Neurological pathological feature analysis method and system

Abstract

The invention discloses a neurological pathological feature analysis method and system. The method comprises the following steps: collecting multi-modal time sequence physiological information and a symptom text diary of a target individual; preprocessing and feature fusion are carried out on the multi-modal time sequence physiological information, time domain, frequency domain and nonlinear entropy features are extracted, and cross-modal physiological state feature vectors are obtained; performing natural language processing on the symptom text diary to generate a symptom intensity sequence and an emotion label sequence; constructing a supervised and unsupervised two-way analysis model based on the cross-modal physiological state feature vector, and respectively outputting an individualized symptom risk prediction value and a population deviation degree; performing time sequence consistency adjustment and feature attribution analysis on double-path output through a gating circulation unit; and finally generating an integrated neural state evaluation report. According to the method, individual-group dual verification under multi-modal data fusion is realized, and the robustness, interpretation consistency and personalized early warning precision of neural state evaluation are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical artificial intelligence and neurological data analysis, and particularly relates to a neurological pathological feature analysis method and system. BACKGROUND

[0002] Neurological pathological feature analysis is an important direction for evaluating and monitoring the functional state of the nervous system, especially in the early identification and long-term management of common psychosomatic diseases such as anxiety and depression. At present, the research and practice in this field mainly rely on a single type of data source, such as obtaining the subjective symptom description of the patient through a scale, or collecting fragmentary physiological indicators using medical equipment. Although this method can reflect part of the state information, it is difficult to fully capture the dynamic interaction between physiological and psychological elements in the development process of the disease. In addition, due to significant individual differences, simply relying on group benchmarks or general models for analysis often cannot accurately adapt to the specific performance of different users, resulting in insufficient sensitivity and relevance of the evaluation results. In practical applications, how to effectively integrate multi-source, heterogeneous time series data and build a robust analysis model that can take into account individualized prediction and group distribution reference is still a major challenge for current technology. Therefore, it is urgent to explore new technical approaches to improve the accuracy and practicality of neurological health status management. SUMMARY

[0003] In view of the above problems, the present application provides a neurological pathological feature analysis method and system, which realizes the complementary enhancement of individualized prediction and group benchmark verification by fusing multi-modal time series data and building a supervised and unsupervised dual-path collaborative analysis model, and solves the problem of insufficient reliability of existing neurological state evaluation results.

[0004] To achieve the above purpose, in a first aspect, the present application provides a neurological pathological feature analysis method, comprising: Collecting multi-modal time series physiological information of a target individual, the multi-modal time series physiological information including autonomous nervous activity indicators and sleep structure indicators obtained through wearable devices, and simultaneously collecting symptom text diaries actively reported by the user as subjective annotation information; Preprocessing and feature fusion of the multi-modal time series physiological information, using a dynamic time warping algorithm to align heterogeneous time series signals, and extracting time domain, frequency domain and nonlinear entropy features to obtain a cross-modal physiological state feature vector in units of individuals; Natural language processing of the symptom text diary, using a pre-training model based on Transformer for symptom entity recognition and sentiment polarity analysis to generate a structured symptom intensity sequence and a sentiment label sequence; Based on the cross-modal physiological state feature vector, a supervised and unsupervised dual-channel analysis model is constructed, the dual-channel analysis model includes a first channel and a second channel, the first channel adopts a long short-term memory network to establish a dynamic mapping relationship between physiological indicators and symptom labels, and outputs an individualized symptom risk prediction value, which is recorded as a first channel prediction result; the second channel adopts a one-class support vector machine to construct an abnormality detection model based on the baseline distribution of the group, and calculates the deviation degree of the current physiological mode relative to the group norm, which is recorded as a second channel prediction result; The output results of the dual-channel analysis model are adjusted for time sequence consistency by a gated recurrent unit. When the first channel prediction result and the second channel prediction result significantly differ, a feature attribution analysis based on attention weights is performed to identify the core physiological dimension causing the difference, and an attribution analysis result is output. An integrated neural state evaluation report is generated according to the first channel prediction result, the second channel prediction result, and the attribution analysis result.

[0005] In some embodiments, the multi-modal time-series physiological information is pre-processed and feature-fused, a dynamic time warping algorithm is used to align heterogeneous time-series signals, and time-domain, frequency-domain and nonlinear entropy features are extracted to obtain cross-modal physiological state feature vectors in units of individuals, including: The autonomic nervous activity indicators are subjected to heart rate variability analysis, and the SDNN and RMSSD indicators in the time-domain features, the low-frequency and high-frequency power ratio indicators in the frequency-domain features, and the autonomic nervous regulation stability indicators in the nonlinear features are extracted by a sample entropy algorithm; The sleep structure indicators are subjected to sleep staging feature extraction, the duration proportions of deep sleep and rapid eye movement sleep are obtained, and the sleep efficiency and the number of awakenings are calculated to obtain sleep quantitative indicators; The dynamic time warping algorithm is used to time-align the above-mentioned autonomic nervous feature sequence and sleep feature sequence to eliminate time sequence deviation caused by differences in collection time points; The aligned multi-modal features and electroencephalogram rhythm power features are spliced through feature-level fusion, and principal component analysis is used for dimension reduction processing to finally generate a cross-modal physiological state feature vector.

[0006] In some embodiments, the symptom text diary is subjected to natural language processing, a pre-training model based on Transformer is used for symptom entity recognition and sentiment polarity analysis to generate structured symptom intensity sequences and sentiment label sequences, including: An independent dictionary for the field of neurological symptoms is constructed, the independent dictionary includes a somatization symptom description word set, an emotion state keyword set, and a degree adverb mapping table; A pre-trained language model based on the Transformer architecture is used as a basic feature extractor to encode the context semantics of the symptom text diary and obtain a deep semantic representation vector for each word. Based on the encoded deep semantic representation vector, a conditional random field model is used for named entity recognition to extract specific symptom entities and their modification relationships, including palpitations, headaches, and gastrointestinal discomfort. The identified symptom entities and their modification relationships are subjected to sentiment polarity analysis, and the text sentiment tendency score is calculated through a multi-head attention mechanism, and the quantified symptom intensity value is generated in combination with the degree adverb mapping table. Based on the timing information, the symptom intensity values are sorted, and the symptom intensity values of the same time period are aggregated into structured records to generate a symptom intensity sequence and a sentiment label sequence containing symptom type, intensity value, and sentiment polarity.

[0007] In some embodiments, based on the cross-modal physiological state feature vector, a supervised and unsupervised dual-path analysis model is constructed, including: A supervised analysis model framework is constructed, taking the cross-modal physiological state feature vector as the model input, and taking the symptom intensity sequence and the sentiment label sequence as the supervised signal. A deep neural network architecture is used to establish a mapping relationship from physiological features to symptom manifestations. An unsupervised analysis model framework is constructed, taking the cross-modal physiological state feature vector of a large-scale population sample as training data, and using a density clustering algorithm to establish a feature distribution space of the population norm, and defining an abnormal score mechanism based on distance measurement. The numerical range output by the supervised analysis model and the unsupervised analysis model is standardized and calibrated to ensure the comparability of the results of the two analysis paths. When new multi-modal time-series physiological information is collected, an online learning algorithm is used to update the parameters of the supervised analysis model, and an incremental learning algorithm is used to adjust the feature distribution space of the population norm for the unsupervised analysis model.

[0008] In some embodiments, the first path uses a long short-term memory network to establish a dynamic mapping relationship between physiological indicators and symptom labels, outputting individualized symptom risk prediction values, denoted as the first path prediction result, including: Taking the cross-modal physiological state feature vector as the input sequence, a long short-term memory network model based on the gating mechanism is constructed, which captures the long-term temporal dependence relationship between physiological indicators through the coordinated control of the forget gate, input gate and output gate. Taking the symptom intensity sequence and the sentiment label sequence as the supervised signal, a time-distributed fully connected layer is used to map the hidden state output by the long short-term memory network to the symptom risk space, generating individualized symptom risk prediction values synchronized with the input sequence. Adjust the weight distribution of the symptom intensity prediction task and the emotion polarity classification task, and update the long short-term memory network parameters through the back-propagation algorithm; A sliding window mechanism is used to process continuous time series data, and the latest individualized symptom risk prediction value is output in real time, which is recorded as the first prediction result.

[0009] In some embodiments, the second path uses a first-class support vector machine to construct an anomaly detection model based on the group baseline distribution, calculates the deviation of the current physiological pattern from the group norm, and records it as the second path prediction result, including: Collect cross-modal physiological state feature vectors of samples from a preset group as a training dataset, and construct a physiological norm representing the relatively stable state of the preset group. The preset group is configured as a group diagnosed with mild anxiety or somatization disorder through clinical assessment and in a clinically stable period; A first-class support vector machine algorithm is used to estimate the distribution density of the cross-modal physiological state feature vectors of the samples, and a decision hyperplane is found through nonlinear kernel function mapping to define the main feature distribution space of the stable preset group. Input the current user's cross-modal physiological state feature vector into the trained anomaly detection model, calculate the relative distance between the current user's cross-modal physiological state feature vector and the decision hyperplane, and quantify the relative distance as the deviation of the current physiological mode from the group stable state norm; The deviations are normalized according to the distribution characteristics of the group steady-state norm to generate a second-path prediction result that can be compared across individuals.

[0010] In some embodiments, adjusting the output of the two-way analysis model for timing consistency through a gated recurrent unit includes: Align and concatenate the first-path prediction results and the second-path prediction results in time series to form a dual-path prediction result sequence as input features; Construct a gated recurrent unit network and learn the temporal dependencies and complementary characteristics between the two-way prediction results through the update gate and reset gate mechanism; The attention mechanism is used to calculate the weight distribution of historical prediction results, enhancing the influence weight of historical information that is highly relevant to the current state; Output the two-way prediction result sequence after temporal consistency adjustment.

[0011] In some embodiments, when there is a significant divergence between the first-path prediction result and the second-path prediction result, a feature attribution analysis based on attention weights is performed to identify the core physiological dimension causing the divergence, and the attribution analysis results are output, including: Based on the two-path prediction result sequence adjusted for timing consistency, the relative entropy value of the two-path prediction results in adjacent time windows is calculated, and when the relative entropy of three consecutive time windows all exceeds the adaptive threshold, it is determined as a significant divergence state; A multi-level feature attribution analysis is started, and at the first level, an integrated gradient algorithm is used to calculate the contribution of each input feature dimension in the long short-term memory network to the current prediction result, to generate a first path feature importance distribution; At the second level, a Shapley value estimation algorithm based on a kernel function is used to analyze the marginal contribution of each feature dimension in the support vector machine model to the abnormal score, to generate a second path feature importance distribution; A feature contribution difference matrix is constructed to compare the first path feature importance distribution and the second path feature importance distribution, and the top K feature dimensions with the largest contribution difference are identified as core divergence features, denoted as the identification result; An attribution analysis result is generated based on the identification result, including the divergence time period, the core feature dimension list, the contribution difference value, and the clinical significance explanation.

[0012] In some embodiments, an integrated neural state evaluation report is generated based on the first path prediction result, the second path prediction result, and the attribution analysis result, including: The symptom risk trend of the first path prediction result, the physiological abnormality degree of the second path prediction result, and the core feature dimension of the attribution analysis result are weighted and fused; According to the consistency degree of the two-path prediction results, a corresponding evaluation mode is selected, including: When the two-path prediction results are consistent, a cooperative confirmation mode is used; When the two-path prediction results are divergent, an attribution explanation mode is used; A structured evaluation result is generated, including symptom risk level, physiological abnormality index, key influencing factor analysis, and confidence level; According to the structured evaluation result, personalized intervention suggestions are generated, including early warning prompts, lifestyle adjustment suggestions, and professional medical guidance; According to the structured evaluation result and the personalized intervention suggestions, an integrated neural state evaluation report is generated.

[0013] In the second aspect, the present application also provides a neurological pathology feature analysis system suitable for the method of the first aspect.

[0014] Differing from the prior art, the technical scheme has the following advantages: the multi-modal time-series physiological information and symptom text diary of a target individual are collected; the multi-modal time-series physiological information is preprocessed and feature fusion is performed to extract time domain, frequency domain and nonlinear entropy features, thereby obtaining a cross-modal physiological state feature vector; the symptom text diary is subjected to natural language processing to generate a symptom intensity sequence and an emotion label sequence; a supervised and unsupervised dual-path analysis model is constructed based on the cross-modal physiological state feature vector, and individualized symptom risk prediction values and group deviation degrees are respectively output; the dual-path outputs are subjected to time-series consistency adjustment and feature attribution analysis through a gated recurrent unit; and finally, an integrated neural state evaluation report is generated. Through the collaborative analysis and time-series consistency adjustment of the supervised and unsupervised dual-path models, the individual-group dual verification under multi-modal data fusion is realized, and the robustness, explanation consistency and personalized early warning accuracy of the neural state evaluation are significantly improved.

[0015] The above summary of the invention is only a summary of the technical scheme of the present application, in order to enable those skilled in the art to more clearly understand the technical scheme of the present application, and then can be implemented according to the content of the description and the drawings, and in order to make the above-mentioned purpose and other purposes, characteristics and advantages of the present application can be more easily understood, the following is described in combination with the specific embodiments of the present application and the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0016] The drawings only serve to illustrate the principles, implementation modes, applications, characteristics and effects of the specific embodiments and other related contents of the present application, and cannot be considered as limiting the present application.

[0017] In the drawings of the specification: Figure 1 Step schematic diagram of steps S101 to S106 of the analysis method described in the specific embodiment; Figure 2 Step schematic diagram of steps S201 to S204 of the analysis method described in the specific embodiment; Figure 3 Step schematic diagram of steps S301 to S305 of the analysis method described in the specific embodiment. DETAILED DESCRIPTION

[0018] In order to explain the possible application scenarios, technical principles, specific schemes that can be implemented, and the purposes and effects that can be achieved, the following will be described in detail in combination with the specific examples listed and the drawings. The examples described herein are only used to more clearly illustrate the technical scheme of the present application, and therefore only serve as examples, and cannot limit the protection scope of the present application.

[0019] The term "embodiment" is mentioned herein means that the specific features, structures, or characteristics described in connection with the embodiment can be included in at least one embodiment of the present application. The term "embodiment" appearing in various places in the specification does not necessarily refer to the same embodiment, nor does it particularly limit the independence or association between other embodiments. In principle, in the present application, as long as there is no technical contradiction or conflict, the technical features mentioned in each embodiment can be combined in any way to form a corresponding implementable technical solution.

[0020] Unless otherwise defined, the meanings of the technical terms used herein are the same as those commonly understood by those skilled in the art to which the present application belongs; the use of related terms herein is only for the purpose of describing specific embodiments, and is not intended to limit the present application.

[0021] In the description of the present application, the phrase "and / or" is a description of the logical relationship between the objects, which means that there can be three relationships, for example, A and / or B, which means that there are three cases: A exists, B exists, and A and B exist at the same time. In addition, the character " / " herein generally represents that the associated objects before and after are a "or" logical relationship.

[0022] In the present application, the terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual quantity, primary and secondary or order relationship between the entities or operations.

[0023] In the present application, without more limitation, the "includes", "contains", "has" or other similar open expressions used in the sentence are intended to cover non-exclusive inclusion, and these expressions do not exclude the presence of other elements in the process, method or product including the elements, so that the process, method or product including a series of elements can not only include those limited elements, but also include other elements not explicitly listed, or also include the elements inherent to such process, method or product.

[0024] As the same as the understanding in the "Guidelines for Examination", in the present application, the expressions such as "greater than", "less than", "exceed" are understood as not including the number; the expressions such as "above", "below", "within" are understood as including the number. In addition, in the description of the embodiments of the present application, the meaning of "multiple" is more than two (including two), and similar expressions related to "multiple" are also understood in this way, for example, "multiple groups", "multiple times", etc., unless otherwise explicitly limited.

[0025] In the description of the embodiments of the present application, the spatial relative expressions used, such as "central", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "vertical", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential", and the like, indicate the orientation or positional relationship shown in the specific embodiments or the drawings, and are only used for the convenience of describing the specific embodiments of the present application or for the convenience of the reader to understand, and do not indicate or imply that the indicated device or component must have a particular position, a particular orientation, or be constructed or operated in a particular orientation, and therefore cannot be understood as a limitation on the embodiments of the present application.

[0026] The processor of the embodiments of the present application can be implemented by hardware, firmware, software or a combination thereof, and can use at least one of circuit, single or multiple Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), Central Processing Unit (CPU), controller, microcontroller, microprocessor, and other physical, biological or chemical structures that can realize the same or equivalent functions as the above-mentioned processors, such as biological neurons, quantum computing units, DNA computing units, etc., so that the processor can execute part or all of the steps or any combination of the steps mentioned in the computer programs or methods of the embodiments of the present application.

[0027] The computer program involved in the embodiments can be stored in a computer device readable storage medium, including but not limited to magnetic disk, magnetic tape, magnetic card, floppy disk, flash memory, optical disc, optical card, read-only memory (ROM), random access memory (RAM), erasable programmable ROM (EPROM) and electrically erasable programmable ROM (EEPROM), etc., and also includes other biological, physical or chemical structures that can realize similar or equivalent functions as the above-mentioned storage media, such as DNA, RNA, protein and other units with information storage ability, etc. In specific embodiments, the storage medium can be one of the above-mentioned medium types, or a combination of the above-mentioned medium types. In different embodiments, the computer program involved in the embodiments can be stored centrally in a single medium, or distributedly in multiple media. The storage medium containing the computer device readable storage medium can be a non-volatile memory or a random access memory. These computer device readable storage media can be built-in in the device, or connected with the device as an external device or part of the external device. In some embodiments, the storage medium with the computer device readable storage medium is deployed locally; in other embodiments, the storage medium can also be deployed remotely from the processor, such as network-attached storage accessed via RF circuit or external port and communication network, wherein the communication network can be the Internet, one or more intranets, a local area network (LAN), a wide area network (WAN), a storage area network (SAN) and the like, or a suitable combination thereof, as long as the computer device can access the storage medium. In addition, the computer program involved in the embodiments can be stored in plaintext / encrypted form, or can be designed as training data and integrated and reorganized by model training to be implicitly saved in the parameter state of a deep neural network or other machine learning model.

[0028] Please refer to Figure 1 In the first aspect, the embodiment provides a neurological pathology feature analysis method, comprising: S101, collecting multi-modal time-series physiological information of a target individual, the multi-modal time-series physiological information comprising autonomic nervous activity indicators and sleep structure indicators obtained by a wearable device, and simultaneously collecting a symptom text diary actively reported by the user as subjective annotation information; S102, pre-processing and feature fusion of the multi-modal time-series physiological information, aligning heterogeneous time-series signals by using a dynamic time warping algorithm, and extracting time domain, frequency domain and nonlinear entropy features to obtain a cross-modal physiological state feature vector in units of individuals; S103, natural language processing of the symptom text diary, performing symptom entity recognition and sentiment polarity analysis by using a pre-training model based on Transformer to generate a structured symptom intensity sequence and a sentiment label sequence; S104, based on the cross-modal physiological state feature vector, a supervised and unsupervised dual-channel analysis model is constructed respectively, the dual-channel analysis model includes a first channel and a second channel, the first channel adopts a long short-term memory network to establish a dynamic mapping relationship between physiological indicators and symptom labels, and outputs an individualized symptom risk prediction value, which is recorded as a first channel prediction result; the second channel adopts a one-class support vector machine to construct an abnormality detection model based on a population baseline distribution, and calculates a deviation degree of a current physiological mode relative to a population norm, which is recorded as a second channel prediction result; S105, the output results of the dual-channel analysis model are adjusted for time sequence consistency through a gated recurrent unit, when the first channel prediction result and the second channel prediction result significantly diverge, a feature attribution analysis based on attention weights is performed, a core physiological dimension causing the divergence is identified, and an attribution analysis result is output; S106, an integrated neural state evaluation report is generated according to the first channel prediction result, the second channel prediction result and the attribution analysis result.

[0029] In step S101, the target individual refers to a specific user who needs to be evaluated for the neural state. Multimodal time series physiological information is continuously collected by a wearable device, wherein the autonomic nervous activity indicators can include parameters such as heart rate variability reflecting the balance of sympathetic and parasympathetic nerves, and the sleep structure indicators can include parameters such as sleep stage duration and wake-up frequency quantifying sleep quality. The wearable device is preferably a portable electronic device such as a smart bracelet, a smart watch, etc. that can continuously monitor physiological parameters such as heart rate and sleep. The symptom text diary is actively recorded by the user through a mobile terminal application, forming subjective annotation information synchronized in time with the physiological data.

[0030] In step S102, the dynamic time warping algorithm is used to solve the problem of time axis misalignment caused by different sampling frequencies of heterogeneous time series signals, and the time mapping between signals is realized by finding the optimal path. Time domain feature extraction can include statistical quantities such as mean and variance, frequency domain features can be obtained by Fourier transform to get the energy distribution of each frequency band, and nonlinear entropy features can be quantified by sample entropy algorithm to quantify the complexity of the signal. Optionally, the multimodal time series physiological information can also include electroencephalogram rhythm power features obtained by electroencephalogram acquisition devices to further improve the completeness of the state representation. These features are fused to form a cross-modal physiological state feature vector, representing the comprehensive physiological state of the individual in a specific time period.

[0031] In step S103, a pre-training model based on Transformer is used for semantic understanding of the symptom text, symptom entity recognition is used to extract keywords describing specific symptoms, and sentiment polarity analysis is used to judge the emotional tendency expressed in the text. By combining degree adverbs and other modifiers, subjective descriptions are converted into quantitative symptom intensity values, and finally standardized symptom records with time sequence structure are generated.

[0032] In the dual-path analysis model constructed in step S104, the first path long short-term memory network learns the time sequence dependency between physiological features and symptoms through its gating mechanism to realize individualized prediction; the second path one-class support vector machine maps the features to a high-dimensional space through a kernel function to establish an abnormality detection boundary based on population data. The two analysis paths are independent of and complementary to each other, the first path focuses on the internal change law of an individual, and the second path provides a population reference benchmark.

[0033] In step S105, the gating recurrent unit learns the time sequence dependency between the prediction results of the two paths through the update gate and the reset gate mechanism, and starts feature attribution analysis when a significant divergence is detected. The attribution method based on attention weights can identify the physiological dimension that contributes most to the prediction divergence, providing an explanation basis for subsequent analysis. Preferably, the attention weights are calculated based on the correlation between historical prediction results and the current state, which can be realized through a query key-value attention mechanism or a correlation score algorithm.

[0034] The integrated neural state assessment report generated in step S106 integrates the prediction results of the two paths and the attribution analysis results, and preferably forms the final assessment result through weighted fusion. The integrated neural state assessment report not only contains the risk level judgment, but also provides key factor analysis leading to the assessment, enhancing the explainability and clinical practicability of the result.

[0035] The embodiment fuses multi-modal time sequence physiological information and symptom text diaries to construct a supervised and unsupervised dual-path analysis model, realizes the collaborative analysis of individualized symptom risk prediction and population abnormality detection, adjusts the time sequence consistency using a gating recurrent unit, and identifies the key physiological dimension through feature attribution analysis when the results diverge, finally generating an integrated neural state assessment report. The dual-path analysis model of the embodiment allows the system to consider individual specificity and population regularity at the same time, confirms each other when the conclusions of the two paths are consistent, and provides a reasonable explanation through attribution analysis when there is a divergence, effectively improving the accuracy, robustness and explainability of the neural state assessment, and providing reliable technical support for personalized health management.

[0036] Please refer to Figure 2 In some embodiments, the multi-modal time sequence physiological information is preprocessed and feature fused, the dynamic time warping algorithm is used to align the heterogeneous time sequence signals, and the time domain, frequency domain and nonlinear entropy features are extracted to obtain the cross-modal physiological state feature vector in units of individuals, including: S201, performing heart rate variability analysis on autonomic nervous activity indicators to extract SDNN and RMSSD indicators in time domain features, low frequency and high frequency power ratio indicators in frequency domain features, and autonomic nervous regulation stability indicators in nonlinear features calculated by a sample entropy algorithm; S202, sleep staging feature extraction is performed on the sleep structure index, the duration ratio of deep sleep period and rapid eye movement period is obtained, and sleep efficiency and awakening frequency are calculated to obtain sleep quantification index; S203, the dynamic time warping algorithm is used to time align the above autonomic nervous feature sequence and sleep feature sequence, and eliminate the time sequence deviation caused by the difference in collection time points; S204, the aligned multi-modal features and electroencephalogram rhythm power features are spliced through feature-level fusion, and principal component analysis is used for dimension reduction processing, and finally a cross-modal physiological state feature vector is generated.

[0037] In step S201, heart rate variability analysis is realized by R-wave detection and segment division on electrocardiogram signal. SDNN index is obtained by calculating the standard deviation of all normal sinus interval, reflecting the overall regulation ability of autonomic nervous system; RMSSD index is obtained by calculating the root mean square of adjacent heart rate interval difference, mainly used for evaluating parasympathetic nerve activity. Frequency domain analysis uses fast Fourier transform to decompose the signal into different frequency bands, low frequency power (0.04-0.15Hz) is related to sympathetic nerve activity, high frequency power (0.15-0.4Hz) reflects parasympathetic nerve activity, and the ratio of the two can quantify the sympathetic-parasympathetic nerve balance state. Sample entropy algorithm assesses the complexity of the signal by calculating the probability of new patterns in the time series, and the higher the entropy value, the more flexible the autonomic nervous regulation.

[0038] In step S202, sleep staging feature extraction is based on the international sleep medicine society standard, and machine learning algorithm is used for automatic staging of sleep polygraph signal. The duration ratio of deep sleep period is obtained by calculating the ratio of the sleep time in this period to the total sleep time, reflecting the recovery function of sleep; the ratio of rapid eye movement period is obtained by calculating the ratio of the duration of this period to the total sleep time, which is related to emotional regulation and memory consolidation. Sleep efficiency is calculated by the percentage of total sleep time to total bed time, and the number of awakenings can be quantified by detecting the number of micro-awakening events in electroencephalogram to quantify the continuity of sleep.

[0039] In step S203, the dynamic time warping algorithm constructs a distance matrix between the autonomic nervous feature sequence and the sleep feature sequence, uses dynamic programming to find the minimum cumulative distance path, and realizes the nonlinear alignment of heterogeneous time series signals. This algorithm can effectively handle the time sequence deviation problem caused by different measurement devices or different time scales of physiological processes, providing time synchronization guarantee for subsequent multi-modal feature fusion.

[0040] The feature-level fusion of step S204 combines the aligned autonomic nervous system features, sleep quantitative indicators, and electroencephalogram rhythm power features into a high-dimensional feature vector using vector splicing. Preferably, principal component analysis selects the principal component corresponding to the largest eigenvalue as the new feature dimension by calculating the covariance matrix and its eigenvectors of the features, realizes dimension reduction processing under the premise of retaining more than 90% of the original data variance, and finally generates a cross-modal physiological state feature vector suitable for machine learning models.

[0041] The embodiment realizes deep processing of multi-modal time-series physiological information through systematic feature extraction and fusion methods. Time-domain indicators such as SDNN and RMSSD and frequency-domain indicators such as low-frequency-to-high-frequency power ratio are extracted using heart rate variability analysis, and autonomic nervous regulation stability is quantified using a sample entropy algorithm. Sleep quantitative indicators such as deep sleep and rapid eye movement period proportions are extracted through sleep staging. Heterogeneous time-series signals are aligned using dynamic time warping algorithm, and finally cross-modal physiological state feature vectors are generated through feature-level fusion and principal component analysis, effectively solving the problems of multi-source physiological data heterogeneity and time-series asynchronization, providing comprehensive, accurate, and interpretable feature basis for neural state evaluation, and significantly improving the input quality and reliability of subsequent analysis models.

[0042] Please refer to Figure 3 In some embodiments, natural language processing is performed on the symptom text diary, and a pre-trained model based on Transformer is used for symptom entity recognition and sentiment polarity analysis to generate structured symptom intensity sequences and sentiment label sequences, including: S301, an independent dictionary for the field of neurological symptoms is constructed, which includes a set of somatization symptom descriptors, a set of emotional state keywords, and a degree adverb mapping table, wherein the degree adverb mapping table is used to quantify descriptors such as "mild" and "severe" into numerical intensity levels; S302, a pre-trained language model based on Transformer architecture is used as a basic feature extractor to encode the context semantics of the symptom text diary and obtain deep semantic representation vectors for each word; S303, based on the encoded deep semantic representation vectors, a conditional random field model is used for named entity recognition to extract specific symptom entities and their modification relationships, including palpitations, headaches, and gastrointestinal discomfort; S304, sentiment polarity analysis is performed on the identified symptom entities and their modification relationships, a text sentiment tendency score is calculated through a multi-head attention mechanism, and quantitative symptom intensity values are generated in combination with the degree adverb mapping table; S305, based on time-series information, the symptom intensity values are sorted, and the symptom intensity values of the same time period are aggregated into structured records to generate symptom intensity sequences and sentiment label sequences containing symptom types, intensity values, and sentiment polarities.

[0043] In step S301, an independent dictionary is specially constructed for the field of neurological symptoms. The somatization symptom description word set contains common somatic symptom expressions such as palpitations and headaches. The emotional state keyword set covers emotion-related words such as anxiety and depression. The degree adverb mapping table maps descriptive words such as “mild”, “moderate”, and “severe” to numerical intensity levels, achieving the objective quantification of subjective descriptions.

[0044] In step S302, the pre-trained language model learns medical semantic representations by performing self-supervised task learning such as mask language modeling on large medical text corpus. The model calculates the correlation weights between words through self-attention mechanism, dynamically aggregates context information, and encodes each word into a deep representation vector containing rich semantic relationships, providing a high-quality semantic feature basis for subsequent symptom analysis.

[0045] In step S303, the conditional random field model preferably uses the BIO tagging strategy, uses deep semantic representation vectors as feature input, and accurately identifies the boundary position of symptom entities through joint optimization of transition probability and state feature function, and learns the grammatical dependency relationship between symptom words and modifier words, achieving accurate symptom entity extraction.

[0046] In step S304, the multi-head attention mechanism can calculate the sentiment features of different levels of text in parallel, obtain the comprehensive sentiment tendency score by weighting and fusing the outputs of each attention head, and then convert the descriptive symptom intensity into a quantifiable numerical indicator by combining the degree adverb mapping table.

[0047] In step S305, the quantified symptom intensity values are time-sequentially sorted based on timestamp information, and multiple symptom records within the same time period are aggregated into unified structured data, and finally a standardized output sequence containing symptom type, intensity value and emotional polarity is generated.

[0048] This embodiment realizes the structured processing of subjective symptom text by combining professional dictionary construction and deep learning technology, which not only retains the richness of natural language expression, but also converts it into machine-processable quantitative indicators, providing standardized text feature input for subsequent multi-modal data analysis, effectively solving the semantic gap problem between subjective reports and objective physiological data.

[0049] In some embodiments, based on the cross-modal physiological state feature vector, a supervised and unsupervised dual-channel analysis model is constructed, including: A supervised analysis model framework is constructed, taking the cross-modal physiological state feature vector as the model input, taking the symptom intensity sequence and emotional label sequence as the supervision signal, and using a deep neural network architecture to establish the mapping relationship from physiological features to symptom manifestations; An unsupervised analysis model framework is constructed, taking the cross-modal physiological state feature vectors of large-scale population samples as training data, and using a density clustering algorithm to establish the feature distribution space of the population norm, and defining an abnormal score mechanism based on distance measurement; The numerical ranges output by the supervised analysis model and the unsupervised analysis model are standardized and calibrated to ensure the comparability of the results of the two analysis paths; And when new multi-modal time-series physiological information is collected, the supervised analysis model uses an online learning algorithm to update the parameters, and the unsupervised analysis model uses an incremental learning algorithm to adjust the feature distribution space of the population norm.

[0050] In this embodiment, the supervised analysis model framework learns the complex mapping relationship between the cross-modal physiological state feature vectors and the symptom intensity sequence and the emotion label sequence through a deep neural network architecture. This framework can use recurrent neural networks or time convolutional networks to capture the temporal dependence of physiological features and establish a prediction model from multi-modal physiological data to subjective symptoms.

[0051] The unsupervised analysis model framework uses a density clustering algorithm to model the distribution of cross-modal physiological state feature vectors of large-scale population samples, and uses Gaussian mixture model or DBSCAN algorithm to establish the feature distribution space of the population norm. The abnormal score mechanism is based on Mahalanobis distance or local outlier factor algorithm to calculate the deviation of individual feature vector from the population distribution center.

[0052] The standardization and calibration process uses min-max scaling or z-score standardization method to map the prediction probability value output by the supervised analysis model and the abnormal score output by the unsupervised analysis model to a unified numerical interval, ensuring the direct comparability of the output results of the two analysis paths.

[0053] The online learning algorithm can use stochastic gradient descent or adaptive moment estimation optimization method, so that the supervised analysis model can dynamically adjust the model parameters according to the newly collected multi-modal time-series physiological information; the incremental learning algorithm updates the clustering center or adjusts the distribution parameters, so that the unsupervised analysis model can adapt to the gradual changes of the population norm features.

[0054] This embodiment realizes the organic combination of individualized prediction and population benchmark analysis through the collaborative construction of supervised and unsupervised dual-path analysis models. The two analysis paths are independent and complementary to each other, the supervised path focuses on learning the individual-specific physiological-symptom mapping rules, and the unsupervised path provides an objective reference benchmark based on population distribution, providing a reliable technical foundation for the subsequent fusion and decision-making of the dual-path results.

[0055] In some embodiments, the first path adopts a long short-term memory network to establish a dynamic mapping relationship between the physiological indicators and the symptom labels, and outputs an individualized symptom risk prediction value, denoted as a first path prediction result, including: The long short-term memory network model based on the gating mechanism is constructed by taking the cross-modal physiological state feature vector as an input sequence, and the long-term time sequence dependency relationship between the physiological indicators is captured through the collaborative control of the forgetting gate, the input gate and the output gate. The symptom intensity sequence and the emotion label sequence are taken as a supervision signal, the hidden state output by the long short-term memory network is mapped to the symptom risk space by using a time distribution fully connected layer, and an individualized symptom risk prediction value synchronized with the input sequence is generated. The weight distribution of the symptom intensity prediction task and the emotion polarity classification task is adjusted, and the long short-term memory network parameters are updated through the back propagation algorithm. The sliding window mechanism is adopted to process the continuous time sequence data, and the latest individualized symptom risk prediction value is output in real time, denoted as the first path prediction result.

[0056] In this embodiment, the long short-term memory network processes the cross-modal physiological state feature vector sequence through its unique gating mechanism. The forgetting gate controls the retention degree of historical information, the input gate adjusts the input amount of new information, and the output gate determines the output of the current hidden state. The three work together to capture the long-term time sequence dependency relationship between the physiological indicators.

[0057] The time distribution fully connected layer independently maps the hidden state of each time step of the long short-term memory network, converts the high-dimensional features to the symptom risk space, generates an individualized symptom risk prediction value synchronized with the input physiological data in time, and realizes end-to-end time sequence prediction.

[0058] In the multi-task learning framework, the loss contribution proportion of the two tasks is automatically adjusted through the learnable weight parameters. Specifically, the system calculates the mean square error loss of the symptom intensity prediction and the cross-entropy loss of the emotion polarity classification, and obtains the total loss function by weighted summation. The back propagation algorithm calculates the gradient layer by layer from the output layer to the input layer based on the chain rule. First, the partial derivative of the total loss with respect to the network output node is calculated, then the gradient is transmitted along the time dimension through the time backward propagation algorithm, and finally the weight parameters of all the gating units and fully connected layers in the long short-term memory network are updated by using the optimizer, realizing the collaborative optimization of the two tasks and improving the representation ability of the model for complex symptom patterns.

[0059] The sliding window mechanism adopts a fixed length time window to slide on the continuous time sequence data, processes only the data segment within the window each time, ensures that the model can process streaming data in real time and output the latest prediction result, and meets the needs of clinical real-time monitoring.

[0060] The embodiment effectively captures the dynamic mapping relationship between physiological indicators and symptom manifestations through the excellent time sequence modeling capability of the long short-term memory network. Combined with multi-task learning and real-time processing mechanism, the accuracy of individualized prediction is improved, and the timeliness of the system in actual application is ensured, thereby providing reliable technical support for personalized health management.

[0061] In some embodiments, the second path adopts a one-class support vector machine to build an abnormality detection model based on a group baseline distribution, calculates a deviation degree of the current physiological mode relative to the group norm, and records the deviation degree as a second path prediction result, including: The sample cross-modal physiological state feature vector of a preset group is collected as a training data set, and a physiological norm representing the relative stable state of the preset group is constructed. The preset group is configured to be a group of patients diagnosed as mild anxiety or somatization disorder through clinical evaluation and in a clinical stable period; A one-class support vector machine algorithm is used to estimate the distribution density of the sample cross-modal physiological state feature vector, and a decision hyperplane is found through a nonlinear kernel function mapping to define the main feature distribution space of the stable preset group; The cross-modal physiological state feature vector of the current user is input into the trained abnormality detection model, the relative distance of the cross-modal physiological state feature vector of the current user to the decision hyperplane is calculated, and the relative distance is quantified as the deviation degree of the current physiological mode relative to the group stable state norm; The deviation degree is normalized according to the distribution characteristics of the group stable state norm to generate a second path prediction result that can be compared across individuals.

[0062] In the embodiment, by continuously collecting multi-modal physiological data of a preset group in a stable state, after feature extraction and fusion processing, a clustering algorithm (such as K-means) or a Gaussian mixture model is used to model the feature distribution, and a physiological norm representing the stable state of the group is constructed. The norm can accurately depict the typical physiological feature distribution pattern of patients in the stable period.

[0063] The one-class support vector machine maps the samples to a high-dimensional feature space through a kernel trick, and finds an optimal hyperplane in the space. Specifically, the similarity between samples is calculated by using a Gaussian kernel function, and the normal data boundary is defined by maximizing the distance from the hyperplane to the origin. The decision hyperplane divides the feature space into normal and abnormal regions, and the normal samples are contained in the hyperplane as much as possible, while the abnormal samples fall outside the hyperplane, thereby accurately defining the feature distribution space of the stable group.

[0064] Relative distance calculation is based on the support vector machine decision function. After mapping the current user's features into a high-dimensional space using a kernel function, the perpendicular distance between the mapped point and the optimal hyperplane is calculated. This distance value is processed using a sign function, with a positive sign indicating the normal side of the hyperplane and a negative sign indicating the abnormal side. A larger absolute value of the distance indicates a greater deviation from the population norm, thus achieving a quantitative conversion from geometric distance to abnormality.

[0065] Normalization processing uses a standardization method based on the statistical characteristics of group distribution, such as z-score normalization or quantile-based normalization, to convert the original distance value into a deviation score with a uniform dimension and comparability, ensuring that the results between different individuals can be fairly compared.

[0066] This example uses an unsupervised anomaly detection model constructed using a class of support vector machines, providing the system with an objective evaluation dimension based on a population benchmark. This model does not rely on individually labeled data, but instead identifies anomalies by learning the characteristic distribution patterns of a stable population. This effectively mitigates the potential individual bias inherent in supervised approaches and enhances the robustness and reliability of the overall system.

[0067] In some embodiments, adjusting the output of the two-way analysis model for timing consistency through a gated recurrent unit includes: Align and concatenate the first-path prediction results and the second-path prediction results in time series to form a dual-path prediction result sequence as input features; Construct a gated recurrent unit network and learn the temporal dependencies and complementary characteristics between the two-way prediction results through the update gate and reset gate mechanism; The attention mechanism is used to calculate the weight distribution of historical prediction results, enhancing the influence weight of historical information that is highly relevant to the current state; Output the two-way prediction result sequence after temporal consistency adjustment.

[0068] In this embodiment, the dual-path prediction result sequence is formed by aligning the timestamps of the first-path prediction result and the second-path prediction result, ensuring that the outputs of the two analysis paths are completely synchronized in the time dimension, providing an accurate data basis for subsequent time series analysis.

[0069] The gated recurrent unit network controls the degree of retention of historical information by updating the gate and adjusts the input ratio of new information by resetting the gate. The two work together to learn the complex temporal dependencies between the dual-path prediction results, and can effectively capture the dynamic correlation patterns and complementary characteristics between the outputs of the two analysis pathways.

[0070] The attention mechanism dynamically evaluates the importance of the prediction results at different historical time points to the current state by calculating the attention weight of each historical time step. The higher the weight, the greater the contribution of the historical information to the current adjustment process, thereby focusing and enhancing the key historical information.

[0071] The timing consistency adjustment process integrates the network state of the gated recurrent unit and the attention weight, smoothens and corrects the two-way prediction results, eliminates abnormal values caused by data noise or model fluctuations, and finally outputs the adjusted sequence with better timing consistency and reliability.

[0072] This embodiment realizes intelligent timing integration of two-way prediction results through the combination of gated recurrent units and attention mechanisms, not only retains the advantages of each analysis path, but also enhances the stability and reliability of the results through timing consistency processing, providing more accurate and consistent input basis for subsequent decision analysis.

[0073] In some embodiments, when the first prediction result and the second prediction result have a significant divergence, a feature attribution analysis based on the attention weight is performed to identify the core physiological dimension causing the divergence, and the attribution analysis result is output, including: Based on the two-way prediction result sequence after timing consistency adjustment, the relative entropy value of the two-way prediction result in the adjacent time window is calculated, and when the relative entropy of the continuous three time windows exceeds the adaptive threshold, it is determined as a significant divergence state; Start multi-level feature attribution analysis, use integrated gradient algorithm at the first level to calculate the contribution of each input feature dimension in the long short-term memory network to the current prediction result, and generate the first feature importance distribution; At the second level, use the kernel function-based Shapley value estimation algorithm to analyze the marginal contribution of each feature dimension in the support vector machine model to the anomaly score, and generate the second feature importance distribution; Construct a feature contribution difference matrix to compare the first feature importance distribution and the second feature importance distribution, identify the top K feature dimensions with the largest contribution difference as the core divergence features, and record them as the identification result; Generate attribution analysis results based on the identification result, including divergence time period, core feature dimension list, contribution difference value, and clinical significance explanation.

[0074] In this embodiment, the two-way prediction results of adjacent time windows are first converted into probability distributions, and then the relative entropy value between the two distributions is calculated. The adaptive threshold is dynamically adjusted according to the statistical characteristics of the historical relative entropy value. When the entropy value of the continuous three time windows exceeds the threshold, it indicates that there is a persistent significant divergence in the two-way prediction results.

[0075] The integrated gradient algorithm first constructs a linear interpolation path between the input features and the baseline, then calculates the prediction output gradient of each point on the path, and finally integrates the gradients to obtain the contribution score of each feature dimension. The integrated gradient algorithm quantifies the contribution of each physiological feature dimension to the long short-term memory network prediction result by calculating the gradient integral of the input features with respect to the prediction output, eliminates the gradient saturation problem through path integration, and ensures the accuracy of feature importance evaluation.

[0076] The Shapley value estimation algorithm based on kernel function generates feature subsets by random sampling, calculates the marginal contribution of each feature value using kernel function weighting, and finally obtains the Shapley value estimation by weighted averaging. Through the kernel technique, the exponential computational complexity is reduced to polynomial level, realizing efficient calculation in large-scale feature space.

[0077] Preferably, the construction of the feature contribution difference matrix adopts the Euclidean distance measurement, respectively calculates the difference between the first path and the feature importance distribution vector, and then normalizes the difference vector. The first K features with the largest difference are selected by sorting, which represent the most significant divergence sources between the two paths.

[0078] Clinical significance interpretation can be achieved through knowledge graph, which associates and maps the identified core features with the pathophysiological mechanisms in the medical knowledge base, and generates targeted explanations combined with clinical guidelines and expert experience, providing data-driven analysis results while ensuring the clinical rationality and operability of the conclusions.

[0079] The present embodiment uses a multi-level feature attribution analysis method to deeply analyze the reasons for the divergence between the two paths from the dimensions of mathematical calculation and clinical interpretation, not only identifying the key feature dimensions that cause the divergence, but also providing clinical significance interpretation, enhancing the interpretability and practicality, and providing important support for precision medicine.

[0080] In some embodiments, an integrated neural state assessment report is generated according to the first path prediction result, the second path prediction result, and the attribution analysis result, including: The symptom risk trend of the first path prediction result, the physiological abnormality degree of the second path prediction result, and the core feature dimension of the attribution analysis result are weighted and fused; According to the consistency degree of the two-path prediction results, a corresponding assessment mode is selected, including: When the two-path prediction results are consistent, a collaborative confirmation mode is adopted; When the two-path prediction results are divergent, an attribution explanation mode is adopted; A structured assessment result is generated, including symptom risk level, physiological abnormality index, key influencing factor analysis, and confidence level; generate personalized intervention suggestions based on the structured assessment results, the personalized intervention suggestions including early warning prompts, lifestyle adjustment suggestions, and professional medical guidance; generate an integrated neurological state assessment report based on the structured assessment results and the personalized intervention suggestions.

[0081] In this embodiment, the weighted fusion process adopts an adaptive weight distribution mechanism, dynamically adjusting the weight proportion of each input item according to the confidence of the dual-path prediction results and the historical accuracy rate. The symptom risk trend mainly reflects the change of subjective feeling, the physiological abnormality degree embodies the deviation of objective indicators, and the core feature dimension provides in-depth analysis of the reasons for the divergence, which form the basis for comprehensive evaluation through weighted fusion.

[0082] The evaluation mode selection is based on the consistency measurement results of the dual-path prediction results. When the dual-path prediction results are consistent at a statistical significance level, the collaborative confirmation mode is adopted, emphasizing the credibility and stability of the results. When there is a significant divergence, the attribution explanation mode is enabled, focusing on analyzing the reasons for the divergence and providing reasonable explanations.

[0083] The generation of structured assessment results adopts a multi-dimensional scoring system. The symptom risk level is determined according to the size and trend of the prediction value, the physiological abnormality index is calculated based on the degree of deviation from the population norm, the key influence factor analysis extracts the core features from the attribution results, and the confidence considers factors such as data quality, model performance, and result consistency.

[0084] The generation of personalized intervention suggestions combines clinical guidelines and individual characteristics. Early warning prompts provide immediate alerts for high-risk states, lifestyle adjustment suggestions provide specific improvement plans based on identified influencing factors, and professional medical guidance recommends appropriate medical intervention measures according to the severity and clinical indications.

[0085] The integrated neurological state assessment report adopts a standardized template, organically integrating the structured assessment results and personalized suggestions, ensuring the scientificity and integrity of the report, and ensuring its clinical practicability and operability. During the report generation process, the understanding ability and information needs of different user groups are also considered, and a hierarchical presentation method is used to enhance readability.

[0086] This embodiment realizes the complete transformation from multi-source data to personalized report through systematic information integration and intelligent decision-making mechanism, not only provides comprehensive state assessment, but also gives specific action guidelines, truly realizes the seamless connection of evaluation and intervention, and provides effective technical support for precision health management.

[0087] In a second aspect, the embodiment also provides a neurological pathology feature analysis system suitable for the method of the first aspect.

[0088] By adopting the above technical scheme, the present application is distinguished from the prior art and has the following beneficial effects: by fusing multi-modal time-series physiological information and symptom text diary, a supervised and unsupervised dual-path analysis model is constructed, the supervised analysis path uses a long short-term memory network to establish individualized symptom risk prediction, the unsupervised analysis path uses a support vector machine to construct a group anomaly detection model, and the dual paths cooperate to realize double consideration of individual specificity and group regularity; time-series consistency adjustment is performed through a gated recurrent unit, and when the results are different, feature attribution analysis based on attention weight is started, which significantly improves the accuracy and robustness of state evaluation. The finally generated integrated neural state evaluation report deeply fuses symptom risk trend, physiological abnormality degree and attribution analysis result, and provides individualized intervention suggestions, realizes a complete technical closed loop from multi-source data acquisition to precise health management, and provides reliable technical support for early screening and individualized intervention of neurological diseases.

[0089] Finally, it should be noted that although the above embodiments have been described in the specification and drawings of the present application, the patent protection scope of the present application should not be limited. Any technical solution obtained by replacing or modifying the equivalent structure or equivalent process based on the essential concept of the present application, using the content described in the specification and drawings of the present application, and directly or indirectly implementing the technical solutions of the above embodiments in other related technical fields, etc. are all included in the patent protection scope of the present application.

Claims

1. A method for analyzing neurological pathological characteristics, characterized in that: include: Collect multimodal temporal physiological information of the target individual, including autonomic nervous system activity indicators and sleep structure indicators obtained through wearable devices, and collect text diaries of symptoms actively reported by users as subjective annotation information; Preprocessing and feature fusion of multimodal temporal physiological information are performed, and a dynamic time warping algorithm is used to align heterogeneous temporal signals. Time domain, frequency domain, and nonlinear entropy features are extracted to obtain cross-modal physiological state feature vectors based on individuals. Perform natural language processing on the symptom text diary, use a Transformer-based pre-trained model to perform symptom entity recognition and sentiment polarity analysis, and generate structured symptom intensity sequences and sentiment label sequences; Based on the cross-modal physiological state feature vector, supervised and unsupervised two-way analysis models are constructed. The two-way analysis model includes a first path and a second path. The first path uses a long short-term memory network to establish a dynamic mapping relationship between physiological indicators and symptom labels, and outputs an individualized symptom risk prediction value, which is recorded as the first path prediction result. The second path uses a first-class support vector machine to build an anomaly detection model based on the group baseline distribution, and calculates the deviation of the current physiological pattern from the group norm, which is recorded as the second path prediction result; The output results of the dual-path analysis model are adjusted for temporal consistency through a gated recurrent unit. When there is a significant divergence between the first-path prediction result and the second-path prediction result, a feature attribution analysis based on attention weights is performed to identify the core physiological dimension that causes the divergence and output the attribution analysis result. Generate an integrated neural status assessment report based on the first-channel prediction results, the second-channel prediction results and the attribution analysis results.

2. The neurological pathology feature analysis method according to claim 1, characterized in that: Multimodal temporal physiological information is preprocessed and feature fused. The dynamic time warping algorithm is used to align heterogeneous temporal signals, and time domain, frequency domain, and nonlinear entropy features are extracted to obtain cross-modal physiological state feature vectors based on individuals, including: Heart rate variability analysis was performed on autonomic nervous system activity indicators, including SDNN and RMSSD indicators from the time domain features, low-frequency to high-frequency power ratio indicators from the frequency domain features, and autonomic nervous system regulation stability indicators from the nonlinear features calculated using the sample entropy algorithm. Extract sleep stage features from sleep structure indicators, obtain the duration ratio of deep sleep and rapid eye movement, and calculate sleep efficiency and number of awakenings to obtain sleep quantitative indicators; The dynamic time warping algorithm is used to align the autonomic nervous system feature sequence with the sleep feature sequence to eliminate the timing offset caused by the difference in acquisition time points. The aligned multimodal features and EEG rhythm power features are spliced ​​together through feature-level fusion, and principal component analysis is used to perform dimensionality reduction, ultimately generating the cross-modal physiological state feature vector.

3. The neurological pathology feature analysis method according to claim 1, characterized in that: Perform natural language processing on the symptom text diary, use a Transformer-based pre-trained model for symptom entity recognition and sentiment polarity analysis, and generate structured symptom intensity sequences and sentiment label sequences, including: Constructing an independent dictionary for the field of neurological symptoms, the independent dictionary includes a somatization symptom description word set, an emotional state keyword set, and a degree adverb mapping table; A pre-trained language model based on the Transformer architecture is used as the basic feature extractor to perform contextual semantic encoding on the symptom text diary and obtain the deep semantic representation vector of each word; Based on the encoded deep semantic representation vector, a conditional random field model is used to perform named entity recognition to extract entities describing specific symptoms and their modification relationships. The symptom entities include palpitations, headaches, and gastrointestinal discomfort. Perform sentiment polarity analysis on the identified symptom entities and their modification relationships, calculate the text sentiment tendency score through a multi-head attention mechanism, and generate a quantitative symptom intensity value by combining it with the degree adverb mapping table; The symptom intensity values ​​are sorted based on time series information, and the symptom intensity values ​​of the same time period are aggregated into structured records to generate a symptom intensity sequence and an emotion label sequence including symptom type, intensity value and emotion polarity.

4. The neurological pathology feature analysis method according to claim 1, characterized in that: Based on the cross-modal physiological state feature vector, supervised and unsupervised dual-path analysis models are constructed, including: A supervised analysis model framework was constructed, using cross-modal physiological state feature vectors as model input, symptom intensity sequences and emotion label sequences as supervisory signals, and a deep neural network architecture to establish a mapping relationship from physiological features to symptom manifestations. An unsupervised analysis model framework is constructed, using the cross-modal physiological state feature vectors of large-scale group samples as training data, a density clustering algorithm is used to establish the feature distribution space of the group norm, and an abnormality scoring mechanism based on distance measurement is defined; Standardize and calibrate the numerical ranges of the outputs of the supervised and unsupervised analysis models to ensure that the results of the two analysis pathways are comparable; Furthermore, when new multimodal temporal physiological information is collected, an online learning algorithm is used to update the parameters of the supervised analysis model, and an incremental learning algorithm is used to adjust the characteristic distribution space of the group norm for the unsupervised analysis model.

5. The neurological pathology feature analysis method according to claim 1, characterized in that: The first path uses a long short-term memory network to establish a dynamic mapping relationship between physiological indicators and symptom labels, and outputs an individualized symptom risk prediction value, which is recorded as the first path prediction result, including: Using cross-modal physiological state feature vectors as input sequences, a long short-term memory network model based on a gating mechanism is constructed. Through the coordinated control of the forget gate, input gate, and output gate, the long-term temporal dependency between physiological indicators is captured. The symptom intensity sequence and the emotion label sequence are used as supervisory signals, and a time-distributed fully connected layer is used to map the hidden state output by the long short-term memory network to the symptom risk space, thereby generating an individualized symptom risk prediction value synchronized with the input sequence; Adjust the weight distribution of the symptom intensity prediction task and the emotion polarity classification task, and update the long short-term memory network parameters through the back-propagation algorithm; A sliding window mechanism is used to process continuous time series data, and the latest individualized symptom risk prediction value is output in real time, which is recorded as the first prediction result.

6. The neurological pathology feature analysis method according to claim 1, characterized in that: The second path uses a first-class support vector machine to build an anomaly detection model based on the group baseline distribution, and calculates the deviation of the current physiological pattern from the group norm, which is recorded as the second-path prediction result, including: Collecting cross-modal physiological state feature vectors of samples from a preset group as a training dataset and constructing a physiological norm representing the relatively stable state of the preset group. The preset group is configured as a group diagnosed with mild anxiety or somatization disorder through clinical assessment and in a clinically stable period; A first-class support vector machine algorithm is used to estimate the distribution density of the cross-modal physiological state feature vectors of the samples, and a decision hyperplane is found through nonlinear kernel function mapping to define the main feature distribution space of the stable preset group; Inputting the current user's cross-modal physiological state feature vector into the trained anomaly detection model, calculating the relative distance between the current user's cross-modal physiological state feature vector and the decision hyperplane, and quantifying the relative distance as the deviation of the current physiological mode from the group stable state norm; The deviations are normalized according to the distribution characteristics of the group steady-state norm to generate a second-path prediction result that can be compared across individuals.

7. The neurological pathology feature analysis method according to claim 1, characterized in that: The output of the dual-path analysis model is adjusted for timing consistency through the gated recurrent unit, including: Align and concatenate the first-path prediction results and the second-path prediction results in time series to form a dual-path prediction result sequence as input features; Construct a gated recurrent unit network and learn the temporal dependencies and complementary characteristics between the two-way prediction results through the update gate and reset gate mechanism; The attention mechanism is used to calculate the weight distribution of historical prediction results, enhancing the influence weight of historical information that is highly relevant to the current state; Output the two-way prediction result sequence after temporal consistency adjustment.

8. The neurological pathology feature analysis method according to claim 1, characterized in that: When there is a significant divergence between the first-pass prediction result and the second-pass prediction result, a feature attribution analysis based on attention weight is performed to identify the core physiological dimension that causes the divergence and output the attribution analysis results, including: Based on the dual-path prediction result sequence after time series consistency adjustment, the relative entropy value of the dual-path prediction results in adjacent time windows is calculated. When the relative entropy of three consecutive time windows exceeds the adaptive threshold, it is determined to be a significant divergence state; Start multi-level feature attribution analysis. At the first level, use the integrated gradient algorithm to calculate the contribution of each input feature dimension in the long short-term memory network to the current prediction result, and generate the first-path feature importance distribution. At the second level, the kernel function-based Shapley value estimation algorithm is used to analyze the marginal contribution of each feature dimension to the anomaly score in a class of support vector machine models and generate the second-path feature importance distribution; Construct a feature contribution difference matrix, compare the first-path feature importance distribution with the second-path feature importance distribution, identify the top K feature dimensions with the largest contribution difference as the core divergent features, and record them as the recognition results; Attribution analysis results are generated based on the recognition results, including divergence time periods, a list of core feature dimensions, contribution differences, and clinical significance explanations.

9. The neurological pathology feature analysis method according to claim 1, characterized in that: Generate an integrated neural status assessment report based on the first-path prediction results, the second-path prediction results, and the attribution analysis results, including: The symptom risk trend of the first-path prediction result, the physiological abnormality degree of the second-path prediction result, and the core characteristic dimensions of the attribution analysis result are weighted and integrated; The corresponding evaluation mode is selected according to the consistency of the two-way prediction results, including: When the prediction results of the two channels are consistent, the collaborative confirmation mode is adopted; When the two-way prediction results diverge, the attribution explanation model is adopted; Generating structured assessment results, including symptom risk level, physiological abnormality index, key influencing factor analysis, and confidence level; Generating personalized intervention recommendations based on the structured assessment results, the personalized intervention recommendations including early warning prompts, lifestyle adjustment recommendations, and professional medical guidance; The integrated neurological status assessment report is generated based on the structured assessment results and personalized intervention recommendations.

10. A neurological pathology feature analysis system, characterized in that: The method according to any one of claims 1 to 9.

Images