Postoperative cognitive quantitative evaluation system and method based on fusion of electroencephalogram and near-infrared spectrum
By constructing a cross-modal manifold matrix and generating topological feature vectors, combined with a closed-loop intervention mechanism, the problem of pathological feature loss caused by neurovascular coupling nonlinear distortion in traditional assessment systems was solved, thereby improving the accuracy of postoperative cognitive impairment identification and the effectiveness of intervention.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHANJIANG CENT PEOPLES HOSPITAL
- Filing Date
- 2026-06-08
- Publication Date
- 2026-07-03
AI Technical Summary
Existing technologies, when processing multimodal physiological signals such as EEG and near-infrared spectroscopy, cannot capture nonlinear topological distortions of neurovascular coupling by using linear splicing methods. This leads to the loss of deep pathological features and inaccurate assessment, which is especially prone to missed diagnosis or misjudgment of postoperative cognitive impairment in elderly patients.
By constructing a cross-modal manifold matrix, projecting the signal onto the tangent space using geodesic distance and logarithmic mapping, generating a topological feature vector, and combining it with the softmax function to output a continuous cognitive assessment index, a closed-loop intervention mechanism is established to reset the time window and ensure the accuracy of the assessment.
It improves upon the problem of lost deep pathological features in traditional assessment systems, enhances the early identification of postoperative cognitive impairment, reduces missed diagnoses and misjudgments, and ensures the accuracy and continuity of intervention.
Smart Images
Figure CN122320488A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of perioperative neurological monitoring technology, and in particular to a postoperative cognitive quantitative assessment system and method that integrates electroencephalography and near-infrared spectroscopy. Background Technology
[0002] Postoperative cognitive impairment (POCD) is a common and serious central nervous system complication in elderly patients after anesthesia and major surgery. Clinically, scalp electroencephalography (EEG) is typically used to track high-frequency neuronal electrical activity, supplemented by near-infrared spectroscopy to obtain low-frequency changes in local cerebral blood oxygen metabolism, in order to assess the neurovascular coupling status of the patient's brain. During the pathological evolution of POCD, the combined impact of anesthetic drug blockade and surgical trauma stress leads to a deep pathological decoupling of the neurovascular coupling mechanism in the patient's brain.
[0003] However, existing technologies suffer from a major technical deficiency when processing multimodal physiological signals such as EEG and near-infrared spectroscopy: most current systems perform mechanical linear splicing or shallow data fusion of EEG and near-infrared features within Euclidean geometry. This linear processing completely ignores the fact that the covariance matrix formed between cross-modal physiological features actually resides in a high-dimensional symmetric positive definite Riemannian manifold space. Traditional Euclidean distance and linear networks cannot resolve the nonlinear distortions and curvature collapses that occur in the high-dimensional topological space when neurovascular coupling mechanisms disintegrate. As a result, when facing elderly patients with significant differences in physiological baselines, the system inevitably loses deep decoupled pathological features, leading to missed diagnoses of early, subtle cognitive impairments or serious inaccuracies in state assessment. Summary of the Invention
[0004] To overcome the above shortcomings, this invention provides a postoperative cognitive quantitative assessment system and method that integrates electroencephalography (EEG) and near-infrared spectroscopy. This system aims to improve upon the problem that traditional multimodal assessments, which mostly rely on linear feature splicing, cannot capture the nonlinear topological distortions of neurovascular couplings, resulting in the loss of deep pathological features and inaccurate assessment.
[0005] In a first aspect, the present invention provides the following technical solution: a postoperative cognitive quantitative assessment system integrating electroencephalography (EEG) and near-infrared spectroscopy, comprising the following modules:
[0006] The baseline anchoring module receives preoperative scalp EEG signals and preoperative near-infrared spectral signals, extracts features and splices them into a first vector, and calculates the covariance of the first vector to generate the preoperative baseline manifold.
[0007] The matrix construction module receives real-time scalp EEG signals and real-time near-infrared spectral signals within a time window, extracts features and concatenates them into a second vector, and calculates the covariance of the second vector to generate a cross-modal manifold matrix.
[0008] The feature extraction module connects the baseline anchoring module and the matrix construction module, calculates the geodesic distance between the cross-modal manifold matrix and the preoperative baseline manifold, projects the cross-modal manifold matrix onto the tangent space of the preoperative baseline manifold using logarithmic mapping to generate a tangent matrix, and combines the tangent matrix with the geodesic distance to generate a topological feature vector.
[0009] The quantitative evaluation module is connected to the feature extraction module. It inputs the topological feature vector into the softmax function layer to output a probability distribution and calculates the cognitive evaluation index based on the probability distribution.
[0010] The closed-loop intervention module is connected to the quantitative assessment module. When the cognitive assessment index meets the preset conditions, it outputs an intervention instruction and feeds back the timestamp to the matrix construction module to reset the time window.
[0011] By adopting the above technical solution, the tangent space topological features of the cross-modal manifold matrix are extracted to calculate the evaluation index, and then the intervention command is output in a closed loop and the time window is reset. This improves the problem that traditional multimodal evaluation mostly uses linear feature splicing, which cannot capture the nonlinear topological distortion of neurovascular coupling, resulting in the loss of deep pathological features and inaccurate evaluation.
[0012] Secondly, this invention provides the following technical solution: a postoperative cognitive quantitative assessment method based on the fusion of electroencephalography (EEG) and near-infrared spectroscopy, comprising the following steps:
[0013] S1. Receive preoperative scalp EEG signals and preoperative near-infrared spectral signals, extract features and splice them into a first vector, calculate the covariance of the first vector to generate the preoperative baseline manifold;
[0014] S2. Receive real-time scalp EEG signals and real-time near-infrared spectral signals within the time window, extract features and splice them into a second vector, calculate the covariance of the second vector to generate a cross-modal manifold matrix;
[0015] S3. Calculate the geodesic distance between the transmodal manifold matrix and the preoperative baseline manifold, and use logarithmic mapping to project the transmodal manifold matrix onto the tangent space of the preoperative baseline manifold to generate a tangent matrix. Combine the tangent matrix with the geodesic distance to generate a topological feature vector.
[0016] S4. Input the topological feature vector into the softmax function layer to output the probability distribution, and calculate the cognitive evaluation index based on the probability distribution;
[0017] S5. When the cognitive assessment index meets the preset conditions, an intervention instruction is output, and the timestamp is fed back to the matrix construction module to reset the time window.
[0018] The present invention has the following beneficial effects:
[0019] 1. In this invention, the evaluation index is calculated by extracting the tangent space topological features of the cross-modal manifold matrix, and then the intervention command is output in a closed loop and the time window is reset. This improves the problem that traditional multimodal evaluation mostly uses linear feature splicing, which cannot capture the nonlinear topological distortion of neurovascular coupling, resulting in the loss of deep pathological features and inaccurate evaluation.
[0020] 2. In this invention, by multiplying the multi-class probability distribution with the corresponding severity penalty constant and summing the products, an evaluation index in the form of a continuous scalar is output. This improves the problem that traditional state assessments mostly use the maximum probability to output a unique discrete label, which is difficult to represent the continuous and gradual changes in physiological state, thus causing high-risk cases at the boundary to be easily missed.
[0021] 3. In this invention, by capturing the system timestamp that triggers the intervention command and feeding it back, the historical cache data is cleared and the time window is reinitialized. This improves the problem that traditional monitoring systems mostly use fixed time windows for continuous calculation, which causes mathematical aliasing of abnormal pathological data before and after the intervention, resulting in distorted subsequent evaluation of the intervention efficacy. Attached Figure Description
[0022] Figure 1 This is a diagram of the architecture of the postoperative cognitive quantitative assessment system that integrates electroencephalography and near-infrared spectroscopy proposed in this invention.
[0023] Figure 2 This invention presents a multimodal data low-level feature processing and evaluation control flow diagram for a postoperative cognitive quantitative assessment system that integrates EEG and near-infrared spectroscopy.
[0024] Figure 3 This is a flowchart of the postoperative cognitive quantitative assessment method based on the fusion of electroencephalography and near-infrared spectroscopy proposed in this invention. Detailed Implementation
[0025] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0026] Example 1:
[0027] In a first embodiment of the present invention, the present invention provides a postoperative cognitive quantitative assessment system that integrates electroencephalography (EEG) and near-infrared spectroscopy, such as... Figures 1-2 As shown, it includes the following steps:
[0028] The baseline anchoring module receives preoperative scalp EEG signals and preoperative near-infrared spectral signals, extracts features and splices them into a first vector, and calculates the covariance of the first vector to generate the preoperative baseline manifold.
[0029] Furthermore, the baseline anchoring module is used to perform the steps of extracting features and concatenating them into a first vector:
[0030] Extract the power spectral density sequence of preoperative scalp EEG signals in multiple preset frequency bands;
[0031] Extract the relative change sequences of oxyhemoglobin concentration and deoxyhemoglobin concentration from the preoperative near-infrared spectral signals;
[0032] The EEG power spectral density sequence, the relative change sequence of oxyhemoglobin concentration, and the relative change sequence of deoxyhemoglobin concentration were timestamped based on a unified clock synchronization benchmark.
[0033] The signal features of the same sampling point after time-stamp alignment are merged and spliced to construct the first vector.
[0034] The baseline anchoring module performs the steps of calculating the covariance of the first vector to generate the preoperative baseline manifold.
[0035] Obtain all first vectors contained within the first preset resting time period, and calculate the mean vector of all first vectors;
[0036] Calculate the centered difference vector between each first vector and the mean vector;
[0037] Perform matrix multiplication on the centered difference vector and its transpose to generate an intermediate feature matrix.
[0038] All intermediate feature matrices within the first preset resting time period are summed to generate a symmetric positive definite covariance matrix, and the symmetric positive definite covariance matrix is solidified into the preoperative baseline manifold.
[0039] Specifically, the input data consists of preoperative scalp electroencephalogram (EEG) signals and preoperative near-infrared spectral signals from the patient at rest. The output data is a symmetric positive definite preoperative baseline manifold. The data flow path is as follows: The underlying raw physiological signals are extracted using frequency band energy and blood oxygen concentration features, aligned and concatenated based on a unified clock source to generate a first vector. The discrete-time series containing this first vector undergoes mean calculation, centering difference operation, and matrix outer product summation, ultimately outputting the preoperative baseline manifold representing the individual's resting neurovascular coupling state.
[0040] When the baseline anchoring module performs feature concatenation, the specific mathematical expression of the first vector is:
[0041] ;
[0042] in Represents the first preset rest period. The first vector constructed at each sampling point. Representing the The EEG power spectral density sequence extracted at each sampling point. Representing the The combined sequence of the relative change sequence of oxyhemoglobin concentration and the relative change sequence of deoxyhemoglobin concentration extracted at each sampling point. This represents the dimension merging and concatenation operator. This represents the matrix transpose operator.
[0043] The specific formulas for calculating the intermediate feature matrix and covariance matrix when generating the preoperative baseline manifold are as follows:
[0044] ;
[0045] ;
[0046] in The mean vector of all first vectors. This represents the total number of discrete sampling points included within the first preset rest period, with a value set between 1000 and 5000. This represents the centered difference vector generated by subtracting the mean vector from the first vector. The transpose of the centered difference vector. This represents the symmetric positive definite covariance matrix generated by the final summation process, i.e., the preoperative baseline manifold.
[0047] By establishing a unified clock synchronization benchmark for processing EEG and near-infrared spectral signals, the sampling rate differences and physical delays inherent in heterogeneous physiological acquisition hardware are eliminated. The dimensional merging and splicing action of the first vector forces electrophysiological and hemodynamic features of different scales into the same high-dimensional vector space, providing a mathematical basis for joint computation of cross-modal data.
[0048] Extracting a centered difference vector from the first vector and calculating the covariance matrix essentially extracts the second-order statistical correlation between various physiological characteristic variables from the time series. This covariance operation transforms linear discrete physiological data into a symmetric positive definite manifold residing in a non-Euclidean geometric space. A preoperative cross-modal covariance baseline manifold is established for each individual using a baseline anchoring module, serving as the absolute origin coordinates for subsequent postoperative monitoring algorithms. This improves upon the traditional technical shortcomings of using uniform fixed judgment thresholds for state classification in conventional assessment methods, thereby avoiding the specific problems of false or missed triggering of postoperative abnormal states due to the patient's age range or underlying cerebrovascular diseases.
[0049] The matrix construction module receives real-time scalp EEG signals and real-time near-infrared spectral signals within a time window, extracts features and concatenates them into a second vector, and calculates the covariance of the second vector to generate a cross-modal manifold matrix.
[0050] Furthermore, the matrix construction module performs the steps of extracting features and concatenating them into a second vector, and calculating the covariance of the second vector to generate a cross-modal manifold matrix.
[0051] Within a real-time sliding time window, the real-time frequency band power spectral density of the real-time scalp EEG signal and the real-time hemoglobin concentration change of the real-time near-infrared spectral signal are extracted synchronously, and the real-time frequency band power spectral density and the real-time hemoglobin concentration change are combined and assembled into a second vector according to the time correspondence.
[0052] Extract all second vectors within the time window and calculate the real-time average vector;
[0053] The second vector within the time window is subtracted from the real-time average vector to generate the real-time residual vector.
[0054] The outer product operation is performed on all real-time residual vectors to generate a real-time updated symmetric positive definite matrix, and the real-time updated symmetric positive definite matrix is used as the cross-modal manifold matrix.
[0055] Specifically, the input data consists of scalp EEG signals and near-infrared spectral signals synchronously acquired within a preset sliding time window. The output data is a symmetric positive definite cross-modal manifold matrix. The data flow path is as follows: The module reads the original multimodal signals within the time window and extracts the frequency band power spectral density and hemoglobin concentration changes, combining them according to time correspondence to form a second vector sequence; the second vector sequence is then subjected to mean calculation and subtraction operations to obtain a real-time residual vector set; and the outer product operation is performed on all real-time residual vector sets to sum them, outputting a real-time updated cross-modal manifold matrix.
[0056] When the matrix construction module performs feature extraction and concatenation, the mathematical expression of the second vector is:
[0057] ;
[0058] in Represents the time window within which the real-time sliding occurs. The second vector is formed by combining and assembling the sampling points at each moment. Representing the The real-time frequency band power spectral density sequence extracted at each sampling point. Representing the The sequence of real-time changes in hemoglobin concentration extracted at each sampling point. This represents the assembly / combination operator. This represents the matrix transpose operator.
[0059] The specific formula for calculating the covariance of the second vector to generate the cross-modal manifold matrix is as follows:
[0060] ;
[0061] ;
[0062] in This represents the real-time average vector within the time window. This represents the total number of sampling points contained within the real-time sliding time window, with a value set between 500 and 2000. This represents the real-time residual vector generated by the subtraction operation. This represents the transpose of the real-time residual vector. This represents a real-time updated symmetric positive definite matrix generated by the summation of the outer product operation, i.e., a cross-modal manifold matrix.
[0063] The matrix construction module uses a real-time sliding time window as the data extraction boundary, continuously reading in the latest acquired multimodal physiological data. Scalp EEG signals reflect instantaneous changes in neuronal electrical activity, while near-infrared spectral signals reflect the hysteretic response of hemodynamics. By assembling heterogeneous data into a second vector according to temporal correspondence, the physical dimensional isolation of single signal modes is broken.
[0064] Subtraction is performed between the second vector and the real-time average vector to remove the inherent static DC bias component in the real-time signal sequence, retaining the dynamic fluctuation characteristics representing state changes. The outer product of all extracted real-time residual vectors is calculated and summed, essentially extracting the dynamic covariance between EEG activity and local cerebral blood flow metabolism within the current time window.
[0065] The cross-modal manifold matrix generated by the outer product summation step transforms one-dimensional time-series features into a non-Euclidean space, forming a symmetric positive definite topological structure. This matrix structure directly maps the tightness of neurovascular coupling mechanisms during the monitoring phase. Relying on the covariance matrix to carry multimodal fusion features, it eliminates data interference caused by drastic fluctuations in basic physiological indicators during postoperative stress, accurately captures the phase separation pathological features between heterogeneous signals, and overcomes the technical deficiency of traditional classification models that are prone to overfitting when directly receiving spliced time-series features.
[0066] The feature extraction module connects the baseline anchoring module and the matrix construction module. It calculates the geodesic distance between the cross-modal manifold matrix and the preoperative baseline manifold, and uses logarithmic mapping to project the cross-modal manifold matrix onto the tangent space of the preoperative baseline manifold to generate a tangent matrix. It then combines the tangent matrix with the geodesic distance to generate a topological feature vector.
[0067] Furthermore, the feature extraction module performs the step of calculating the geodesic distance between the cross-modal manifold matrix and the preoperative baseline manifold:
[0068] Calculate the inverse square root matrix of the preoperative baseline manifold;
[0069] The inverse square root matrix is multiplied by the cross-modal manifold matrix to generate the projection transition matrix.
[0070] The projection transition matrix is multiplied by the inverse square root matrix to generate the affine invariant Riemannian metric matrix.
[0071] Perform principal logarithm operation on the affine invariant Riemannian metric matrix, calculate the matrix norm of the result of the principal logarithm operation, and output the calculated matrix norm as geodesic distance.
[0072] The feature extraction module performs the step of projecting the cross-modal manifold matrix onto the tangent space of the preoperative baseline manifold using a logarithmic mapping to generate the tangent matrix.
[0073] The preoperative baseline manifold is set as a fixed Riemann tangent point in the manifold topological space, and the square root matrix of the fixed Riemann tangent point is calculated.
[0074] Multiply the square root matrix, the affine invariant Riemannian metric matrix after performing the principal logarithm operation, and the square root matrix in sequence.
[0075] The resulting matrix from the sequential multiplication operations is used as the set of eigenvectors projected onto the Euclidean geometric plane to generate the tangent matrix.
[0076] The feature extraction module is used to perform the step of generating topological feature vectors by combining the tangent matrix and the geodesic distance:
[0077] Retrieve all upper triangular matrix elements and diagonal matrix elements of the tangent matrix;
[0078] The elements of the upper triangular matrix and the diagonal matrix are arranged in a dimension-reduced manner to generate a one-dimensional structural feature sequence.
[0079] Extract each element from the one-dimensional structural feature sequence, and then multiply each element of the one-dimensional structural feature sequence with the geodesic distance for weighting.
[0080] The final topological feature vector is obtained by combining and arranging all the elements after multiplication and weighting.
[0081] Specifically, the input data consists of the preoperative baseline manifold and the cross-modal manifold matrix. The output data is the dimensionality-reduced topological feature vector. The data flow path is as follows: After reading the underlying manifold matrix, inverse square root decomposition and matrix multiplication are performed to generate an affine invariant Riemannian metric matrix. Then, geodesic distance data is output through a branch of principal logarithmic operations and norm solving. Another branch projects the tangent matrix by multiplying the square root matrix through logarithmic mapping. The tangent matrix undergoes upper triangular and diagonal element extraction and recombination, and is then multiplied and weighted by the geodesic distance. Finally, a topological feature vector adapted to Euclidean geometric space is output.
[0082] When calculating the geodesic distance in the feature extraction module, the specific calculation formula is defined as follows:
[0083] ;
[0084] in This represents the geodesic distance calculated and output. The inverse square root matrix represents the preoperative baseline manifold. The cross-modal manifold matrix represents the input to real-time computation. This represents the principal logarithm operator for matrices. The Frobenius norm solver for a matrix.
[0085] The specific formula for generating the tangent matrix using the logarithmic mapping is defined as follows:
[0086] ;
[0087] in This represents the tangent matrix generated by projecting onto the Euclidean geometric plane. The matrix representing the square root of a fixed Riemann tangent point.
[0088] The specific formula for extracting features and generating the final topological feature vector is defined as follows:
[0089] ;
[0090] in This represents the final topological feature vector output after multiplication and weighting. This represents a matrix vectorization operator used to extract all upper triangular matrix elements and diagonal matrix elements from a tangent matrix.
[0091] The covariance matrix formed by cross-modal physiological characteristics resides in a non-Euclidean topological space. Directly using conventional distance formulas to calculate state deviations can lead to spatial metric inflation and distort the true degree of physiological deviation. By performing a multiplication operation between the inverse square root matrix and the cross-modal manifold matrix, and calculating the principal log-norm within the Riemannian geometric framework, the shortest geodesic between two points on the manifold surface can be accurately found. The calculated geodesic distance physically quantifies the overall absolute deterioration of the patient's postoperative neurovascular decoupling relative to their preoperative resting state.
[0092] Since deep learning evaluation networks cannot directly analyze curved manifold data, the system establishes the preoperative baseline manifold as a fixed tangent point in space. Through logarithmic mapping, the distorted real-time cross-modal manifold matrix is locally expanded to the flat Euclidean tangent space containing this tangent point. This expansion process preserves the local spatial geometric topology of the covariance characteristics between heterogeneous physiological signals, transforming the incalculable curved surface matrix into a directly quantifiable tangent matrix.
[0093] The tangent matrix contains symmetric redundancy. Vectorization is used to extract the upper triangular and diagonal elements to construct a one-dimensional sequence, significantly reducing the computational burden on the evaluation network. Each element of the local structural features is multiplied by the geodesic distance representing the overall deterioration level, physically embedding the global deviation weights into the local spatial structural features. The final output topological feature vector bridges the dimensionality gap between high-dimensional non-Euclidean physiological representations and conventional classification models, resolving the technical deficiency of traditional systems that discard deep pathological features due to spatial incompatibility.
[0094] The quantitative evaluation module connects to the feature extraction module, inputs the topological feature vector into the softmax function layer to output the probability distribution, and calculates the cognitive evaluation index based on the probability distribution.
[0095] Furthermore, the quantification evaluation module is used to perform the steps of inputting the topological feature vector into the softmax function layer to output the probability distribution, and calculating the cognitive evaluation index based on the probability distribution:
[0096] The topological feature vector is multiplied by the classification weight matrix in the preset multi-classification model, and the category bias term is added to generate the initial logical output value.
[0097] The initial logic output value is input into the softmax function layer for exponential operation and normalization to generate probability distributions corresponding to multiple preset cognitive state categories;
[0098] Obtain the penalty quantization constant pre-configured for each preset cognitive state category;
[0099] The probability distribution of each preset cognitive state category is multiplied and summed with the corresponding penalty quantification constant to generate a cognitive assessment index in scalar numerical form.
[0100] Specifically, the input data is a topological feature vector generated after dimensionality reduction in the preceding steps. The output data is a cognitive assessment index in scalar numerical form. The data flow path is as follows: The quantization assessment module reads the input topological feature vector, performs a spatial inner product operation, and adds a bias term to generate an initial logical output value. The initial logical output value is fed into a pre-configured activation function channel for exponentialization and normalization calculations, outputting a probability distribution for each level of cognitive state. Finally, the probability distributions at each level are extracted and weighted summed with an externally configured penalty constant to output a single-value cognitive assessment index representing the degree of current cognitive impairment.
[0101] When the quantization evaluation module performs inner product calculations to generate the initial logic output value, the specific linear mapping formula is defined as follows:
[0102] ;
[0103] in Represents the corresponding number The initial logical output value of a preset cognitive state category. This represents the corresponding class established during training in the pre-defined multi-class classification model. A classification weight matrix for each preset cognitive state category. This represents the received topological feature vector. Representing the Category bias terms for each preset cognitive state category. This represents the matrix transpose operator.
[0104] The specific formula for generating the probability distribution by performing exponential operations and normalization is defined as follows:
[0105] ;
[0106] in This indicates that the current topological feature vector is determined to belong to the first... The probability distribution of a preset cognitive state category. This represents the total number of preset cognitive state categories, and is set to an integer between 3 and 5 according to the clinical anesthesia grading standards. Represented by the natural constant Exponential operators with base 0. Represents the corresponding number The initial logical output value of a preset cognitive state category.
[0107] The specific formula for calculating the cognitive assessment index by performing product summation is defined as follows:
[0108] ;
[0109] in The cognitive assessment index represents the final output of the system. The representative is the first Each preset cognitive state category has a pre-configured penalty quantification constant. The deeper the clinically defined level of cognitive impairment, the larger the value of this penalty quantification constant.
[0110] The topological feature vectors, after spatial mapping and dimensionality reduction, are essentially a set of high-dimensional floating-point arrays lacking direct physical interpretation. The quantization evaluation module establishes a conversion channel from high-dimensional feature manifolds to clinical diagnostic indicators. Inner product operations are performed using the classification weight matrix to construct hyperplane boundaries for different pathological states. Simple linear operations lack boundary constraints. These results are input into a softmax function layer for nonlinear mapping, forcing discrete logical truth values into confidence percentages that sum to one. This transformation outputs the system's probabilistic tendency to classify the current neural coupling blockade state as mild disorientation, moderate delirium, or severe impairment.
[0111] Conventional multi-class diagnostic models mostly employ the maximum probability principle to directly output a unique classification label. Simply outputting the label with the highest probability masks potential physiological deterioration trends. During the postoperative recovery period, with subtle fluctuations in neurological characteristics, if the patient is in the boundary region between two cognitive states, rigid classification is prone to conclusion shifts. This paper utilizes penalty quantization constants set for different severity levels to fuse all probability distributions, transforming the classification task into a regression task. The output scalar index comprehensively considers the weight cost of being in a low-probability but extremely high-risk state, overcoming the technical shortcomings of single discrete labels in representing continuous and gradual changes in physiological state and the potential for missed diagnoses of borderline risks. This provides a smooth and continuous regulatory benchmark variable for subsequent closed-loop intervention and control.
[0112] The closed-loop intervention module connects to the quantitative assessment module. When the cognitive assessment index meets the preset conditions, it outputs an intervention instruction and feeds back the timestamp to the matrix construction module to reset the time window.
[0113] Furthermore, the closed-loop intervention module is used to execute the steps of outputting intervention instructions when the cognitive assessment index meets preset conditions, and feeding back the timestamp to the matrix construction module to reset the time window:
[0114] The cognitive assessment index is compared with a pre-set state alert threshold.
[0115] When the cognitive assessment index exceeds the state warning threshold and the duration of the state exceeding the threshold reaches the preset tolerance time limit, the preset conditions are determined to be met, and an intervention command to trigger external devices is generated.
[0116] The system clock data at the moment the intervention command is generated is recorded as a timestamp and sent to the data receiving end of the matrix construction module;
[0117] The trigger matrix construction module clears the historical multimodal cache data within the current time window based on the timestamp and reinitializes the integration starting point of the time window.
[0118] Specifically, the input data includes: the cognitive assessment index output by the dimensionality reduction and quantification assessment module, a pre-set state alert threshold, and a preset tolerance time limit. The output data includes: intervention commands for external medical devices and system-level control feedback timestamps. The data flow path is as follows: the closed-loop intervention module continuously reads the cognitive assessment index and inputs it to the internal timing logic comparator. Upon a value exceeding the limit and the continuous state meeting the timing requirements, the module's output sends a trigger level command to the external intervention device. The main control program captures the absolute clock data generated by this trigger action, generates a control feedback timestamp, and transmits it in reverse through the system's internal bus to the preceding matrix construction module, forcibly erasing the underlying physiological parameters in the current buffer queue and resetting the integral timer.
[0119] When the closed-loop intervention module performs state comparison and command triggering, the specific mathematical expression of the internal control logic is defined as follows:
[0120] ;
[0121] in This represents the command trigger status variable. A value of 1 indicates that an intervention command has been generated to trigger external devices, while a value of 0 indicates that the system maintains a normal silent monitoring state. Representative at Cognitive assessment index in scalar numerical form received at any given time. This represents a pre-set state warning threshold, used to define the numerical boundary between safe physiological states and dangerous decoupled states. This represents the duration during which the cognitive assessment index continuously exceeds the state alert threshold. This represents the preset tolerance time limit, which physically means the width of the time filter set to prevent transient interference. The value range is set between 3 and 10 seconds.
[0122] When generating the control feedback timestamp and resetting the time window, the integral interval of the new time window is defined by the following formula:
[0123] ;
[0124] Postoperative monitoring is complex, and even slight patient positional movements or momentary loosening of sensor probes can trigger violent pulsatile oscillations in underlying EEG or hemodynamic signals. These non-physiological mutations, transmitted to the subsequent assessment network, can cause extreme spikes in the assessment index. A dual-judgment mechanism, combining a state alert threshold with a preset tolerance time limit, utilizes the hysteresis effect of the time dimension to filter out false alarm spikes caused by transient physical interference. External intervention devices are only activated to perform stimulation when the patient experiences genuine irreversible neurovascular decoupling and their condition continues to deteriorate.
[0125] Once the external device releases the intervention stimulus, the distribution of neurotransmitters and the local blood flow in the patient's brain undergo a forced physiological remodeling. The old time window cache sequence is filled with abnormal pathological data from before the intervention. If this old data is allowed to continue participating in covariance calculations, the newly generated manifold matrix will conflate two distinct physiological states, leading to severe distortion in the system's assessment of the intervention's effectiveness. By using precisely captured control feedback timestamps to reverse-clear the historical multimodal cache data of the matrix construction module, the mathematical correlation between physiological characteristics before and after the intervention is severed at the algorithm's underlying level. The newly initialized integral time window forces the system to reconstruct the neurovascular coupling manifold using only the pure data after physical awakening, ensuring that subsequent monitoring indicators fully map the actual efficacy of the medical intervention, thus constructing a complete hardware and software feedback loop from state assessment to underlying data self-calibration.
[0126] Example 2:
[0127] In the post-anesthesia recovery room or intensive care unit setting for elderly patients undergoing major surgery, medical staff need to continuously track the patient's neurological function recovery trajectory to prevent postoperative cognitive impairment and delirium. Clinically, scalp electroencephalography (EEG) is routinely used to track high-frequency neuronal electrical activity, supplemented by near-infrared spectroscopy to obtain low-frequency changes in local cerebral blood oxygen metabolism. However, a significant technical bottleneck exists in such complex vital sign monitoring scenarios: the combined impact of anesthetic drug blockade and surgical trauma stress causes deep pathological decoupling of the neurovascular coupling mechanisms in the patient's brain. Traditional monitoring systems, when processing this heterogeneous data, often perform mechanical feature splicing within Euclidean geometry or apply absolute empirical thresholds for alarms, ignoring the fact that the covariance matrix formed by cross-modal physiological features actually resides in a high-dimensional symmetric positive definite Riemannian manifold space. Traditional linear computation methods cannot resolve the nonlinear distortions and curvature collapses caused by the disintegration of coupling mechanisms in high-dimensional topological spaces. This leads to a high risk of manifold data misalignment and loss of deep-seated pathological features when dealing with elderly patients with significant individual physiological baseline differences. Consequently, this can result in missed diagnoses or misjudgments of early, subtle cognitive impairments, leading to missed opportunities for optimal physical or pharmacological intervention. To address these issues, this invention employs a postoperative cognitive quantitative assessment method that fuses electroencephalography (EEG) and near-infrared spectroscopy. Its structure is as follows: Figure 3As shown. The specific implementation process of this method is as follows:
[0128] By extracting preoperative and real-time EEG and near-infrared features to construct a covariance matrix, heterogeneous physiological signals are mapped to manifold matrices, thereby quantifying the deep topological state of neurovascular coupling. A preoperative baseline manifold is introduced to establish an individualized absolute reference. Geodesic distances are calculated in non-Euclidean space, and the distorted cross-modal manifold matrix is expanded to the tangent space through logarithmic mapping to generate topological feature vectors. This overcomes the spatial metric expansion and pathological feature loss defects caused by linear splicing in traditional assessment systems. The probability distribution output by the classification network is aggregated into a continuous scalar cognitive assessment index, resolving the risk of missed diagnoses due to the inability of a single discrete label to characterize the gradual change in physiological state. When an intervention command is triggered, the system provides reverse feedback of the timestamp to forcibly clear historical buffer data and reset the integration time window, severing the mathematical aliasing chain of abnormal physiological representations before and after intervention. This constructs a complete regulatory feedback loop from manifold space dimensionality reduction diagnosis to front-end data time axis self-calibration.
[0129] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A postoperative cognitive quantitative assessment system integrating electroencephalography (EEG) and near-infrared spectroscopy, characterized in that, Includes the following modules: The baseline anchoring module receives preoperative scalp EEG signals and preoperative near-infrared spectral signals, extracts features and splices them into a first vector, and calculates the covariance of the first vector to generate the preoperative baseline manifold. The matrix construction module receives real-time scalp EEG signals and real-time near-infrared spectral signals within a time window, extracts features and concatenates them into a second vector, and calculates the covariance of the second vector to generate a cross-modal manifold matrix. The feature extraction module connects the baseline anchoring module and the matrix construction module, calculates the geodesic distance between the cross-modal manifold matrix and the preoperative baseline manifold, projects the cross-modal manifold matrix onto the tangent space of the preoperative baseline manifold using logarithmic mapping to generate a tangent matrix, and combines the tangent matrix with the geodesic distance to generate a topological feature vector. The quantitative evaluation module is connected to the feature extraction module. It inputs the topological feature vector into the softmax function layer to output a probability distribution and calculates the cognitive evaluation index based on the probability distribution. The closed-loop intervention module is connected to the quantitative assessment module. When the cognitive assessment index meets the preset conditions, it outputs an intervention instruction and feeds back the timestamp to the matrix construction module to reset the time window.
2. The postoperative cognitive quantitative assessment system based on the fusion of electroencephalography and near-infrared spectroscopy according to claim 1, characterized in that, The baseline anchoring module is used to perform the steps of extracting features and concatenating them into a first vector: Extract the power spectral density sequence of the preoperative scalp EEG signal in multiple preset frequency bands; Extract the relative change sequences of oxyhemoglobin concentration and deoxyhemoglobin concentration from the preoperative near-infrared spectral signals; The EEG power spectral density sequence, the relative change sequence of oxyhemoglobin concentration, and the relative change sequence of deoxyhemoglobin concentration are timestamped and aligned based on a unified clock synchronization reference. The signal features of the same sampling point after time-stamp alignment are merged and spliced to construct the first vector.
3. The postoperative cognitive quantitative assessment system based on the fusion of electroencephalography and near-infrared spectroscopy according to claim 2, characterized in that, The baseline anchoring module is used to perform the step of calculating the covariance of the first vector to generate the preoperative baseline manifold: Obtain all the first vectors contained within the first preset resting time period, and calculate the mean vector of all the first vectors; Calculate the centered difference vector between each of the first vectors and the mean vector; Perform matrix multiplication on the centered difference vector and its transpose to generate an intermediate feature matrix; All intermediate feature matrices within the first preset resting time period are summed to generate a symmetric positive definite covariance matrix, and the symmetric positive definite covariance matrix is solidified into the preoperative baseline manifold.
4. The postoperative cognitive quantitative assessment system based on the fusion of electroencephalography and near-infrared spectroscopy according to claim 1, characterized in that, The matrix construction module is used to perform the following steps: extracting features and concatenating them into a second vector, and calculating the covariance of the second vector to generate a cross-modal manifold matrix: Within the real-time sliding time window, the real-time frequency band power spectral density of the real-time scalp EEG signal and the real-time hemoglobin concentration change of the real-time near-infrared spectral signal are extracted synchronously, and the real-time frequency band power spectral density and the real-time hemoglobin concentration change are combined and assembled into the second vector according to the time correspondence. Extract all the second vectors within the time window and calculate the real-time average vector; The second vector within the time window is subtracted from the real-time average vector to generate a real-time residual vector. The outer product operation is performed on all the real-time residual vectors to generate a real-time updated symmetric positive definite matrix, and the real-time updated symmetric positive definite matrix is used as the cross-modal manifold matrix.
5. The postoperative cognitive quantitative assessment system based on the fusion of electroencephalography and near-infrared spectroscopy according to claim 1, characterized in that, The feature extraction module is used to perform the step of calculating the geodesic distance between the cross-modal manifold matrix and the preoperative baseline manifold: Calculate the inverse square root matrix of the preoperative baseline manifold; The inverse square root matrix is multiplied by the transmodal manifold matrix to generate the projection transition matrix. The projection transition matrix is multiplied by the inverse square root matrix to generate an affine invariant Riemann metric matrix; Perform a principal logarithm operation on the affine invariant Riemannian metric matrix, calculate the matrix norm of the result of the principal logarithm operation, and output the calculated matrix norm as the geodesic distance.
6. The postoperative cognitive quantitative assessment system based on the fusion of electroencephalography and near-infrared spectroscopy according to claim 5, characterized in that, The feature extraction module is used to perform the step of projecting the cross-modal manifold matrix onto the tangent space of the preoperative baseline manifold using a logarithmic mapping to generate a tangent matrix: The preoperative baseline manifold is set as a fixed Riemann tangent point in the manifold topological space, and the square root matrix of the fixed Riemann tangent point is calculated. The square root matrix, the affine invariant Riemann metric matrix after performing the principal logarithm operation, and the square root matrix are multiplied sequentially. The resulting matrix from the sequential multiplication operations is used as the set of eigenvectors projected onto the Euclidean geometric plane to generate the tangent matrix.
7. The postoperative cognitive quantitative assessment system based on the fusion of electroencephalography and near-infrared spectroscopy according to claim 6, characterized in that, The feature extraction module is used to perform the step of generating a topological feature vector by combining the tangent matrix and the geodesic distance: Obtain all upper triangular matrix elements and diagonal matrix elements of the tangent matrix; The elements of the upper triangular matrix and the elements of the diagonal matrix are arranged in a dimension-reduced manner to generate a one-dimensional structural feature sequence. Extract each element from the one-dimensional structural feature sequence, and multiply each element of the one-dimensional structural feature sequence with the geodesic distance for weighting. The elements after multiplication and weighting are combined and arranged to output the final topological feature vector.
8. The postoperative cognitive quantitative assessment system based on the fusion of electroencephalography and near-infrared spectroscopy according to claim 1, characterized in that, The quantitative evaluation module is used to perform the following steps: inputting the topological feature vector into the softmax function layer to output a probability distribution, and calculating the cognitive evaluation index based on the probability distribution. The topological feature vector is multiplied by the classification weight matrix in the preset multi-classification model, and the category bias term is added to generate the initial logical output value. The initial logic output value is input into the softmax function layer for exponential operation and normalization to generate the probability distribution corresponding to multiple preset cognitive state categories; Obtain the penalty quantization constant pre-configured for each preset cognitive state category; The probability distribution of each preset cognitive state category is multiplied and summed with the corresponding penalty quantification constant to generate the cognitive evaluation index in scalar numerical form.
9. The postoperative cognitive quantitative assessment system based on the fusion of electroencephalography and near-infrared spectroscopy according to claim 1, characterized in that, The closed-loop intervention module is used to execute the following steps: when the cognitive assessment index meets the preset conditions, output an intervention instruction and feed back the timestamp to the matrix construction module to reset the time window: The cognitive assessment index is compared with a pre-set state alert threshold. When the cognitive assessment index exceeds the state alert threshold and the duration of the state exceeding the threshold reaches a preset tolerance time limit, it is determined that the preset condition is met, and an intervention command for triggering the external device is generated. The system clock data at the instant the intervention command is generated is recorded as the timestamp, and the timestamp is sent to the data receiving end of the matrix construction module; The matrix construction module is triggered to clear the historical multimodal cache data within the current time window based on the timestamp, and to reinitialize the integration starting point of the time window.
10. A method for quantitative postoperative cognitive assessment by fusing electroencephalography (EEG) and near-infrared spectroscopy, characterized in that, The postoperative cognitive quantitative assessment system for EEG and near-infrared spectroscopy fusion as described in any one of claims 1-9 includes the following steps: S1. Receive preoperative scalp EEG signals and preoperative near-infrared spectral signals, extract features and splice them into a first vector, calculate the covariance of the first vector to generate the preoperative baseline manifold; S2. Receive real-time scalp EEG signals and real-time near-infrared spectral signals within the time window, extract features and splice them into a second vector, calculate the covariance of the second vector to generate a cross-modal manifold matrix; S3. Calculate the geodesic distance between the transmodal manifold matrix and the preoperative baseline manifold, and use logarithmic mapping to project the transmodal manifold matrix onto the tangent space of the preoperative baseline manifold to generate a tangent matrix. Combine the tangent matrix with the geodesic distance to generate a topological feature vector. S4. Input the topological feature vector into the softmax function layer to output the probability distribution, and calculate the cognitive evaluation index based on the probability distribution; S5. When the cognitive assessment index meets the preset conditions, an intervention instruction is output, and the timestamp is fed back to the matrix construction module to reset the time window.