Risk prediction and personalized intervention recommendation system for dysphagia in stroke patients

Abstract

This invention relates to the field of medical intelligent assisted diagnostic technology, specifically a risk prediction and personalized intervention recommendation system for dysphagia in stroke patients. The system includes: synchronously acquiring multimodal physiological signals under a standard swallowing protocol; dynamically segmenting and labeling the signals by analyzing quasi-periodic rhythmic perturbation patterns; filtering signal segments based on a synergistic disintegration index; reconstructing them into characteristic trajectories and mapping them to anomalous attractors in a high-dimensional phase space; and extracting their geometric invariants as risk features. A cascaded decision forest model is used to achieve risk stratification, and multi-layered progressive rule retrieval is triggered based on the stratification results to generate a time-adaptive personalized intervention plan. This invention achieves precise quantification and mechanistic risk stratification of the dynamic vulnerability of swallowing function, and can automatically match individual deficit patterns to generate customized rehabilitation strategies.

Description

Technical Field

[0001] This invention relates to the field of medical intelligent auxiliary diagnostic technology, and in particular to a system for predicting the risk of dysphagia in stroke patients and recommending personalized interventions. Background Technology

[0002] Currently, risk prediction for dysphagia after stroke mainly relies on clinical scale assessments or observations based on medical imaging. While video fluorescein angiography is considered the gold standard, it only provides static or discrete dynamic images of anatomical structures during swallowing, failing to continuously quantify the temporal dynamics of neuromuscular coordination. Some studies have attempted to use surface electromyography or laryngeal acceleration signals for auxiliary analysis, but most remain at the level of simple statistics on signal amplitude and timing, or feature extraction within fixed time windows. These methods struggle to capture the subtle rhythmic disturbances and instantaneous breakdown of multi-muscle group coordination in this quasi-cyclic physiological activity of swallowing, thus failing to effectively identify early, dynamic markers of swallowing function decline.

[0003] Another drawback of existing technologies lies in the disconnect between risk assessment and intervention recommendations. Most predictive models output a single risk level or probability, lacking a fine-grained stratification of the underlying physiological mechanisms of risk. Intervention recommendations often rely on clinical experience or simple decision trees, failing to dynamically adjust based on specific synergistic disintegration patterns reflected in real-time patient assessments. This results in interventions lacking sufficient individualization and temporal adaptability, making it difficult to match the dynamic trajectory of changes in the patient's functional state. Therefore, a technological approach is needed that can deeply analyze the dynamic rhythms and synergies of swallowing and automatically map quantitative risk characteristics into a tiered, personalized intervention plan. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a risk prediction and personalized intervention recommendation system for dysphagia in stroke patients.

[0005] To achieve the above objectives, the present invention employs the following technical solution: a risk prediction and personalized intervention recommendation system for dysphagia in stroke patients, comprising:

[0006] The multimodal synchronous acquisition unit is used to capture the raw multimodal physiological signal stream when a stroke patient performs standard swallowing actions under the guidance of a specific swallowing protocol.

[0007] The rhythm perturbation parsing unit is used to parse the quasi-periodic rhythm components in the original multimodal physiological signal stream, and to segment the signal according to the rhythm perturbation mode to generate a set of physiological signal segments carrying rhythm tags.

[0008] A collaborative disintegration quantification unit is used to quantify the collaborative disintegration index of each segment in the set of physiological signal segments;

[0009] The trajectory construction and mapping unit is used to reconstruct physiological signal fragments that conform to a specific disintegration pattern into characteristic trajectories that reflect the vulnerability of swallowing function, and to map the characteristic trajectories to abnormal attractors in a high-dimensional phase space;

[0010] A geometric feature extraction unit is used to extract the geometric invariants of the abnormal attractor as potential risk features of dysphagia.

[0011] The risk stratification decision unit is used to input the potential risk characteristics into the risk stratification model based on cascaded decision forest to obtain individualized risk prediction labels.

[0012] The personalized solution generation unit is used to trigger a multi-level progressive retrieval of the structured intervention rule base based on the risk prediction tags, and generate a personalized intervention solution with time-adaptive characteristics.

[0013] Preferably, the capture of the raw multimodal physiological signal stream when a stroke patient performs standard swallowing actions under the guidance of a specific swallowing protocol specifically includes:

[0014] Configure a standardized assessment process that includes swallowing tasks with foods of different textures, and simultaneously activate the cortical EEG acquisition device, the laryngeal and cervical electromyography sensor array, and the swallowing sound intensity detector.

[0015] The cortical electroencephalogram (EEG) acquisition device records the cortical potential fluctuation sequence in specific brain regions before and after the swallowing command is issued;

[0016] The laryngeal and cervical electromyography sensor array captures the electromyographic activation waveform sequence of the suprahyoid muscle group and the pharyngeal constriction muscle group.

[0017] The swallowing sound intensity detector collects the sound vibration signal sequence generated by the swallowing action;

[0018] The cortical potential fluctuation sequence, electromyographic activation waveform sequence, and sound vibration signal sequence are stamped and aligned according to a unified time reference, and integrated to form the original multimodal physiological signal stream.

[0019] Preferably, the step of parsing the quasi-periodic rhythmic components in the original multimodal physiological signal stream and fragmenting the signal according to the rhythmic perturbation pattern includes:

[0020] An adaptive spectral estimation algorithm is applied to the original multimodal physiological signal stream to identify rhythmic frequency bands with significant power in the signal;

[0021] The quasi-periodic oscillation signal component driven by the swallowing center pattern generator is separated from the rhythm frequency band;

[0022] The instantaneous frequency jump point and amplitude collapse point of the quasi-periodic oscillation signal component are detected, and the signal interval between every two adjacent collapse points is defined as a rhythm unit.

[0023] Based on the frequency stability and amplitude decay rate of the rhythm unit, a rhythm tag reflecting the severity of its disturbance is assigned to each rhythm unit;

[0024] All rhythm units carrying rhythm tags are gathered together to form the set of physiological signal segments.

[0025] Preferably, the quantification index for the cooperative disintegration of each segment in the set of physiological signal segments includes:

[0026] For each physiological signal segment in the set of physiological signal segments, the phase-locking value between its cortical potential fluctuation sequence and electromyographic activation waveform sequence is calculated.

[0027] Calculate the degree of disorder in the activation sequence among the channels of its electromyographic activation waveform sequence;

[0028] Calculate the spectral distortion distance between the time spectrum of its sound vibration signal sequence and the spectrum of the normal swallowing template;

[0029] The phase-locked value, activation order disorder degree, and spectral distortion distance are weighted and fused to generate a comprehensive collaborative disintegration index value;

[0030] When the value of the collaborative disintegration index exceeds the preset disintegration threshold, the physiological signal fragment is determined to conform to the specific disintegration pattern.

[0031] Preferably, the reconstructing of physiological signal fragments conforming to a specific disintegration pattern into characteristic trajectories reflecting the vulnerability of swallowing function includes:

[0032] The synergistic disintegration index values ​​of all physiological signal fragments that conform to a specific disintegration pattern are extracted in chronological order to form a disintegration index sequence that evolves over time.

[0033] The disintegration index sequence is smoothed by interpolation to construct a continuous feature trajectory curve;

[0034] Calculate the curvature extrema and inflection points of the characteristic trajectory curve, and use the curvature extrema and inflection points as state representation points of the characteristic trajectory;

[0035] The step of smoothing and interpolating the disintegration index sequence to construct a continuous feature trajectory curve specifically includes:

[0036] An adaptive kernel density estimation algorithm is used to model the probability density distribution of discrete data points in the collapse index sequence, generating a local density estimate for each data point.

[0037] A Gaussian process regression model is constructed based on the local density estimate, with the timestamps of the disintegration index sequence as input variables and the values ​​of the disintegration indexes as output variables.

[0038] The posterior predicted distribution at any time point is calculated using the Gaussian process regression model, generating a smooth mean function curve.

[0039] The mean function curve is sampled at equal intervals using a cubic spline interpolation algorithm to obtain a characteristic trajectory curve with continuous first and second derivatives;

[0040] The characteristic trajectory curve is normalized so that its numerical range is mapped to between zero and one, which facilitates subsequent phase space mapping operations.

[0041] Preferably, mapping the feature trajectory to an anomalous attractor in a high-dimensional phase space includes:

[0042] Using the state representation points of the feature trajectory as the reference points for delayed reconstruction, the one-dimensional feature trajectory is reconstructed into a set of point clouds in a high-dimensional phase space using the time delay embedding method.

[0043] Density clustering is performed on the point cloud in the high-dimensional phase space to identify regions where the point cloud distribution is dense and stable.

[0044] The region is defined as an abnormal attractor that characterizes the dynamics of abnormal swallowing function.

[0045] Preferably, extracting the geometric invariants of the abnormal attractor as potential risk features of dysphagia includes:

[0046] Calculate the correlation dimension of the anomalous attractor to measure its structural complexity;

[0047] Calculate the Lyapunov exponent spectrum of the anomalous attractor to quantify its dynamic instability;

[0048] The variance distribution ratio of the anomalous attractor along its principal axis in phase space is calculated to describe its morphological anisotropy.

[0049] The correlation dimension, Lyapunov index spectrum, and variance distribution ratio are used together as the potential risk features.

[0050] Preferably, the step of inputting the potential risk features into a risk hierarchical model based on cascaded decision forests to obtain individualized risk prediction labels includes:

[0051] The risk hierarchical model based on cascaded decision forests consists of multiple sequentially connected decision forests, each of which focuses on assessing risk in a specific dimension.

[0052] The output confidence vector of the previous decision forest will be used as part of the input features of the next decision forest;

[0053] The final decision forest output layer will generate a multi-dimensional risk prediction label, which simultaneously encodes the obstacle level, the main physiological dimensions of damage, and the risk evolution trend.

[0054] Preferably, the step of triggering a multi-level progressive retrieval of the structured intervention rule base based on the risk prediction tag to generate a personalized intervention plan with time-adaptive characteristics includes:

[0055] The structured intervention rule base is organized in the form of a directed hypergraph, where nodes represent intervention measures or physiological states, and hyperedges represent multidimensional triggering and transformation rules.

[0056] Based on the obstacle level code in the risk prediction label, locate the initial intervention node cluster in the intervention rule base;

[0057] Based on the coding of the main damaged physiological dimensions in the risk prediction labels, a subset of targeted intervention measures is selected from the initial intervention node cluster;

[0058] Based on the risk evolution trend code in the risk prediction label, the adjustment path of the execution intensity and duration of each measure in the subset of targeted intervention measures is dynamically planned;

[0059] All interventions with dynamically adjustable pathways are combined to form the time-adaptive personalized intervention program.

[0060] Preferably, the step of performing density clustering on the point cloud in the high-dimensional phase space to identify regions with dense and stable point cloud distribution includes:

[0061] A density-based noise-based spatial clustering algorithm is used to perform density reachability analysis on the point cloud in the high-dimensional phase space, and the density connectivity of each point within its neighborhood radius is calculated.

[0062] Hierarchical density clustering method is used to aggregate mutually density-reachable points into initial clusters and remove isolated point clouds formed by noise points;

[0063] Stability assessment is performed on each initial cluster, and the centroid shift variance and intra-cluster point density index of the cluster in phase space are calculated.

[0064] Stable clusters are selected based on centroid movement variance being lower than the variance threshold and intra-cluster point density being higher than the density threshold.

[0065] The centroid of the stable cluster is used as the core anchor point of the anomalous attractor, and the region with dense and stable point cloud distribution is expanded around the anchor point.

[0066] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0067] By analyzing the quasi-periodic rhythmic components in multimodal physiological signals and dynamically segmenting and labeling the signals according to their perturbation patterns, this method can directly capture microscopic anomalies in the temporal coordination of swallowing movements. Signal segments with specific cooperative disintegration patterns are reconstructed into continuous trajectories and mapped to a high-dimensional phase space to construct anomalous attractors, thereby extracting their geometric invariants. This technique transforms non-stationary physiological dynamics into quantitative features characterizing system stability, enabling a direct measurement of the potential instability of swallowing function and overcoming the limitation of traditional fixed-window features being insensitive to instantaneous state changes.

[0068] A synergistic disintegration index for neuromuscular activity within multimodal signal fragments is defined, directly quantifying the loss of synchronization and coordination among multiple subsystems during swallowing. This index, along with kinetic invariants, is input into a cascaded decision forest model. The model achieves layer-by-layer fine-grained identification from ubiquitous risk to specific risk patterns through sequential decision nodes. Based on this hierarchical result, a progressive retrieval of a structured intervention rule base is triggered. Each level of retrieval corresponds to different risk mechanisms and intervention intensities, enabling the generated plan to accurately match the individual's specific physiological deficit pattern. Furthermore, a strategy path that is dynamically adjusted as the rehabilitation process progresses is pre-set, achieving mechanistic customization and forward-looking planning of the intervention strategy. Attached Figure Description

[0069] Figure 1 This is a timeline diagram of the dysphagia risk prediction and personalized intervention recommendation system for stroke patients described in this invention.

[0070] Figure 2 A flowchart for analyzing and fragmenting physiological signal rhythm perturbations;

[0071] Figure 3 A flowchart for constructing a feature trajectory of swallowing function vulnerability;

[0072] Figure 4 A risk confidence distribution map for each layer of a cascaded decision forest;

[0073] Figure 5 This is a graph illustrating the index analysis of the synergistic breakdown of physiological signal fragments. Detailed Implementation

[0074] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0075] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0076] See Figure 1 The multimodal synchronous acquisition unit, guided by a specific swallowing protocol, captures the raw multimodal physiological signal stream of stroke patients performing standard swallowing actions. The rhythmic perturbation analysis unit analyzes the quasi-periodic rhythmic components in the raw multimodal physiological signal stream and fragments the signal according to rhythmic perturbation patterns, generating a set of physiological signal segments carrying rhythmic labels. The collaborative disintegration quantification unit quantifies the collaborative disintegration index of each segment in the physiological signal segment set. The trajectory construction and mapping unit reassembles physiological signal segments conforming to specific disintegration patterns into feature trajectories reflecting swallowing function vulnerability and maps these feature trajectories to anomalous attractors in a high-dimensional phase space. The geometric feature extraction unit extracts the geometric invariants of the anomalous attractors as potential risk features of swallowing disorders. The risk stratification decision unit inputs the potential risk features into a risk stratification model based on a cascaded decision forest to obtain individualized risk prediction labels. The personalized treatment plan generation unit, based on the risk prediction labels, triggers a multi-layered progressive retrieval of a structured intervention rule base to generate a time-adaptive personalized intervention plan.

[0077] In one embodiment of the present invention, see [reference] Figure 2A standardized assessment process was configured, including swallowing tasks involving foods of varying consistencies. Simultaneously, an EEG acquisition device, a laryngeal and cervical electromyography (EMG) sensor array, and a swallowing sound intensity detector were activated. The EEG acquisition device recorded the cortical potential fluctuation sequence in specific brain regions before and after the swallowing command was issued. The laryngeal and cervical EMG sensor array captured the EMG activation waveform sequence of the suprahyoid and pharyngeal constriction muscles. The swallowing sound intensity detector acquired the sound vibration signal sequence generated by the swallowing action. The cortical potential fluctuation sequence, EMG activation waveform sequence, and sound vibration signal sequence were stamped and aligned according to a unified time reference, integrating them to form a raw multimodal physiological signal stream. An adaptive spectral estimation algorithm was applied to the raw multimodal physiological signal stream to identify rhythmic frequency bands with significant power. Quasi-periodic oscillation signal components driven by the swallowing center pattern generator were separated from the rhythmic frequency bands. The instantaneous frequency jump points and amplitude collapse points of the quasi-periodic oscillation signal components were detected, and the signal interval between every two adjacent collapse points was defined as a rhythmic unit. Based on the frequency stability and amplitude decay rate of the rhythm units, each rhythm unit is assigned a rhythm tag reflecting the severity of its perturbation. All rhythm units carrying rhythm tags are then collected to form a set of physiological signal segments.

[0078] In practice, the standardized swallowing assessment process is executed concurrently with the simultaneous acquisition of multimodal physiological signals. The acquired raw multimodal physiological signal stream is then analyzed for rhythmic components and fragmented. An exemplary scenario involves assessing the swallowing function of a stroke patient in recovery. The assessment process involves sequentially swallowing foods of different consistencies, specifically 5 ml of a pureed food, 150 ml of a pureed food, and 5 ml of a liquid. Each swallowing task begins after a clear auditory or visual instruction. During this standardized assessment process, the EEG acquisition device, the laryngeal and cervical electromyography (EMG) sensor array, and the swallowing sound intensity detector are simultaneously activated. The electrodes of the EEG acquisition device are placed on the scalp covering the primary motor and sensory cortices according to the international 10-20 system standard to record the cortical potential fluctuation sequence within a specific time window from the onset of the swallowing instruction to the completion of the swallowing action. The laryngeal and cervical electromyography (EMG) sensor array consists of multiple surface EMG sensors, which are precisely attached to the bilateral submental muscles, suprahyoid muscles, and the pharyngeal surface projection area to capture the EMG activation waveform sequence of swallowing-related muscles. The swallowing sound intensity detector is a highly sensitive neck contact microphone fixed next to the thyroid cartilage to collect the sound vibration signal sequence triggered by swallowing. It can be understood that at the beginning of data acquisition, the above three devices are injected with a synchronized timestamp by a unified time base generator, ensuring that the subsequent cortical potential fluctuation sequence, EMG activation waveform sequence, and sound vibration signal sequence can be precisely aligned according to a unified time base, thereby integrating to form a strictly synchronized original multimodal physiological signal stream.

[0079] In specific implementations, the original multimodal physiological signal stream obtained from the above integration undergoes rhythmic perturbation analysis. In some embodiments, an autoregressive model is used for adaptive spectral estimation to identify rhythmic frequency bands with significant power in the signal, such as specific frequency bands related to the swallowing center pattern generator rhythm that may be identified in electromyographic activation waveform sequences. From the identified rhythmic frequency bands, the quasi-periodic oscillation signal component driven by the swallowing center pattern generator is separated using bandpass filtering and Hilbert transform. Subsequently, the instantaneous phase and envelope of this quasi-periodic oscillation signal component are detected; points of abrupt change in instantaneous frequency are marked as frequency jump points, and points of rapid decrease in signal envelope are marked as amplitude collapse points. It can be understood that the signal interval between every two adjacent amplitude collapse points is defined as a complete rhythmic unit. Based on the characteristics of the rhythmic unit, a rhythmic label reflecting the severity of the perturbation is assigned to it. For example, the assignment of a rhythmic label can be calculated using a quantization formula based on frequency stability and amplitude decay rate.

[0080]

[0081] in: The numerical value representing the assigned rhythm label. The reciprocal of the standard deviation of the instantaneous frequency within this rhythmic unit. The absolute value of the slope of the linear fit representing the amplitude envelope of the rhythmic unit. and These are preset weighting coefficients used to balance the contributions of frequency stability and amplitude decay rate in rhythm tag evaluation. In practice, all rhythm units that have undergone the above processing and carry rhythm tags are collected to form a set of physiological signal segments for subsequent analysis.

[0082] In some embodiments, for rhythm analysis of cortical potential fluctuation sequences, the adaptive spectral estimation algorithm may focus on the μ-rhythm and β-rhythm frequency bands related to motor preparation and execution. The separated quasi-periodic oscillation signal components reflect the cortical modulation rhythm of swallowing commands. Optionally, when detecting amplitude collapse points, the algorithm sets a dynamic threshold, which is dynamically calculated based on the local mean and standard deviation of the signal envelope. Only when the envelope decreases beyond this dynamic threshold and the rate of decrease reaches a preset threshold is it confirmed as a valid amplitude collapse point. In specific implementations, for sound vibration signal sequences, rhythm analysis may focus on the quasi-periodic fluctuations in their amplitude envelope that reflect the transitions between swallowing phases. The corresponding amplitude collapse points may correspond to abnormal interruptions in swallowing phase transitions. Optionally, the numerical range of the rhythm label can be preset from 0 to 1, where a higher value indicates a milder rhythmic disturbance in the rhythm unit, and a lower value indicates a more severe rhythmic disturbance, providing a grading basis for subsequent analysis.

[0083] In one embodiment of the present invention, see [reference] Figure 3 For each physiological signal segment in the set of physiological signal segments, the phase-lock value between its cortical potential fluctuation sequence and electromyographic activation waveform sequence is calculated. The activation sequence disorder degree between each channel of its electromyographic activation waveform sequence is calculated. The spectral distortion distance between the time spectrum of its sound vibration signal sequence and the normal swallowing template spectrum is calculated. The phase-lock value, activation sequence disorder degree, and spectral distortion distance are weighted and fused to generate a comprehensive collaborative disintegration index value. When the collaborative disintegration index value exceeds a preset disintegration threshold, the physiological signal segment is determined to conform to a specific disintegration pattern. The collaborative disintegration index values ​​of all physiological signal segments conforming to the specific disintegration pattern are extracted in chronological order to form a disintegration index sequence that evolves over time. The disintegration index sequence is smoothed and interpolated to construct a continuous feature trajectory curve. An adaptive kernel density estimation algorithm is used to model the probability density distribution of discrete data points in the disintegration index sequence, generating a local density estimate for each data point. Based on the local density estimate, a Gaussian process regression model is constructed, using the timestamp of the disintegration index sequence as the input variable and the disintegration index value as the output variable. A Gaussian process regression model is used to calculate the posterior predicted distribution at any time point, generating a smooth mean function curve. A cubic spline interpolation algorithm is then used to sample the mean function curve at equal intervals, yielding a characteristic trajectory curve with continuous first and second derivatives. The characteristic trajectory curve is normalized to map its numerical range to zero and one, facilitating subsequent phase space mapping operations. The curvature extrema and inflection points of the characteristic trajectory curve are calculated and used as state representation points of the characteristic trajectory.

[0084] In practice, a set of physiological signal segments from a subject is analyzed. This set contains 20 consecutive rhythmic unit segments, each containing an aligned sequence of cortical potential fluctuations, an electromyographic activation waveform sequence, and a sound vibration signal sequence. Specifically, for the first physiological signal segment, the phase-locking value between its cortical potential fluctuation sequence and the electromyographic activation waveform sequence is calculated. The phase-locking value can be calculated by using Hilbert transform to extract the instantaneous phase of the two signals, followed by calculating the ring statistic of the phase difference sequence. The degree of activation sequence disorder among the channels of its electromyographic activation waveform sequence is calculated. This disorder can be quantified by comparing the temporal order of the peak appearance of the electromyographic signal envelope in each channel with the standard swallowing muscle activation order. The spectral distortion distance between the time spectrum of its sound vibration signal sequence and a pre-stored normal swallowing template spectrum is calculated. This spectral distortion distance can be measured using a dynamic time warping algorithm or by calculating a correlation coefficient. It is understandable that a comprehensive collaborative disintegration index value can be generated by weighting and fusing the phase-locked value, activation sequence disorder degree, and spectral distortion distance obtained from the above calculations. The formula for weighted fusion can be expressed as follows:

[0085]

[0086] in: The numerical value represents the index of the breakdown of synergy. Represents the phase-locked loop value. Represents the degree of disorder in the activation sequence. Represents the spectral distortion distance. , , The weighting coefficients are preset and satisfy the following conditions: In practice, when the calculated collaborative disintegration index value exceeds a preset disintegration threshold, the current physiological signal segment is determined to conform to a specific disintegration pattern. This process iterates through all segments in the physiological signal segment set, assuming that ultimately 8 segments are determined to conform to the specific disintegration pattern.

[0087] In practical implementation, the collaborative disintegration index values ​​of eight physiological signal segments conforming to a specific disintegration pattern are extracted in chronological order to form a disintegration index sequence. In some embodiments, the disintegration index sequence is discrete in time, with each point corresponding to the median time of a segment and its collaborative disintegration index value. The disintegration index sequence is smoothed and interpolated to construct a continuous feature trajectory curve. An adaptive kernel density estimation algorithm is used to model the probability density distribution of discrete data points in the disintegration index sequence, calculating the local density estimate based on the distance between each data point and its neighbors. It can be understood that a Gaussian process regression model is constructed based on the local density estimate, using the timestamp of the disintegration index sequence as the input variable and the disintegration index value as the output variable. The local density estimate is used to adjust the noise level or weight of each data point in the Gaussian process regression model; regions with higher density are considered to have higher reliability. The posterior predicted distribution at any time point is calculated using the Gaussian process regression model, generating a smooth mean function curve as the initial trajectory. A cubic spline interpolation algorithm is used to sample the mean function curve at equal intervals to obtain a feature trajectory curve with continuous first and second derivatives. In practice, the feature trajectory curve is normalized by using a min-max normalization method to map the numerical range of the curve to between zero and one. Finally, the curvature of the normalized feature trajectory curve is calculated, and the local maximum points of curvature are identified as curvature extrema, and the points where the sign of curvature changes are identified as inflection points. These curvature extrema and inflection points are defined as the state representation points of the feature trajectory.

[0088] In some embodiments, when calculating the degree of disorder in the activation sequence, the standard swallowing muscle activation sequence can be defined as the submental muscles activating before the suprahyoid muscles, and the suprahyoid muscles activating before the pharyngeal muscles. The degree of disorder is quantified by calculating the Kendall's rank correlation coefficient between the actual peak time series and the standard sequence. Optionally, a preset weighting coefficient can be used. , , The values ​​can be set to 0.5, 0.3, and 0.2 respectively to emphasize the relative importance of cortical-muscle coupling in synergy assessment. In practice, the bandwidth parameter in the adaptive kernel density estimation algorithm can be adaptively adjusted according to the dispersion of the disintegration index sequence. A larger bandwidth is used when the dispersion is large to obtain a smooth density estimate, while a smaller bandwidth is used when the dispersion is small to preserve details. Optionally, the Gaussian process regression model can choose a squared exponential kernel as the covariance function, and the model's hyperparameters are optimized by maximizing the marginal likelihood function.

[0089] In one embodiment of the present invention, the state characterization points of the feature trajectory are used as reference points for delayed reconstruction. A time-delay embedding method is employed to reconstruct a set of point clouds in a high-dimensional phase space from the one-dimensional feature trajectory. Density clustering is performed on the point clouds in the high-dimensional phase space to identify densely distributed and stable regions. A density-based noise-applied spatial clustering algorithm is used to perform density reachability analysis on the point clouds in the high-dimensional phase space, calculating the density connectivity of each point within its neighborhood radius. A hierarchical density clustering method is used to aggregate mutually density-reachable points into initial clusters, and isolated point clouds formed by noise points are removed. The stability of each initial cluster is evaluated by calculating the centroid movement variance and the density index of points within the cluster in the phase space. Stable clusters with centroid movement variance below the variance threshold and point density within the cluster above the density threshold are selected. The centroid of the stable cluster is used as the core anchor point of the abnormal attractor, and a densely distributed and stable region of point cloud is formed by expanding around the anchor point. This region is defined as the abnormal attractor characterizing the abnormal dynamics of swallowing function.

[0090] In practice, the feature trajectory is a normalized curve containing 150 time points, from which 5 curvature extrema and 3 inflection points are identified as state representation points. Using these state representation points as the reference points for delayed reconstruction, a time-delay embedding method is employed to reconstruct the one-dimensional feature trajectory into a set of point clouds in a high-dimensional phase space. Time-delay embedding requires determining the embedding dimension. and time delay For example, the first local minimum can be determined using mutual information as a time delay. The embedding dimension was determined using the spurious nearest neighbor method. Assumptions determined One sampling point, For each point on the characteristic trajectory, Its corresponding point in high-dimensional phase space It consists of the following coordinate vectors: Traversing all available points on the feature trajectory generates a point cloud set containing 145 3D points. It can be understood that the essence of the time-delay embedding method is to reconstruct the phase space geometry of the entire dynamical system using historical information from a single observed variable.

[0091] In practical implementation, density clustering is performed on point clouds in high-dimensional phase space, and a density-based noise-based applied spatial clustering algorithm is used to analyze the density reachability of the point clouds. The density-based noise-based applied spatial clustering algorithm requires setting a neighborhood radius. and minimum points For example, setting (Under normalized phase space distance) and Density-based noise-based spatial clustering algorithms are used to calculate the density of each point within its range. The number of neighbors within the neighborhood radius, if the number of neighbors of a point is greater than or equal to If a point is density-reachable, it is marked as a core point. Hierarchical density clustering is used to aggregate mutually density-reachable points into initial clusters. Specifically, starting from a core point, all density-reachable points (including other core points and boundary points) are found and grouped into the same cluster. This process is repeated until all core points have been visited. Points not assigned to any cluster are marked as noise points and removed. In some embodiments, the stability of each generated initial cluster is evaluated by calculating the centroid movement variance and the cluster density index in phase space. The centroid movement variance is obtained by randomly dividing the cluster points into subsets, calculating the centroid positions of each subset, and then calculating the variance of these centroid positions relative to the overall cluster centroid. The cluster density index can be quantified by calculating the reciprocal of the average distance from all points within the cluster to their centroids, using the formula:

[0092]

[0093] in: This represents an index indicating the density of point distribution within a cluster. This represents the number of points in the cluster. Represents the first in the cluster The coordinate vector of each point The coordinate vector representing the centroid of this cluster. This represents the Euclidean norm. In practice, stable clusters are selected based on their centroid movement variance being below a preset variance threshold and their intra-cluster point density index being above a preset density threshold. Assume that after selection, two clusters meeting the stability criteria are identified from the point cloud. The centroids of these two stable clusters are used as the core anchor points of the anomalous attractor. Centered on each anchor point, [the following is a process / method / mechanism] is used... All points within the neighborhood are included, thus expanding to form a dense and stable point cloud distribution area.

[0094] In some embodiments, time delay Alternatively, the autocorrelation function can be used to decay the value to its initial value. The corresponding delay. Optionally, in the aggregation process, hierarchical density clustering methods define the boundary points as belonging to the cluster of the first core point that discovered them. In specific implementations, when calculating the centroid movement variance, the subset partitioning can use a bootstrap method for multiple random sampling to obtain a stable variance estimate. It can be understood that the parameters of density-based noise-based spatial clustering algorithms... and The adjustment needs to be made based on the specific distribution scale of the phase space point cloud. In some cases where the data distribution is relatively sparse, it may be appropriate to increase the scale. or reduce To capture meaningful cluster structures.

[0095] In one embodiment of the invention, the correlation dimension of the anomalous attractor is calculated to measure its structural complexity. The Lyapunov exponent spectrum of the anomalous attractor is calculated to quantify its dynamic instability. The variance distribution ratio of the anomalous attractor along its principal axis in phase space is calculated to describe its morphological anisotropy. The correlation dimension, Lyapunov exponent spectrum, and variance distribution ratio are used together as potential risk features. The potential risk features are input into a risk stratification model based on cascaded decision forests to obtain individualized risk prediction labels. The risk stratification model based on cascaded decision forests consists of multiple sequentially connected decision forests, each focusing on assessing risk in a specific dimension. The output confidence vector of the previous decision forest is used as part of the input features of the next decision forest. The final decision forest output layer produces a multidimensional risk prediction label, which simultaneously encodes the obstacle level, the main physiological dimensions of damage, and the risk evolution trend.

[0096] In practical implementation, the correlation dimension of the anomalous attractor is calculated to measure its structural complexity. The correlation dimension reflects the fractal characteristics of the attractor's point distribution. It is estimated by calculating the slope of the logarithm of the number of point pairs with a distance less than a certain radius in the phase space versus the logarithm of the radius. The correlation dimension value is obtained by fitting the slope of a linear region in a double logarithmic coordinate system. The Lyapunov exponent spectrum of the anomalous attractor is calculated to quantify its dynamic instability. The Lyapunov exponent spectrum describes the average divergence or convergence rate of neighboring trajectories in the phase space. For a phase space with an embedding dimension of m, m Lyapunov exponents can be estimated. The calculation of the maximum Lyapunov exponent can be achieved by tracking the evolution of neighboring point pairs in the phase space over time using the Wolf algorithm. If the maximum Lyapunov exponent is positive, it indicates that the system has chaotic characteristics. To describe the morphological anisotropy of anomalous attractors, the variance distribution ratio along their principal axes in phase space is calculated. First, principal component analysis is performed on the anomalous attractor point cloud to obtain eigenvectors (principal axes) and their corresponding eigenvalues ​​(variance). The variance distribution ratio can be quantified by the ratio of the largest to the smallest eigenvalue, expressed by the following formula:

[0097]

[0098] in: Represents the variance distribution ratio, This represents the largest eigenvalue obtained after principal component analysis. This represents the smallest eigenvalue obtained after principal component analysis. It can be understood that the correlation dimension, Lyapunov index spectrum, and variance distribution ratio are collectively used as potential risk features. These features constitute a multi-dimensional characterization of the dynamics of swallowing dysfunction. In some cases, refer to Table 1 for the calculated values ​​of geometric invariants.

[0099] Table 1: Calculated values ​​of geometric invariants for example anomalous attractors

[0100]

[0101] In practice, potential risk features are input into a risk stratification model based on cascaded decision forests to obtain individualized risk prediction labels. This model consists of multiple sequentially connected decision forests, each focusing on assessing a specific dimension of risk. For example, the first forest might focus on assessing risk levels based on structural complexity, the second on risk dimensions based on dynamic instability, and the third on risk dimensions based on morphological heterogeneity. The output confidence vector of the preceding decision forest serves as part of the input features for the following decision forest. For instance, after processing the input features (association dimension, Lyapunov index spectrum, variance distribution ratio), the first decision forest outputs not only a class prediction but also a confidence vector belonging to each risk level. This confidence vector is concatenated with the original geometric invariant features and used as input to the second decision forest. It is understandable that the final output layer of the decision forest will generate a multi-dimensional risk prediction label. The risk prediction label contains the encoding of the obstacle level, the main physiological dimension of damage, and the risk evolution trend. For example, a specific risk prediction label may be encoded as a triple (obstacle level: 2, main physiological dimension of damage: [cortical-muscle coupling], risk evolution trend: rising). The obstacle level can be represented by an integer to indicate the severity, the main physiological dimension of damage can be represented by binary encoding to indicate the damage status of multiple physiological dimensions, and the risk evolution trend can be represented by classification labels such as "stable", "rising", and "falling".

[0102] In some embodiments, the calculation of the Lyapunov index spectrum may be performed on the entire time series of the anomalous attractor point cloud, requiring the time series to be long enough to meet the algorithm's stability requirements. Optionally, the anomalous attractor point cloud can be centered before principal component analysis, i.e., the mean vector of the point cloud is subtracted to ensure that the principal axis passes through the centroid of the point cloud. In specific implementations, each decision forest in the cascaded decision forest can consist of multiple decision trees, for example, each forest contains 100 decision trees, and training is performed using the bagging method to improve generalization ability. In some embodiments, the obstacle level encoding in the risk prediction label may be divided into multiple levels, such as level 0 (no risk), level 1 (low risk), level 2 (medium risk), and level 3 (high risk), based on clinical standards. Optionally, the encoding of the main impaired physiological dimension may be a binary vector of length 3, corresponding to the three dimensions of "cortical rhythm," "muscle coordination," and "swallowing sound," with each bit being 1 indicating impairment in that dimension and 0 indicating normality. It is understandable that the generation of risk evolution trend encoding may rely on a comparison of geometric invariant features between the current assessment and historical assessments. If the correlation dimension or the maximum Lyapunov exponent shows an upward trend, it is encoded as "rising". In practice, the training of the cascaded decision forest requires the use of labeled clinical datasets, where each sample contains geometric invariant features and a true risk label assessed by experts.

[0103] See Figure 4 This is a risk confidence distribution chart for each layer of a cascaded decision forest. This grouped bar chart clearly shows the risk confidence output of the three-level cascaded decision forest under different assessment dimensions in the stroke dysphagia risk prediction system, reflecting the model's hierarchical and progressive risk assessment logic. As the decision level advances, the confidence of medium risk continuously rises from 0.40 to 0.48, indicating that the model's judgment of medium risk is continuously strengthened under multi-dimensional validation. This is highly consistent with the clinical characteristics that "dysphagia is mostly progressive, with medium risk accounting for the highest proportion." The risk-free confidence decreases from 0.12 in the first layer to 0.05 in the third layer, reflecting that the model continuously eliminates "false negatives" and reduces the risk of missed diagnosis through dynamic instability and morphological bias assessment in subsequent layers.

[0104] In one embodiment of the present invention, a multi-layered progressive retrieval of a structured intervention rule base is triggered based on risk prediction tags to generate a time-adaptive personalized intervention plan. The structured intervention rule base is organized in the form of a directed hypergraph, where nodes represent intervention measures or physiological states, and hyperedges represent multi-dimensional triggering and transformation rules. Based on the obstacle level code in the risk prediction tags, an initial cluster of intervention nodes in the intervention rule base is located. Based on the main damaged physiological dimension code in the risk prediction tags, a subset of targeted intervention measures is selected from the initial cluster of intervention nodes. Based on the risk evolution trend code in the risk prediction tags, an adjustment path for the execution intensity and duration of each measure in the subset of targeted intervention measures is dynamically planned. All intervention measures with dynamically adjusted paths are combined to form a time-adaptive personalized intervention plan.

[0105] In practical implementation, the implementation of the example relies on the risk prediction labels output by the risk stratification decision unit to trigger the retrieval and reasoning of the structured intervention rule base to generate personalized intervention plans. In a specific example scenario, the risk prediction labels output by the risk stratification decision unit are (Disorder Level: 2, Main Affected Physiological Dimension: [Corticomuscular Coupling], Risk Evolution Trend: Increasing). In practical implementation, the structured intervention rule base is organized in the form of a directed hypergraph. The nodes of the directed hypergraph represent specific intervention measures or physiological states, and the hyperedges of the directed hypergraph represent multi-dimensional triggering and transformation rules. A hyperedge can connect a predecessor node set and a successor node set, indicating that when all the conditions defined by the predecessor node set are met, the intervention measures or state transitions defined by the successor node set will be activated or recommended. It can be understood that the construction of the directed hypergraph is based on the domain knowledge base and clinical guidelines, formally representing intervention measures, physiological states, and their complex logical relationships.

[0106] In specific implementation, based on the obstacle level code in the risk prediction label, the initial intervention node cluster in the intervention rule base is located. The obstacle level code "2" corresponds to the preset "moderate swallowing disorder intervention cluster" in the rule base. This initial intervention node cluster may include a series of intervention measures such as "basic compensatory strategy teaching," "moderate intensity pharyngeal cold stimulation," and "neuromuscular electrical stimulation (parameter group A)," as well as physiological state nodes such as "moderate delayed pharyngeal initiation." Based on the main damaged physiological dimension code in the risk prediction label, a subset of targeted intervention measures is selected from the initial intervention node cluster. The main damaged physiological dimension code "[cortical-muscle coupling]" indicates that intervention measures aimed at improving the synergistic function between the cortex and muscles need to be selected. In some embodiments, by traversing the attribute labels of all intervention measure nodes in the "moderate swallowing disorder intervention cluster," nodes containing the labels "enhancing cortical-muscle coupling" or "improving neural drive" are matched. For example, nodes such as "transcranial magnetic stimulation at a specific frequency" and "electromyokinetic biofeedback training combined with motor imagery" are selected to form a subset of targeted intervention measures.

[0107] In practical implementation, based on the risk evolution trend code in the risk prediction label, the adjustment path for the execution intensity and duration of each measure in the targeted intervention measure subset is dynamically planned. An "rising" risk evolution trend code indicates that the current risk is worsening, and the intervention plan needs to be designed as an adaptive adjustment mode with increasing intensity or frequency. For example, for "electromyography biofeedback training combined with motor imagery" in the targeted intervention measure subset, the dynamic planning adjustment path is as follows: Week 1, once a day, 20 minutes each time, feedback threshold set to 70% of the baseline level; Week 2, once a day, 25 minutes each time, feedback threshold adjusted to 80% of the baseline level; Week 3, twice a day, 25 minutes each time, feedback threshold adjusted to 90% of the baseline level. The generation of the dynamic planning adjustment path relies on a predefined adjustment rule base, which associates the trend code with specific adjustment parameters. The formula for calculating the adjustment intensity can be expressed as:

[0108]

[0109] in: This represents the new intensity or parameter value of the plan. The intensity parameter represents the current cycle or the initial period. It is a gain coefficient related to the obstacle level. It is a numerical mapping of trend encoding. Optionally, the duration can be adjusted according to another independent rule, such as increasing the number of training sessions per week when the trend is "rising".

[0110] In some embodiments, node attributes in the directed hypergraph may include structured metadata in addition to text labels, such as the range of intervention intensity, a list of applicable obstacle levels, and a list of target physiological dimensions, to facilitate procedural screening. In practical implementation, when dynamically planning adjustment paths, not only trend coding is considered, but obstacle level coding may also be cross-referenced. For example, for high-risk situations with an upward trend, the adjustment step size and frequency increment will be larger. All interventions with dynamically adjusted paths are combined to form a time-adaptive personalized intervention plan. The combination process generates a structured plan document containing the specific content of each intervention, initial execution parameters, and a parameter adjustment plan arranged chronologically. It can be understood that the generated personalized intervention plan will serve as a clinical recommendation output, guiding rehabilitation therapists to conduct phased and moduloable interventions for patients.

[0111] See Figure 5This is a bar chart analyzing the synergistic disintegration index of physiological signal segments. The bar chart displays the synergistic disintegration index values ​​for eight physiological signal segments, with a red dashed line marking the disintegration threshold (0.6), used to determine whether swallowing function is abnormal. The synergistic disintegration indices for segments 4, 5, and 6 all exceed the threshold of 0.6, indicating a significant disintegration of the synergistic relationship between cortical potentials, electromyographic signals, and swallowing sounds during this continuous time period, suggesting an abnormality in swallowing function. The index values ​​of abnormal segments show a trend of "first rising and then falling," peaking at segment 5 (0.82), reflecting the most severe degree of synergistic disintegration of swallowing function at this moment. It subsequently eased but remained within the abnormal range. The synergistic disintegration index is a core indicator for quantifying the vulnerability of swallowing function. Segments exceeding the threshold are reconstructed into characteristic trajectories, further mapped to a high-dimensional phase space to identify abnormal attractors, ultimately providing a basis for risk stratification.

[0112] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A system for predicting the risk of dysphagia in stroke patients and recommending personalized interventions, characterized in that: The system includes: The multimodal synchronous acquisition unit is used to capture the raw multimodal physiological signal stream when a stroke patient performs standard swallowing actions under the guidance of a specific swallowing protocol. The rhythm perturbation parsing unit is used to parse the quasi-periodic rhythm components in the original multimodal physiological signal stream, and to segment the signal according to the rhythm perturbation mode to generate a set of physiological signal segments carrying rhythm tags. A collaborative disintegration quantification unit is used to quantify the collaborative disintegration index of each segment in the set of physiological signal segments; The trajectory construction and mapping unit is used to reconstruct physiological signal fragments that conform to a specific disintegration pattern into characteristic trajectories that reflect the vulnerability of swallowing function, and to map the characteristic trajectories to abnormal attractors in a high-dimensional phase space; A geometric feature extraction unit is used to extract the geometric invariants of the abnormal attractor as potential risk features of dysphagia. The risk stratification decision unit is used to input the potential risk characteristics into the risk stratification model based on cascaded decision forest to obtain individualized risk prediction labels. The personalized solution generation unit is used to trigger a multi-level progressive retrieval of the structured intervention rule base based on the risk prediction tags, and generate a personalized intervention solution with time-adaptive characteristics.

2. The system for predicting the risk of dysphagia in stroke patients and recommending personalized interventions according to claim 1, characterized in that, The capture of raw multimodal physiological signal streams during standard swallowing maneuvers performed by stroke patients under the guidance of a specific swallowing protocol specifically includes: Configure a standardized assessment process that includes swallowing tasks with foods of different textures, and simultaneously activate the cortical EEG acquisition device, the laryngeal and cervical electromyography sensor array, and the swallowing sound intensity detector. The cortical electroencephalogram (EEG) acquisition device records the cortical potential fluctuation sequence in specific brain regions before and after the swallowing command is issued; The laryngeal and cervical electromyography sensor array captures the electromyographic activation waveform sequence of the suprahyoid muscle group and the pharyngeal constriction muscle group. The swallowing sound intensity detector collects the sound vibration signal sequence generated by the swallowing action; The cortical potential fluctuation sequence, electromyographic activation waveform sequence, and sound vibration signal sequence are stamped and aligned according to a unified time reference, and integrated to form the original multimodal physiological signal stream.

3. The system for predicting the risk of dysphagia in stroke patients and recommending personalized interventions according to claim 2, characterized in that, The process of analyzing the quasi-periodic rhythmic components in the original multimodal physiological signal stream and fragmenting the signal according to the rhythmic perturbation pattern includes: An adaptive spectral estimation algorithm is applied to the original multimodal physiological signal stream to identify rhythmic frequency bands with significant power in the signal; The quasi-periodic oscillation signal component driven by the swallowing center pattern generator is separated from the rhythm frequency band; The instantaneous frequency jump point and amplitude collapse point of the quasi-periodic oscillation signal component are detected, and the signal interval between every two adjacent collapse points is defined as a rhythm unit. Based on the frequency stability and amplitude decay rate of the rhythm unit, a rhythm tag reflecting the severity of its disturbance is assigned to each rhythm unit; All rhythm units carrying rhythm tags are gathered together to form the set of physiological signal segments.

4. The system for predicting the risk of dysphagia in stroke patients and recommending personalized interventions according to claim 3, characterized in that, The quantification index for the cooperative disintegration of each segment in the set of physiological signal segments includes: For each physiological signal segment in the set of physiological signal segments, the phase-locking value between its cortical potential fluctuation sequence and electromyographic activation waveform sequence is calculated. Calculate the degree of disorder in the activation sequence among the channels of its electromyographic activation waveform sequence; Calculate the spectral distortion distance between the time spectrum of its sound vibration signal sequence and the spectrum of the normal swallowing template; The phase-locked value, activation order disorder degree, and spectral distortion distance are weighted and fused to generate a comprehensive collaborative disintegration index value; When the value of the collaborative disintegration index exceeds the preset disintegration threshold, the physiological signal fragment is determined to conform to the specific disintegration pattern.

5. The system for predicting the risk of dysphagia in stroke patients and recommending personalized interventions according to claim 4, characterized in that, The reconstructing of physiological signal fragments conforming to specific disintegration patterns into characteristic trajectories reflecting swallowing function vulnerability includes: The synergistic disintegration index values ​​of all physiological signal fragments that conform to a specific disintegration pattern are extracted in chronological order to form a disintegration index sequence that evolves over time. The disintegration index sequence is smoothed by interpolation to construct a continuous feature trajectory curve; Calculate the curvature extrema and inflection points of the characteristic trajectory curve, and use the curvature extrema and inflection points as state representation points of the characteristic trajectory; The step of smoothing and interpolating the disintegration index sequence to construct a continuous feature trajectory curve specifically includes: An adaptive kernel density estimation algorithm is used to model the probability density distribution of discrete data points in the collapse index sequence, generating a local density estimate for each data point. A Gaussian process regression model is constructed based on the local density estimate, with the timestamps of the disintegration index sequence as input variables and the values ​​of the disintegration indexes as output variables. The posterior predicted distribution at any time point is calculated using the Gaussian process regression model, generating a smooth mean function curve. The mean function curve is sampled at equal intervals using a cubic spline interpolation algorithm to obtain a characteristic trajectory curve with continuous first and second derivatives; The characteristic trajectory curve is normalized so that its numerical range is mapped to between zero and one, which facilitates subsequent phase space mapping operations.

6. The system for predicting the risk of dysphagia in stroke patients and recommending personalized interventions according to claim 5, characterized in that, The step of mapping the feature trajectory to an anomalous attractor in a high-dimensional phase space includes: Using the state representation points of the feature trajectory as the reference points for delayed reconstruction, the one-dimensional feature trajectory is reconstructed into a set of point clouds in a high-dimensional phase space using the time delay embedding method. Density clustering is performed on the point cloud in the high-dimensional phase space to identify regions where the point cloud distribution is dense and stable. The region is defined as an abnormal attractor that characterizes the dynamics of abnormal swallowing function.

7. The system for predicting the risk of dysphagia in stroke patients and recommending personalized interventions according to claim 6, characterized in that, The extraction of geometric invariants of the abnormal attractors as potential risk features of dysphagia includes: Calculate the correlation dimension of the anomalous attractor to measure its structural complexity; Calculate the Lyapunov exponent spectrum of the anomalous attractor to quantify its dynamic instability; The variance distribution ratio of the anomalous attractor along its principal axis in phase space is calculated to describe its morphological anisotropy. The correlation dimension, Lyapunov index spectrum, and variance distribution ratio are used together as the potential risk features.

8. The system for predicting the risk of dysphagia in stroke patients and recommending personalized interventions according to claim 1, characterized in that, The step of inputting the potential risk characteristics into a risk hierarchical model based on cascaded decision forests to obtain individualized risk prediction labels includes: The risk hierarchical model based on cascaded decision forests consists of multiple sequentially connected decision forests, each of which focuses on assessing risk in a specific dimension. The output confidence vector of the previous decision forest will be used as part of the input features of the next decision forest; The final decision forest output layer will generate a multi-dimensional risk prediction label, which simultaneously encodes the obstacle level, the main physiological dimensions of damage, and the risk evolution trend.

9. The system for predicting the risk of dysphagia in stroke patients and recommending personalized interventions according to claim 8, characterized in that, Based on the risk prediction tags, a multi-layered progressive retrieval of the structured intervention rule base is triggered to generate a personalized intervention plan with time-adaptive characteristics, including: The structured intervention rule base is organized in the form of a directed hypergraph, where nodes represent intervention measures or physiological states, and hyperedges represent multidimensional triggering and transformation rules. Based on the obstacle level code in the risk prediction label, locate the initial intervention node cluster in the intervention rule base; Based on the coding of the main damaged physiological dimensions in the risk prediction labels, a subset of targeted intervention measures is selected from the initial intervention node cluster; Based on the risk evolution trend code in the risk prediction label, the adjustment path of the execution intensity and duration of each measure in the subset of targeted intervention measures is dynamically planned; All interventions with dynamically adjustable pathways are combined to form the time-adaptive personalized intervention program.

10. The risk prediction and personalized intervention recommendation system for dysphagia in stroke patients according to claim 6, characterized in that, The step of performing density clustering on the point cloud in the high-dimensional phase space to identify densely distributed and stable regions of the point cloud includes: A density-based noise-based spatial clustering algorithm is used to perform density reachability analysis on the point cloud in the high-dimensional phase space, and the density connectivity of each point within its neighborhood radius is calculated. Hierarchical density clustering method is used to aggregate mutually density-reachable points into initial clusters and remove isolated point clouds formed by noise points; Stability assessment is performed on each initial cluster, and the centroid shift variance and intra-cluster point density index of the cluster in phase space are calculated. Stable clusters are selected based on centroid movement variance being lower than the variance threshold and intra-cluster point density being higher than the density threshold. The centroid of the stable cluster is used as the core anchor point of the anomalous attractor, and the region with dense and stable point cloud distribution is expanded around the anchor point.

Images