Postoperative oral function rehabilitation training analysis method for oral maxillofacial surgery

Abstract

The invention discloses a post-operation oral function rehabilitation training analysis method for oral maxillofacial surgery, and relates to the technical field of biomedical engineering, which comprises the following steps of: constructing a characteristic relation matrix based on a time sequence embedding characteristic sequence, extracting a single characteristic change rate of the time sequence embedding characteristic sequence by adopting a discrete Riesz fractional derivative transformation method, and calculating a single characteristic change rate of the time sequence embedding characteristic sequence; fusing the feature relation matrix and the single feature change rate to generate a preliminary rehabilitation fingerprint prototype sequence; decomposing and reconstructing the rehabilitation fingerprint curve to generate a multi-scale sub-band, comparing the multi-scale sub-band with the rehabilitation fingerprint curve to generate an abnormal trajectory labeling graph, extracting fractal dimensions of abnormal timestamps according to the abnormal trajectory labeling graph, fusing the fractal dimensions with energy field changes to generate future trend prediction, and screening intervention nodes to generate an active intervention plan list; through the discrete Riesz fractional derivative transformation method, the fractional order characteristics of the non-stationary signals are effectively captured, and therefore the sensitivity of rehabilitation trend analysis is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biomedical engineering, and particularly relates to an oral cavity function rehabilitation training analysis method after oral and maxillofacial surgery. BACKGROUND

[0002] In the biomedical engineering industry, the oral cavity function rehabilitation training analysis after oral and maxillofacial surgery is an important research direction in the intersection field of clinical medicine and rehabilitation engineering. At present, in this field, a method based on multi-modal data acquisition and statistical modeling is mainly used, through integrating postoperative movement function, physiological indexes and subjective feedback data of a patient, a linear or nonlinear regression model is constructed to evaluate the rehabilitation progress. A conventional technical scheme usually depends on time sequence feature extraction and machine learning classification algorithm, such as using dynamic time warping (DTW) to align data, combining principal component analysis (PCA) or support vector machine (SVM) to realize feature dimension reduction and classification. In addition, some researches introduce a frequency domain analysis method, through fast Fourier transform (FFT) to extract signal frequency spectrum features to assist in rehabilitation state evaluation. Such methods have formed a standardized process in the biomedical engineering industry, and provide a basic technical support for postoperative rehabilitation management.

[0003] The existing method still has limitations in feature fusion and dynamic prediction: firstly, the heterogeneity of cross-modal features makes it difficult to accurately depict the nonlinear correlation between features by using traditional cosine similarity or Euclidean distance measurement, which affects the construction accuracy of the relationship matrix; secondly, the time-varying characteristics of the rehabilitation progress require a higher-order mathematical representation of the feature change rate, and the conventional integer-order derivative transformation cannot fully capture the dynamic evolution law in the fractional-order domain. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides an oral cavity function rehabilitation training analysis method after oral and maxillofacial surgery to solve the problems of insufficient cross-modal feature fusion accuracy and limited dynamic change representation capability.

[0006] To solve the above technical problems, the present application provides the following technical scheme: The present application provides an oral cavity function rehabilitation training analysis method after oral and maxillofacial surgery, which comprises: collecting rehabilitation data for time alignment and standardization processing, generating an actual rehabilitation fingerprint, obtaining cross-modal embedded features and high-order spectrum features and splicing to generate a time sequence embedded feature sequence; Based on the time sequence embedded feature sequence, a feature relationship matrix is constructed, and a discrete Riesz fractional derivative transformation method is used to extract a single feature change rate of the time sequence embedded feature sequence, the feature relationship matrix and the single feature change rate are fused to generate a preliminary rehabilitation fingerprint prototype sequence; Based on the preliminary rehabilitation fingerprint prototype sequence, a spatiotemporal tensor sequence is constructed and multidimensional indicators are extracted. Multi-feature fusion of the multidimensional indicators is then performed to generate a rehabilitation fingerprint curve. The rehabilitation fingerprint curve is decomposed and reconstructed to generate multi-scale sub-bands, and compared with the rehabilitation fingerprint curve to generate an abnormal trajectory annotation map. The fractal dimension of the abnormal timestamp and the energy field change are extracted from the abnormal trajectory annotation map and fused to generate future trend prediction. At the same time, intervention nodes are selected to generate a list of proactive intervention plans. The active intervention plan list is analyzed to generate a target information list. The target information list is compared with the actual rehabilitation fingerprint to generate a training effect analysis report.

[0007] As a preferred embodiment of the oral and maxillofacial surgery postoperative oral function rehabilitation training analysis method of the present invention, the steps of collecting rehabilitation data, performing time alignment and standardization processing, and generating actual rehabilitation fingerprints are as follows. Each rehabilitation data point is assigned a rehabilitation timestamp, which maps the rehabilitation data to a unified point in time, generating time-aligned rehabilitation data. The time-aligned rehabilitation data is standardized to generate actual rehabilitation fingerprints.

[0008] As a preferred embodiment of the oral function rehabilitation training analysis method after oral and maxillofacial surgery described in this invention, the step of obtaining cross-modal embedding features and higher-order spectral features and splicing them to generate a temporal embedding feature sequence refers to performing column-level standardization on the cross-modal embedding features and higher-order spectral features respectively to generate column-level standardized cross-modal embedding features and column-level standardized higher-order spectral features, and then cascading and splicing them along the feature dimensions to generate a temporal embedding feature sequence.

[0009] As a preferred embodiment of the oral functional rehabilitation training analysis method after oral and maxillofacial surgery described in this invention, the step of constructing a feature relationship matrix based on time-series embedded feature sequences is as follows: The cosine similarity between the column-level standardized cross-modal embedding features and the column-level standardized higher-order spectral features is calculated, and the absolute value of the first-order difference between the cross-modal embedding features and the higher-order spectral features is taken. The weighted fusion is then performed to generate the original relation value matrix. The original relation value matrix is ​​normalized to its maximum value to generate the feature relation matrix.

[0010] As a preferred embodiment of the oral and maxillofacial surgery postoperative oral function rehabilitation training analysis method of the present invention, the method of extracting the single feature change rate of the temporal embedded feature sequence using the discrete Riesz fractional derivative transform method, and fusing the feature relationship matrix and the single feature change rate to generate a preliminary rehabilitation fingerprint prototype sequence is as follows. The discrete Riesz fractional derivative transform method is used to map the time-series embedded feature sequence to the frequency domain, generate a fractional rate of change sequence, perform time-domain restoration, generate a single feature rate of change of the time-series embedded feature sequence, and perform normalization processing. The feature relationship matrix and the single feature change rate of the temporal embedded feature sequence are fused to generate a preliminary rehabilitation fingerprint prototype sequence.

[0011] As a preferred embodiment of the oral functional rehabilitation training analysis method after oral and maxillofacial surgery described in this invention, the steps for constructing the spatiotemporal tensor sequence and extracting multidimensional indicators are as follows: A spatiotemporal tensor sequence was constructed based on the preliminary rehabilitation fingerprint prototype sequence, and the fractal dimension was generated by box counting. The spatiotemporal energy field is obtained, and the frequency domain energy distribution complexity of the spatiotemporal energy field is quantified by Tsallis spectral entropy to generate spectral complexity entropy. The multidimensional indicators include spatiotemporal tensor sequence, fractal dimension, and spectral complexity entropy.

[0012] As a preferred embodiment of the oral function rehabilitation training analysis method after oral and maxillofacial surgery described in this invention, the step of generating a rehabilitation fingerprint curve by fractal dimension, spatiotemporal energy field and spectral complexity entropy is to generate multidimensional indicators and perform standardization processing, and then perform multi-feature fusion through principal component analysis to generate a rehabilitation fingerprint curve.

[0013] As a preferred embodiment of the oral and maxillofacial surgery postoperative oral function rehabilitation training analysis method of the present invention, the steps of decomposing and reconstructing the rehabilitation fingerprint curve to generate multi-scale sub-bands and comparing them with the rehabilitation fingerprint curve to generate an abnormal trajectory annotation map are as follows. The rehabilitation fingerprint curve is decomposed and reconstructed to generate multi-scale sub-bands; Calculate the absolute value and standard deviation of the residuals between the multi-scale subbands of the rehabilitation timestamp and the rehabilitation fingerprint curve; Compare the absolute value of the residuals with the standard deviation of the residuals to mark the timestamps of the anomalies, and generate an anomaly trajectory annotation map.

[0014] As a preferred embodiment of the oral and maxillofacial surgery postoperative oral function rehabilitation training analysis method of the present invention, the steps of extracting the fractal dimension and energy field changes of abnormal timestamps from the abnormal trajectory annotation map to generate future trend predictions, and simultaneously screening intervention nodes to generate an active intervention plan list, are as follows: Based on the abnormal trajectory annotation map, the fractal dimension and spatiotemporal energy field corresponding to each abnormal timestamp are read and weighted and fused to extrapolate the rehabilitation fingerprint trend and generate a future trend prediction curve. By comparing the future trend prediction curve with the rehabilitation fingerprint curve, a difference sequence is generated, all local extreme points are identified, and the rehabilitation timestamps corresponding to the local extreme points are used as intervention nodes. All intervention nodes are integrated to generate an active intervention plan list.

[0015] As a preferred embodiment of the oral and maxillofacial surgery postoperative oral function rehabilitation training analysis method of the present invention, the steps of parsing the active intervention plan list, generating a target information list, comparing the target information list with the actual rehabilitation fingerprint, and generating a training effect analysis report are as follows. Analyze the list of proactive intervention plans, generate intervention parameters and summarize them in chronological order to generate a list of target information; A difference analysis is performed on the target information list and the actual rehabilitation fingerprints to generate a difference statistics table. Based on the difference statistics table, a training effect analysis report with both text and graphics is generated.

[0016] The beneficial effects of this invention are as follows: By using the discrete Riesz fractional derivative transform method, the time-series embedded feature sequence is mapped to the frequency domain and the time-domain rate of change is restored, thus achieving a precise characterization of the dynamic evolution of rehabilitation data and effectively capturing the fractional-order characteristics of non-stationary signals, thereby improving the sensitivity of rehabilitation trend analysis. Based on the preliminary rehabilitation fingerprint prototype sequence, a spatiotemporal tensor sequence is constructed. Through multi-dimensional joint extraction of fractal dimension, spatiotemporal energy field, and spectral complexity entropy, a three-dimensional representation model of the rehabilitation process is established, achieving deep coupling between physiological function and motor behavior characteristics. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Fig. 1 This is a flowchart illustrating the analysis method for oral functional rehabilitation training after oral and maxillofacial surgery.

[0019] Fig. 2 This is a flowchart for cross-modal feature fusion and rehabilitation fingerprint generation.

[0020] Fig. 3 A flowchart generated for anomaly detection and intervention plans.

[0021] Fig. 4 A flowchart for analyzing training effectiveness. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figs. 1-4 This is one embodiment of the present invention, which provides a method for analyzing oral functional rehabilitation training after oral and maxillofacial surgery, including the following steps: S1. Collect rehabilitation data, perform time alignment and standardization processing, generate actual rehabilitation fingerprints, obtain cross-modal embedding features and high-order spectral features and splice them to generate a temporal embedding feature sequence. Rehabilitation data includes oral occlusal force, speech audio, swallowing electromyography, mouth opening degree, and the use of nursing assistive devices.

[0026] Each rehabilitation data point is assigned a rehabilitation timestamp. The rehabilitation data is then mapped to a unified time point using linear interpolation to generate time-aligned rehabilitation data. Furthermore, each rehabilitation data point is labeled with a rehabilitation timestamp from the time of surgery. The rehabilitation cycle (e.g., 30 days, 90 days) is determined according to the postoperative medical advice. The sampling interval (e.g., 1 minute) is divided according to the rehabilitation needs and the sampling time points are preset. The rehabilitation data is mapped to the sampling time points according to the rehabilitation timestamp using linear interpolation. The interpolated rehabilitation data is smoothed and corrected using the moving average method to generate continuous and aligned rehabilitation time series data. It should be noted that the moving average method calculates the time interval between all adjacent peaks of the actual rehabilitation fingerprint, takes half of the median as the target smoothing duration, divides the target smoothing duration by the sampling interval to generate an initial value for the number of window samples, sets the odd number closest to the initial value of the number of window samples as the sliding window W (e.g., 5), takes W consecutive sampling time points to calculate the average value of the rehabilitation data, generates the data average value, replaces the rehabilitation data at the sampling time point in the middle position with the data average value, and performs smoothing correction processing.

[0027] The time-aligned rehabilitation data is standardized by outlier imputation and noise removal to generate actual rehabilitation fingerprints; Furthermore, the rehabilitation mean and rehabilitation standard deviation of the rehabilitation data are calculated. Based on Gaussian empirical rule, the rehabilitation mean ± 3 rehabilitation standard deviation is set as the rehabilitation abnormal interval, and rehabilitation data falling outside the rehabilitation abnormal interval are rehabilitation outliers. The outlier imputation method involves averaging the rehabilitation data from the previous and next sampling time points of the outlier sampling time point and then filling the outlier sampling time point with the average. Calculate the difference between rehabilitation data at adjacent sampling time points and take the absolute value to generate an absolute value sequence. According to Gauss's rule of thumb, take 3 times the median of the absolute value sequence as the noise threshold. Rehabilitation data greater than the noise threshold is considered noise data, and rehabilitation data less than the noise threshold is considered non-noise data. Based on the sampling time point of the noise data, locate the sampling time points of adjacent non-noise data, calculate new rehabilitation data by interpolation based on the linear ratio of the relative time distance, and replace the noise data.

[0028] The rehabilitation data is standardized by outlier imputation and noise reduction to generate actual rehabilitation fingerprints.

[0029] Based on actual rehabilitation fingerprints, cross-modal embedding features are generated through spectral embedding, while high-order spectral features are generated through modal transient frequency analysis. Furthermore, the actual rehabilitation fingerprints and rehabilitation data are concatenated through vector concatenation to generate rehabilitation feature vectors. The Euclidean distance between each pair of rehabilitation feature vectors is calculated and substituted into a Gaussian kernel function to generate similarity weights. The similarity weights are then mapped to rows and columns according to the rehabilitation timestamps to generate a similarity matrix. A normalized Laplacian matrix is ​​constructed from the similarity matrix. The normalized Laplacian matrix is ​​then decomposed using a spectral embedding function to generate Laplacian spectral feature vectors. The first X Laplacian spectral feature vectors are extracted to generate cross-modal embedding features. The actual rehabilitation fingerprint was separated into intrinsic modal functions using a modal decomposition method. The modal amplitudes of the intrinsic modal functions were read sequentially at all sampling time points to generate a list of modal amplitudes. Based on the list of modal amplitudes, the modal amplitudes were paired with the sampling time points one by one to generate time-amplitude pairs. The modal amplitudes were arranged and connected according to the sampling time points of the time-amplitude pairs to form a continuous curve to generate the intrinsic modal time series. The Hilbert-Ma transform is applied to the intrinsic mode time series to obtain the imaginary part signal. The intrinsic mode time series is then paired with the imaginary part signal as the real part signal to generate an analytic signal. Vector synthesis is performed on the real and imaginary parts of the analytic signal to generate an instantaneous amplitude curve. The third-order cross-spectral analysis method is used to calculate the instantaneous amplitude curve to generate the third-order spectral moment. The instantaneous phase of the analytic signal is extracted and differential operation is performed to generate the instantaneous rate of change. The instantaneous rate of change is arranged in the order of recovery timestamps to generate an instantaneous frequency curve. The short-time Fourier entropy method is used to calculate the instantaneous frequency to generate a spectral entropy sequence. The third-order spectrum and the spectral entropy sequence are summarized to generate a higher-order spectral feature. It should be noted that higher-order spectral features refer to frequency domain features extracted using third-order or higher-order statistics of the signal.

[0030] Column-level normalization is performed on the cross-modal embedding features and the higher-order spectral features to generate column-level normalized cross-modal embedding features and column-level normalized higher-order spectral features. These are then concatenated along the feature dimensions to generate a temporal embedding feature sequence. Furthermore, the maximum and minimum values ​​of each column of the cross-modal embedding feature are recorded, and column-level standardized cross-modal embedding features are generated using the min-max normalization method; the maximum and minimum values ​​of each column of the higher-order spectral feature are recorded, and column-level standardized higher-order spectral features are generated using the min-max normalization method; the column-level standardized cross-modal embedding features and column-level standardized higher-order spectral features are time-aligned, and the number of rows of the column-level standardized cross-modal embedding features and column-level standardized higher-order spectral features correspond to the same sampling time point. At each sampling time point, the elements of the column-level standardized higher-order spectral features are embedded into the elements of the column-level standardized cross-modal embedding features, and then sorted according to the rehabilitation timestamp to generate a time-series embedding feature sequence.

[0031] S2. Based on the temporal embedded feature sequence, construct the feature relationship matrix, and use the discrete Riesz fractional derivative transform method to extract the single feature change rate of the temporal embedded feature sequence. Then, fuse the feature relationship matrix and the single feature change rate to generate a preliminary rehabilitation fingerprint prototype sequence. At each recovery timestamp, the cosine similarity between the column-level standardized cross-modal embedding features and the column-level standardized higher-order spectral features is calculated, and the absolute value of the first-order difference between the cross-modal embedding features and the higher-order spectral features is taken. The results are then weighted and fused to generate the original relation value matrix. Furthermore, column-level normalized cross-modal embedding features and column-level normalized higher-order spectral features of the temporal embedding feature sequence are extracted, and cosine similarity is calculated to generate a cosine similarity list; The column-level standardized cross-modal embedding features and the column-level standardized higher-order spectral features are calculated by difference based on the two acquisition time points and the absolute value is taken to generate the column-level standardized cross-modal embedding feature difference absolute value sequence and the column-level standardized higher-order spectral feature difference absolute value sequence. The cosine similarity list, the column-level standardized cross-modal embedding feature difference absolute value sequence, and the column-level standardized high-order spectral feature difference absolute value sequence are normalized and aligned according to the recovery timestamp. The relationship weights are calculated using the entropy weight method, and the cosine similarity list, the column-level standardized cross-modal embedding feature difference absolute value sequence, and the column-level standardized high-order spectral feature difference absolute value sequence are weighted and superimposed to generate the original relationship value matrix.

[0032] It should be noted that cosine similarity is an index that measures the degree of similarity between two vectors in a direction. It reflects the similarity by calculating the cosine value of the angle between them.

[0033] The original relation value matrix is ​​normalized to its maximum value to generate a feature relation matrix; Furthermore, the maximum value of the original relation value matrix is ​​identified, the maximum relation value is generated, and the feature relation matrix is ​​generated by dividing the original relation matrix by the maximum relation value.

[0034] The discrete Riesz fractional derivative transform method is adopted to map the time-series embedded feature sequence to the frequency domain through discrete Fourier transform, and perform weighted calculation to generate a fractional rate of change sequence. Then, the time domain is restored through inverse Fourier transform to generate a single feature rate of change of the time-series embedded feature sequence and perform normalization processing. Furthermore, a discrete Fourier transform is performed on the time-series embedded feature sequence to generate a frequency domain representation matrix; a weighted calculation is then performed on the frequency domain representation matrix to generate a fractional-order rate of change sequence. The fractional rate of change sequence is restored to a time-domain waveform by using inverse Fourier transform, generating a single feature rate of change of the time-series embedded feature sequence, and then normalized.

[0035] By fusing the single feature change rate of the feature relationship matrix and the temporal embedded feature sequence through weighted superposition based on the entropy weight method, a preliminary rehabilitation fingerprint prototype sequence is generated.

[0036] Furthermore, the feature relationship matrix and the rate of change of a single feature are aligned by the recovery timestamp and merged into an index matrix by column. The information entropy is calculated for each column of the index matrix, and the complement of the information entropy is taken to obtain the information entropy redundancy. The information entropy redundancy is normalized to obtain the entropy weight coefficient, and a list of entropy weight coefficients is generated. The index matrix and the entropy weight coefficient list are sequentially weighted and merged using entropy weights to generate a fused value sequence. The fusion value sequence is sorted according to the recovery timestamp to generate a preliminary recovery fingerprint prototype sequence.

[0037] S3. Based on the preliminary rehabilitation fingerprint prototype sequence, construct a spatiotemporal tensor sequence and extract multidimensional indicators. Perform multi-feature fusion on the multidimensional indicators to generate a rehabilitation fingerprint curve. Based on the preliminary rehabilitation fingerprint prototype sequence, a spatiotemporal tensor sequence is constructed using the sliding window tensor method, and the fractal dimension is generated using the box counting method. Furthermore, the preliminary rehabilitation fingerprint prototype sequence is stacked according to the rehabilitation timestamp to form a multidimensional data block by using the sliding window tensor method. The multidimensional data blocks are then arranged and summarized according to the sliding window order to generate a spatiotemporal tensor sequence. It should be noted that the spatiotemporal tensor sequence includes three dimensions: the time dimension, the first spatial dimension, and the second spatial dimension. The fractal dimension is generated using box counting, expressed as follows: ; in, for The fractal dimension of time; For measurement accuracy; for At any given moment, the minimum number of boxes required to arrange the boxes along the time dimension to cover the signal time series; for At any given moment, the minimum number of boxes required to cover the signal time series by arranging the boxes along one spatial dimension; for At any given time, the minimum number of boxes required to cover the signal time series by arranging the boxes along the second spatial dimension; For recovery timestamps; It should be noted that the time dimension corresponds to the recovery timestamp, the first spatial dimension corresponds to the sampling time point, and the second spatial dimension corresponds to the preliminary recovery fingerprint prototype sequence. Based on the spatiotemporal tensor sequence, the spatiotemporal energy field is obtained through the graph Laplacian spectral energy transfer algorithm, and the frequency domain energy distribution complexity of the spatiotemporal energy field is quantified by the Tsallis spectral entropy to generate the spectral complexity entropy. Furthermore, the spatiotemporal energy field is obtained using the Graph Laplace spectral energy transfer algorithm, expressed as follows: ; in, for The spacetime energy field at any given moment For a moment In the feature relation matrix, the first Line 1 Liede Elements of the characteristic relationship matrix of the channel; For the first Channel, sampling time point The weights of the plot Laplace spectrum at sampling time point y; For at any time , No. The difference between the elements of each node in the feature relation matrix and the elements reconstructed by the weighted Laplacian spectrum of the graph; This is the residual penalty coefficient (typically ranging from 0 to 1). This refers to the row index node in the feature relation matrix; For column index nodes in the feature relation matrix; For feature channel index; is the dimension of the feature relation matrix; For the first There are 1 sampling time points; y is the y-th sampling time point; The frequency domain energy distribution complexity of the spatiotemporal energy field is quantified using Tsallis spectral entropy, generating a spectral complexity entropy, expressed as follows: ; in, The complexity is spectral entropy. In order to be in Time, Number The first in the passage The proportion of energy at each node to the overall energy field; This is the Tsallis entropy index.

[0038] It should be noted that the spatiotemporal energy field is a single numerical index output for each recovery timestamp based on the spatiotemporal tensor sequence. It is a numerical sequence that changes over time. The larger the value, the stronger the coupling between the sampling time points of the preliminary recovery fingerprint prototype sequence within that window, and the more significant the energy concentration and transmission.

[0039] It should be noted that weighted reconstruction refers to obtaining new matrix element values ​​by weighting the elements of the feature relation matrix using the graph Laplace spectral weights. Statistical fractal dimension, spatiotemporal energy field and spectral complexity entropy are used to generate a multidimensional index set and standardize it. Principal component analysis is used to fuse multiple features to generate a rehabilitation fingerprint curve. Furthermore, the fractal dimension, spatiotemporal energy field, and spectral complexity entropy are time-aligned according to the recovery timestamp, and the fractal dimension, spatiotemporal energy field, and spectral complexity entropy are combined at each recovery timestamp to generate a three-dimensional index. The three-dimensional indexes are then summarized according to the order of the recovery timestamps to generate a multi-dimensional index matrix. The multidimensional indicator matrix is ​​subjected to min-max normalization to generate a standardized multidimensional indicator matrix. Principal component analysis is performed on the standardized multidimensional indicator matrix, and the time series of the first principal component is extracted and sorted according to the rehabilitation timestamp to generate a rehabilitation fingerprint curve.

[0040] S4. Decompose and reconstruct the rehabilitation fingerprint curve to generate multi-scale sub-bands, and compare them with the rehabilitation fingerprint curve to generate an abnormal trajectory annotation map. Extract the fractal dimension of the abnormal timestamp and the energy field change from the abnormal trajectory annotation map and fuse them to generate a future trend prediction curve. At the same time, select intervention nodes to generate an active intervention plan list. The rehabilitation fingerprint curve was decomposed and reconstructed using Daubechies-4 wavelets to generate multi-scale sub-bands. Furthermore, the rehabilitation fingerprint curve is decomposed using a Daubechies-4 wavelet filter. The total number of sampling points of the rehabilitation fingerprint curve is counted, and the maximum number of levels is determined by combining the minimum support length of the wavelet filter. The maximum number of levels is rounded down to generate a compromise number of levels as the predetermined number of decomposition levels. The first-level approximation component and detail component are split by low-pass and high-pass filters. The approximation component is split repeatedly until the predetermined number of decomposition levels is reached. The approximation component and detail component of each level are output. The approximation component and the corresponding detail component are paired and arranged in order from low frequency to high frequency to generate multi-scale sub-bands.

[0041] The absolute value and standard deviation of the residuals between the multi-scale sub-bands of the rehabilitation timestamp and the rehabilitation fingerprint curve were calculated using the sample standard deviation method. Furthermore, the multi-scale sub-band sequence is compared with the rehabilitation fingerprint curve to generate a residual sequence, and the residual absolute value sequence and residual standard deviation are generated based on the residual sequence.

[0042] Anomaly timestamps are marked by comparing the absolute value of the residuals with the standard deviation of the residuals, and anomaly trajectory annotation map is generated by scatter plot annotation method; Furthermore, the absolute value of the residuals is compared with the standard deviation of the residuals. Recovery timestamps where the absolute value of the residuals exceeds the standard deviation of the residuals are marked as abnormal timestamps, and an abnormal timestamp list is generated. Based on the abnormal timestamp list, abnormal timestamps are extracted from the recovery fingerprint curve, and the recovery values ​​corresponding to the recovery fingerprint curve are obtained. The abnormal timestamps are used as the x-axis of the scatter plot, and the recovery values ​​corresponding to the recovery fingerprint curve are used as the y-axis of the scatter plot. A line graph of the recovery fingerprint curve is plotted, and abnormal coordinate points are marked according to the x-axis and y-axis of the scatter plot, generating an abnormal trajectory annotation map.

[0043] Based on the abnormal trajectory annotation map, the fractal dimension and spatiotemporal energy field corresponding to each abnormal timestamp are read and weighted and fused. The recovery fingerprint trend is extrapolated through local polynomial regression to generate a future trend prediction curve. Furthermore, the fractal dimension and spatiotemporal energy field of the abnormal timestamps are extracted from the abnormal trajectory annotation map, and an abnormal time index table is generated. The fractal dimension and spatiotemporal energy field of the abnormal timestamps are sorted according to the recovery timestamp to generate fractal dimension sequence and spatiotemporal energy field sequence. The ratio of the sample standard deviation to the sample mean of the fractal dimension sequence and spatiotemporal energy field sequence are calculated respectively to generate the fractal dimension coefficient of variation and the spatiotemporal energy field coefficient of variation. The fractal dimension coefficient of variation is divided by the sum of the fractal dimension coefficients of variation to generate the fractal dimension weight ratio. The spatiotemporal energy field coefficient of variation is divided by the sum of the spatiotemporal energy field coefficients of variation to generate the spatiotemporal energy field weight ratio. Based on the weight ratio of fractal dimension and the weight ratio of spatiotemporal energy field, a weighted sum of fractal dimension and spatiotemporal energy field is generated to produce an abnormal driving force sequence. Local multinomial regression is applied to the abnormal driving force sequence based on the recovery time stamp, and a neighborhood window is used for smooth fitting. Then, the fitted function is extrapolated to future sampling points along the time axis to output the future trend prediction curve.

[0044] By comparing the future trend prediction curve with the rehabilitation fingerprint curve, a difference sequence is generated. All local extreme points are identified by the first-order difference zero crossover of the difference sequence. The rehabilitation timestamps corresponding to the local extreme points are used as intervention nodes. All intervention nodes are integrated to generate an active intervention plan list. Furthermore, the difference between the future trend prediction curve and the rehabilitation fingerprint curve is calculated based on the rehabilitation timestamp to generate a difference sequence. The difference between each consecutive difference sequence is calculated based on the rehabilitation timestamp to generate a first-order difference sequence. The sign change of the values ​​of adjacent elements in the first-order difference sequence is determined. The rehabilitation timestamp position of the next element of the adjacent element of the first-order difference sequence with the sign change is recorded as a local extreme point, and a local extreme point index list is generated. The rehabilitation timestamps corresponding to the local extreme points are used as intervention nodes. All intervention nodes are summarized to generate an active intervention plan list. It should be noted that a sign change means that when the current difference is less than or equal to zero and the next difference is greater than zero, or when the current difference is greater than zero and the next difference is less than or equal to zero, the sign of the two differences is considered to have changed.

[0045] S5. Analyze the active intervention plan list, generate a target information list, compare the target information list with the actual rehabilitation fingerprint, and generate a training effect analysis report.

[0046] The active intervention plan list is parsed by traversing the table rows, intervention parameters are generated and summarized in chronological order to generate a target information list. Furthermore, the active intervention plan list is read sequentially by traversing the table rows to generate intervention parameter records, which are then arranged and merged in order of rehabilitation timestamps to generate a target information list; It should be noted that the target information list includes intervention time, action type, intensity coefficient, and frequency coefficient.

[0047] Perform difference analysis on the target information list and the actual rehabilitation fingerprint, generate a difference statistics table, and generate a training effect analysis report with both text and graphics based on the difference statistics table; Furthermore, each rehabilitation timestamp in the target information list is read one by one, and the same rehabilitation timestamp is located in the actual rehabilitation fingerprint. The average of the three rehabilitation timestamps before this rehabilitation timestamp is calculated to generate the pre-intervention average. The average of the three rehabilitation timestamps after this rehabilitation timestamp is calculated to generate the post-intervention average. The difference between the mean after intervention and the mean before intervention is calculated to generate the actual response value with a rehabilitation timestamp. The actual response value is then normalized and summarized in chronological order to generate the actual response sequence. Calculate the difference between the actual response sequence and the target information list, record the absolute value of the difference and the direction label (positive for improvement, negative for decline), and generate a difference statistics table. Extract the mean values ​​before and after intervention from the difference statistics table and combine them with the rehabilitation timestamp to generate line comparison data; extract the actual response value and the absolute value of the actual response value and combine them with the rehabilitation timestamp to generate difference bar data; perform statistical analysis on the line comparison data and the difference bar data to generate a visualization dataset; Based on the visualization dataset, a line comparison chart is generated from the line comparison data, and a difference bar chart is generated from the difference bar data; Based on the difference statistics table, line comparison chart, and difference bar chart, the rehabilitation training situation is analyzed, and a rehabilitation training situation analysis report is generated. The difference statistics table, line comparison chart, difference bar chart, and rehabilitation training situation analysis report are arranged in the order of "charts - text interpretation - conclusion" to generate a training effect analysis report with both text and graphics. It should be noted that the difference statistics table includes rehabilitation timestamp, intensity coefficient, frequency coefficient, actual response value, target impact value, difference, absolute value of difference, and direction marker.

[0048] This embodiment also provides a computer device applicable to the postoperative oral function rehabilitation training analysis method for oral and maxillofacial surgery, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the postoperative oral function rehabilitation training analysis method for oral and maxillofacial surgery as proposed in the above embodiment.

[0049] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0050] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the method for analyzing oral function rehabilitation after oral and maxillofacial surgery as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0051] In summary, this invention utilizes the discrete Riesz fractional derivative transform method to map the time-series embedded feature sequence to the frequency domain and restore the time-domain rate of change, thereby achieving a precise characterization of the dynamic evolution of rehabilitation data and effectively capturing the fractional-order characteristics of non-stationary signals, thus improving the sensitivity of rehabilitation trend analysis. Based on the preliminary rehabilitation fingerprint prototype sequence, a spatiotemporal tensor sequence is constructed. Through multi-dimensional joint extraction of fractal dimension, spatiotemporal energy field, and spectral complexity entropy, a three-dimensional representation model of the rehabilitation process is established, achieving deep coupling between physiological function and motor behavior characteristics.

[0052] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for analyzing oral functional rehabilitation training after oral and maxillofacial surgery, characterized in that: include, Collect rehabilitation data, perform time alignment and standardization processing to generate actual rehabilitation fingerprints, obtain cross-modal embedding features and high-order spectral features and splice them to generate a temporal embedding feature sequence; Based on the temporal embedded feature sequence, a feature relationship matrix is ​​constructed, and the discrete Riesz fractional derivative transform method is used to extract the single feature change rate of the temporal embedded feature sequence. The feature relationship matrix and the single feature change rate are fused to generate a preliminary rehabilitation fingerprint prototype sequence. Based on the preliminary rehabilitation fingerprint prototype sequence, a spatiotemporal tensor sequence is constructed and multidimensional indicators are extracted. Multi-feature fusion of the multidimensional indicators is then performed to generate a rehabilitation fingerprint curve. The rehabilitation fingerprint curve is decomposed and reconstructed to generate multi-scale sub-bands, and compared with the rehabilitation fingerprint curve to generate an abnormal trajectory annotation map. The fractal dimension of the abnormal timestamp and the energy field change are extracted from the abnormal trajectory annotation map and fused to generate future trend prediction. At the same time, intervention nodes are selected to generate a list of proactive intervention plans. The active intervention plan list is analyzed to generate a target information list. The target information list is compared with the actual rehabilitation fingerprint to generate a training effect analysis report.

2. The method for analyzing oral functional rehabilitation training after oral and maxillofacial surgery as described in claim 1, characterized in that: The collected rehabilitation data undergoes time alignment and standardization processing to generate actual rehabilitation fingerprints. The steps are as follows: Each rehabilitation data point is assigned a rehabilitation timestamp, which maps the rehabilitation data to a unified point in time, generating time-aligned rehabilitation data. The time-aligned rehabilitation data is standardized to generate actual rehabilitation fingerprints.

3. The method for analyzing oral functional rehabilitation training after oral and maxillofacial surgery as described in claim 2, characterized in that: The step of acquiring and concatenating cross-modal embedding features and higher-order spectral features to generate a temporal embedding feature sequence refers to performing column-level normalization on the cross-modal embedding features and higher-order spectral features respectively, generating column-level normalized cross-modal embedding features and column-level normalized higher-order spectral features. Concatenate and splice along the feature dimension to generate a temporal embedded feature sequence.

4. The method for analyzing oral functional rehabilitation training after oral and maxillofacial surgery as described in claim 3, characterized in that: The steps for constructing a feature relationship matrix based on temporal embedded feature sequences are as follows. The cosine similarity between the column-level standardized cross-modal embedding features and the column-level standardized higher-order spectral features is calculated, and the absolute value of the first-order difference between the cross-modal embedding features and the higher-order spectral features is taken. The weighted fusion is then performed to generate the original relation value matrix. The original relation value matrix is ​​normalized to its maximum value to generate the feature relation matrix.

5. The method for analyzing oral functional rehabilitation training after oral and maxillofacial surgery as described in claim 4, characterized in that: The discrete Riesz fractional derivative transform method is used to extract the single feature change rate of the temporal embedded feature sequence. The feature relationship matrix and the single feature change rate are then fused to generate a preliminary rehabilitation fingerprint prototype sequence. The steps are as follows. The discrete Riesz fractional derivative transform method is used to map the time-series embedded feature sequence to the frequency domain, generate a fractional rate of change sequence, perform time-domain restoration, generate a single feature rate of change of the time-series embedded feature sequence, and perform normalization processing. The feature relationship matrix and the single feature change rate of the temporal embedded feature sequence are fused to generate a preliminary rehabilitation fingerprint prototype sequence.

6. The method for analyzing oral functional rehabilitation training after oral and maxillofacial surgery as described in claim 5, characterized in that: The steps for constructing the spatiotemporal tensor sequence and extracting multidimensional indicators are as follows: A spatiotemporal tensor sequence was constructed based on the preliminary rehabilitation fingerprint prototype sequence, and the fractal dimension was generated by box counting. The spatiotemporal energy field is obtained, and the frequency domain energy distribution complexity of the spatiotemporal energy field is quantified by Tsallis spectral entropy to generate spectral complexity entropy. The multidimensional indicators include spatiotemporal tensor sequence, fractal dimension, and spectral complexity entropy.

7. The method for analyzing oral functional rehabilitation training after oral and maxillofacial surgery as described in claim 6, characterized in that: The process of generating a rehabilitation fingerprint curve by fusing multiple features of multidimensional indicators refers to generating multidimensional indicators by statistical fractal dimension, spatiotemporal energy field, and spectral complexity entropy, and then standardizing them. Principal component analysis is then used to fuse multiple features to generate the rehabilitation fingerprint curve.

8. The method for analyzing oral functional rehabilitation training after oral and maxillofacial surgery as described in claim 7, characterized in that: The steps for decomposing and reconstructing the rehabilitation fingerprint curve to generate multi-scale sub-bands, and comparing them with the rehabilitation fingerprint curve to generate an abnormal trajectory annotation map, are as follows: The rehabilitation fingerprint curve is decomposed and reconstructed to generate multi-scale sub-bands; Calculate the absolute value and standard deviation of the residuals between the multi-scale subbands of the rehabilitation timestamp and the rehabilitation fingerprint curve; Compare the absolute value of the residuals with the standard deviation of the residuals to mark the timestamps of the anomalies, and generate an anomaly trajectory annotation map.

9. The method for analyzing oral functional rehabilitation training after oral and maxillofacial surgery as described in claim 8, characterized in that: The steps are as follows: extracting the fractal dimension of abnormal timestamps from the abnormal trajectory annotation map and fusing them with energy field changes to generate future trend predictions; simultaneously, selecting intervention nodes to generate a list of proactive intervention plans. Based on the abnormal trajectory annotation map, the fractal dimension and spatiotemporal energy field corresponding to each abnormal timestamp are read and weighted and fused to extrapolate the rehabilitation fingerprint trend and generate a future trend prediction curve. By comparing the future trend prediction curve with the rehabilitation fingerprint curve, a difference sequence is generated, all local extreme points are identified, and the rehabilitation timestamps corresponding to the local extreme points are used as intervention nodes. All intervention nodes are integrated to generate an active intervention plan list.

10. The method for analyzing oral functional rehabilitation training after oral and maxillofacial surgery as described in claim 9, characterized in that: The process of analyzing the active intervention plan list, generating a target information list, comparing the target information list with the actual rehabilitation fingerprints, and generating a training effect analysis report follows. Analyze the list of proactive intervention plans, generate intervention parameters and summarize them in chronological order to generate a list of target information; A difference analysis is performed on the target information list and the actual rehabilitation fingerprints to generate a difference statistics table. Based on the difference statistics table, a training effect analysis report with both text and graphics is generated.

Images