline intensity is calculated to obtain the compensation ratio sequence. The compensation ratio reflects the degree of muscle compensation during postural changes; a higher ratio indicates stronger muscle compensation activity.

[0034] This model combines the compensation ratio sequence with an instability feature set and inputs it into a risk assessment neural network model to generate patient risk level assessment results. The risk assessment neural network model employs a recurrent neural network structure, specifically a long short-term memory network, which effectively captures the time dependency of the compensation ratio sequence. The network input layer receives displacement amplitude data from the compensation ratio sequence and the instability feature set, the hidden layer contains 64 neurons, and the output layer provides the patient's risk level classification. Risk levels are divided into three categories: low risk, medium risk, and high risk, each corresponding to different nursing plans and interventions. This network model was trained on a large amount of clinical case data and achieved a high level of accuracy.

[0035] This invention achieves precise risk grading for patients after vertebral fracture surgery through deep fusion analysis of muscle activation signals and vertebral instability characteristics. This invention not only considers the instability of the vertebral body itself but also introduces the assessment of muscle compensation mechanisms, making risk grading more comprehensive and accurate. By extracting the spectral characteristics of muscle activation signals through neural networks and calculating the compensation ratio sequence, the contribution of muscles to maintaining spinal stability can be objectively quantified. This allows for early detection of potential risks caused by insufficient muscle compensation capacity, providing medical staff with more accurate risk assessment data, facilitating the development of personalized postoperative care plans, improving the effectiveness of vertebral fracture surgery, reducing the incidence of complications, and accelerating patient recovery.

[0036] like Figure 2 As shown, Figure 2 This diagram illustrates a comparative analysis of muscle compensation indices under different postural change events in this embodiment. It shows a comparison between the compensation ratio sequence calculated by this technical solution and the normalized intensity calculated by the traditional RMS method (Root Mean Square Amplitude Analysis) over six consecutive postural change events. The horizontal axis represents the sequence of different postural change events, and the vertical axis represents the corresponding index values. Data points marked with diamonds and connected by solid lines represent the compensation ratios of this technical solution. These ratios fluctuate significantly between different events. For example, in event 1, the calculated compensation ratio is as high as 8.13, and in events 3 and 6, it reaches high levels of 9.20 and 9.00 respectively, keenly reflecting the strong compensatory activity of muscles during these specific moments of instability. In contrast, the data points marked with circles and connected by dashed lines, representing the "traditional RMS method," maintain a relatively low level of 3.0 to 3.25 across various events, with very little variation. This fully demonstrates the insufficient sensitivity of traditional methods in capturing the dynamic and transient muscle compensatory responses caused by postural changes. This technical solution, by deeply mining the differences in spectral energy distribution, can provide more discriminative and clinically significant quantitative indicators, providing high-quality input data for subsequent risk assessment neural network models.

[0037] The time points of body position changes are time-aligned with the time points in the instability feature set. Based on the time alignment results, signal segments corresponding to the time points are selected, including: The raw signals from motion sensing data are collected, and the time points of body position change are constructed into a body position change time series. The raw signals are segmented according to the body position change time series to obtain the initial signal segments. Calculate the time difference between adjacent time points in the body position change time series and generate a time difference sequence; The temporal distance between the time difference sequence and the time point sequence in the instability feature set is calculated based on the dynamic programming algorithm, a temporal distance matrix is ​​constructed, and a time alignment matching path is extracted from the temporal distance matrix. A time-to-time correspondence table is generated based on the time alignment matching path, and time alignment position information is extracted from the correspondence table. The initial signal segment is decomposed into a multi-scale decomposition matrix using wavelet transform. The coefficient values ​​at the corresponding time points are located and extracted from the decomposition coefficient matrix based on the time alignment information. The coefficient values ​​are reconstructed using the inverse wavelet transform to obtain the signal segment corresponding to the time point. The signal segment is then subjected to piecewise smoothing to obtain the final signal segment.

[0038] Raw signals from motion sensing data were collected using wearable bioelectric sensors. These raw signals included electromyography (EMG), acceleration, and angular velocity (Angular velocity) signals, reflecting muscle activation and postural changes during daily activities. Postural change time points were constructed into a postural change time series, identified based on abrupt changes in acceleration and angular velocity signals. A postural change time point was marked when the composite value of the acceleration signal exceeded a threshold of 1.5 and its duration exceeded 0.5 seconds. For example, in 24-hour monitoring data of a patient after vertebral fracture surgery, 42 postural change time points were identified, forming a postural change time series. The raw signals were segmented according to the postural change time series, with each segment starting 10 seconds before the next postural change time point and ending 10 seconds after, thus obtaining initial signal segments.

[0039] The time difference between adjacent time points in the body position change time series is calculated to generate a time difference sequence, which reflects the time interval pattern of the patient's body position changes. For the 42 body position change time points mentioned above, 41 time differences were calculated, in minutes. These time differences are distributed between 5 minutes and 120 minutes, forming a time difference sequence. The temporal distance between the time difference sequence and the time point sequence in the instability feature set is calculated based on a dynamic programming algorithm. The time point sequence in the instability feature set is obtained through previous vertebral body image analysis and includes the time points of abrupt changes in vertebral displacement direction. The dynamic programming algorithm achieves temporal alignment by constructing a distance accumulation matrix. This algorithm defines a distance function to measure the similarity of corresponding points in two time series and then finds the path with the minimum cumulative distance as the optimal alignment method. The distance function is Euclidean distance, and the window constraint parameter is set to 10% of the sequence length to limit the search range of the alignment path. A temporal distance matrix is ​​constructed, where the rows of the matrix represent time points in the body position change time series, the columns represent time points in the instability feature set, and the matrix element values ​​represent the temporal distance between corresponding time points. Extract the time alignment matching path from the time distance matrix. The matching path is a path from the top left corner to the bottom right corner of the matrix. The path with the minimum sum of element values ​​represents the optimal time alignment.

[0040] A time-alignment matching path is used to generate a correspondence table between time points and moment points. This table contains three columns: postural change time points, instability feature moment points, and matching confidence scores. Matching confidence scores are calculated based on temporal distance; the smaller the distance, the higher the confidence score. For example, if 7 out of 8 instability feature moment points are successfully matched with corresponding postural change time points, the matching confidence score is between 0.75 and 0.95. Time alignment position information is extracted from the table. This position information includes the index position of the successfully matched postural change time points in the original signal, and the index position of the corresponding instability feature moment points in the instability feature set. This position information will be used for accurate extraction of subsequent signal segments.

[0041] The initial signal segment was decomposed into a multi-scale decomposition using wavelet transform, yielding a decomposition coefficient matrix. The Daubechies-4 wavelet basis was used, with a decomposition level of 5. Wavelet transform decomposes the signal into different frequency components; the high-frequency component reflects the signal's detailed features, while the low-frequency component reflects its approximate contour. The decomposition process is completed through a series of filters and downsampling operations, ultimately resulting in a decomposition coefficient matrix containing approximate and detail coefficients at different scales. Based on the previously obtained time alignment information, the coefficient values ​​at the corresponding time points were located and extracted from the decomposition coefficient matrix. The precise location of the instability point in the original signal was determined based on the time alignment information, and the corresponding coefficient values ​​were extracted from the decomposition coefficient matrix. For each successfully matched time point, the extracted coefficients included all coefficient values ​​within a 30-second time window surrounding that time point.

[0042] The extracted coefficient values ​​are reconstructed using the inverse wavelet transform to obtain signal segments corresponding to specific time points. The inverse transform is the reverse operation of the decomposition process, reconstructing the original signal through upsampling and filtering. The reconstructed signal segments precisely correspond to muscle activity near the instability characteristic time points, reflecting the compensatory responses of muscles during vertebral instability. Piecewise smoothing is then applied to the reconstructed signal segments to obtain the final signal segments. The piecewise smoothing employs a sliding window mid-range filtering method with a window length of 15 sampling points. This process effectively suppresses high-frequency noise that may be introduced during reconstruction while preserving the main characteristics of the signal. The final signal segments are uniformly 60 seconds long, centered at the instability characteristic time points; these signal segments will serve as the basis for subsequent muscle activation signal analysis.

[0043] The final signal fragments were analyzed to extract muscle activation features. Feature extraction included time-domain and frequency-domain features. Time-domain features, such as mean, variance, and peak value, reflected the statistical characteristics of the signal, while frequency-domain features, such as dominant frequency and band energy ratio, reflected the frequency composition of the signal. These features were combined with displacement direction and amplitude information from the instability feature set and input into a pre-trained neural network model for risk assessment. The neural network model adopted a three-layer structure: the input layer received feature vectors, the hidden layer contained 32 neurons using the ReLU activation function, and the output layer provided the patient's risk level. Model training employed the backpropagation algorithm, the cross-entropy function as the loss function, and the Adam algorithm as the optimizer. Through training on a large amount of postoperative data from patients with vertebral fractures, the model was able to accurately assess the patient's risk level, providing support for clinical decision-making.

[0044] This invention achieves precise matching between vertebral instability features and muscle activation signals through time alignment technology, providing a novel technical solution for risk stratification of patients after vertebral fracture surgery. Utilizing dynamic programming algorithms and wavelet transform, it overcomes the problem of inconsistent temporal sequences in multi-source heterogeneous data, ensuring the accuracy of the correspondence between time points and signal segments. Through multi-scale analysis and segmented smoothing, useful signal components are enhanced, noise interference is suppressed, and the quality of feature extraction is improved. It can capture the temporal correlation between vertebral instability and muscle compensatory responses, thereby more comprehensively assessing the patient's stability and recovery, providing precise evidence for medical staff to develop personalized rehabilitation plans, effectively reducing the risk of postoperative complications, and improving the patient's quality of life.

[0045] The process of inputting the time-frequency feature matrix into a neural network to extract the spectral energy distribution includes: A neural network structure is constructed, which includes a feature extraction module and a fully connected module. The feature extraction module consists of convolutional layers and pooling layers, and the fully connected module consists of multiple layers of connection units. The time-frequency feature matrix is ​​divided into training data matrices according to time windows, and the training data matrices are normalized by maximum and minimum values ​​to obtain normalized feature matrices. The normalized feature matrix is ​​input into the feature extraction module, and frequency domain feature vectors are obtained through convolution and pooling operations. The frequency domain feature vector is input into the fully connected module, and the energy mapping vector is obtained by mapping through the activation function. Perform inverse normalization on the energy mapping vector and construct a spectral energy curve by arranging them in chronological order. Extract the peak and trough position coordinates from the spectral energy curve and generate a spectral energy distribution based on the peak and trough position coordinates.

[0046] A neural network structure is constructed to extract spectral energy distribution. This structure consists of two main parts: a feature extraction module and a fully connected module. The feature extraction module comprises convolutional layers and pooling layers, with three convolutional layers and three pooling layers arranged alternately. The first convolutional layer contains 32 kernels, each 3×3 in size, with a stride of 1 and zero padding. The second convolutional layer contains 64 kernels, each 3×3 in size, with a stride of 1 and zero padding. The third convolutional layer contains 128 kernels, each 3×3 in size, with a stride of 1 and zero padding. Each convolutional layer is followed by a pooling layer using max pooling with a 2×2 pooling window and a stride of 2. The fully connected module consists of three layers of connection units: the first layer contains 256 neurons, the second layer contains 128 neurons, and the third layer contains 64 neurons. The fully connected layers use the ReLU activation function, and the last layer uses the Sigmoid activation function to map the output to between 0 and 1.

[0047] The time-frequency feature matrix is ​​divided into training data matrices according to time windows. The size of the time-frequency feature matrix is ​​frequency dimension × time dimension. For electromyography (EMG) signals from patients after vertebral fracture surgery, the frequency dimension is usually 128, representing a frequency range of 0-500Hz; the time dimension depends on the length of the signal segment. A sliding window method is used for time window division, with a window size of 64 time points and a sliding step size of 8 time points. For a signal segment with a length of 256 time points, 25 training data matrices can be obtained, each with a size of 128×64. Max-min normalization is performed on the training data matrices to compress the data range to between 0 and 1. The maximum and minimum values ​​of the original data are recorded during normalization for later inverse normalization. Taking the EMG signal of a patient with a vertebral fracture as an example, the value range of the original training data matrix is ​​0.013 to 0.893. After normalization, the normalized feature matrix is ​​obtained, with a value range of 0 to 1.

[0048] The normalized feature matrix is ​​input into the feature extraction module, where convolution and pooling operations are used to obtain a frequency domain feature vector. Convolution involves sliding the convolution kernel across the input feature map, performing weighted summations on local regions to extract local feature patterns. Pooling reduces the feature map size through downsampling, preserving salient features and reducing computational complexity. Using a normalized feature matrix (128×64) as input, after the first convolutional and pooling layers, the feature map size becomes 64×32×32; after the second convolutional and pooling layers, it becomes 32×16×64; and after the third convolutional and pooling layers, it becomes 16×8×128. Finally, the feature map is flattened into a one-dimensional vector, resulting in a frequency domain feature vector of length 16384. This frequency domain feature vector contains rich features of electromyography (EMG) signals in the time-frequency domain, reflecting the frequency characteristics and temporal evolution of muscle activation in patients after vertebral fracture surgery.

[0049] The frequency domain feature vector is input into a fully connected module, and an energy mapping vector is obtained through an activation function. The fully connected module implements a nonlinear mapping from the high-dimensional feature space to the energy mapping space. The frequency domain feature vector (16384 dimensions) is reduced to 256 dimensions through the first fully connected layer, and nonlinearity is introduced by the ReLU activation function; it is further reduced to 128 dimensions through the second fully connected layer, and then activated by the ReLU function; the third fully connected layer reduces the dimension to 64 dimensions, and then the energy mapping vector with values ​​ranging from 0 to 1 is obtained by the Sigmoid activation function. For the above frequency domain feature vector, the 64-dimensional energy mapping vector is obtained after processing by the fully connected module, with a value range of 0.027 to 0.981. The energy mapping vector represents the energy distribution at 64 frequency points, which can intuitively reflect the energy change pattern during muscle activation.

[0050] The energy mapping vectors are denormalized, and then arranged in chronological order to construct a spectral energy curve. The denormalization process uses the previously recorded maximum and minimum values ​​of the original data to restore the normalized energy mapping vectors to the scale range of the original data. For the above energy mapping vectors, the denormalized values ​​range from 0.023 to 0.876. The energy mapping vectors corresponding to multiple time windows are arranged in chronological order to form a spectral energy curve. The horizontal axis of the spectral energy curve represents time, and the vertical axis represents frequency; the curve value represents the energy magnitude at the corresponding time and frequency point. For a signal segment with 256 time points, 25 energy mapping vectors can be obtained, which, after arrangement, form a complete spectral energy curve. The peak and trough coordinates are extracted from the spectral energy curve; peaks represent local maximum energy points, and troughs represent local minimum energy points. The spectral energy curve is smoothed to remove high-frequency noise; the first derivative of the curve is calculated, with points where the derivative changes from positive to negative being peaks and points where the derivative changes from negative to positive being troughs; the time and frequency coordinates of all peaks and troughs are recorded. The distribution characteristics of these peaks and troughs reflect the temporal patterns of muscle activation and areas of energy concentration in patients.

[0051] The spectral energy distribution is generated based on the coordinates of the peak and trough positions. This distribution is represented by a heatmap, with time on the horizontal axis and frequency on the vertical axis, and color intensity indicating energy levels. In the heatmap, peaks represent high-energy regions, and troughs represent low-energy regions. For patients after vertebral fracture surgery, the spectral energy distribution reflects the energy concentration characteristics of muscles in an activated state. By analyzing the temporal evolution and frequency distribution characteristics of the spectral energy distribution, signal characteristics in resting and contracted states can be identified. In the resting state, energy is mainly distributed in the low-frequency range (5-15Hz), with relatively low energy values; in the contracted state, the energy distribution extends to the mid-to-high frequency range (20-150Hz), with significantly increased energy values. Based on this characteristic, the signal intensity in the resting state can be identified as the baseline intensity, and the signal intensity in the contracted state can be identified as the peak intensity. The ratio of peak intensity to baseline intensity is then calculated to obtain a compensation ratio sequence reflecting muscle compensatory ability.

[0052] This invention provides a precise technical means for risk stratification of patients after vertebral fracture surgery by constructing a deep neural network structure to extract spectral energy distribution. This invention can automatically extract deep features from electromyography (EMG) signals without the need for manually designed feature extraction algorithms. Through the combination of convolutional and pooling layers, it effectively captures the local patterns and global structure of EMG signals in the time-frequency domain, improving the accuracy and robustness of feature extraction. Spectral energy distribution, as a precise representation of muscle activation state, can objectively reflect the muscle compensatory ability of patients after vertebral fracture surgery, providing a scientific basis for clinicians to assess patient recovery and develop personalized treatment plans.

[0053] The frequency of abrupt changes in displacement direction within the statistical instability feature set, combined with the numerical growth rate of the compensation ratio sequence, is used to calculate the risk transmission intensity value, including: Extract the moment points of sudden displacement direction changes from the set of instability features, arrange the moment points of sudden displacement direction changes in chronological order, and calculate the time interval between adjacent moment points of sudden displacement direction changes. Based on the time interval, a sampling interval is set, and the number of times the displacement direction change occurs within the sampling interval is counted to obtain the frequency of displacement direction change. The compensation ratio sequence is truncated according to the sampling interval, and the numerical change of the compensation ratio within the sampling interval is calculated to obtain the numerical growth rate of the compensation ratio sequence. The risk transmission intensity value is obtained by weighting the frequency of abrupt changes in displacement direction with the numerical growth rate of the compensation ratio sequence.

[0054] The system extracts abrupt changes in displacement direction from the set of instability features. By calculating the change in the angle between two adjacent time points, a change in displacement direction is identified when the angle exceeds a preset threshold, and the corresponding time point is recorded. The preset threshold can be set based on clinical experience, typically between 30 and 45 degrees. These abrupt changes in displacement direction are then arranged chronologically to form a time series. For the sorted time series, the time interval between adjacent abrupt changes in displacement direction is calculated. The unit of calculation for the time interval can be hours or days, depending on the monitoring frequency.

[0055] Sampling intervals are set based on calculated time intervals. The length of the sampling interval can be determined according to the characteristics of the patient's recovery cycle, typically set to 7 or 14 days. For patients requiring long-term monitoring, a sliding window method can be used to set the sampling interval, with a fixed window length that slides forward a certain time unit each time. Within each sampling interval, the number of times the displacement direction change occurs is counted, and then divided by the length of the sampling interval to obtain the frequency of displacement direction change. This frequency reflects the degree of postural instability of the patient within a specific time period.

[0056] The compensation ratio sequence is extracted using the same sampling intervals. The compensation ratio sequence refers to the data on changes in the strength of the patient's compensatory mechanism during the monitoring period. Within each sampling interval, the compensation ratio at the beginning and end of the interval is taken, the difference is calculated, and then divided by the length of the sampling interval to obtain the numerical growth rate of the compensation ratio sequence. This growth rate reflects the speed at which the patient's compensatory ability changes; a positive value indicates a strengthening of the compensatory mechanism, and a negative value indicates a weakening of the compensatory mechanism.

[0057] The risk transmission strength value is obtained by weighting the frequency of abrupt changes in displacement direction with the numerical growth rate of the compensation ratio sequence. The weighting calculation uses a linear combination method: the frequency of abrupt changes in displacement direction is multiplied by a first weighting coefficient, and the numerical growth rate of the compensation ratio sequence is multiplied by a second weighting coefficient. The two weighting coefficients are then added together to obtain the risk transmission strength value. The weighting coefficients range from 0 to 1, and the sum of the two weighting coefficients equals 1. The weighting coefficients are determined through learning using a neural network model. This neural network model uses historical patient data as a training set and clinically confirmed risk levels as labels, optimizing the weighting parameters through a backpropagation algorithm.

[0058] Different weighting coefficients can be set for different types of vertebral fracture patients. For elderly patients, the weight of the frequency of abrupt changes in displacement direction can be increased; for patients with osteoporosis, the weight of the rate of increase in the compensation ratio can be increased. After the risk transmission intensity value is calculated, it is mapped to a predefined risk level interval to complete the patient risk classification.

[0059] To improve the accuracy of risk assessment, a time decay factor can be introduced, making recent abrupt changes in displacement direction and compensation ratios have a greater impact on the risk transmission intensity value. The time decay factor can be set to decrease over time, so that the influence of historical data gradually weakens over time.

[0060] In calculating the risk transmission intensity value, basic clinical information of the patient, such as age, gender, bone density, and past medical history, can be combined. A neural network model can then be used to correct the risk transmission intensity value, resulting in a risk assessment result that better reflects individual characteristics. The neural network model employs a multilayer perceptron structure. The input layer receives the risk transmission intensity value and the patient's basic clinical information, the hidden layer performs feature extraction and nonlinear transformation, and the output layer provides the corrected risk level.

[0061] The risk transmission intensity value calculated using the above method can reflect the recovery status and potential risks of patients after vertebral fracture surgery in real time. A higher risk transmission intensity value indicates that the patient's recovery is poor, requiring adjustments to the treatment plan or increased rehabilitation training; a lower risk transmission intensity value indicates that the patient's recovery is good, and the frequency of follow-up examinations can be reduced.

[0062] This invention accurately calculates the risk transmission intensity value by statistically analyzing the frequency of abrupt changes in displacement direction and combining it with the numerical growth rate of the compensation ratio sequence, thus achieving precise risk grading for patients after vertebral fracture surgery. It can effectively identify potentially high-risk patients, providing clinicians with a scientific basis for decision-making and developing personalized follow-up and intervention plans. Through the fusion analysis of multi-source data using a neural network model, the accuracy and reliability of risk assessment are improved, reducing the incidence of postoperative complications after vertebral fracture surgery. This method supports continuous dynamic monitoring, enabling timely detection of abnormal changes during patient rehabilitation, achieving early warning and intervention, improving rehabilitation outcomes, reducing medical resource consumption, and providing an effective tool for the long-term management of patients after vertebral fracture surgery.

[0063] Based on the numerical range of the risk transmission intensity value, combined with abnormal fluctuations in physiological indicator data, the level of postoperative complication risk for patients is determined and the grading results are output, including: Collect historical patient risk transmission intensity values ​​and sort them to obtain the numerical range of historical risk transmission intensity values; Based on the numerical range, a dividing benchmark is set, and the numerical range is divided into multiple risk level ranges using the dividing benchmark, and a correspondence between the risk level ranges and the risk levels is established. The fluctuation range of blood pressure and heart rate values ​​in physiological indicator data is statistically analyzed, the fluctuation range is divided into multiple fluctuation level intervals, and the correspondence between fluctuation level intervals and risk correction values ​​is established. The initial risk level is obtained by positioning the risk transmission intensity value of the patient to be evaluated into the risk level range, and the risk correction value is obtained by positioning the physiological index data of the patient to be evaluated into the fluctuation level range. The initial risk level is combined with the risk correction value to calculate and determine the level of postoperative complication risk for the patient, and the classification result is output.

[0064] Risk transmission intensity data should be collected from patients who have undergone historical vertebral fracture surgery. This data should contain a sufficient sample size to ensure the reliability of subsequent analyses. The collected historical patient risk transmission intensity values ​​should be sorted in ascending order to obtain a complete range of historical risk transmission intensity values. The sorting can be implemented using algorithms such as quicksort or mergesort to ensure the accuracy and efficiency of the sorting results. After sorting, the resulting range is represented as a continuous interval from the minimum to the maximum risk transmission intensity value.

[0065] Based on the obtained historical risk transmission intensity values, reasonable segmentation benchmarks are set. The setting of segmentation benchmarks can be determined based on clinical experience or statistical methods; commonly used methods include equal-frequency segmentation, equal-interval segmentation, or segmentation based on clustering algorithms. In risk assessment of patients after vertebral fracture surgery, an adaptive segmentation method based on neural networks can be used. By learning the correspondence between historical patient risk transmission intensity values ​​and actual postoperative complication rates, the optimal segmentation point location is automatically determined. Using a multilayer perceptron neural network, with risk transmission intensity values ​​as input and complication probability as output, the network parameters are optimized using gradient descent, ultimately setting segmentation benchmarks at locations where probability changes are significant.

[0066] The numerical range is divided into multiple risk level intervals using a defined segmentation benchmark. The risk levels are categorized into five levels: low risk, low-to-medium risk, medium risk, medium-to-high risk, and high risk, each corresponding to one of the five risk level intervals. When establishing the correspondence between risk level intervals and risk levels, considering the differences in various clinical environments, hospital-specific adjustment factors can be introduced during the neural network training process to make the grading results more consistent with actual clinical situations.

[0067] For physiological data processing, the focus is primarily on two key indicators: blood pressure and heart rate. Continuous monitoring data of blood pressure and heart rate were collected during the patient's postoperative recovery period, and the fluctuation range was statistically analyzed. A sliding window technique was used, with a window size set to 24 hours. The difference between the maximum and minimum values ​​of the indicators within the window was calculated to obtain the fluctuation amplitude. For blood pressure, fluctuations in both systolic and diastolic blood pressure were considered; for heart rate, the variation amplitude under different activity states was observed.

[0068] The fluctuation range of blood pressure and heart rate is divided into three fluctuation levels: slight fluctuation, moderate fluctuation, and severe fluctuation. The criteria for dividing the fluctuation level range can be determined according to medical guidelines or derived through statistical analysis of historical patient data. A correspondence is established between the fluctuation level range and risk correction values, with slight fluctuations corresponding to smaller risk correction values ​​and severe fluctuations corresponding to larger risk correction values. The specific values ​​of the risk correction values ​​can be obtained through neural network learning. This neural network takes blood pressure and heart rate fluctuation data as input and the actual increase in the risk of complications as output, optimizing the correction value parameters through a backpropagation algorithm.

[0069] For patients undergoing vertebral fracture surgery to be evaluated, their risk transmission intensity value is calculated and positioned within a pre-defined risk level range to obtain an initial risk level. The patient's blood pressure and heart rate fluctuations are analyzed and positioned within corresponding fluctuation level ranges to obtain corresponding risk correction values. Weighting coefficients can be set to adjust the influence of blood pressure and heart rate on the risk correction values, allowing for personalized adjustments based on different types of vertebral fracture surgery and the patient's underlying medical conditions.

[0070] The initial risk level and risk correction value are combined to determine the final risk level of postoperative complications for the patient. The combined calculation uses a weighted adjustment method, meaning that the initial risk level is adjusted upwards or downwards based on the risk correction value. If physiological indicators fluctuate significantly and the risk correction value is positive, the risk level is increased; if physiological indicators fluctuate slightly and the risk correction value is negative, the risk level is decreased. The magnitude of the risk level adjustment is proportional to the size of the risk correction value, ensuring that the adjustment result accurately reflects the patient's actual risk status.

[0071] Once the risk level is determined, the classification results are output, including the risk level identifier and corresponding risk description. For high-risk patients, risk warnings and intervention suggestions can be provided simultaneously. The risk classification results can be directly pushed to medical staff through the hospital information system, or integrated into the patient's electronic medical record for easy access and decision-making reference by medical staff.

[0072] This invention achieves precise risk grading of postoperative complications in patients with vertebral fractures by combining risk transmission intensity values ​​with abnormal fluctuations in physiological indicators, significantly improving the accuracy and reliability of risk assessment. This method overcomes the limitations of traditional single-indicator assessments, comprehensively considering the patient's physiological state and compensatory capacity, enabling early identification of high-risk patients and providing a scientific basis for clinical intervention. Adaptive segmentation and personalized adjustments are achieved through neural network technology, making the grading results more consistent with individual patient characteristics and improving the sensitivity and specificity of risk warning.

[0073] Figure 3 A flowchart of a neural network-based risk stratification method for postoperative vertebral fracture patients provided in this invention is included, comprising the following steps: Collect imaging data, physiological data, and motion sensor data of patients after vertebral fracture surgery; Based on the vertebral spatial coordinates at continuous time points in the image data, the three-dimensional displacement trajectory of the vertebral body during the change of body position is calculated, and the time points and displacement amplitudes of the abrupt change in displacement direction in the three-dimensional displacement trajectory are extracted to obtain the set of instability features. The time sequence of body position changes in motion sensing data is time-aligned with the time points in the set of instability features to obtain the muscle activation signals corresponding to the time points. The ratio of the peak intensity to the baseline intensity of the muscle activation signals is calculated by a neural network to obtain the compensation ratio sequence. The frequency of abrupt changes in displacement direction in the statistical instability feature set is used to calculate the risk transmission intensity value by combining the numerical growth rate of the compensation ratio sequence. Based on the numerical range of the risk transmission intensity value and the abnormal fluctuations in physiological indicator data, the level of postoperative complication risk for patients is determined and the classification result is output.

[0074] One technical solution provided in this embodiment of the invention is an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the aforementioned systems.

[0075] One technical solution provided in this embodiment of the invention is a computer-readable storage medium storing a computer program, wherein the processor executes the computer program to implement the steps in any of the aforementioned systems.

[0076] The specific embodiments described above are preferred embodiments of the present invention and are not intended to limit the specific scope of the present invention. The scope of the present invention includes, but is not limited to, these specific embodiments. All equivalent changes made in accordance with the shape and structure of the present invention are within the protection scope of the present invention.

Claims

1. A risk stratification system for postoperative vertebral fracture patients based on neural networks, characterized in that, The system includes: The data acquisition module is used to collect imaging data, physiological index data, and motion sensor data of patients after vertebral fracture surgery; The displacement trajectory analysis module is used to calculate the three-dimensional displacement trajectory of the vertebrae during the change of body position based on the vertebral spatial coordinates of continuous time points in the image data, extract the time points and displacement amplitudes of the abrupt change in displacement direction in the three-dimensional displacement trajectory, and obtain the set of instability features. The neural network processing module is used to time-align the body position change time sequence in motion sensing data with the time points in the instability feature set, obtain the muscle activation signal corresponding to the time point, and calculate the ratio of the peak intensity to the baseline intensity of the muscle activation signal through the neural network to obtain the compensation ratio sequence. The risk intensity calculation module is used to statistically analyze the frequency of sudden changes in displacement direction in the set of instability characteristics, and calculate the risk transmission intensity value by combining the numerical growth rate of the compensation ratio sequence. The risk grading module is used to determine the level of postoperative complication risk for patients based on the numerical range of the risk transmission intensity value and the abnormal fluctuations in physiological indicator data, and output the grading results.

2. The system according to claim 1, characterized in that, The operation of the displacement trajectory analysis module includes the following steps: Collect vertebral body image data at continuous time points and extract the grayscale matrix from the vertebral body image data; Region segmentation is performed based on the gray value matrix to obtain the vertebral body boundary point set, and the pedicle position coordinates and vertebral endplate position coordinates are extracted from the vertebral body boundary point set. Calculate the coordinates of the centroid of the vertebral body based on the coordinates of the pedicle and the coordinates of the endplate of the vertebral body. Construct a local coordinate system with the coordinates of the centroid of the vertebral body as the origin and the direction of the pedicle coordinates as the principal axis. Project the pedicle and vertebral endplate position coordinates onto the local coordinate system and calculate the displacement direction and displacement amplitude between adjacent time points. A three-dimensional displacement trajectory is constructed based on the displacement direction and displacement amplitude. The moment point of sudden change in displacement direction is identified by calculating the gradient value of the three-dimensional displacement trajectory. The moment point is combined with the corresponding displacement amplitude to obtain the set of instability features.

3. The system according to claim 1, characterized in that, The operation of the neural network processing module includes the following steps: Muscle activation signals and body position change time sequences are collected from motion sensing data, and the muscle activation signals are filtered to obtain signal waveforms. Extract the time points of body position changes from the body position change time sequence, and divide the signal waveform into multiple signal segments according to the time points of body position changes; The time points of body position changes are time-aligned with the time points in the set of instability characteristics, and the signal segments corresponding to the time points are selected based on the time alignment results. The signal segment corresponding to the time point is subjected to time-frequency transformation to generate a time-frequency feature matrix, and the time-frequency feature matrix is ​​input into a neural network to extract the spectral energy distribution; The signal intensity at rest is identified as the baseline intensity based on the spectral energy distribution, and the signal intensity at contraction is identified as the peak intensity. The ratio of the peak intensity to the baseline intensity is calculated to obtain a compensation ratio sequence.

4. The system according to claim 3, characterized in that, The time points of body position changes are time-aligned with the time points in the instability feature set. Based on the time alignment results, signal segments corresponding to the time points are selected, including: The raw signals from motion sensing data are collected, and the time points of body position change are constructed into a body position change time series. The raw signals are segmented according to the body position change time series to obtain the initial signal segments. Calculate the time difference between adjacent time points in the body position change time series and generate a time difference sequence; The temporal distance between the time difference sequence and the time point sequence in the instability feature set is calculated based on the dynamic programming algorithm, a temporal distance matrix is ​​constructed, and a time alignment matching path is extracted from the temporal distance matrix. A time-to-time correspondence table is generated based on the time alignment matching path, and time alignment position information is extracted from the correspondence table. The initial signal segment is decomposed into a multi-scale decomposition matrix using wavelet transform. The coefficient values ​​at the corresponding time points are located and extracted from the decomposition coefficient matrix based on the time alignment information. The coefficient values ​​are reconstructed using the inverse wavelet transform to obtain the signal segment corresponding to the time point. The signal segment is then subjected to piecewise smoothing to obtain the final signal segment.

5. The system according to claim 3, characterized in that, The process of inputting the time-frequency feature matrix into a neural network to extract the spectral energy distribution includes: A neural network structure is constructed, which includes a feature extraction module and a fully connected module. The feature extraction module consists of convolutional layers and pooling layers, and the fully connected module consists of multiple layers of connection units. The time-frequency feature matrix is ​​divided into training data matrices according to time windows, and the training data matrices are normalized by maximum and minimum values ​​to obtain normalized feature matrices. The normalized feature matrix is ​​input into the feature extraction module, and frequency domain feature vectors are obtained through convolution and pooling operations. The frequency domain feature vector is input into the fully connected module, and the energy mapping vector is obtained by mapping through the activation function. Perform inverse normalization on the energy mapping vector and construct a spectral energy curve by arranging them in chronological order. Extract the peak and trough position coordinates from the spectral energy curve and generate a spectral energy distribution based on the peak and trough position coordinates.

6. The system according to claim 1, characterized in that, The operation of the risk intensity calculation module includes the following steps: Extract the moment points of sudden displacement direction changes from the set of instability features, arrange the moment points of sudden displacement direction changes in chronological order, and calculate the time interval between adjacent moment points of sudden displacement direction changes. Based on the time interval, a sampling interval is set, and the number of times the displacement direction change occurs within the sampling interval is counted to obtain the frequency of displacement direction change. The compensation ratio sequence is truncated according to the sampling interval, and the numerical change of the compensation ratio within the sampling interval is calculated to obtain the numerical growth rate of the compensation ratio sequence. The risk transmission intensity value is obtained by weighting the frequency of abrupt changes in displacement direction with the numerical growth rate of the compensation ratio sequence.

7. The system according to claim 1, characterized in that, The operation of the risk classification module includes the following steps: Collect historical patient risk transmission intensity values ​​and sort them to obtain the numerical range of historical risk transmission intensity values; Based on the numerical range, a dividing benchmark is set, and the numerical range is divided into multiple risk level ranges using the dividing benchmark, and a correspondence between the risk level ranges and the risk levels is established. The fluctuation range of blood pressure and heart rate values ​​in physiological indicator data is statistically analyzed, the fluctuation range is divided into multiple fluctuation level intervals, and the correspondence between fluctuation level intervals and risk correction values ​​is established. The initial risk level is obtained by positioning the risk transmission intensity value of the patient to be evaluated into the risk level range, and the risk correction value is obtained by positioning the physiological index data of the patient to be evaluated into the fluctuation level range. The initial risk level is combined with the risk correction value to calculate and determine the level of postoperative complication risk for the patient, and the classification result is output.

8. A method for risk stratification of patients after vertebral fracture surgery based on neural networks, characterized in that, The method includes: Collect imaging data, physiological data, and motion sensor data of patients after vertebral fracture surgery; Based on the vertebral spatial coordinates at continuous time points in the image data, the three-dimensional displacement trajectory of the vertebral body during the change of body position is calculated, and the time points and displacement amplitudes of the abrupt change in displacement direction in the three-dimensional displacement trajectory are extracted to obtain the set of instability features. The time sequence of body position changes in motion sensing data is time-aligned with the time points in the set of instability features to obtain the muscle activation signals corresponding to the time points. The ratio of the peak intensity to the baseline intensity of the muscle activation signals is calculated by a neural network to obtain the compensation ratio sequence. The frequency of abrupt changes in displacement direction in the statistical instability feature set is used to calculate the risk transmission intensity value by combining the numerical growth rate of the compensation ratio sequence. Based on the numerical range of the risk transmission intensity value and the abnormal fluctuations in physiological indicator data, the level of postoperative complication risk for patients is determined and the classification result is output.

9. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps in the system as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions that, when executed by a processor, implement the steps in the system as described in any one of claims 1 to 7.

Images