A postoperative brain function state evaluation method and system based on multi-modal fusion

Abstract

The application provides a postoperative brain function state evaluation method and system based on multi-modal fusion, relates to the technical field of postoperative brain function state evaluation, and comprises the following steps: acquiring multi-source original data of a postoperative patient, obtaining standardized multi-source data, and establishing an immediate postoperative brain function baseline; forming a multi-modal feature subset, and then obtaining a postoperative-preoperative function offset; generating a fusion feature vector and an implicit inhibition index through fusion processing, obtaining the implicit inhibition index; outputting a brain function state preliminary classification result and a new brain response feature after intervention through classification processing and intervention processing, driving an acousto-optic-electric closed-loop intervention; finally obtaining a postoperative brain function state evaluation result, and iteratively calculating again until the final evaluation result consistent with clinical diagnosis and passing the baseline is obtained. The application has the beneficial effect of realizing rapid and accurate evaluation and individualized intervention of the postoperative brain function state, and further improving the efficiency and quality of postoperative brain function management.

Description

Technical Field

[0001] This invention relates to the field of postoperative brain function status assessment technology, and more specifically, to a method and system for postoperative brain function status assessment based on multimodal fusion. Background Technology

[0002] In the medical field, postoperative brain function assessment (especially for patients undergoing major surgeries such as neurosurgery and cardiac surgery, assessing cognitive function, consciousness recovery, and monitoring of occult brain injury) has become a crucial step in ensuring the quality of postoperative rehabilitation. Current technological advancements show a trend towards "widespread multimodal data acquisition and diversified assessment indicators." Clinically, methods such as brain oxygen saturation monitors, electroencephalography (EEG) devices, event-related potential (ERP) recorders, and functional magnetic resonance imaging (fMRI) can be used to acquire multi-source raw data on brain oxygen metabolism, brain electrical activity, cognitive potentials, and brain network connectivity, providing a data foundation for postoperative brain function assessment.

[0003] However, existing technologies still have significant shortcomings: First, the problem of "fragmented processing" of multimodal data is prominent. Most assessment methods only analyze one type of data (such as judging oxygen supply status solely through brain oxygen saturation, or judging the degree of brain electrical inhibition solely through EEG), failing to achieve the synergistic integration of multi-dimensional features such as brain oxygen metabolism, cognitive potentials, and brain network connectivity. This makes the assessment results susceptible to interference from single data noise and unable to comprehensively reflect the overall state of brain function. Second, "baseline comparison lacks individualization." Existing technologies mostly use a general baseline of healthy individuals as a reference, without considering the individual differences in patients' preoperative brain function characteristics (such as preoperative cognitive impairment and baseline cerebral blood flow). This leads to large deviations in the calculation of postoperative functional deviations, making it difficult to accurately identify hidden brain function inhibition. Third, "assessment-intervention closed loop is broken." Existing technologies can only output preliminary classification results of brain function status, lacking a dynamic intervention parameter matching mechanism based on assessment results, and have not established an iterative verification process for intervention effects, failing to achieve a complete closed loop from "assessment" to "precise intervention" and then to "effect verification." To address the aforementioned issues, existing technologies often employ the outdated method of "integrating manual experience," which relies on clinicians manually compiling various data reports, combining subjective experience to assess brain function status, and formulating intervention plans. This method is not only inefficient (assessment of a single patient can take 1-2 hours), but also highly susceptible to differences in physician experience, making it difficult to guarantee the accuracy of assessment and the precision of intervention. It cannot meet the clinical needs for rapid, precise, and individualized brain function management after surgery. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for assessing postoperative brain function based on multimodal fusion, in order to improve the aforementioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows:

[0005] In a first aspect, this application provides a method for assessing postoperative brain function based on multimodal fusion, including:

[0006] We acquired multi-source raw data from postoperative patients, including immediate postoperative near-infrared tissue oxygenation, ultrasound microflow spectrum, event-related potentials from customized high-frequency chirped stimulation, as well as resting-state functional magnetic resonance imaging data and high-density electroencephalogram data within a preset time interval after surgery. After targeted preprocessing, we obtained standardized multi-source data and established an immediate postoperative brain function baseline.

[0007] Key brain network functional connections and power spectral density features of preset time intervals are extracted from multi-source raw data to form a multimodal feature subset. Based on the multimodal feature subset, nonlinear manifold alignment processing is performed on the postoperative immediate brain function baseline and the preoperative awake baseline under the same stimulation conditions to obtain the postoperative-preoperative functional offset.

[0008] Using a subset of multimodal features and postoperative-preoperative functional shift, a fusion feature vector and a concealment inhibition index are generated through fusion processing. The fusion processing includes determining feature weights by combining cross-validation with random forest, generating a weighted fusion feature vector, and simultaneously coupling compressed oxygen blood flow and latency information through a dual-channel gating network to obtain the concealment inhibition index.

[0009] Based on the fusion feature vector and the concealment inhibition index, the preliminary classification results of brain functional status and the response characteristics of newborn brain after intervention are output through classification and intervention. The classification process uses a pre-trained deep neural network model to output multi-level functional impairment results, and the intervention process drives a closed-loop intervention of sound, light, electricity and electricity by comparing the concealment inhibition index with the stepwise awakening strategy table.

[0010] Combining the preliminary classification results with the post-intervention neonatal brain response characteristics, and after validation and matching processes, the final postoperative brain function status assessment results were obtained. The validation process included combining the preliminary results of images, electrophysiological data, and clinical diagnosis within a preset postoperative time interval. The matching process included re-aligning the post-intervention response characteristics with the immediate postoperative baseline and matching them with the preoperative baseline. If the validation was consistent and the matching was successful, the preliminary classification result was output as the final assessment result. If any step failed, the iterative calculation was repeated until a final assessment result consistent with the clinical diagnosis and with a successful baseline match was obtained.

[0011] Preferably, the key brain network functional connections and power spectral density features of the preset time interval frequency band are extracted from the multi-source raw data to form a multimodal feature subset. Based on the multimodal feature subset, nonlinear manifold alignment processing is performed on the postoperative immediate brain function baseline and the preoperative awake baseline under the same stimulation conditions to obtain the postoperative-preoperative functional shift, including:

[0012] Core features were extracted from standardized multi-source data. This included extracting time series of key brain network nodes from corrected resting-state oxygenation level-dependent data using a brain network analysis toolkit and calculating the functional connectivity strength between nodes; obtaining the power spectral density and EEG complexity index for a preset time interval using an EEG analysis toolkit for denoised continuous EEG data; and integrating the calculated functional connectivity strength, power spectral density, and EEG complexity index with the obtained brain oxygen saturation coefficient of variation, blood flow velocity pulsatility index, and P300 parameters to form a multimodal functional feature subset.

[0013] The patient's conscious brain function data were retrieved within a pre-defined time window and under the same stimulation conditions as after surgery. The conscious brain function data was preprocessed to obtain new conscious brain function data. At the same time, the corresponding coefficient of variation of brain oxygen saturation, blood flow velocity pulsatility index and P300 parameters were calculated to obtain the preoperative individualized conscious baseline. The preoperative individualized conscious baseline includes 24 features, corresponding to the same brain regions and parameter types.

[0014] Cross-time point data matching was performed between the pre-individualized awake baseline and the immediate postoperative brain function baseline. This included constructing a Riemannian space framework, mapping the two types of baseline data to Riemannian manifold space, and introducing a symmetric KL divergence constraint during the mapping process. This constraint minimized the symmetric difference in probability distribution between preoperative and postoperative data in the manifold space, while preserving the original brain function feature association structure of the two types of data, thus obtaining the postoperative and preoperative manifold coordinates. Based on a subset of multimodal functional features, the geodesic distance weighted sum of the corresponding brain function dimensions in the Riemannian manifold space was calculated. This weighted sum was used to quantify the degree of difference between postoperative and preoperative brain function states, resulting in the postoperative-preoperative brain function offset.

[0015] Preferably, the method of generating a fused feature vector and a concealment inhibition index by utilizing a multimodal feature subset and postoperative-preoperative functional shift through fusion processing includes determining feature weights through cross-validation combined with random forest, generating a weighted fused feature vector, and simultaneously obtaining the concealment inhibition index by coupling compressed oxygen blood flow and latency information through a dual-channel gating network, including:

[0016] The dynamic weights of each feature in the multimodal functional feature subset were determined using a cross-validation combined with a multi-model fusion algorithm. This included grouping the multimodal functional feature subset according to patient identification, selecting a subset of groups as the training set, and using the remaining groups as the validation set; constructing multiple different types of machine learning models, using the functional scores of historical patients within a preset time window after surgery as labels, and calculating the permutation importance of each feature through the validation set; and taking a weighted average and normalizing the feature permutation importance output by different models to obtain the weight coefficients corresponding to each feature.

[0017] Based on the weight coefficients corresponding to each feature, the multimodal functional feature subset is subjected to spatiotemporal joint fusion processing: in the spatiotemporal dimension, the mean of each feature within a preset time window and the trend slope within a preset time period are weighted according to the set weight ratio; in the spatial dimension, features of different brain regions are grouped according to functional correlation, the features within each group are summed first, and then weighted according to the feature weight coefficients to finally generate a fusion feature vector with unified dimensions.

[0018] Two key offset information types were extracted from the postoperative-preoperative brain function offset. These included brain oxygen metabolism offset derived from changes in the coefficient of variation of brain oxygen saturation and cognitive potential offset derived from changes in P300 latency. Simultaneously, a fusion feature vector with uniform dimensions was retrieved, and feature weights corresponding to the coefficient of variation of brain oxygen saturation and P300 parameters were extracted and used as priority coefficients for the offset information. A dual-channel gated attention network was used to couple and compress the two key offset information types to obtain the latent brain function inhibition index.

[0019] Preferably, based on the fused feature vector and the concealment inhibition index, the preliminary classification results of brain functional status and the response characteristics of the newborn brain after intervention are output through classification and intervention processing. The classification processing uses a pre-trained deep neural network model to output multi-level functional impairment results, and the intervention processing drives a closed-loop acoustic-electric intervention by comparing the concealment inhibition index with a stepwise awakening strategy table, including:

[0020] A pre-trained dual-branch deep neural network model is loaded. The dual-branch deep neural network model includes a feature adaptation branch and a classification branch. The feature adaptation branch maps the fused feature vector into a feature map that meets the input requirements of the classification branch through a dimension transformation operation within a preset time interval. The classification branch adopts a multi-layer fully connected network structure, using the fused feature vectors of historical patients and the corresponding postoperative brain function status labels as training data. An optimizer is used to train the model according to the set parameters, and the model parameters are maximized through preset evaluation indicators. The postoperative brain function status labels include categories such as normal, mild cognitive impairment, moderate cognitive impairment, and severe functional impairment.

[0021] The fused feature vector is input into the trained dual-branch deep neural network model. After feature adaptation branch dimension expansion and enhancement, the classification branch calculates the probability of each brain functional state through the Softmax function. The category with the highest probability is taken as the preliminary classification result, and the confidence score is output simultaneously. When the confidence score is lower than the set threshold, the classification result is marked as pending verification.

[0022] Based on the preliminary classification results, the individualized stepwise awakening strategy table developed for the patient was retrieved. The concealment inhibition index was compared with the individualized stepwise awakening strategy table to determine the intervention parameters. The target-controlled intervention device was then activated to perform the intervention. After the intervention lasted for a period of time, the patient's P300 potential and brain oxygen data were collected again. After preprocessing, iterative calculations were performed to obtain the neonatal brain response characteristics after the intervention.

[0023] Preferably, the combination of preliminary classification results and post-intervention neonatal brain response characteristics, after verification and matching processes, yields the final postoperative brain function status assessment results, including:

[0024] Dynamic imaging data, neurophysiological monitoring data, and clinical verification data were collected from patients within a preset time window after surgery. The Kappa coefficient was used to verify the consistency between the preliminary classification results of postoperative brain function status and the clinical verification data. If the coefficient met the standard, the results were considered consistent; otherwise, they were inconsistent.

[0025] The Riemannian manifold mapping method was used to align the neonatal brain response characteristics after intervention with the immediate postoperative brain function baseline and calculate the post-intervention to immediate postoperative offset. Then, the neonatal brain response characteristics after intervention were matched with the preoperative individualized awake baseline, and the matching degree was calculated. If the matching degree reached the set standard, it was considered passed; otherwise, it was not passed.

[0026] The consistency check and matching results are evaluated comprehensively: if both pass, the preliminary classification result is output as the final postoperative brain function status assessment result; if either fails, the response characteristics of the newborn brain after intervention are used as new inputs, and the calculation is iterated again until both pass, and the final postoperative brain function status assessment result is output.

[0027] Secondly, this application also provides a postoperative brain function status assessment system based on multimodal fusion, including:

[0028] Acquisition module: used to acquire multi-source raw data of postoperative patients. The multi-source raw data includes near-infrared tissue oxygenation map, ultrasound micro-blood flow spectrum, event-related potentials of customized high-frequency chirped sound stimulation, as well as resting-state functional magnetic resonance imaging data and high-density electroencephalogram data within a preset time interval after surgery. After targeted preprocessing, standardized multi-source data are obtained and a baseline of brain function is established immediately after surgery.

[0029] Extraction module: used to extract key brain network functional connections and power spectral density features of preset time interval frequency bands from multi-source raw data to form a multimodal feature subset. Based on the multimodal feature subset, nonlinear manifold alignment processing is performed on the postoperative immediate brain function baseline and the preoperative awake baseline under the same stimulation conditions to obtain the postoperative-preoperative functional offset.

[0030] Fusion module: Used to generate fusion feature vector and concealment inhibition index by using multimodal feature subsets and postoperative-preoperative functional offset. The fusion processing includes determining feature weights by cross-validation combined with random forest, generating fusion feature vector by weighting, and obtaining concealment inhibition index by coupling compressed oxygen blood flow and latency information through a dual-channel gating network.

[0031] Processing module: Based on the fused feature vector and the concealment inhibition index, it outputs the preliminary classification results of brain functional status and the response characteristics of newborn brain after intervention through classification and intervention processing. The classification processing uses a pre-trained deep neural network model to output multi-level functional impairment results, and the intervention processing drives the closed-loop intervention of sound, light, electricity and electricity by comparing the concealment inhibition index with the ladder awakening strategy table.

[0032] The assessment module combines the preliminary classification results with the post-intervention neonatal brain response characteristics. After verification and matching processes, it finally obtains the postoperative brain function status assessment results. The verification process includes combining the preliminary results of images, electrophysiological data, and clinical diagnosis within a preset postoperative time interval. The matching process includes re-aligning the post-intervention response characteristics with the immediate postoperative baseline and matching them with the preoperative baseline. If the verification is consistent and the matching is successful, the preliminary classification result is output as the final assessment result. If any step fails, the iterative calculation is repeated until a final assessment result consistent with the clinical diagnosis and with a successful baseline match is obtained.

[0033] Thirdly, this application also provides a postoperative brain function status assessment device based on multimodal fusion, comprising:

[0034] Memory, used to store computer programs;

[0035] A processor is used to implement the steps of the postoperative brain function status assessment method based on multimodal fusion when executing the computer program.

[0036] Fourthly, this application also provides a readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method for assessing postoperative brain function status based on multimodal fusion.

[0037] The beneficial effects of this invention are as follows:

[0038] This invention addresses the problems of insufficient multimodal data fusion, lack of individualized baseline comparison, broken assessment-intervention closed loop, and low efficiency and accuracy caused by reliance on human experience in existing postoperative brain function assessment techniques. It proposes a postoperative brain function assessment method based on multimodal feature fusion and individualized closed-loop intervention. This method first extracts key brain network functional connections and power spectral density of preset time intervals from multi-source raw data (including spatiotemporal distribution maps of brain oxygen saturation, microflow spectra of the middle cerebral artery, P300 component potentials, resting-state oxygenation level-dependent signals, and continuous EEG signals) to construct a multimodal feature subset. Second, using nonlinear manifold alignment technology, it individually compares the immediate postoperative brain function baseline with the patient's preoperative awake baseline under the same stimulation conditions to calculate a precise postoperative-preoperative functional shift. Then, based on cross-validation and multi-model fusion algorithms, it determines the dynamic weights of the features, generates a unified fusion feature vector through spatiotemporal dimensional fusion, and inputs it into a dual-branch deep neural network model to achieve preliminary classification of brain functional states. Simultaneously, it combines the hidden brain function inhibition index extracted from the multimodal feature subset with an individualized stepwise awakening strategy table to initiate target-controlled device intervention. Finally, it verifies the intervention through clinical validation (combining imaging, neurophysiological data, and physician diagnosis) and post-intervention feature matching, forming a "feature extraction - baseline alignment - classification intervention - iterative validation" process. This complete closed loop enables rapid and accurate assessment and individualized intervention of postoperative brain function status, effectively addressing the shortcomings of existing technologies and improving the efficiency and quality of postoperative brain function management.

[0039] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a schematic diagram of the postoperative brain function status assessment method based on multimodal fusion as described in an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of the postoperative brain function status assessment system based on multimodal fusion as described in an embodiment of the present invention;

[0043] Figure 3 This is a schematic diagram of the postoperative brain function status assessment device based on multimodal fusion as described in an embodiment of the present invention.

[0044] In the diagram: 701, Acquisition module; 702, Extraction module; 703, Fusion module; 704, Processing module; 705, Evaluation module; 800, Postoperative brain function status evaluation device based on multimodal fusion; 801, Processor; 802, Memory; 803, Multimedia component; 804, I / O interface; 805, Communication component. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of 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, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0046] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0047] Example 1:

[0048] This embodiment provides a method for assessing postoperative brain function based on multimodal fusion.

[0049] See Figure 1 The figure shows that the method includes steps S100, S200, S300, S400 and S500.

[0050] S100. Acquire multi-source raw data from postoperative patients, including immediate postoperative near-infrared tissue oxygenation, ultrasound micro-blood flow spectrum, event-related potentials from customized high-frequency chirped sound stimulation, resting-state functional magnetic resonance imaging data and high-density electroencephalogram data within a preset time interval after surgery. After targeted preprocessing, standardized multi-source data are obtained and an immediate postoperative brain function baseline is established.

[0051] It should be noted that the preprocessing includes spatial smoothing of near-infrared tissue oxygenation maps, high-pass ultrasound microflow spectroscopy, and potential bandpassing of event-related potentials. At the same time, head motion correction, time-level correction, and spatial standardization are performed on resting-state functional magnetic resonance imaging data, and filtering, noise reduction, and artifact removal are performed on high-density electroencephalogram data.

[0052] It is understood that step S100 includes S101, S102, and S103, wherein:

[0053] S101. Collect multi-source raw data from patients who are not fully awake in the postoperative recovery room. Specifically, this includes the spatiotemporal distribution map of brain oxygen saturation obtained immediately postoperatively via non-invasive near-infrared spectroscopy, the micro-blood flow spectrum of the middle cerebral artery obtained via transcranial Doppler ultrasound, the P300 component potential induced by a customized high-frequency chirp obtained via multi-channel event-related potential analyzer, and the resting-state oxygenation level-dependent signal obtained via functional magnetic resonance imaging (fMRI) and continuous EEG signals obtained via high-density electroencephalography (EEG) within a preset time interval postoperatively. The spatiotemporal distribution map of brain oxygen saturation is denoted as O(x,y,t), where x and y represent the two-dimensional detection coordinates of the frontotemporal lobe region, and t represents the data acquisition time point. The micro-blood flow spectrum of the middle cerebral artery is denoted as F(r,f,t), where r represents the depth gradient point detected by ultrasound, and f represents the frequency domain component of blood flow velocity. The P300 component potential is denoted as P(ch,t), where ch represents the electrode channel for potential acquisition. The resting-state oxygenation level-dependent signal is denoted as M... (v,t), where v represents the voxel number of the whole brain; continuous EEG signals are denoted as E(ch,t), where ch represents the electrode channel for EEG acquisition.

[0054] S102. Targeted preprocessing operations are performed on the collected multi-source raw data: Spatial smoothing of the spatiotemporal distribution map of brain oxygen saturation is performed using a Gaussian filter of appropriate size, and the coefficient of variation of brain oxygen saturation is calculated; the micro-blood flow spectrum of the middle cerebral artery is processed using a high-pass filter with a preset time interval cutoff frequency, and the blood flow velocity pulsatility index is calculated; the P300 component potentials are processed using a bandpass filter with a preset time interval frequency band combined with baseline correction, and the P300 latency and amplitude are extracted; the resting-state oxygen level-dependent signal is processed using a professional brain imaging toolkit to perform head motion correction, spatial standardization, delinearization, and low-frequency filtering; continuous EEG signals are processed using independent component analysis to remove artifacts, and notch filtering is superimposed to eliminate power frequency interference; the key formula for calculating the coefficient of variation of brain oxygen saturation is:

[0055] In the formula, This represents the smoothed brain oxygen saturation data. std (·) represents the standard deviation calculation function, and mean (·) represents the mean calculation function. This formula can quantify the degree of fluctuation in brain oxygen saturation.

[0056] S103. Integrate the smoothed brain oxygen saturation data, high-pass data of microvascular flow in the middle cerebral artery, and bandpass data of P300 component potentials after preprocessing in step S102. Calculate the mean values ​​of each parameter within the preset time window immediately after surgery, and remove data points with fluctuations exceeding a reasonable range to form the baseline of brain function immediately after surgery. Simultaneously, normalize the resting-state oxygen level-dependent correction data and continuous EEG denoising data after preprocessing in step 1.2 using a standardization method. Combine these with the aforementioned integrated data to construct a unified multimodal spatiotemporal aligned data matrix, ultimately forming standardized multi-source data. This normalization process eliminates dimensional differences between different data dimensions.

[0057] S200: Extract key brain network functional connections and preset time interval frequency band power spectral density features from multi-source raw data to form a multimodal feature subset. Based on the multimodal feature subset, perform nonlinear manifold alignment processing on the postoperative immediate brain function baseline and the preoperative awake baseline under the same stimulation conditions to obtain the postoperative-preoperative functional offset.

[0058] It is understood that step S200 includes S201, S202, and S203, wherein:

[0059] S201. Extract core features from standardized multi-source data, including extracting time series of key brain network nodes from corrected resting-state oxygenation level-dependent data using a brain network analysis toolkit and calculating the functional connectivity strength between nodes; obtaining the power spectral density and EEG complexity index for a preset time interval using an EEG analysis toolkit for denoised continuous EEG data; integrating the calculated functional connectivity strength, power spectral density, and EEG complexity index with the obtained brain oxygen saturation coefficient of variation, blood flow velocity pulsatility index, and P300 parameters to form a multimodal functional feature subset, wherein the formula for calculating functional connectivity strength is as follows:

[0060]

[0061] In the formula, This represents the degree of functional connection between two voxel nodes in the brain. This is the first voxel node number in the whole-brain voxels corresponding to the resting-state blood oxygenation level-dependent signal. t represents the second voxel node number in the whole-brain voxels corresponding to the resting-state oxygenation level-dependent signal, t represents the data acquisition time point number corresponding to the resting-state oxygenation level-dependent signal, and T represents the total number of data acquisition time points for the resting-state oxygenation level-dependent signal. For the preprocessed first The corrected resting-state oxygenation level at time t, dependent on the signal value. For the preprocessed first The corrected resting-state oxygenation level at time t, dependent on the signal value. For the first Individualized resting oxygenation levels at all T time points were calculated as a signal-mean value. For the first The signal-dependent mean of resting oxygen levels at all T time points after correction for individual values;

[0062] It should be noted that the coefficient of variation of cerebral oxygen saturation, blood flow velocity pulsatility index, and P300 parameter are all derived from the multi-source raw data collected in step S100, and obtained after targeted preprocessing in step S102. Specifically: The coefficient of variation for brain oxygen saturation was obtained by smoothing the spatiotemporal distribution map O(x,y,t) of brain oxygen saturation acquired immediately post-surgery using a non-invasive near-infrared spectrometer with a Gaussian filter of appropriate size. The pulsatility index was calculated from the micro-blood flow spectrum F(r,f,t) of the middle cerebral artery acquired by transcranial Doppler ultrasound, processed by a high-pass filter at a preset time interval cutoff frequency, and then calculated according to the definition formula of the pulsatility index (peak systolic velocity minus end-diastolic velocity, divided by the average velocity). In transcranial Doppler ultrasound examination, this index reflects the vascular compliance and resistance of the middle cerebral artery. The P300 parameters were obtained by acquiring P300 component potentials evoked by customized high-frequency chirps using a multi-channel event-related potential analyzer. After bandpass filtering and baseline correction within a preset time interval, the P300 latency and amplitude were extracted as key parameters. These parameters reflect changes in the brain's cognitive processing potentials in response to stimuli within a preset time interval.

[0063] S202. Retrieve the patient's conscious brain function data within the pre-preoperative time window and under the same stimulation conditions as after surgery. Preprocess the conscious brain function data to obtain new conscious brain function data after preprocessing. At the same time, calculate the corresponding brain oxygen saturation variation coefficient, blood flow velocity pulsatility index and P300 parameter to obtain the preoperative individualized conscious baseline. The preoperative individualized conscious baseline includes 24-dimensional features, corresponding to the same brain regions and parameter types.

[0064] It should be noted that these preoperative data are processed according to the same preprocessing procedure as in step S102, including a series of operations such as Gaussian filtering, high-pass filtering, band-pass filtering, and head movement correction. The data structure of the preoperative individualized awake baseline is completely matched with the postoperative immediate brain function baseline obtained in step S103. Both contain multi-dimensional features and correspond to the same brain regions and parameter types. This matching design can effectively eliminate the interference of individual brain structure differences on subsequent evaluation.

[0065] S203. Cross-time point data matching was performed between the pre-individualized awake baseline and the immediate postoperative brain function baseline. This included constructing a Riemannian space framework, mapping the two types of baseline data to the Riemannian manifold space, and introducing a symmetric KL divergence constraint during the mapping process. This constraint minimizes the symmetric difference in probability distribution between preoperative and postoperative data in the manifold space while preserving the original brain function feature association structure of the two types of data, thus obtaining the postoperative and preoperative manifold coordinates. Based on a subset of multimodal functional features, the geodesic distance weighted sum of the corresponding brain function dimensions in the Riemannian manifold space was calculated. This weighted sum quantifies the degree of difference between postoperative and preoperative brain function states, obtaining the postoperative-preoperative brain function offset. The formula for calculating the geodesic distance weighted sum is as follows:

[0066]

[0067] In the formula, ΔB represents the postoperative-preoperative brain function shift. Let be the weight of the k-th dimension in the Riemannian manifold space, and d(·,·) be the geodesic distance calculation function in the Riemannian manifold space. Let be the value of the postoperative manifold coordinates in the k-th dimension. Let be the preoperative manifold coordinates in the k-th dimension, where k is the dimension of the Riemannian manifold space.

[0068] In this embodiment, the t-SNE nonlinear manifold alignment algorithm can also be used to perform cross-time point data matching between the preoperative individualized awake baseline and the immediate postoperative brain function baseline. The 24-dimensional baseline data of the two classes are mapped to a 3-dimensional manifold space that retains 95% of the data variance, obtaining the postoperative manifold coordinates and the preoperative manifold coordinates. Based on the multimodal functional feature subset, the weighted sum of Euclidean distances of the corresponding brain function dimensions in the manifold space is calculated to obtain the postoperative-preoperative brain function offset. The formula for calculating the weighted sum of Euclidean distances is as follows:

[0069]

[0070] In the formula, This refers to the postoperative-preoperative brain function shift. Let the weights be the weights of the k-th dimensional manifold space. Let be the value of the postoperative manifold coordinates in the k-th dimension. The value of the preoperative manifold coordinates in the k-th dimension is given, where k represents the 3-dimensional manifold space dimension.

[0071] Understandably, the multimodal functional feature subset obtained in this step plays a significant role. On one hand, when using a 5-fold cross-validation combined with a random forest-gradient boosting fusion algorithm to determine the dynamic weights of each feature, the multimodal functional feature subset serves as the foundational data, divided into 5 groups based on patient ID for training and validating the random forest and gradient boosting models. Using the postoperative 7-day GOS-E scores of 500 historical patients as labels, the relative importance of each feature in assessing postoperative brain function is determined by calculating the permutation importance of each feature, thus obtaining feature weight coefficients. These weight coefficients directly determine the contribution ratio of each feature during subsequent feature fusion, influencing the generation of the fused feature vector.

[0072] On the other hand, in the process of generating the latent brain function inhibition index, the cerebral oxygen metabolism shift (derived from changes in the coefficient of variation of cerebral oxygen saturation) and cognitive potential shift (derived from changes in P300 latency) in the multimodal functional feature subset are key input factors. Simultaneously, changes in the blood flow velocity pulsatility index are used to calculate the cerebral blood flow compensation coefficient, and changes in the EEG complexity index are used to calculate the complexity attenuation factor; all of these participate in the calculation of the latent brain function inhibition index. By coupling and compressing the cerebral oxygen metabolism shift and cognitive potential shift using a dual-channel gated attention network, the parameters involved in the gating function and compression formula are closely related to the multimodal functional feature subset, ultimately yielding the latent brain function inhibition index, providing a quantitative indicator for assessing the latent inhibition of postoperative brain function.

[0073] S300. Using a multimodal feature subset and postoperative-preoperative functional offset, a fusion feature vector and a concealment inhibition index are generated through fusion processing. The fusion processing includes determining feature weights by combining cross-validation with random forest, generating a weighted fusion feature vector, and simultaneously coupling compressed oxygen blood flow and latency information through a dual-channel gating network to obtain the concealment inhibition index.

[0074] It is understood that step S300 includes S301, S302, and S303, wherein:

[0075] S301. A cross-validation combined with a multi-model fusion algorithm is used to determine the dynamic weights of each feature in the multimodal functional feature subset. This includes grouping the multimodal functional feature subset according to patient identification, selecting a subset as the training set, and the remaining subset as the validation set; constructing multiple different types of machine learning models, using the functional scores of historical patients within a preset postoperative time window as labels, and calculating the permutation importance of each feature through the validation set; taking a weighted average of the feature permutation importance output by different models and normalizing it to obtain the weight coefficients corresponding to each feature. The formula for calculating the permutation importance is as follows:

[0076] In the formula, For the permutation importance of the i-th feature, The accuracy obtained by training the model using the original features. The accuracy obtained by training the model after randomly permuting the i-th feature, where i is the feature number;

[0077] S302. Based on the weight coefficients corresponding to each feature, perform spatiotemporal joint fusion processing on the multimodal functional feature subset: In the spatiotemporal dimension, calculate the weighted average of each feature within a preset time window and the trend slope within a preset time interval according to the set weight ratio; in the spatial dimension, group the features of different brain regions according to their functional correlation, first sum the features within each group, and then calculate the weighted sum according to the feature weight coefficients, finally generating a fusion feature vector with unified dimensions. The calculation formula for the fusion feature vector is as follows:

[0078]

[0079] In the formula, To fuse feature vectors, Let be the weight coefficient of the i-th feature, and a and b be the weight ratios of the feature mean and the trend slope, respectively. Let be the mean of the i-th feature within a preset time interval window. Let i be the trend slope of the i-th feature within a preset time interval, where i is the feature number;

[0080] It should be noted that the core of time dimension fusion is to “take into account both the immediate steady state and dynamic changes of features” – the postoperative brain function state is not static. For example, brain oxygen saturation may fluctuate greatly immediately after surgery and gradually stabilize after 2 hours. If only the value at a single time point is used, it is easy to ignore the dynamic signals of early hidden damage; if only the trend is used, it may be affected by short-term noise interference.

[0081] In practice, the "mean within the preset time interval window" needs to be set based on clinical practice of postoperative brain function monitoring. For example, for patients in the postoperative recovery period, 1-2 hours postoperatively (a critical period for the transition of brain function from anesthesia recovery to stability) is usually selected as the time window. The moving average of the features within this window is calculated (e.g., sampling every 10 minutes, for a total of 12 samples, and then calculating the mean) to reflect the "immediate steady-state level" of the features. This setting refers to the clinical consensus in neuro-intensive care units (NICUs) that "brain function baseline is established 2 hours postoperatively," avoiding feature distortion caused by short-term interference such as residual anesthesia and postoperative stress. The "trend slope within the preset time interval" needs to cover a longer dynamic period (e.g., 2-6 hours postoperatively). The slope of the feature changes over time is calculated using linear regression or smoothed spline functions: if the slope of the theta band power spectral density of EEG is positive (gradually increasing over time), it may indicate worsening brain function inhibition; if it is negative (decreasing over time), it may indicate relief of inhibition. The "setting of weight ratios" here (such as mean weight 0.6 and slope weight 0.4) needs to be verified based on historical data. In actual research, the contribution of mean and slope to "postoperative 7-day cognitive function outcome" can be analyzed through follow-up data of more than 100 postoperative patients to finally determine the weights and ensure that the fusion results in the time dimension are both stable and reflect dynamic trends.

[0082] In this step, the specific process of dimensional unification is as follows: First, the "mean-slope" features fused from the time dimension (each original feature corresponds to 2 dimensions) and the "group-level features" fused from the spatial dimension (each group corresponds to 1 dimension) are integrated into an "intermediate feature matrix" (e.g., original 6 features × 2 time dimensions + 3 spatial group-level features = 15 dimensions); then, dimensionality compression is performed using principal component analysis (PCA) or an autoencoder to retain more than 95% of the feature information, ultimately generating a fused feature vector with fixed dimensions (e.g., 12 or 24 dimensions). The choice of dimensions here needs to balance "information retention" and "model efficiency"—in actual research, if a deep neural network classification is used, the dimension can be appropriately increased (e.g., 24 dimensions) to retain more details; if a traditional machine learning model (e.g., random forest) is used, the dimension needs to be reduced (e.g., 12 dimensions) to avoid overfitting.

[0083] In this embodiment, spatiotemporal fusion is used to transform the originally chaotic and fragmented multi-source features (such as features of 10 different dimensions) into vectors with unified dimensions and focused information, providing high-quality input for subsequent classification models. Spatiotemporal fusion captures the dynamic trend of brain function (time dimension) and network synergy (spatial dimension), making the fused features more consistent with the real changes in brain function after surgery, and improving the individualization and clinical adaptability of the assessment.

[0084] S303. Two types of key offset information are extracted from the postoperative-preoperative brain function offset. The two types of key offset information include brain oxygen metabolism offset caused by changes in the coefficient of variation of brain oxygen saturation and cognitive potential offset caused by changes in P300 latency. At the same time, a fusion feature vector with uniform dimension is retrieved, and feature weights corresponding to the coefficient of variation of brain oxygen saturation and P300 parameters are extracted from it and used as priority coefficients for offset information. A dual-channel gated attention network is used to couple and compress the two types of key offset information to finally obtain the hidden brain function inhibition index.

[0085] It should be noted that the coupling compression process involves introducing a cerebral blood flow compensation coefficient calculated from the change in the blood flow velocity pulsation index in the gating function design to adjust the weights of the two types of offset information; and introducing a complexity attenuation factor calculated from the change in the EEG complexity index in the compression formula design to correct the coupling compression result. The calculation formula for the gating function is as follows:

[0086]

[0087] In the formula, and These are the gate values ​​for the two channels, and σ(·) is the activation function. and These are the learnable weight matrices. This is due to a shift in brain oxygen metabolism. Here, b1 and b2 are the cerebral blood flow compensation coefficients, and b1 and b2 are the bias terms, respectively. This is a cognitive potential shift.

[0088] Therefore, this gating function can be used to dynamically adjust the weights of the two types of offset information. That is, by fusing the feature weights of the feature vectors, the coupling of offset information can be made to better reflect the importance of key brain function features.

[0089] S400. Based on the fusion feature vector and the concealment inhibition index, the preliminary classification results of brain functional status and the response characteristics of newborn brain after intervention are output through classification and intervention processing. The classification processing uses a pre-trained deep neural network model to output multi-level functional impairment results, and the intervention processing drives the closed-loop intervention of sound, light and electricity by comparing the concealment inhibition index with the ladder awakening strategy table.

[0090] It is understood that step S400 includes S401, S402, and S403, wherein:

[0091] S401. Load the pre-trained dual-branch deep neural network model, which includes a feature adaptation branch and a classification branch. The feature adaptation branch maps the fused feature vector into a feature map that meets the input requirements of the classification branch through a dimension transformation operation within a preset time interval. The classification branch adopts a multi-layer fully connected network structure, using the fused feature vectors of historical patients and the corresponding postoperative brain function status labels as training data. An optimizer is used to train the model according to the set parameters, and the model parameters are optimized to the maximum extent through preset evaluation indicators. The postoperative brain function status labels include categories such as normal, mild cognitive impairment, moderate cognitive impairment, and severe functional impairment.

[0092] S402. Input the fused feature vector into the trained dual-branch deep neural network model. After feature adaptation branch dimension expansion and enhancement, the classification branch calculates the probability of each brain functional state through the Softmax function, takes the category with the highest probability as the preliminary classification result, and outputs the confidence score simultaneously. When the confidence score is lower than the set threshold, the classification result is marked as pending verification.

[0093] It should be noted that, firstly, the fusion feature vector is expanded in dimension and enhanced in feature adaptation branch; then, the probability of each brain functional state category is calculated by probability calculation function through the output layer of classification branch; finally, the category with the highest probability is selected as the preliminary classification result of the postoperative brain functional state, and the confidence level corresponding to the classification result is output. When the confidence level is lower than the set threshold, the classification result is marked as pending verification; the probability distribution of each state category can be calculated by probability calculation function.

[0094] In this step, if the preliminary classification result is "severe functional impairment," even if the "latent inhibition index" in step S303 is in the moderate range, the intervention intensity should be appropriately increased (e.g., increasing the duration of acoustic stimulation or the frequency of light stimulation) based on the baseline parameters matched in the strategy table to avoid delaying brain function recovery due to insufficient intervention. If the preliminary classification result is "normal," but the "latent inhibition index" is slightly higher than the safety threshold (e.g., close to 2.5), the intervention interval can be shortened (e.g., from 30 minutes / session to 20 minutes / session) based on the mild intervention parameters matched in the strategy table to achieve "early intervention and mild intervention" and reduce the risk of over-intervention. This connection essentially combines "macro-functional state classification" with "micro-inhibition index," making the intervention in step S403 more aligned with the patient's overall brain function level and avoiding intervention bias caused by relying solely on the inhibition index. In other words, if the initial classification result is "moderate cognitive impairment", then after intervention, the "post-intervention neonatal brain response characteristics" need to be re-entered into the model (subsequent iteration process) to observe whether the classification result transforms into "mild cognitive impairment" or "normal"—if it transforms, it indicates that the intervention is effective; if it remains "moderate cognitive impairment" or worsens, then the intervention parameters need to be adjusted (such as switching to combined audio-visual intervention); even if the result is marked as "to be verified", its corresponding probability distribution (such as "mild cognitive impairment" probability 0.6, "normal" probability 0.55) can also serve as a reference for the intervention effect: if the probability of "normal" in some prognoses increases to 0.7, it indicates that the intervention promotes the development of brain function in a benign direction, providing a trend basis for subsequent iterative classification. If the preliminary classification result is "severe functional impairment", it usually indicates that the patient's brain metabolic capacity is weak. The safe threshold for brain oxygen saturation should be appropriately increased (e.g., from 50% to 55%), and the monitoring interval should be shortened (e.g., from 1 minute / time to 30 seconds / time) to avoid the risk of brain hypoxia. If the preliminary classification result is "normal", then monitoring can be performed at the conventional safe threshold (e.g., 50%) to reduce unnecessary consumption of monitoring resources.

[0095] S403. Based on the preliminary classification results, retrieve the individualized stepwise awakening strategy table formulated for the patient, compare the concealment inhibition index with the individualized stepwise awakening strategy table to determine the intervention parameters, and start the target-controlled intervention device to perform the intervention; after the intervention lasts for a period of time, re-collect the patient's P300 potential and brain oxygen data, and after preprocessing, iteratively calculate to obtain the neonatal brain response characteristics after the intervention.

[0096] It should be noted that this strategy table is formulated based on individual information such as the patient's preoperative underlying diseases, operation duration, and anesthetic drug dosage, and includes intervention parameters corresponding to different intervals of the occult brain function inhibition index. The occult brain function inhibition index obtained in step S303 is compared with this strategy table to determine the matching intervention parameters. The target-controlled intervention device is activated to perform the intervention operation according to the determined intervention parameters. During the intervention, the patient's brain oxygen saturation changes are monitored in real time. When the brain oxygen saturation is lower than the set safety threshold, the intervention operation is paused. After the intervention operation continues for a preset time interval, the patient's P300 potential and brain oxygen saturation spatiotemporal distribution data are re-collected, and these newly collected data are processed according to the same preprocessing procedure as in step S102 to finally obtain the post-intervention neonatal brain response characteristics. The dimension of the post-intervention neonatal brain response characteristics is consistent with the fusion feature vector obtained in step S302 for subsequent matching and evaluation.

[0097] S500: Combining the preliminary classification results with the post-intervention neonatal brain response characteristics, and after verification and matching processes, the final postoperative brain function status assessment results are obtained. The verification process includes combining the preliminary results of images, electrophysiological data, and clinical diagnosis within a preset postoperative time interval. The matching process includes re-aligning the post-intervention response characteristics with the immediate postoperative baseline and matching them with the preoperative baseline. If the verification is consistent and the matching is successful, the preliminary classification result is output as the final assessment result. If any step fails, the iterative calculation is repeated until the final assessment result that is consistent with the clinical diagnosis and the baseline is successfully matched is obtained.

[0098] It is understood that step S500 includes S501, S502, and S503, wherein:

[0099] S501. Collect dynamic imaging follow-up data and neurophysiological monitoring data of the patient within a preset time window after surgery, and combine them with the comprehensive diagnostic opinions of clinicians to form a clinical verification basis; verify the consistency between the obtained preliminary classification results of postoperative brain function status and the clinical verification basis, and use the consistency evaluation coefficient of the preset time interval to quantify the verification results; when the consistency evaluation coefficient reaches the set standard, it is determined that the preliminary classification results are consistent with the clinical verification basis; when the consistency evaluation coefficient does not reach the set standard, it is determined that the preliminary classification results are inconsistent with the clinical verification basis; the key formula for calculating the consistency evaluation coefficient is K=(P0 - Pe) / (1 - Pe), where K represents the consistency evaluation coefficient, P0 represents the observed consistency rate between the preliminary classification results and the clinical verification basis, and Pe represents the expected consistency rate between the preliminary classification results and the clinical verification basis. This formula can quantify the degree of consistency between the two.

[0100] S502. The post-intervention neonatal brain response characteristics and the immediate postoperative brain function baseline obtained above are re-aligned using the same mapping method as in step S203. The post-intervention-immediate postoperative offset is calculated. Subsequently, the post-intervention neonatal brain response characteristics are matched with the obtained preoperative individualized awake baseline. The matching degree between the two is calculated using a preset time interval matching degree calculation method. When the matching degree reaches the set standard, the post-intervention neonatal brain response characteristics and the preoperative individualized awake baseline are considered to have passed the matching. When the matching degree does not reach the set standard, the matching is considered to have failed. The matching degree calculation formula can quantify the degree of matching between the post-intervention brain function state and the preoperative awake state.

[0101] S503. A comprehensive judgment is made on the consistency verification result and the matching result of step S501: If the preliminary classification result of postoperative brain function status is consistent with the clinical verification basis, and the post-intervention neonatal brain response characteristics match the preoperative individualized awake baseline, the obtained preliminary classification result of postoperative brain function status is directly output and used as the final postoperative brain function status assessment result; if any of the above results does not meet the set standard, that is, the preliminary classification result is inconsistent with the clinical verification basis, or the post-intervention neonatal brain response characteristics do not match the preoperative individualized awake baseline, the post-intervention neonatal brain response characteristics obtained in step S403 are used as new input data, and the entire process of steps two to four is iteratively executed again, including feature extraction, functional offset calculation, feature fusion, occultation inhibition index generation, functional classification and closed-loop intervention, until a result that simultaneously satisfies the consistency verification and matching is obtained, and this result is used as the final postoperative brain function status assessment result.

[0102] Example 2:

[0103] like Figure 2 As shown, this embodiment provides a postoperative brain function status assessment system based on multimodal fusion. See [link to documentation]. Figure 2 The system includes:

[0104] Acquisition Module 701: Used to acquire multi-source raw data of postoperative patients, including near-infrared tissue oxygenation map, ultrasound micro-blood flow spectrum, event-related potentials of customized high-frequency chirped sound stimulation immediately after surgery, as well as resting-state functional magnetic resonance imaging data and high-density electroencephalogram data within a preset time interval after surgery. After targeted preprocessing, standardized multi-source data are obtained and a baseline of brain function is established immediately after surgery.

[0105] Extraction module 702: used to extract key brain network functional connections and preset time interval frequency band power spectral density features from multi-source raw data to form a multimodal feature subset. Based on the multimodal feature subset, nonlinear manifold alignment processing is performed on the postoperative immediate brain function baseline and the preoperative awake baseline under the same stimulation conditions to obtain the postoperative-preoperative functional offset.

[0106] Fusion module 703: Used to generate a fusion feature vector and a concealment inhibition index by using a subset of multimodal features and postoperative-preoperative functional offsets. The fusion processing includes determining feature weights by combining cross-validation with random forest, generating a weighted fusion feature vector, and simultaneously coupling compressed oxygen blood flow and latency information through a dual-channel gating network to obtain the concealment inhibition index.

[0107] Processing module 704: Based on the fused feature vector and the concealment inhibition index, it outputs the preliminary classification results of brain functional status and the response characteristics of newborn brain after intervention through classification processing and intervention processing. The classification processing uses a pre-trained deep neural network model to output multi-level functional impairment results, and the intervention processing drives the closed-loop intervention of sound, light, electricity and electricity by comparing the concealment inhibition index with the ladder awakening strategy table.

[0108] Assessment Module 705: This module combines the preliminary classification results with the post-intervention neonatal brain response characteristics, and after verification and matching processes, finally obtains the postoperative brain function status assessment results. The verification process includes combining the preliminary results of images, electrophysiological data, and clinical diagnosis verification within a preset postoperative time interval. The matching process includes re-aligning the post-intervention response characteristics with the immediate postoperative baseline and matching them with the preoperative baseline. If the verification is consistent and the matching is successful, the preliminary classification result is output as the final assessment result. If any step fails, the iterative calculation is repeated until a final assessment result consistent with the clinical diagnosis and with a successful baseline match is obtained.

[0109] Specifically, the extraction module 702 includes:

[0110] Integration Unit: This unit extracts core features from standardized multi-source data. This includes extracting time series data of key brain network nodes from corrected resting-state oxygenation level-dependent data using a brain network analysis toolkit and calculating the functional connectivity strength between nodes; obtaining power spectral density and EEG complexity index for preset time intervals from denoised continuous EEG data using an EEG analysis toolkit; and integrating the calculated functional connectivity strength, power spectral density, and EEG complexity index with the obtained coefficient of variation of brain oxygen saturation, blood flow velocity pulsatility index, and P300 parameters to form a multimodal functional feature subset. The formula for calculating functional connectivity strength is as follows:

[0111]

[0112] In the formula, This represents the degree of functional connection between two voxel nodes in the brain. This is the first voxel node number in the whole-brain voxels corresponding to the resting-state blood oxygenation level-dependent signal. t represents the second voxel node number in the whole-brain voxels corresponding to the resting-state oxygenation level-dependent signal, t represents the data acquisition time point number corresponding to the resting-state oxygenation level-dependent signal, and T represents the total number of data acquisition time points for the resting-state oxygenation level-dependent signal. For the preprocessed first The corrected resting-state oxygenation level at time t, dependent on the signal value. For the preprocessed first The corrected resting-state oxygenation level at time t, dependent on the signal value. For the first Individualized resting oxygenation levels at all T time points were calculated as a signal-mean value. For the first The signal-dependent mean of resting oxygen levels at all T time points after correction for individual values;

[0113] Preprocessing unit: Used to retrieve the patient's conscious brain function data within a pre-set time window and under the same stimulation conditions as after surgery, preprocess the conscious brain function data to obtain new conscious brain function data, and simultaneously calculate the corresponding brain oxygen saturation variation coefficient, blood flow velocity pulsatility index and P300 parameter to obtain the preoperative individualized conscious baseline, which includes 24-dimensional features corresponding to the same brain regions and parameter types;

[0114] The matching unit is used to perform cross-time point data matching between the pre-individualized awake baseline and the immediate postoperative brain function baseline. This includes constructing a Riemannian space framework, mapping the two types of baseline data to the Riemannian manifold space, and introducing a symmetric KL divergence constraint during the mapping process. This constraint minimizes the symmetric difference in probability distribution between preoperative and postoperative data in the manifold space while preserving the original brain function feature association structure of the two types of data, thus obtaining the postoperative and preoperative manifold coordinates. Based on a subset of multimodal functional features, the weighted sum of geodesic distances for the corresponding brain function dimensions in the Riemannian manifold space is calculated. This weighted sum quantifies the degree of difference between postoperative and preoperative brain function states, obtaining the postoperative-preoperative brain function offset. The formula for calculating the weighted sum of geodesic distances is as follows:

[0115]

[0116] In the formula, ΔB represents the postoperative-preoperative brain function shift. Let be the weight of the k-th dimension in the Riemannian manifold space, and d(·,·) be the geodesic distance calculation function in the Riemannian manifold space. Let be the value of the postoperative manifold coordinates in the k-th dimension. Let be the preoperative manifold coordinates in the k-th dimension, where k is the dimension of the Riemannian manifold space.

[0117] Specifically, the fusion module 703 includes:

[0118] The construction unit is used to determine the dynamic weights of each feature in a multimodal functional feature subset using a cross-validation combined with a multi-model fusion algorithm. This includes grouping the multimodal functional feature subset according to patient identification, selecting a subset as the training set, and the remaining subset as the validation set; constructing multiple different types of machine learning models, using the functional scores of historical patients within a preset postoperative time window as labels, and calculating the permutation importance of each feature through the validation set; taking a weighted average of the feature permutation importance output by different models and normalizing it to obtain the weight coefficients corresponding to each feature. The formula for calculating the permutation importance is as follows:

[0119] In the formula, For the permutation importance of the i-th feature, The accuracy obtained by training the model using the original features. The accuracy obtained by training the model after randomly permuting the i-th feature, where i is the feature number;

[0120] Fusion Unit: Used to perform spatiotemporal joint fusion processing on a subset of multimodal functional features based on the weight coefficients corresponding to each feature. In the spatiotemporal dimension, the mean of each feature within a preset time window and the trend slope within a preset time interval are weighted according to a set weight ratio. In the spatial dimension, features from different brain regions are grouped according to functional correlation. The features within each group are first summed, and then weighted according to the feature weight coefficients to generate a unified fusion feature vector. The calculation formula for the fusion feature vector is as follows:

[0121]

[0122] In the formula, To fuse feature vectors, Let be the weight coefficient of the i-th feature, and a and b be the weight ratios of the feature mean and the trend slope, respectively. Let be the mean of the i-th feature within a preset time interval window. Let i be the trend slope of the i-th feature within a preset time interval, where i is the feature number;

[0123] Extraction Unit: This unit extracts two types of key offset information from the postoperative-preoperative brain function offset. These two types of key offset information include brain oxygen metabolism offset derived from changes in the coefficient of variation of brain oxygen saturation, and cognitive potential offset derived from changes in P300 latency. Simultaneously, it retrieves a fusion feature vector with uniform dimensions, extracts feature weights corresponding to the coefficient of variation of brain oxygen saturation and P300 parameters, and uses these weights as priority coefficients for the offset information. A dual-channel gated attention network is used to couple and compress the two types of key offset information to finally obtain the hidden brain function inhibition index.

[0124] Specifically, the processing module 704 includes:

[0125] Training Unit: Used to load a pre-trained dual-branch deep neural network model, which includes a feature adaptation branch and a classification branch. The feature adaptation branch maps the fused feature vectors into feature maps that meet the input requirements of the classification branch through a dimensionality transformation operation over a preset time interval. The classification branch adopts a multi-layer fully connected network structure, using the fused feature vectors of historical patients and their corresponding postoperative brain function status labels as training data. An optimizer is used to train the model according to set parameters, and the model parameters are maximized through preset evaluation indicators. The postoperative brain function status labels include categories such as normal, mild cognitive impairment, moderate cognitive impairment, and severe functional impairment.

[0126] The computing unit is used to input the fused feature vector into the trained dual-branch deep neural network model. After feature adaptation branch dimension expansion and enhancement, the classification branch calculates the probability of each brain functional state through the Softmax function, takes the category with the highest probability as the preliminary classification result, and outputs the confidence score simultaneously. When the confidence score is lower than the set threshold, the classification result is marked as pending verification.

[0127] Iterative Unit: Based on the preliminary classification results, it retrieves the individualized stepwise awakening strategy table formulated for the patient, compares the concealment inhibition index with the individualized stepwise awakening strategy table to determine the intervention parameters, and starts the target-controlled intervention device to perform the intervention; after the intervention lasts for a period of time, the patient's P300 potential and brain oxygen data are collected again, and after preprocessing, iterative calculation is performed to obtain the neonatal brain response characteristics after the intervention.

[0128] Specifically, the evaluation module 705 includes:

[0129] The first judgment unit is used to collect dynamic imaging follow-up data, neurophysiological monitoring data and clinical verification data within a preset time window after the operation of the patient. The Kappa coefficient is used to verify the consistency between the preliminary classification results of the postoperative brain function status and the clinical verification data. If the coefficient reaches the standard, it is judged to be consistent, otherwise it is inconsistent.

[0130] The second judgment unit is used to align the post-intervention neonatal brain response characteristics with the immediate postoperative brain function baseline using the Riemann manifold mapping method and calculate the post-intervention-immediate postoperative offset; then, it matches the post-intervention neonatal brain response characteristics with the preoperative individualized awake baseline and calculates the matching degree. If the matching degree reaches the set standard, it passes; otherwise, it fails.

[0131] Comprehensive Judgment Unit: Used to comprehensively judge the consistency verification and matching results: If both pass, the preliminary classification result is output as the final postoperative brain function status assessment result; if either fails, the new brain response characteristics after intervention are used as new inputs, and the calculation is iterated again until both pass, and the final postoperative brain function status assessment result is output.

[0132] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0133] Example 3:

[0134] Corresponding to the above method embodiments, this embodiment also provides a postoperative brain function status assessment device based on multimodal fusion. The postoperative brain function status assessment device based on multimodal fusion described below and the postoperative brain function status assessment method based on multimodal fusion described above can be referred to in correspondence.

[0135] Figure 3 This is a block diagram illustrating a postoperative brain function status assessment device 800 based on multimodal fusion, according to an exemplary embodiment. Figure 3 As shown, the postoperative brain function assessment device 800 based on multimodal fusion includes a processor 801 and a memory 802. The device also includes one or more of a multimedia component 803, an I / O interface 804, and a communication component 805.

[0136] The processor 801 controls the overall operation of the multimodal fusion-based postoperative brain function assessment device 800 to complete all or part of the steps in the aforementioned multimodal fusion-based postoperative brain function assessment method. The memory 802 stores various types of data to support the operation of the multimodal fusion-based postoperative brain function assessment device 800. This data may include, for example, instructions for any application or method operating on the multimodal fusion-based postoperative brain function assessment device 800, as well as application-related data such as contact data, sent and received messages, images, audio, video, etc. The memory 802 can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in the memory 802 or transmitted via the communication component 805. The audio component also includes at least one speaker for outputting audio signals. I / O interface 804 provides an interface between processor 801 and other interface modules, such as a keyboard, mouse, or buttons. These buttons can be virtual or physical. Communication component 805 is used for wired or wireless communication between the multimodal fusion-based postoperative brain function assessment device 800 and other devices. Wireless communication includes, for example, Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof. Therefore, the corresponding communication component 805 may include a Wi-Fi module, a Bluetooth module, or an NFC module.

[0137] In an exemplary embodiment, the postoperative brain function status assessment device 800 based on multimodal fusion may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described postoperative brain function status assessment method based on multimodal fusion.

[0138] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided. When executed by a processor, these program instructions implement the steps of the multimodal fusion-based postoperative brain function state assessment method described above. For example, the computer-readable storage medium may be the memory 802 including the program instructions described above. These program instructions may be executed by the processor 801 of the multimodal fusion-based postoperative brain function state assessment device 800 to complete the multimodal fusion-based postoperative brain function state assessment method described above.

[0139] Example 4:

[0140] Corresponding to the above method embodiments, this embodiment also provides a readable storage medium. The readable storage medium described below can be referred to in conjunction with the postoperative brain function status assessment method based on multimodal fusion described above.

[0141] A computer program is stored on a readable storage medium, and when the computer program is executed by a processor, it implements the steps of the postoperative brain functional status assessment method based on multimodal fusion in the above method embodiments.

[0142] Specifically, the readable storage medium can be a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, or any other readable storage medium capable of storing program code.

[0143] In summary, this invention extracts key brain network functional connectivity features (such as the default mode network and salient network connectivity strength) and power spectral density features (such as the δ / θ band) from multiple sources of raw data, including spatiotemporal distribution maps of brain oxygen saturation, microflow spectra of the middle cerebral artery, P300 component potentials, resting oxygen level-dependent signals, and continuous EEG signals. Simultaneously, it integrates parameters such as the coefficient of variation of brain oxygen saturation and the pulsatility index of blood flow velocity to construct a multimodal feature subset. This overcomes the limitations of existing technologies that rely solely on single EEG or brain oxygen data, achieving synergistic integration of multidimensional information on brain metabolism, hemodynamics, electrophysiology, and cognitive potentials. This results in more comprehensive feature coverage, effectively reduces the interference of single data noise on the assessment results, and improves the accuracy of postoperative brain function status characterization. Based on the brain function dimension weights determined by the multimodal feature subset, a nonlinear manifold alignment algorithm is used to perform cross-time point data matching between the immediate postoperative brain function baseline and the patient's preoperative awake baseline under the same stimulation conditions, rather than using the universal baseline of healthy individuals used in existing technologies. By mapping two baselines to a low-dimensional manifold space and calculating a distance-weighted sum, a precise postoperative-preoperative functional shift is obtained. This effectively identifies "hidden functional deterioration after surgery in patients with preoperative mild cognitive impairment," which is easily missed by existing technologies, thus improving the individualized accuracy of functional shift calculation. Using historical postoperative functional scores as labels, the importance of feature permutations is calculated and normalized through cross-validation combined with a multi-model fusion algorithm to obtain dynamic weights. Then, the mean and trend slope of features are fused according to weight ratios in the time dimension, and features are fused according to brain function correlation in the spatial dimension. This differs from the simple weighted summation method of existing technologies, assigning core features (such as P300) with dynamic weights. Prioritizing latency and cerebral oxygen metabolism shifts, spatiotemporal fusion further aligns with the physiological laws of "dynamic change + network collaboration" in brain function. The resulting dimensionally unified fusion feature vector retains key information while avoiding dimensional redundancy, providing high-quality input for subsequent classification models and improving the reliability of brain function state classification. The fused feature vector is input into a pre-trained dual-branch deep neural network, achieving preliminary classification through feature adaptation branch enhancement and classification branch probability calculation. Simultaneously, combined with concealment inhibition index matching, an individualized stepwise awakening strategy based on the patient's preoperative underlying diseases and surgical duration is implemented to initiate target-controlled device intervention. Finally, clinical validation and post-intervention feature matching degree verification are performed; if the verification fails, iterative calculation is performed. This breaks through the limitations of existing technologies that "only assess without intervention" or "intervention relies on human experience," constructing a complete closed loop. This achieves rapid and automatic classification of brain function states, dynamically adjusts intervention parameters based on individual patient conditions, and ensures assessment and intervention effectiveness through iterative verification, significantly improving the clinical practicality and safety of postoperative brain function management.

[0144] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0145] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A postoperative brain function status assessment system based on multimodal fusion, characterized in that, include: Acquisition Module: Used to acquire multi-source raw data from postoperative patients, including spatiotemporal distribution map of brain oxygen saturation, microblood flow spectrum of middle cerebral artery, P300 component potential induced by customized high-frequency chirping stimulation, and resting oxygen level-dependent signals and continuous EEG signals acquired within a preset time interval after surgery. After targeted preprocessing, standardized multi-source data are obtained and an immediate postoperative brain function baseline is established. Extraction module: Used to extract the time series of brain network nodes and the functional connectivity strength between nodes from multi-source raw data, the frequency band power spectral density and EEG complexity index in the preset time interval, the coefficient of variation of brain oxygen saturation, the blood flow velocity pulsatility index and the P300 component potential related parameters, forming a multimodal functional feature subset. Based on the multimodal functional feature subset, nonlinear manifold alignment processing is performed on the postoperative immediate brain function baseline and the preoperative awake baseline under the same stimulation conditions to obtain the postoperative-preoperative functional offset. Fusion module: Used to generate fusion feature vector and concealment inhibition index by using a subset of multimodal functional features and postoperative-preoperative functional offset. The fusion processing includes determining feature weights by combining cross-validation with random forest, generating fusion feature vector by weighting, and obtaining concealment inhibition index by coupling compressed oxygen blood flow and latency information through a dual-channel gating network. Processing module: Based on the fused feature vector and the concealment inhibition index, it outputs the preliminary classification results of brain functional status and the response characteristics of newborn brain after intervention through classification and intervention processing. The classification processing uses a pre-trained deep neural network model to output multi-level functional impairment results, and the intervention processing drives the closed-loop intervention of sound, light, electricity and electricity by comparing the concealment inhibition index with the ladder awakening strategy table. The assessment module combines the preliminary classification results of brain function status with the response characteristics of the newborn brain after intervention. After verification and matching, the final assessment result of postoperative brain function status is obtained. The verification process includes combining the preliminary results of images, electrophysiological data and clinical diagnosis verification within a preset time interval after surgery. The matching process includes re-aligning the response characteristics after intervention with the immediate postoperative brain function baseline and matching it with the preoperative awake baseline under the same stimulation conditions. If the verification is consistent and the matching is successful, the preliminary classification result of brain function status is output as the postoperative brain function status assessment result. If either fails, the new brain response characteristics after intervention are used as new inputs, and the calculation is iterated again until both pass, and the final postoperative brain function status assessment result is output. The acquisition module obtains multi-source raw data, including postoperative spatiotemporal distribution maps of brain oxygen saturation, micro-blood flow spectra of the middle cerebral artery, P300 component potentials induced by customized high-frequency chirping stimulation, and resting-state oxygenation level-dependent signals and continuous EEG signals acquired within a preset time interval after surgery. Targeted preprocessing operations are performed on the acquired multi-source raw data: the spatiotemporal distribution maps of brain oxygen saturation are spatially smoothed using a Gaussian filter of the corresponding size to obtain preprocessed smoothed brain oxygen saturation data; simultaneously, the coefficient of variation of brain oxygen saturation is calculated, and the micro-blood flow spectra of the middle cerebral artery are processed using a high-pass filter with a preset time interval cutoff frequency. The data is processed using filters, and the blood flow velocity pulsatility index is calculated to obtain high-pass data of microflow in the middle cerebral artery. For the P300 component potentials, bandpass filtering within a preset time interval frequency band combined with baseline correction is applied, and the latency and amplitude of the P300 component potentials are extracted to obtain P300 component potential bandpass data. For resting-state oxygenation level-dependent signals, a professional brain imaging processing toolkit is used to perform head motion correction, spatial normalization, delinearization, and low-frequency filtering. For continuous EEG signals, independent component analysis is used to remove artifacts, and notch filtering is superimposed to eliminate power line interference. The key formula for calculating the coefficient of variation of brain oxygen saturation is: In the formula, This represents the smoothed brain oxygen saturation data; std (·) represents the standard deviation calculation function; and mean (·) represents the mean calculation function. By integrating preprocessed smoothed cerebral oxygen saturation data, high-pass data of microvascular flow in the middle cerebral artery, and bandpass data of P300 component potentials, the mean values ​​of each parameter within a preset time window after surgery were calculated, and data points with fluctuations exceeding a reasonable range were removed to form an immediate postoperative baseline of brain function. Simultaneously, preprocessed resting-state oxygenation level-dependent correction data and continuous EEG denoising data were normalized using a standardization method, and then combined with the integrated data to construct a unified multimodal spatiotemporal aligned data matrix, ultimately forming standardized multi-source data. The specific steps in the fusion module to obtain the concealment inhibition index by coupling and compressing oxygen blood flow and latency information through a dual-channel gating network include: extracting two types of key offset information from the postoperative-preoperative functional offset; the two types of key offset information include cerebral oxygen metabolism offset originating from changes in the coefficient of variation of cerebral oxygen saturation, and cognitive potential offset originating from changes in latency of P300 component potentials; simultaneously retrieving a fusion feature vector with uniform dimensions, extracting the feature weights corresponding to the parameters of the coefficient of variation of cerebral oxygen saturation and P300 component potentials, and using them as priority coefficients for the offset information; and using a dual-channel gating attention network to couple and compress the two types of key offset information to finally obtain the concealment inhibition index.

2. The postoperative brain function status assessment system based on multimodal fusion according to claim 1, characterized in that, The extraction module includes: Integration Unit: This unit extracts core features from standardized multi-source data. This includes extracting time series data of key brain network nodes from corrected resting-state oxygenation level-dependent data using a brain network analysis toolkit and calculating the functional connectivity strength between nodes; obtaining power spectral density and EEG complexity index for preset time intervals from denoised continuous EEG data using an EEG analysis toolkit; and integrating the calculated functional connectivity strength, power spectral density, and EEG complexity index with parameters such as the obtained coefficient of variation of brain oxygen saturation, blood flow velocity pulsatility index, and P300 component potentials to form a multimodal functional feature subset. The formula for calculating functional connectivity strength is as follows: In the formula, This represents the degree of functional connection between two voxel nodes in the brain. This is the first voxel node number in the whole-brain voxels corresponding to the resting-state blood oxygenation level-dependent signal. t represents the second voxel node number in the whole-brain voxels corresponding to the resting-state oxygenation level-dependent signal, t represents the data acquisition time point number corresponding to the resting-state oxygenation level-dependent signal, and T represents the total number of data acquisition time points for the resting-state oxygenation level-dependent signal. For the preprocessed first The corrected resting-state oxygenation level at time t, dependent on the signal value. For the preprocessed first The corrected resting-state oxygenation level at time t, dependent on the signal value. For the first Individualized resting oxygenation levels at all T time points were calculated as a signal-mean value. For the first The signal-dependent mean of resting oxygen levels at all T time points after correction for individual values; Preprocessing unit: Used to retrieve the patient's conscious brain function data within a pre-set time window and under the same stimulation conditions as after surgery, preprocess the conscious brain function data to obtain new conscious brain function data, and simultaneously calculate the corresponding parameters of brain oxygen saturation variation coefficient, blood flow velocity pulsatility index and P300 component potential, thereby obtaining the conscious baseline under the same stimulation conditions before surgery. The conscious baseline under the same stimulation conditions before surgery includes 24-dimensional features, corresponding to the same brain regions and parameter types. The matching unit is used to perform cross-time point data matching between the preoperative awake baseline under the same stimulus conditions and the immediate postoperative brain function baseline. This includes constructing a Riemannian space framework, mapping the two types of baseline data to the Riemannian manifold space, and introducing a symmetric KL divergence constraint during the mapping process. This constraint minimizes the symmetric difference in probability distribution between preoperative and postoperative data in the manifold space while preserving the original brain function feature association structure of the two types of data, thus obtaining the postoperative and preoperative manifold coordinates. Based on a subset of multimodal functional features, the geodesic distance weighted sum of the corresponding brain function dimensions in the Riemannian manifold space is calculated. This weighted sum quantifies the degree of difference between postoperative and preoperative brain function states, obtaining the postoperative-preoperative functional offset. The formula for calculating the geodesic distance weighted sum is as follows: In the formula, ΔB represents the postoperative-preoperative functional shift. Let be the weight of the k-th dimension in the Riemannian manifold space, and d(·,·) be the geodesic distance calculation function in the Riemannian manifold space. Let be the value of the postoperative manifold coordinates in the k-th dimension. Let be the preoperative manifold coordinates in the k-th dimension, where k is the dimension of the Riemannian manifold space.

3. The postoperative brain function status assessment system based on multimodal fusion according to claim 1, characterized in that, The fusion module includes: The construction unit is used to determine the dynamic weights of each feature in a multimodal functional feature subset using a cross-validation combined with a multi-model fusion algorithm. This includes grouping the multimodal functional feature subset according to patient identification, selecting a subset as the training set, and the remaining subset as the validation set; constructing multiple different types of machine learning models, using the functional scores of historical patients within a preset postoperative time window as labels, and calculating the permutation importance of each feature through the validation set; taking a weighted average of the feature permutation importance output by different models and normalizing it to obtain the weight coefficients corresponding to each feature. The formula for calculating the permutation importance is as follows: In the formula, For the permutation importance of the i-th feature, The accuracy obtained by training the model using the original features. The accuracy obtained by training the model after randomly permuting the i-th feature, where i is the feature number; Fusion Unit: Used to perform spatiotemporal joint fusion processing on a subset of multimodal functional features based on the weight coefficients corresponding to each feature. In the spatiotemporal dimension, the mean of each feature within a preset time window and the trend slope within a preset time interval are weighted according to a set weight ratio. In the spatial dimension, features from different brain regions are grouped according to functional correlation. The features within each group are first summed, and then weighted according to the feature weight coefficients to generate a unified fusion feature vector. The calculation formula for the fusion feature vector is as follows: In the formula, To fuse feature vectors, Let be the weight coefficient of the i-th feature, and a and b be the weight ratios of the feature mean and the trend slope, respectively. Let be the mean of the i-th feature within a preset time interval window. Let i be the trend slope of the i-th feature within a preset time interval, where i is the feature number; Extraction Unit: This unit extracts two types of key offset information from the postoperative-preoperative functional offset. These two types of key offset information include brain oxygen metabolism offset, which originates from changes in the coefficient of variation of brain oxygen saturation, and cognitive potential offset, which originates from changes in the latency of P300 component potentials. Simultaneously, it retrieves a fusion feature vector with uniform dimensions and extracts feature weights corresponding to the parameters of the coefficient of variation of brain oxygen saturation and P300 component potentials, which are then used as priority coefficients for the offset information. A dual-channel gated attention network is used to couple and compress the two types of key offset information to finally obtain the concealment inhibition index.

4. The postoperative brain function status assessment system based on multimodal fusion according to claim 1, characterized in that, The processing module includes: Training Unit: Used to load a pre-trained dual-branch deep neural network model, which includes a feature adaptation branch and a classification branch. The feature adaptation branch maps the fused feature vectors into feature maps that meet the input requirements of the classification branch through a dimensionality transformation operation over a preset time interval. The classification branch adopts a multi-layer fully connected network structure, using the fused feature vectors of historical patients and their corresponding postoperative brain function status labels as training data. An optimizer is used to train the model according to set parameters, and the model parameters are maximized through preset evaluation indicators. The postoperative brain function status labels include categories such as normal, mild cognitive impairment, moderate cognitive impairment, and severe functional impairment. The computing unit is used to input the fused feature vector into the trained dual-branch deep neural network model. After feature adaptation branch dimension expansion and enhancement, the classification branch calculates the probability of each brain functional state through the Softmax function. The category with the highest probability is taken as the preliminary classification result of the brain functional state, and the confidence score is output simultaneously. When the confidence score is lower than the set threshold, the preliminary classification result is marked as pending verification. Iterative Unit: Based on the preliminary classification results of brain functional status, it retrieves the individualized stepwise awakening strategy table formulated for the patient, compares the concealed inhibition index with the individualized stepwise awakening strategy table to determine the intervention parameters, and starts the target-controlled intervention device to perform the intervention; after the intervention lasts for a period of time, the patient's P300 component potential and brain oxygen data are collected again, and after preprocessing, iterative calculation is performed to obtain the neonatal brain response characteristics after the intervention.

5. The postoperative brain function status assessment system based on multimodal fusion according to claim 1, characterized in that, The evaluation module includes: The first judgment unit is used to collect dynamic imaging follow-up data, neurophysiological monitoring data and clinical verification data within a preset time window after surgery. The Kappa coefficient is used to verify the consistency between the preliminary classification results of brain function status and the clinical verification data. If the coefficient meets the standard, it is judged to be consistent; otherwise, it is inconsistent. The second judgment unit is used to align the post-intervention neonatal brain response characteristics with the postoperative brain function baseline using the Riemann manifold mapping method and calculate the post-intervention-postoperative offset; then, it matches the post-intervention neonatal brain response characteristics with the preoperative awake baseline under the same stimulus conditions and calculates the matching degree. If the matching degree reaches the set standard, it passes; otherwise, it fails. Comprehensive Judgment Unit: Used to comprehensively judge the consistency verification and matching results: If both pass, the preliminary classification result of brain functional status is output as the final postoperative brain functional status assessment result; if either fails, the response characteristics of the newborn brain after intervention are used as new inputs, and the calculation is iterated again until both pass, and the final postoperative brain functional status assessment result is output.

Images