A method and system for detecting and analyzing neurological diseases

Abstract

The application discloses a kind of neural disease detection analysis method and system, are acquired bracelet collection signal, sphygmomanometer collection signal, motion behavior image and behavior test data, are extracted tremor intensity characteristics, gait symmetry characteristics and autonomic nervous rhythm characteristics to bracelet collection signal by multi-band decomposition;Analysis motion behavior image and standardized action test obtains motor function score;Carry out heart rate variability analysis and identify abnormal mode of neural function;Neural function state atlas is constructed and feature weight is calculated;Combine historical monitoring data to predict disease progression trend;Through subsequent feedback correction information dynamic adjustment prediction result.The application realizes long-term trend prediction and dynamic correction based on initial data by the way that short-time intensive monitoring is combined with long-period intermittent collection, significantly improves the reliability and practicality of neural disease risk assessment in home scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent medical detection, in particular to a neural disease detection analysis method and system. BACKGROUND

[0002] Early risk assessment of neurodegenerative diseases is of great significance for delaying disease progression. Currently, clinical diagnosis mainly relies on regular hospital examinations, which has the problems of long monitoring interval and data dispersion, making it difficult to capture early changes in the disease in a timely manner. In recent years, the application of wearable devices has made home monitoring possible, but existing technical solutions still have obvious limitations.

[0003] Traditional monitoring methods usually require patients to wear devices for a long time and collect complete data, which faces the challenges of poor compliance and unstable data quality in actual application. At the same time, existing analysis models are mostly based on static evaluation of single complete data set, which cannot derive long-term trends through short-term intensive monitoring, and also lacks the ability to adaptively adjust to subsequent sparse sampling. More importantly, most systems only provide current state evaluation and fail to establish an effective dynamic prediction mechanism to predict disease development trajectory.

[0004] In addition, the prediction model of the existing scheme often uses fixed parameters, which is difficult to adapt to individual differences and nonlinear characteristics of disease evolution. When new monitoring data appears, the system usually needs to retrain the model, which cannot realize rapid online update. This limitation leads to deviation of the prediction result from the actual progress, which is more obvious in the home scene that needs long-term tracking. SUMMARY

[0005] In view of the above problems, the present application provides a neural disease detection analysis method and system, which realizes long-term trend prediction and dynamic correction based on initial data through a data processing mechanism combining short-term intensive collection and long-term intermittent monitoring, solving the problem of being unable to balance monitoring convenience and prediction accuracy.

[0006] To achieve the above purpose, in a first aspect, the present application provides a neural disease detection analysis method, comprising:

[0007] Obtaining wearable device monitoring data and behavior test data of a user, the wearable device monitoring data including bracelet collection signals, sphygmomanometer collection signals and motion behavior images, and the behavior test data including standardized motion test results;

[0008] Performing multi-band decomposition on the bracelet collection signals, and separating tremor components, gait components and autonomic nervous fluctuation components by wavelet transform;

[0009] The tremor component is extracted by a long short-term memory network to obtain a tremor intensity feature, the gait component is extracted by a time convolution network to obtain a gait symmetry feature, and the autonomic nerve fluctuation component is extracted by a Gaussian process regression algorithm to obtain an autonomic nerve rhythm feature;

[0010] The eye movement trajectory feature and the motion coordination feature are extracted from the motion behavior image by using a deep learning model, and the motion function score is calculated in combination with the standardized motion test result;

[0011] The heart rate variability analysis is performed on the sphygmomanometer collected signal and the bracelet collected signal, the diurnal rhythm fluctuation feature is calculated, and the neural function abnormality mode is identified by using a density clustering algorithm;

[0012] Based on the tremor intensity feature, the gait symmetry feature, the autonomic nerve rhythm feature, the motion function score and the neural function abnormality mode, a neural function state atlas is constructed, and the contribution weight of each feature to the disease risk is calculated by using a graph attention mechanism;

[0013] The historical monitoring data of the user is obtained, the continuous time evolution law of the neural function state is learned in the neural function state atlas according to the historical detection data by using a neural differential equation, the disease progression trend in a preset future period is predicted, and a prediction result is generated;

[0014] At least one of the bracelet collected signal, the sphygmomanometer collected signal and the motion behavior image captured by the mobile phone camera in the preset future period is collected as feedback correction information, and the prediction result is corrected based on the feedback correction information by using an adaptive Kalman filter;

[0015] A neural disease risk assessment report is output according to the prediction result.

[0016] In some embodiments, based on the tremor intensity feature, the gait symmetry feature, the autonomic nerve rhythm feature, the motion function score and the neural function abnormality mode, a neural function state atlas is constructed, and the contribution weight of each feature to the disease risk is calculated by using a graph attention mechanism, including:

[0017] The tremor intensity feature, the gait symmetry feature, the autonomic nerve rhythm feature, the motion function score and the neural function abnormality mode are taken as node features of a graph neural network, and an initial edge connection relationship is constructed according to the physiological correlation between each node feature;

[0018] The node features are modeled by using a graph attention mechanism to calculate the motion function correlation weight between the tremor intensity feature and the gait symmetry feature, the autonomic nerve regulation weight between the autonomic nerve rhythm feature and the neural function abnormality mode, and the comprehensive coordination weight between the motion function score and other node features;

[0019] The neighborhood node information is aggregated by a multi-layer graph convolution network, the embedding representation of each node feature is iteratively optimized, and a graph embedding vector reflecting the overall state of the neural function is output.

[0020] Based on the graph embedding vector, the dynamic contribution weight of each node feature to the disease risk is generated by an attention score calculation module. The weight of the tremor intensity feature and the gait symmetry feature is used for motor dysfunction evaluation. The weight of the autonomic rhythm feature and the abnormal pattern of neural function is used for autonomic nervous function evaluation. The weight of the motor function score is used for comprehensive behavior function evaluation.

[0021] In some embodiments, historical monitoring data of the user is obtained, and a neural differential equation is used to learn the continuous time evolution law of the neural function state in the neural function state graph according to the historical detection data, to predict the disease progression trend in a preset future period, and to generate a prediction result, including:

[0022] The historical monitoring data is integrated into the neural function state graph to form a plurality of historical embedding vectors, and the plurality of historical embedding vectors are arranged in time sequence to construct a time evolution sequence of the neural function state;

[0023] The neural differential equation is used to calculate the time evolution sequence, a learnable continuous time dynamics function is used to describe the gradual change process of the neural function state, and formula (1) is used to represent the formula (1) as follows:

[0024] ;

[0025] In formula (1), is a neural differential equation, is a learnable continuous time dynamics function, is a neural function state vector at time t, is a weight parameter of , and is a continuous time variable;

[0026] A gating mechanism is introduced into the neural differential equation framework to dynamically adjust the state change rate at different time scales to capture the short-term fluctuations and long-term degradation characteristics of the neural function, and a neural differential equation model is obtained. Formula (2) is used to represent the formula (2) as follows:

[0027] ;

[0028] In formula (2), is a gating vector corresponding to the gating mechanism, is a residual term, is an element-wise multiplication;

[0029] based on the current neural function state map and the trained neural differential equation model, multi-step state prediction in a preset future period is performed, and a neural function state prediction trajectory containing time continuity is output;

[0030] The neural function state prediction trajectory is probabilistically processed, the disease risk probability distribution at each time point is calculated, and the key time nodes whose risk exceeds a preset threshold are identified;

[0031] The neural function state prediction trajectory is compared and analyzed with the user's individual baseline data to generate a personalized prediction result of disease progression trend, and the prediction result includes a risk level change curve and key time node warning information.

[0032] In some embodiments, at least one of the bracelet acquisition signal, the sphygmomanometer acquisition signal, and the motion behavior image captured by the mobile phone camera is collected in a preset future period, denoted as feedback correction information, and the prediction result is corrected based on the feedback correction information based on adaptive Kalman filtering, including:

[0033] The feedback correction information is input into a feature extraction module, and at least one of the tremor intensity feature, the gait symmetry feature, the autonomic nervous rhythm feature, and the motor function score is updated according to the extracted feedback features, to generate a neural function state observation vector at the current time;

[0034] A state space model is constructed, the prediction trajectory output by the neural differential equation is taken as a state transition prior, and the neural function state observation vector is taken as a measurement value to generate a state estimation result;

[0035] An adaptive Kalman filtering algorithm is used to generate a confidence weight of the state estimation result by dynamically adjusting a process noise covariance matrix and an observation noise covariance matrix;

[0036] And when the deviation between the observation value and the prediction value exceeds a dynamic threshold, a recalibration process of the prediction model is triggered;

[0037] The state estimation result is corrected according to the confidence weight;

[0038] The corrected state estimation result is fed back to the neural differential equation model, the parameters of the kinetic function are adjusted to adapt to the individualized neural function change mode of the user, and a corrected prediction result is output.

[0039] In some embodiments, the method further comprises:

[0040] It is judged whether the parameters in the feedback correction information exceed their corresponding normal fluctuation range, and if so, a re-collection mode is triggered;

[0041] Or, the deviation degree of the corrected prediction result and the initial prediction result is calculated, and when the deviation degree exceeds a preset deviation threshold, a re-collection mode is triggered;

[0042] The re-collection mode is configured to:

[0043] Re-collecting wearable device monitoring data and behavior test data of the user;

[0044] Re-decomposing the signals collected by the bracelet into multiple frequency bands, and separating tremor components, gait components and autonomic nerve fluctuation components by wavelet transform;

[0045] Re-extracting tremor intensity features from the tremor components by a long short-term memory network, extracting gait symmetry features from the gait components by a time convolution network, and extracting autonomic nerve rhythm features from the autonomic nerve fluctuation components by a Gaussian process regression algorithm;

[0046] Re-extracting eye movement trajectory features and motion coordination features from the motion behavior image by a deep learning model, and calculating a motor function score in combination with a standardized motion test result;

[0047] Re-performing heart rate variability analysis on the signals collected by the sphygmomanometer and the bracelet, calculating diurnal rhythm fluctuation features, and identifying a neural function abnormality mode by a density clustering algorithm;

[0048] Re-constructing a neural function state atlas based on the tremor intensity features, the gait symmetry features, the autonomic nerve rhythm features, the motor function score and the neural function abnormality mode, and calculating contribution weights of the features to disease risk by a graph attention mechanism;

[0049] In the neural function state atlas, the continuous time evolution law of the neural function state is learned by a neural differential equation based on historical detection data, a disease progression trend in a preset future period is predicted, and an updated prediction result is generated.

[0050] In some embodiments, the bracelet collection signals include acceleration signals and gyroscope signals, and the bracelet collection signals are decomposed into multiple frequency bands, and the tremor components, the gait components and the autonomic nerve fluctuation components are separated by wavelet transform, including:

[0051] The acceleration signals and the gyroscope signals are decomposed into multiple scales by a wavelet basis function with time-frequency localization characteristics, and the original signals are separated into first frequency band signals containing tremor features, second frequency band signals containing gait cycle features, and third frequency band signals reflecting autonomic nerve regulation according to a preset frequency band;

[0052] An adaptive decomposition algorithm based on signal energy distribution is used for the first frequency band signal, the second frequency band signal and the third frequency band signal, and the number of wavelet decomposition layers is dynamically adjusted according to the signal-to-noise ratio of each frequency band signal, so as to obtain the first sub-band signal, the second sub-band signal and the third sub-band signal after decomposition, and the motion artifact suppression processing is performed on each sub-band signal after decomposition;

[0053] The first sub-band signal after motion artifact suppression processing is combined into a tremor component continuous in time domain through a phase consistency reconstruction algorithm;

[0054] The second sub-band signal after motion artifact suppression processing is subjected to motion cycle phase calibration by using a dynamic time warping algorithm, so as to obtain a gait component;

[0055] The third sub-band signal after motion artifact suppression processing is sequentially subjected to ensemble empirical mode decomposition to remove baseline drift, maximum likelihood estimation to restore physiological rhythm components, and 0.01-0.1 Hz band-pass filtering processing to extract sympathetic-parasympathetic nerve regulation features, so as to obtain an autonomic nerve fluctuation component.

[0056] In some embodiments, the tremor component is subjected to a long short-term memory network to extract tremor intensity features, including:

[0057] A bidirectional long short-term memory network model with an attention mechanism is constructed;

[0058] The tremor component is subjected to time series analysis to obtain amplitude change rate and time series fluctuation features of the tremor component;

[0059] The amplitude change rate and the time series fluctuation features are used to generate tremor intensity features representing tremor severity;

[0060] The gait component is subjected to a time series convolution network to extract gait symmetry features, including:

[0061] The gait component is input into a multi-scale dilated time series convolution network, and different time scale gait cycle features are extracted through a plurality of dilated convolution layers arranged in parallel;

[0062] The cross-correlation functions of left and right limb motion signals in time domain and frequency domain are calculated according to the gait cycle features;

[0063] The Hilbert transform is used to calculate the phase difference, and an asymmetry feature vector representing gait asymmetry is output;

[0064] The autonomic nerve fluctuation component is subjected to a Gaussian process regression algorithm to extract autonomic nerve rhythm features, including:

[0065] The autonomic nerve fluctuation component is input into a Gaussian process regression model to fit its nonlinear change trend;

[0066] and the autocorrelation characteristics of the autonomic nervous fluctuation component are modeled using a square exponential kernel function;

[0067] The autonomic nervous rhythm characteristics including the low frequency oscillation power ratio, the high frequency oscillation power ratio and the ratio of the two are output.

[0068] In some embodiments, a deep learning model is used to extract eye movement trajectory features and motion coordination features from the motor behavior video, and a motor function score is calculated in combination with the standardized motion test results, including:

[0069] The spatiotemporal features of the motor behavior video are encoded by a three-dimensional convolutional neural network, and the spatiotemporal sequence features of the eye movement are extracted by the spatiotemporal convolution layer, and the eye movement trajectory feature vector is output.

[0070] A topological graph model of the joint motion data of the limbs is constructed using a graph convolution network, and an action coordination feature matrix is generated based on the spatial connection relationship of the joints.

[0071] The eye movement trajectory feature vector and the action coordination feature matrix are fused, and the time-frequency features of the electromyographic signals collected in the standardized motion test are simultaneously input, and the cross-modal feature weighting is performed by a multilayer perceptron to output the motor function score value.

[0072] In some embodiments, the blood pressure meter collected signals and the bracelet collected signals are analyzed for heart rate variability, the circadian rhythm fluctuation characteristics are calculated, and the neural function abnormality pattern is identified by a density clustering algorithm, including:

[0073] The blood pressure meter collected signals are analyzed for dynamic blood pressure trend extraction, and the periodic change curve of the systolic pressure / diastolic pressure is obtained.

[0074] The bracelet collected signals are analyzed for R-R interval sequence analysis, and the low frequency power and high frequency power circadian ratio is extracted based on frequency domain transformation.

[0075] The periodic change curve and the circadian ratio are input into the DBSCAN-based density clustering model, and the abnormal clustering cluster representing sympathetic-parasympathetic nerve imbalance is divided in the feature space of the density clustering model through adaptive neighborhood radius parameter optimization.

[0076] The data points corresponding to the periodic change curve and the circadian ratio falling into the abnormal clustering cluster are marked as the neural function abnormality pattern and the abnormal probability is output.

[0077] In a second aspect, the present application also provides a neural disease detection and analysis system suitable for the method of the first aspect, the system comprising: a wearable device, a behavior collection device, a signal processing unit, a disease analysis unit, the wearable device comprising at least one of a bracelet, glasses, a helmet, and earphones; the behavior collection device comprising at least one of a mobile phone, a tablet, and a camera; the signal processing unit being in communication connection with the wearable device and the behavior collection device, the signal processing unit being configured to acquire wearable device monitoring data and behavior test data of a user; performing multi-band decomposition on bracelet collection signals, and separating tremor components, gait components, and autonomic nervous fluctuation components by wavelet transform; extracting tremor intensity features from the tremor components by a long short-term memory network, extracting gait symmetry features from the gait components by a time series convolution network, and extracting autonomic nervous rhythm features from the autonomic nervous fluctuation components by a Gaussian process regression algorithm; extracting eye movement trajectory features and motion coordination features from motion behavior images by a deep learning model, and calculating a motor function score in combination with standardized motion test results; performing heart rate variability analysis on sphygmomanometer collection signals and bracelet collection signals, calculating circadian rhythm fluctuation features, and identifying a neural function abnormality pattern by a density clustering algorithm; the disease analysis unit being configured to construct a neural function state atlas based on the tremor intensity features, the gait symmetry features, the autonomic nervous rhythm features, the motor function score, and the neural function abnormality pattern, calculate contribution weights of the features to disease risk by a graph attention mechanism, acquire historical monitoring data of the user, learn continuous time evolution rules of the neural function state in the neural function state atlas according to the historical monitoring data by a neural differential equation, predict a disease progression trend in a preset future period, and generate a prediction result; acquire at least one of bracelet collection signals, sphygmomanometer collection signals, and motion behavior images captured by a mobile phone camera in the preset future period, record the at least one as feedback correction information, correct the prediction result based on the feedback correction information by an adaptive Kalman filter, and output a neural disease risk assessment report according to the prediction result.

[0078] Unlike the prior art, the above technical solution acquires bracelet collection signals, sphygmomanometer collection signals, motion behavior images, and behavior test data, extracts tremor intensity features, gait symmetry features, and autonomic nervous rhythm features from the bracelet collection signals by multi-band decomposition, analyzes motion behavior images and standardized motion tests to obtain a motor function score, performs heart rate variability analysis to identify a neural function abnormality pattern, constructs a neural function state atlas and calculates feature weights, predicts a disease progression trend in combination with historical monitoring data, and dynamically adjusts the prediction result through subsequent feedback correction information. The present application realizes long-term trend prediction based on initial data and dynamic correction by combining short-term intensive monitoring with long-term intermittent collection, significantly improving the reliability and practicality of neural disease risk assessment in a home environment.

[0079] The above summary related to the invention is only a summary of the technical solutions of the present application. In order to enable those skilled in the art to more clearly understand the technical solutions of the present application, and then implement the same according to the contents of the description and drawings, and in order to enable the above and other purposes, characteristics and advantages of the present application to be more easily understood, the following will be described in combination with the specific embodiments of the present application and the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0080] The accompanying drawings are only used to show the principles, implementation manners, applications, characteristics and effects of the specific embodiments and other related contents of the present application, and cannot be considered as limitations of the present application.

[0081] In the drawings:

[0082] Figure 1 Method step diagram of steps S101 to S109 of the method described in the specific embodiment;

[0083] Figure 2 Method step diagram of steps S201 to S204 of the method described in the specific embodiment. DETAILED DESCRIPTION

[0084] 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 of the present application in detail, the following will be described in combination with the specific embodiments listed and the accompanying drawings. The embodiments described in the present document are only used to more clearly illustrate the technical solutions of the present application, and therefore only serve as examples, and cannot limit the protection scope of the present application.

[0085] In the present document, the term "embodiment" means that the specific features, structures or characteristics described in combination with the embodiment can be included in at least one embodiment of the present application. The term "embodiment" appearing at various positions in the specification does not necessarily refer to the same embodiment, and does not 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, each technical feature mentioned in each embodiment can be combined in any manner to form a corresponding implementable technical solution.

[0086] Unless otherwise defined, the meanings of the technical terms used in the present document are the same as those commonly understood by those skilled in the art to which the present application belongs; the use of related terms in the present document is only for the purpose of describing specific embodiments, and is not intended to limit the present application.

[0087] In the description of the present application, the phrase "and / or" is a description of a logical relationship between 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.

[0088] 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.

[0089] In the present application, without more limitation, the "include", "contain", "have" or other similar open expressions used in the sentence are intended to cover non-exclusive inclusion, and these expressions do not exclude the existence of other elements in the process, method or product including the described 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 in such process, method or product.

[0090] As the same understanding as 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" and the like, unless otherwise explicitly limited.

[0091] In the description of the embodiments of the present application, the spatial-related expressions used, such as "center", "longitudinal", "transverse", "length", "width", "thickness", "up", "down", "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 based on the orientation or positional relationship shown in the specific embodiments or the drawings, and are only for the convenience of describing the specific embodiments of the present application or for the reader to understand, and do not indicate or imply that the indicated device or component must have a specific position, a specific orientation, or be constructed or operated in a specific orientation, and therefore cannot be understood as a limitation on the embodiments of the present application.

[0092] The processor described in 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 various embodiments of the present application.

[0093] 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 disk, 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 distributed 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 into the device, or connected to 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, local area network (LAN), wide area network (WAN), 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 into the parameter state of the deep neural network or other machine learning model through model training.

[0094] Please refer to Figure 1 In the first aspect, the embodiment provides a neural disease detection analysis method, comprising:

[0095] S101, acquiring wearable device monitoring data and behavior test data of a user, wherein the wearable device monitoring data includes bracelet acquisition signals, sphygmomanometer acquisition signals and motion behavior images, and the behavior test data includes standardized motion test results;

[0096] S102, performing multi-frequency segment decomposition on the bracelet acquisition signals, and separating tremor components, gait components and autonomic nerve fluctuation components by wavelet transform;

[0097] S103, extracting tremor intensity features from the tremor components by a long short-term memory network, extracting gait symmetry features from the gait components by a time series convolution network, and extracting autonomic nerve rhythm features from the autonomic nerve fluctuation components by a Gaussian process regression algorithm;

[0098] S104, extracting eye movement trajectory features and motion coordination features from the motion behavior image using a deep learning model, and calculating a motor function score in combination with the standardized motion test results;

[0099] S105, performing heart rate variability analysis on the sphygmomanometer collected signal and the bracelet collected signal, calculating circadian rhythm fluctuation features, and identifying abnormal neural function patterns through density clustering algorithm;

[0100] S106, based on tremor intensity features, gait symmetry features, autonomic nervous rhythm features, motor function scores and abnormal neural function patterns, constructing a neural function state atlas, and calculating the contribution weight of each feature to the disease risk through graph attention mechanism;

[0101] S107, obtaining the user's historical monitoring data, learning the continuous time evolution law of the neural function state in the neural function state atlas according to the historical detection data using neural differential equation, predicting the disease progression trend in the preset future period, and generating a prediction result;

[0102] S108, collecting at least one of the bracelet collected signal, the sphygmomanometer collected signal and the motion behavior image captured by the mobile phone camera in the preset future period, as feedback correction information, and correcting the prediction result based on the adaptive Kalman filter according to the feedback correction information;

[0103] S109, outputting a neural disease risk assessment report according to the prediction result.

[0104] In step S101, the wearable device monitoring data is collected in real time by the wearable device. Preferably, the bracelet collected signal contains acceleration signal and gyroscope signal, which is used to capture limb movement features. The sphygmomanometer collected signal can be obtained by photoelectric sensor, which reflects the cardiovascular regulation function. The motion behavior image is recorded by the mobile phone camera, which contains the space-time information of facial expressions and limb movements. The standardized motion test results are obtained through the preset clinical evaluation program, such as finger tapping test or balance test, which is used to quantify the degree of motor dysfunction. These multi-source data together constitute the basic data set for neural function evaluation.

[0105] In step S102, wavelet transform is used to decompose the bracelet collected signal into multiple frequency bands, where the tremor component corresponds to a specific frequency band signal, which is used to represent the involuntary movement feature; the gait component corresponds to another specific frequency band signal, which reflects the walking period feature; the autonomic nervous fluctuation component is located in a specific low frequency band, which reflects the sympathetic and parasympathetic regulation function. During signal decomposition, the decomposition parameters are dynamically adjusted according to the actual signal quality to ensure that each component retains effective physiological information while suppressing motion artifacts.

[0106] In step S103, the long short-term memory network is used to extract tremor intensity features, the tremor intensity features include an amplitude variability parameter and a duration parameter, the amplitude variability parameter represents the fluctuation intensity of the tremor signal and reflects the degree of abnormal motor control; the duration parameter records the duration of tremor attack and is used to assess the severity of the symptoms. The timing convolution network extracts gait symmetry features by analyzing the correlation between left and right limb signals. The autonomic nervous rhythm features are extracted by a Gaussian process regression algorithm, and the autonomic nervous rhythm features include quantitative indicators such as the ratio of low-frequency power to high-frequency power. The tremor intensity features reflect abnormal patterns of motor control function, the gait symmetry features reflect abnormal patterns of coordination function, and the autonomic nervous rhythm features reflect abnormal patterns of autonomic nervous regulation function.

[0107] In step S104, the motion behavior image data is processed by a deep learning model, wherein the eye movement trajectory features include a saccade speed parameter and a fixation stability parameter; the action coordination features are obtained by analyzing the motion relationship of the limb joint. The standardized action test results and the image features are processed by feature fusion, and the final output of the motor function score is used to comprehensively evaluate the functional status of the motor system.

[0108] In step S105, preferably, heart rate variability analysis is performed by processing R-R interval sequence data to calculate circadian rhythm fluctuation features. The density clustering algorithm is used to identify abnormal patterns of neural function, and the parameters are adaptively adjusted according to the actual data distribution. This step focuses on capturing abnormal features of circadian regulation of autonomic nervous function.

[0109] In step S106, the graph attention network adopts a three-layer structure, the first layer processes motion-related features, the second layer analyzes autonomic nervous regulation features, and the third layer fuses the two types of features through a cross-domain attention mechanism, thereby quantifying the contribution of different feature dimensions to disease risk, for example, in an exemplary implementation, the tremor intensity feature weight is generally higher than the gait symmetry feature.

[0110] In step S107, the neural differential equation modeling preferably uses a continuous time series analysis method, using historical monitoring data as the initial state, learning the evolution law of neural function state through a reversible neural network, predicting the functional change trend in the future within the same time range based on the monitoring data in the past within the same time range. The prediction result includes the change curve of each feature dimension and the comprehensive risk assessment value.

[0111] In step S108, the feedback correction mechanism is implemented through an adaptive Kalman filter, the correction process first calculates the deviation between the prediction result and the new observation data, and then dynamically adjusts the state estimation parameters. Preferably, when the correction amplitude exceeds a preset threshold for three consecutive times, the incremental learning of the neural function state atlas is automatically triggered.

[0112] The neurological disease risk assessment report output in step S109 includes clinical information such as current state assessment, trend prediction, and intervention suggestion.

[0113] The embodiment fuses wearable device monitoring data and behavior test data to construct a dynamically evolving neurological function evaluation system. Through multi-band decomposition and feature extraction technology, tremor intensity features, gait symmetry features, and autonomic nervous rhythm features are obtained from the bracelet collected signals. In combination with the eye movement trajectory features and motion coordination features obtained through motion behavior image analysis, a multi-dimensional motor function score is formed. Further, through heart rate variability analysis and density clustering, a neurological function abnormality pattern is identified, and a neurological function state graph reflecting the weight of each feature is constructed using a graph attention mechanism. Based on historical monitoring data, a neural differential equation model is used to predict disease progression trends, and adaptive Kalman filtering is used to dynamically correct the prediction results. The embodiment realizes the organic combination of short-term comprehensive data collection and long-term local monitoring, which not only guarantees the accuracy of the initial evaluation, but also maintains the long-term prediction accuracy through continuous feedback correction. The final output of the neurological disease risk assessment report provides a reliable basis for clinical decision-making, and relies on common devices to realize a convenient detection method at home.

[0114] Please refer to Figure 2 In some embodiments, based on the tremor intensity features, gait symmetry features, autonomic nervous rhythm features, motor function scores, and neurological function abnormality patterns, a neurological function state graph is constructed, and the contribution weight of each feature to the disease risk is calculated through a graph attention mechanism, including:

[0115] S201, the tremor intensity features, gait symmetry features, autonomic nervous rhythm features, motor function scores, and neurological function abnormality patterns are used as node features of a graph neural network, and an initial edge connection relationship is constructed according to the physiological correlation between each node feature;

[0116] S202, a graph attention mechanism is used to model the spatial dependence of the node features, calculate the motor function correlation weight between the tremor intensity features and the gait symmetry features, the autonomic nervous regulation weight between the autonomic nervous rhythm features and the neurological function abnormality pattern, and the comprehensive coordination weight between the motor function score and other node features;

[0117] S203, through a multi-layer graph convolution network, the neighborhood node information is aggregated, the embedding representation of each node feature is iteratively optimized, and a graph embedding vector reflecting the overall state of the neurological function is output;

[0118] S204, based on the atlas embedding vector, generate the dynamic contribution weight of each node feature to the disease risk through the attention score calculation module, and the weight of the tremor intensity feature and the gait symmetry feature is used for motor dysfunction evaluation, the weight of the autonomic nervous rhythm feature and the abnormal mode of neural function is used for autonomic nervous function evaluation, and the weight of the motor function score is used for comprehensive behavior function evaluation.

[0119] In step S201, the construction of node features follows the actual correlation of physiological systems, wherein the tremor intensity feature and the gait symmetry feature are connected through the motor control neural network, the autonomic nervous rhythm feature and the abnormal mode of neural function are connected through the brainstem regulation pathway, and the motor function score as a comprehensive index is connected with each node feature. The weight distribution of the initial edge connection adopts a semi-supervised method based on medical prior knowledge, for example, the initial weight of the connection between the motion-related features is set to be higher than the weight of the cross-system connection, so as to ensure that the graph structure has clinical interpretability at the beginning of training.

[0120] In step S202, the spatial dependence modeling is realized by using a multi-head graph attention mechanism, each attention head focusing on a specific type of feature interaction. The calculation of the motor function correlation weight introduces a time alignment module to ensure that the interaction between the tremor intensity feature and the gait symmetry feature in the same physiological cycle is accurately captured; the autonomic nervous regulation weight is enhanced through frequency domain coherence analysis, focusing on extracting the phase synchronization features of low-frequency signals; and the comprehensive coordination weight adopts a cross-modal attention mechanism, so that the motor function score can dynamically adjust the influence strength of different node features.

[0121] In step S203, each layer of the multi-layer graph convolution network includes two stages of feature space projection and neighborhood information aggregation. The feature space projection is realized by a learnable affine transformation, which maps node features of different dimensions to a unified embedding space; and the neighborhood information aggregation adopts a gating mechanism to dynamically control the information propagation strength according to the edge weight. In the iterative optimization process, the node embedding representation gradually fuses local details and global topological features, and the finally generated atlas embedding vector contains multi-scale representations from microscopic feature fluctuations to macroscopic functional states.

[0122] In step S204, preferably, the dynamic contribution weight calculation introduces a differentiable weight constraint mechanism to ensure that the weight distribution of the motor dysfunction evaluation, the autonomic nervous function evaluation and the comprehensive behavior function evaluation conforms to the cognitive logic of clinical medicine. In specific implementation, the motor dysfunction evaluation focuses on the abnormal degree of feature amplitude, the autonomic nervous function evaluation focuses on the duration of rhythm disorder, and the comprehensive behavior function evaluation integrates the coordination index of spatio-temporal features. Further, the output weight can be calibrated by clinical expert knowledge to ensure its consistency with existing medical evaluation standards.

[0123] This embodiment constructs a neural function state atlas, using multivariate physiological indicators such as tremor intensity and gait symmetry as graph nodes. Initial edge connections are established based on physiological correlations, and graph attention mechanisms are used to calculate feature interaction relationships such as motor function correlation weights and autonomic nervous system modulation weights. Finally, dynamic contribution weights are output for multidimensional assessment. This embodiment combines medical prior knowledge with graph neural networks, capturing nonlinear correlations between features through spatial dependency modeling. This allows short-term collected multivariate data to reflect the overall state of neural function, enabling long-term disease risk prediction based on fragmented monitoring data in a home environment. Furthermore, iteratively optimized graph embedding vectors support subsequent dynamic corrections, improving the comprehensiveness and timeliness of the assessment while ensuring clinical interpretability.

[0124] In some embodiments, historical monitoring data of the user is acquired, and the continuous-time evolution law of the neural functional state is learned in the neural functional state atlas based on the historical monitoring data using neural differential equations to predict the disease progression trend in a preset future time period, generating prediction results, including:

[0125] Historical monitoring data is integrated into a neural functional state atlas to form multiple historical embedding vectors. These vectors are then arranged in a time series to construct a temporal evolution sequence of neural functional states.

[0126] The time evolution sequence is calculated using neural differential equations. The gradual change process of neural functional state is described by a learnable continuous-time dynamic function, which is expressed by formula (1), as follows:

[0127] ;

[0128] In formula (1), For neural differential equations, It is a learnable continuous-time dynamic function. for The neural functional state vector at time t, for The weight parameters, As a continuous-time variable, a learnable continuous-time dynamic function is defined in the neural differential equation framework, expressed by formula (3), which is as follows:

[0129] ;

[0130] in, It is the sigmoid activation function. Weight parameters The first weight matrix, Weight parameters The second weight matrix, Weight parameters The third weight matrix, This refers to the nonlinear activation function in the neural differential equation. For the first bias term parameter, For the second bias term parameter, For the third bias term parameter, It is the hyperbolic tangent function;

[0131] A gating mechanism is introduced into the neural differential equation framework to dynamically adjust the rate of state change at different time scales in order to capture the short-term fluctuations and long-term degenerative characteristics of neural function. The neural differential equation model is obtained and expressed by formula (2), which is as follows:

[0132] ;

[0133] In formula (2), This is the gate vector corresponding to the gating mechanism, with a value range of [0,1]. It is generated by the sigmoid function and controls the state update ratio. For residual terms, capture long-term degradation trends. This is element-wise multiplication;

[0134] Based on the current neural functional state map and the trained neural differential equation model, multi-step state prediction is performed within a preset future time period, and the neural functional state prediction trajectory containing temporal continuity is output.

[0135] The predicted trajectory of neurological function status is probabilistically processed to calculate the disease risk probability distribution at each time point and identify key time points where the risk exceeds a preset threshold.

[0136] By comparing and analyzing the predicted trajectory of neurological function status with the user's individual baseline data, personalized predictions of disease progression trends are generated. The predictions include risk level change curves and early warning information for key time points.

[0137] In this embodiment, historical monitoring data is integrated into a neural functional state atlas through a time alignment module. The resulting historical embedding vector preserves the spatial topological relationships of the original features, ensuring that the temporal evolution sequence contains both temporal information and maintains the atlas's structural characteristics. The neural differential equation framework employs an adaptive step-size numerical solution method (such as the dopri5 algorithm) to handle continuous-time modeling. The learnable continuous-time dynamic function is constructed using a multilayer perceptron, and its weight parameters... During training, gradient optimization is performed using the accompanying sensitivity method, enabling the system to automatically learn dynamic patterns across multiple scales, from minute-level tremor fluctuations to month-level functional decline.

[0138] The implementation of the gating mechanism adopts a physiological time scale division, preferably, short-term fluctuation gating Focus on the state change within 0-6 hours, extract transient features through LSTM network; long-term degradation gating Then associate the gradual functional decline in units of weeks, residual term A linear transformation with an attenuation factor is used to simulate the irreversible damage of neural function. This double-time-scale modeling method is particularly important in the prediction of Parkinson's disease and other degenerative diseases, and can distinguish between temporary symptom fluctuations and substantial functional deterioration.

[0139] The generation of the gating vector is self-adaptively calibrated by the time-frequency analysis results of historical monitoring data, and the division of the short-term fluctuation time scale (0-6 hours) is determined according to the power spectral density peak value of the tremor signal within the user's daily activity cycle; the attenuation factor of the residual term is dynamically adjusted according to the slope of the historical degradation trend, ensuring that the long-term degradation modeling matches the decline rate of the neurological function score recorded in the clinical follow-up.

[0140] The probability processing stage preferably simulates prediction uncertainty through Monte Carlo dropout, and the calculation of the disease risk probability distribution integrates three key factors: the Euclidean distance between the neurological function state prediction trajectory and the preset threshold, the confidence interval of the historical prediction error, and the dynamic offset of the individualized baseline data. The identification of key time nodes can use a change point detection algorithm, which triggers an early warning when the first derivative of the disease risk probability exceeds a preset threshold, preferably set by a specialist consensus to a safe boundary of no more than 5% weekly risk growth.

[0141] The generation of personalized prediction results introduces a transfer learning strategy, with user individual baseline data as source domain features, and the distribution matching degree of the neurological function state prediction trajectory is adjusted through a maximum mean difference loss function. The risk level change curve uses dynamic color encoding, mapping the prediction values of motor dysfunction assessment, autonomic nervous function assessment, and comprehensive behavior function assessment in three dimensions to a unified risk index, and the key time node warning information is integrated through a Bayesian network Multiple modal prediction results, finally output a structured report that meets the requirements of a clinical decision support system.

[0142] This embodiment models the continuous-time evolution of neural functional states using neural differential equations. A time-evolutionary sequence constructed from historical monitoring data is input into a dynamic system with a gating mechanism. A learnable continuous-time dynamic function is used to simultaneously capture short-term fluctuations and long-term degenerative characteristics, ultimately outputting a time-continuous predictive trajectory. This embodiment integrates the temporal scale division of physiological constraints into the gating mechanism design. Through the synergistic effect of short-term fluctuation gating and long-term degenerative gating, the prediction results based on fragmented monitoring data can reflect both immediate symptom changes and track disease progression trends. Combined with probabilistic processing and individual baseline comparison, it achieves cross-scale disease risk assessment from short-term observation to long-term prediction. Simultaneously, a dynamically updated neural functional state atlas supports continuous correction of the prediction results, significantly improving the timeliness and personalization of the prediction while ensuring clinical applicability.

[0143] In some embodiments, at least one of the following is collected within a preset future time period: a signal collected by a wristband, a signal collected by a blood pressure monitor, and motion behavior images captured by a mobile phone camera, denoted as feedback correction information. The prediction result is corrected based on the feedback correction information using an adaptive Kalman filter, including:

[0144] The feedback correction information is input into the feature extraction module. Based on the extracted feedback features, at least one of the following is updated: tremor intensity feature, gait symmetry feature, autonomic rhythm feature, and motor function score, to generate the current neurological function state observation vector.

[0145] A state-space model is constructed, using the predicted trajectory output by the neural differential equation as the state transition prior and the neural functional state observation vector as the measurement value to generate the state estimation result. The constructed state-space model is represented by formula (4), which is as follows:

[0146] ;

[0147] Define observations To pass through the observation matrix The predicted values ​​are obtained by mapping them, and can be expressed by the formula as follows: When the observation matrix When the matrix is ​​identity, the observation equation simplifies to , in formula (4), For the first The neural functional state observation vector at time t, i.e., the measured value. The predicted trajectory output by the neural differential equation is in the th... The neural functional state vector at time t, i.e., the predicted value. The state transition matrix is ​​obtained by discretizing the continuous-time differential equation. To observe the noise, This is process noise;

[0148] Adopting the adaptive Kalman filtering algorithm, the confidence weight of the state estimation result is generated by dynamically adjusting the process noise covariance matrix and the observation noise covariance matrix;

[0149] And when the deviation between the observation value and the prediction value exceeds the dynamic threshold, the recalibration process of the prediction model is triggered, and the noise parameters are dynamically adjusted based on the difference between the prediction value And the measured value The process noise reflects the model uncertainty, and the observation noise reflects the data reliability. The process noise covariance matrix is represented by formula (5):

[0150] ;

[0151] Wherein, is the process noise covariance matrix, is the model uncertainty coefficient, is the baseline noise level, is the Jacobian matrix of the continuous-time dynamic function in formula (3) with respect to the weight parameter , and is the unit matrix;

[0152] The observation noise covariance matrix is represented by formula (6):

[0153] ;

[0154] Wherein, is the observation noise covariance matrix, is the scaling factor of the dynamic threshold, is the observation value, is the standard deviation of the normal fluctuation range, is the diagonal matrix;

[0155] According to the confidence weight, the state estimation result is corrected;

[0156] The corrected state estimation result is fed back to the neural differential equation model, the parameters of the dynamic function are adjusted to adapt to the individualized neural function change mode of the user, and the corrected prediction result is output.

[0157] In this embodiment, the feature extraction module of the feedback correction information adopts the same processing flow as the historical monitoring data to ensure the consistency of the feature space. Tremor intensity feature update is realized through a long short-term memory network to retain the time dependence of the original signal; gait symmetry feature update introduces a phase calibration module to eliminate the time misalignment caused by intermittent collection; autonomic nervous rhythm feature update maintains a Gaussian process regression framework and increases sliding window smoothing to adapt to dynamic adjustment requirements. The motor function score update adopts a weighted fusion strategy, and the weights of new data and historical data are automatically adjusted according to the collection time length to ensure that data under different monitoring frequencies can be effectively integrated.

[0158] In the construction of the state space model, the dimension of the observation matrix is determined by the feature dimension of the neural function state vector, and is initially set as a unit matrix to reflect the characteristic that each dimension feature can be directly observed, and is automatically converted into a mask matrix when there is partial data missing. Preferably, the state transition matrix is obtained through the discretization processing of the neural differential equation, and a fixed step Runge-Kutta method is used to realize the conversion from continuous time domain to discrete time domain, and the time step is adaptively determined according to the average collection interval of the user's historical monitoring data. The covariance matrices of the observation noise and the process noise are dynamically adjusted through an adaptive mechanism, wherein the construction of the process noise covariance matrix is ensured to match the dimension of the Jacobian matrix and the unit matrix through the flattening processing of the weight parameter , and the exponential term of the observation noise covariance matrix is designed to be an element-wise operation and to be element-wise multiplied with a diagonal matrix.

[0159] Preferably, the value of the dynamic threshold is obtained by rolling standard deviation calculation of the neural function state vector in the user's historical data, and the re-calibration process is triggered when the prediction deviation exceeds three times the standard deviation; in the calibration process, the transfer learning strategy is preferentially used to adjust the dynamic function parameters, and the gradient information of the weight parameter is obtained in real time through automatic differentiation technology to ensure strict synchronization with the noise estimation of the adaptive filtering process. The corrected state estimation is subjected to feature weight rebalancing through a graph attention network, and is compatible with the overall topological structure of the neural function state atlas. The discrete time index and the corresponding continuous time variable establish a clear mapping relationship.

[0160] The model uncertainty coefficient in the process noise covariance matrix is inversely proportional to the root mean square error of the prediction error of the neural differential equation training set, and the baseline noise level takes the median of the fluctuation amplitude of the user's historical state vector; the observation noise scaling factor is linked with the dynamic threshold, and is automatically increased to reduce the weight of abnormal observations when the number of consecutive corrections exceeds a preset value.

[0161] The embodiment realizes dynamic optimization of neurofunctional degeneration prediction through real-time feedback correction mechanism of multi-source physiological signals. The feature extraction module maintains the consistency of the feature space of historical data and real-time signals. The state space model dynamically balances the credibility of the prediction trajectory and the observation data using adaptive Kalman filtering. The intelligent switching mechanism of the observation matrix effectively handles data missing situations. The adaptive adjustment of the noise covariance matrix accurately distinguishes model errors and measurement disturbances. When the prediction deviation exceeds the threshold based on historical fluctuation statistics, the parameter recalibration process is triggered, and individual adaptation is achieved through transfer learning of neural differential equation parameters. The embodiment significantly improves the stability of long-term prediction, ensuring the convenience of home use while allowing short-term high-precision data collection to continuously correct the prediction results under long-term sparse monitoring, effectively overcoming the evaluation bias caused by discontinuous data in the home environment.

[0162] In some embodiments, the method further comprises:

[0163] determining whether the parameters in the feedback correction information exceed their corresponding normal fluctuation range, and if so, triggering a re-collection mode;

[0164] Alternatively, calculating the deviation of the corrected prediction result from the initial prediction result, and when the deviation exceeds a preset deviation threshold, triggering the re-collection mode;

[0165] The re-collection mode is configured to:

[0166] re-collecting wearable device monitoring data and behavior test data of the user;

[0167] re-performing multi-band decomposition on the wristband collected signals, using wavelet transform to separate tremor components, gait components, and autonomic nervous fluctuation components;

[0168] re-extracting tremor intensity features from the tremor components through a long short-term memory network, gait symmetry features from the gait components through a time series convolution network, and autonomic nervous rhythm features from the autonomic nervous fluctuation components through a Gaussian process regression algorithm;

[0169] re-extracting eye movement trajectory features and motion coordination features from the motion behavior images using a deep learning model, and calculating motor function scores in combination with standardized motion test results;

[0170] re-performing heart rate variability analysis on the sphygmomanometer collected signals and the wristband collected signals, calculating circadian rhythm fluctuation features, and identifying neurofunctional abnormality patterns through a density clustering algorithm;

[0171] Reconstruct the neurological function state atlas based on tremor intensity features, gait symmetry features, autonomic nervous rhythm features, motor function scores, and neurological dysfunction patterns, and calculate the contribution weight of each feature to the disease risk through graph attention mechanism;

[0172] In the neurological function state atlas, the continuous time evolution law of the neurological function state is learned from the historical detection data using a neural differential equation, the disease progression trend in the preset future period is predicted, and an updated prediction result is generated.

[0173] In this embodiment, the trigger mechanism of the re-collection mode is realized by multi-dimensional verification to ensure reliability. The parameters in the feedback correction information include not only the tremor component amplitude collected by the bracelet, the gait cycle coefficient of variation and other time domain features, but also the smoothness index of eye movement trajectory obtained through motion behavior image analysis and other spatial features. The normal fluctuation range of these parameters is determined by the dynamic percentile method, that is, the distribution interval of the feature value is calculated according to the user's recent several effective monitoring results, and a proper safety margin is retained. When it is detected that the parameter value exceeds the interval for several times in succession, it indicates that there may be device abnormalities or user state mutations, and the re-collection mode triggered at this time will preferentially collect the original signal corresponding to the controversial parameter.

[0174] Preferably, the calculation window length of the distribution interval is set to the user's recent 30 effective monitoring periods by default, and the safety margin is set differently according to the feature type: the tremor component takes 1.5 times the standard deviation in the window, the gait feature takes 1.2 times, and the eye movement feature uses the 95% percentile interval of the non-parametric method. The window length and safety margin parameters can be manually adjusted by the clinician interface according to the disease course stage.

[0175] The re-collection mode implements key monitoring strategies for abnormal parameters on the basis of maintaining the standard processing flow. For the case of abnormal tremor component, the collection time is extended and the sampling frequency is increased; when the gait feature deviates, test scenarios under different walking states are added. This targeted collection strategy optimizes the collection efficiency while ensuring data quality by adjusting the combination of collection time, sampling frequency, and test scenarios.

[0176] In the feature extraction link, the model parameters are adaptively adjusted according to the characteristics of the re-collection data. For example, when the signal quality is detected to be reduced, the denoising strength of wavelet transform is automatically enhanced; if the feature distribution is found to be shifted, the input normalization parameters of the neural network model are adjusted accordingly. This dynamic adaptation mechanism effectively improves the robustness of feature extraction under abnormal conditions.

[0177] Preferably, the atlas construction stage introduces a feature credibility evaluation mechanism to dynamically adjust the initial weights of each feature in the graph attention network according to the quality indicators of the reacquired data. For parameters with less than ideal acquisition conditions, their contribution weights are appropriately reduced, while the decision-making proportion of other reliable features is increased. This adaptive feature fusion strategy ensures the stability of the final prediction results.

[0178] Preferably, the prediction model update adopts a gradual adjustment strategy to control the update amplitude of the model parameters by comparing the difference between the new and old prediction results. When the difference is small, fine-tuning is used to retain existing knowledge, and when the difference is significant, a complete model retraining process is started. This hierarchical model update mechanism ensures the timeliness of the prediction while avoiding model oscillation caused by single abnormal data.

[0179] This embodiment realizes the self-correction ability of the prediction system by establishing a dynamic feedback correction mechanism that automatically triggers the reacquisition mode when parameter abnormalities or prediction results deviate. This mechanism first judges the data reliability based on the personalized normal fluctuation range or deviation threshold, and then performs a complete reacquisition and processing process, including multi-source data acquisition, signal decomposition, feature extraction, atlas construction, and trend prediction. By preferentially acquiring controversial parameters, adaptively adjusting the processing strategy, and gradually updating the model, the system can continuously provide accurate assessment of the neurological status and prediction of disease progression, significantly improving the reliability and practicality of long-term monitoring.

[0180] In some embodiments, the wristband acquires signals including acceleration signals and gyroscope signals, and the wristband acquires signals are subjected to multi-band decomposition, and wavelet transform is used to separate tremor components, gait components, and autonomic nervous fluctuation components, including:

[0181] The acceleration signals and gyroscope signals are subjected to multi-scale decomposition using a wavelet basis function with time-frequency localization characteristics, and the original signals are separated into a first frequency band signal containing tremor characteristics, a second frequency band signal containing gait cycle characteristics, and a third frequency band signal reflecting autonomic nervous regulation according to a pre-set frequency band;

[0182] An adaptive decomposition algorithm based on signal energy distribution is used for the first frequency band signal, the second frequency band signal, and the third frequency band signal, and the number of wavelet decomposition layers is dynamically adjusted according to the signal-to-noise ratio of each frequency band signal to obtain decomposed first sub-band signals, second sub-band signals, and third sub-band signals, and motion artifact suppression processing is performed on the decomposed sub-band signals;

[0183] The first sub-band signal subjected to motion artifact suppression processing is combined into a time-domain continuous tremor component by a phase consistency reconstruction algorithm;

[0184] The second sub-band signal subjected to the motion artifact suppression processing is subjected to motion cycle phase calibration using a dynamic time warping algorithm to obtain a gait component;

[0185] The third sub-band signal subjected to the motion artifact suppression processing is sequentially subjected to ensemble empirical mode decomposition to remove baseline drift, maximum likelihood estimation to restore physiological rhythm components, and 0.01-0.1 Hz band-pass filtering processing to extract sympathetic-parasympathetic nerve regulation features to obtain an autonomic nerve fluctuation component.

[0186] In the embodiment, the wavelet base function with time-frequency localization characteristics is preferably a db4 wavelet, and the compact support characteristic is suitable for capturing the transient characteristics of limb movement. The acceleration signal and the gyroscope signal are ensured to be spatio-temporally aligned through synchronous acquisition, and the division of the preset frequency band is determined according to the typical physiological frequency bands of clinical tremor, gait and autonomic nerve fluctuation.

[0187] The adaptive decomposition algorithm calculates the short-time energy entropy of each frequency band signal through a sliding window, automatically increases the decomposition layer number when the signal-to-noise ratio is lower than a preset signal-to-noise ratio threshold to enhance the feature separation effect, and otherwise reduces the layer number to avoid over-decomposition. Preferably, the preset signal-to-noise ratio threshold is determined by energy distribution statistics of user historical monitoring data, and the decomposition layer number adjustment is triggered when the signal quality evaluation index is lower than the threshold. The motion artifact suppression processing uses independent component analysis combined with a motion sensor reference signal to eliminate interference components introduced by non-physiological limb activity through blind source separation.

[0188] The phase consistency reconstruction algorithm calculates the instantaneous phase gradient of the first sub-band signal, selects the frequency band components with a phase synchronicity higher than a preset value for superposition to ensure the time-domain continuity of the tremor component. Preferably, the phase synchronicity preset value is dynamically determined by analyzing the phase stability characteristics of the user's historical tremor signal, and the frequency band components matching the typical pathological tremor pattern are preferentially retained. The dynamic time warping algorithm is used for the second sub-band signal, and the user's historical gait cycle is used as a template for nonlinear alignment to eliminate phase shifts caused by changes in walking speed, so that the gait component retains cycle stability. In the processing of the third sub-band signal, the ensemble empirical mode decomposition eliminates low-frequency drift caused by breathing or device movement through noise-assisted decomposition; the maximum likelihood estimation restores the rhythm components hidden by noise based on the Markov characteristics of physiological rhythms; and the 0.01-0.1 Hz band-pass filtering finally extracts the frequency band reflecting the balance of sympathetic-parasympathetic nerves, and the cutoff frequency strictly corresponds to the very low frequency band and the low frequency band of heart rate variability analysis.

[0189] The embodiment separates the mixed acceleration / gyroscope signal into tremor component, gait component and autonomic nervous fluctuation component through multi-band decomposition and adaptive signal processing of wavelet transform, wherein the number of decomposition layers is dynamically adjusted to ensure the robustness of feature extraction, and the phase consistency reconstruction and dynamic time warping solve the problems of fragmentation of tremor signal and asynchronous gait phase, and the reliable extraction of sympathetic-parasympathetic nerve features is realized by combining physiological rhythm recovery and band-pass filtering, which provides high-fidelity multi-modal physiological feature input for predicting long-term disease risk based on short-time data.

[0190] In some embodiments, the tremor component is used to extract tremor intensity features through a long short-term memory network, including:

[0191] A bidirectional long short-term memory network model with attention mechanism is constructed;

[0192] The tremor component is subjected to time series analysis to obtain amplitude change rate and time series fluctuation features of the tremor component;

[0193] The amplitude change rate and the time series fluctuation features are used to generate tremor intensity features representing the severity of tremor;

[0194] The gait component is used to extract gait symmetry features through a time series convolution network, including:

[0195] The gait component is input into a multi-scale dilated time series convolution network, and different time scale gait period features are extracted through multiple dilated convolution layers arranged in parallel;

[0196] The cross-correlation functions of left and right limb movement signals in time and frequency domains are calculated according to the gait period features;

[0197] The Hilbert transform is used to calculate the phase difference, and an asymmetry feature vector representing gait asymmetry is output;

[0198] The autonomic nervous fluctuation component is used to extract autonomic nervous rhythm features through a Gaussian process regression algorithm, including:

[0199] The autonomic nervous fluctuation component is input into a Gaussian process regression model to fit its nonlinear change trend;

[0200] And a square exponential kernel function is used to model the autocorrelation characteristics of the autonomic nervous fluctuation component;

[0201] Autonomic nervous rhythm features including low frequency oscillation power ratio, high frequency oscillation power ratio and the ratio of the two are output.

[0202] In the embodiment, the bidirectional long short-term memory network model with attention mechanism captures the long-range dependence of tremor components through a gating unit, where the attention mechanism dynamically weights the amplitude mutation information of key time nodes. The bidirectional structure synchronously analyzes historical and future tremor trends, thereby improving the sensitivity of tremor intensity features to disease progression. Preferably, the amplitude change rate quantifies the fluctuation rate of tremor energy through a difference operation, and the time series fluctuation feature represents the non-stationarity of the tremor signal using the Hurst index. The generated tremor intensity feature after fusion can distinguish between physiological tremor and pathological tremor deterioration patterns.

[0203] In the multi-scale dilated temporal convolution network, the dilated convolution layers arranged in parallel cover multi-granularity analysis from a single step to a complete gait cycle through differentiated dilation coefficients, and the hollow structure expands the receptive field while maintaining the temporal resolution. The cross-correlation function calculation uses the sliding window covariance method combined with the instantaneous phase difference detection of the Hilbert transform, so that the symmetry feature vector simultaneously reflects the amplitude difference and time delay of the left and right limb movements, accurately quantifying the gait abnormalities caused by neurological diseases such as Parkinson's disease.

[0204] In the Gaussian process regression algorithm, the square exponential kernel function adaptively matches the quasi-periodicity of the autonomic nervous fluctuation component through the length scale parameter, and the fitted nonlinear trend preserves the physiological delay characteristics of sympathetic-parasympathetic nerve regulation. The extraction of the low-frequency to high-frequency oscillation power ratio strictly follows the frequency band division standard of heart rate variability analysis, where the low-frequency band (0.04-0.15Hz) reflects sympathetic nerve tension, and the high-frequency band (0.15-0.4Hz) represents parasympathetic nerve activity. The ratio of the two is directly corresponding to the autonomic nervous balance index and the clinical evaluation standard.

[0205] The embodiment captures the temporal evolution of tremor components through a bidirectional long short-term memory network, and enhances key pathological feature extraction through an attention mechanism, so that the tremor intensity feature can sensitively reflect the disease development trend. The multi-scale dilated temporal convolution network analyzes the gait cycle from different time granularities, and cooperates with cross-correlation analysis and phase difference detection to accurately quantify gait asymmetry. The Gaussian process regression accurately extracts autonomic nervous rhythm features through nonlinear modeling. The synergistic effect of the three feature extraction methods provides high-precision multi-dimensional quantitative basis for predicting the progression of neurological diseases based on short-term monitoring data.

[0206] In some embodiments, deep learning models are used to extract eye movement trajectory features and motion coordination features from motor behavior videos, and motion function scores are calculated based on standardized motion test results, including:

[0207] The three-dimensional convolutional neural network encodes the spatiotemporal features of the motor behavior video, and the spatiotemporal convolution layer extracts the spatiotemporal sequence features of the eye movement, outputting an eye movement trajectory feature vector;

[0208] A topological graph model of limb joint motion data is constructed using a graph convolution network, and an action coordination feature matrix is generated based on the spatial connection relationship of the joint space.

[0209] The eye movement trajectory feature vector and the action coordination feature matrix are fused, and the electromyogram time-frequency features collected in the standardized action test are also input, and a multi-layer perceptron is used for cross-modal feature weighting to output the motor function score.

[0210] In this embodiment, the spatio-temporal convolution layer of the three-dimensional convolutional neural network adopts a separated 3D convolution architecture, in which the spatial convolution kernel extracts the eyeball position features in a single frame of image, and the temporal convolution kernel captures the eye movement trajectory change pattern between consecutive frames, and the two work together to realize the spatio-temporal coding of eye movement behavior. The eye movement trajectory feature vector contains quantitative indicators such as saccade speed and fixation point dwell time, and its dimension is aligned with the clinical parameters of the standardized eye movement database, ensuring the interpretability of the features.

[0211] The topological graph model constructed by the graph convolution network takes the anatomical connection relationship of the human body as the initial edge weight, and dynamically adjusts the functional connection strength between the joints through adaptive graph learning. The action coordination feature matrix is generated through a spatio-temporal graph attention mechanism, which not only preserves the motion dominance of large joints such as shoulder-elbow-wrist, but also captures the fine coordination pattern of finger micro-motions, thereby comprehensively reflecting the detailed features of motor dysfunction.

[0212] In the feature fusion stage, a gated attention unit is used to realize cross-modal interaction. Preferably, the electromyogram time-frequency features extract the activation timing and intensity information of each muscle group through short-time Fourier transform, which is complementary to the visual behavior features; the weighting coefficients of the multi-layer perceptron are determined through pre-training comparison, so that the motor function score meets the discrimination standard of the clinical scale and the sensitivity requirement of home monitoring, providing standardized input for the transformation of short-term evaluation results to long-term prognosis prediction.

[0213] This embodiment accurately extracts the spatio-temporal features of eye movement trajectory through a three-dimensional convolutional neural network, models the topological relationship of limb joint motion through a graph convolution network, and realizes multi-dimensional analysis of motor behavior; by fusing eye movement features, action coordination features and standardized electromyogram features, a comprehensive motor function score is generated using a cross-modal weighting mechanism. This scheme converts complex clinical motor function evaluation into quantifiable objective indicators, providing a high-precision evaluation basis for predicting the progression of neurodegenerative diseases based on short-term behavior data, while also considering the implementation convenience of home scenarios.

[0214] In some embodiments, the blood pressure meter collected signal and the bracelet collected signal are analyzed for heart rate variability, the circadian rhythm fluctuation features are calculated, and the neural function abnormality pattern is identified through a density clustering algorithm, including:

[0215] The dynamic blood pressure trend extraction is performed on the blood pressure meter collected signals to obtain the periodical change curve of systolic pressure / diastolic pressure;

[0216] The R-R interval sequence analysis is performed on the wristband collected signals, and the day-night ratio of low frequency power and high frequency power is extracted based on frequency domain transformation;

[0217] The periodical change curve and the day-night ratio are input into the DBSCAN-based density clustering model, and the adaptive neighborhood radius parameter is optimized, and the abnormal clustering cluster representing the sympathetic-parasympathetic nerve imbalance is divided in the feature space of the density clustering model;

[0218] The data points corresponding to the periodical change curve and the day-night ratio falling into the abnormal clustering cluster are marked as the abnormal nerve function mode and the abnormal probability is output.

[0219] In the embodiment, the dynamic blood pressure trend extraction adopts the sliding window smoothing algorithm to process the original blood pressure signal, and the periodical change curve of systolic pressure / diastolic pressure is obtained by 24-hour segmentation fitting, and the least square fitting coefficient of each period is used to represent the blood pressure circadian rhythm characteristics. Preferably, the R-R interval sequence analysis adopts the Lomb-Scargle periodogram method to process the non-uniform sampling data, and the day-night ratio calculation of the low frequency power (0.04-0.15 Hz) and the high frequency power (0.15-0.4 Hz) adopts the segmented integral method to ensure the analysis reliability in the home discontinuous monitoring scene.

[0220] The DBSCAN-based density clustering model optimizes the neighborhood radius parameter through the dynamic core point determination strategy, and the best neighborhood range can be determined by the k-distance diagram method, so that the feature space can accommodate the multi-dimensional features of blood pressure variability and heart rate variability. The determination standard of the abnormal clustering cluster combines the threshold boundary set by the clinical guidelines, and when the clustering center deviates from the normal physiological reference range, the abnormal nerve function marking is automatically triggered. The output abnormal probability calculates the membership degree of the data point and the abnormal cluster center through the kernel density estimation, and the numerical range corresponds to the grading standard of the existing nerve function evaluation scale, which is convenient for clinical result interpretation.

[0221] The embodiment obtains the blood pressure periodical change curve and the heart rate variability day-night ratio through the dynamic blood pressure trend extraction and the R-R interval analysis, and automatically identifies the sympathetic-parasympathetic nerve imbalance mode by using the density clustering model. This method converts the discrete monitoring data into continuous nerve function evaluation indicators, realizes the objective identification of early nerve function abnormalities through intelligent division of abnormal clustering clusters, provides a quantitative basis for predicting autonomic nervous system function degradation based on short-term physiological data, and meets the convenient monitoring demand of home scene.

[0222] In a second aspect, the embodiment also provides a neural disease detection and analysis system, which is suitable for the method of the first aspect. The system comprises a wearable device, a behavior collection device, a signal processing unit, and a disease analysis unit. The wearable device comprises at least one of a bracelet, glasses, a helmet, and a headset. The behavior collection device comprises at least one of a mobile phone, a tablet, and a camera. The signal processing unit is in communication connection with the wearable device and the behavior collection device. The signal processing unit is used to obtain wearable device monitoring data and behavior test data of a user. The bracelet collection signal is subjected to multi-band decomposition, and a tremor component, a gait component, and an autonomic nerve fluctuation component are separated by wavelet transform. The tremor component is subjected to long short-term memory network extraction to extract tremor intensity features, the gait component is subjected to time series convolution network extraction to extract gait symmetry features, and the autonomic nerve fluctuation component is subjected to Gaussian process regression algorithm extraction to extract autonomic nerve rhythm features. A deep learning model is used to extract eye movement trajectory features and motion coordination features from the motion behavior image, and a motion function score is calculated in combination with a standardized motion test result. Heart rate variability analysis is performed on the sphygmomanometer collection signal and the bracelet collection signal, diurnal rhythm fluctuation features are calculated, and a neural function abnormality mode is identified by a density clustering algorithm. The disease analysis unit is used to construct a neural function state atlas based on the tremor intensity features, the gait symmetry features, the autonomic nerve rhythm features, the motion function score, and the neural function abnormality mode. The contribution weight of each feature to the disease risk is calculated by a graph attention mechanism. Historical monitoring data of the user is obtained, and a neural differential equation is used to learn the continuous time evolution law of the neural function state in the neural function state atlas according to the historical detection data, so as to predict the disease progression trend in a preset future period and generate a prediction result. At least one of the bracelet collection signal, the sphygmomanometer collection signal, and the motion behavior image captured by the mobile phone camera is collected in the preset future period, which is recorded as feedback correction information. The prediction result is corrected based on the adaptive Kalman filter according to the feedback correction information. A neural disease risk assessment report is output according to the prediction result.

[0223] In the embodiment, the bracelet in the wearable device is used to collect physiological signals such as heart rate and blood pressure, the glasses and the helmet are used to track eye movement trajectories, and the headset can be used to collect voice tremor features. The mobile phone and the tablet of the behavior collection device obtain gait data through built-in sensors, and the camera is used to record motion behavior images. The signal processing unit comprises a multi-channel signal synchronization module to ensure that the data from different devices are time-aligned.

[0224] The neural function state atlas in the disease analysis unit is constructed by a graph neural network, and the node represents each type of feature, and the edge weight is dynamically adjusted by a graph attention mechanism. The neural differential equation adopts a neural ordinary differential equation architecture to model the continuous dynamic process of disease progression by differentiating the hidden state. The adaptive Kalman filter is updated online through a noise covariance matrix to realize incremental correction of the prediction result.

[0225] The system obtains multi-modal data through wearable devices and behavior acquisition devices, extracts key features such as tremor intensity features, gait symmetry features and autonomic nervous rhythm features through a signal processing unit, and then a disease analysis unit constructs a dynamic neural function state atlas, combines historical data and real-time feedback, and uses a neural differential equation to predict disease progression trends. The system realizes comprehensive assessment and long-term prediction of neural function in a home environment, and provides an objective basis for early intervention.

[0226] By adopting the above technical solutions, the present application is different from the prior art and has the following beneficial effects: short-term comprehensive monitoring data is obtained through wearable devices and behavior acquisition devices, multi-band signal decomposition technology is used to separate tremor intensity features, gait symmetry features and autonomic nervous rhythm features, a deep learning model is used to extract eye movement trajectory and motion coordination features, and a multi-dimensional neural function state atlas is constructed. Based on the neural differential equation, the continuous evolution law of neural function is modeled, and long-term disease progression trends are predicted from short-term data. The system obtains local monitoring data through a real-time feedback mechanism, dynamically corrects the prediction results using an adaptive filtering algorithm, and automatically triggers a complete data reacquisition and processing process when data anomalies or prediction deviations are detected, ensuring the accuracy of the prediction. This combination of short-term comprehensive acquisition and long-term local monitoring effectively solves the problem of discontinuous data in long-term prediction of neurological diseases in a home environment, ensuring the comprehensiveness of initial assessment and maintaining the reliability of long-term prediction through dynamic correction, and providing an objective basis for early intervention of neurological diseases.

[0227] Finally, it should be noted that although the above embodiments have been described in the specification and drawings of the present application, they do not limit the patent protection scope of the present application. Any equivalent structure or equivalent process substitution or modification based on the essential concept of the present application, using the content described in the specification and drawings, and directly or indirectly implementing the technical solutions of the above embodiments in other related technical fields, are all included in the patent protection scope of the present application.

Claims

1. A method for detecting and analyzing neurological diseases, characterized in that, include: Acquire user wearable device monitoring data and behavioral test data. Wearable device monitoring data includes signals collected by wristbands, signals collected by blood pressure monitors, and motion behavior images. Behavioral test data includes standardized movement test results. The wristband-acquired signal was decomposed into multiple frequency bands, and wavelet transform was used to separate the tremor component, gait component and autonomic nerve fluctuation component. Tremor intensity features were extracted from the tremor component using a long short-term memory network; gait symmetry features were extracted from the gait component using a temporal convolutional network; and autonomic rhythm features were extracted from the autonomic undulation component using a Gaussian process regression algorithm. Deep learning models were used to extract eye-tracking and motion coordination features from motion behavior images, and motor function scores were calculated by combining them with standardized motion test results. Heart rate variability analysis was performed on the signals collected by the blood pressure monitor and wristband to calculate the diurnal rhythm fluctuation characteristics, and the abnormal patterns of neurological function were identified by density clustering algorithm. Based on tremor intensity characteristics, gait symmetry characteristics, autonomic rhythm characteristics, motor function scores, and abnormal neurological function patterns, a neurological function state atlas was constructed, and the contribution weight of each characteristic to disease risk was calculated through graph attention mechanism. The system acquires users' historical monitoring data, uses neural differential equations to learn the continuous time evolution of neural function states in the neural function state atlas based on the historical monitoring data, predicts the disease progression trend in a preset future period, and generates prediction results. Collect at least one of the following within a preset future time period: a signal collected by a wristband, a signal collected by a blood pressure monitor, and motion behavior images captured by a mobile phone camera. Record this as feedback correction information. Correct the prediction result based on adaptive Kalman filtering according to the feedback correction information. A risk assessment report for neurological diseases is generated based on the prediction results.

2. The method for detecting and analyzing neurological diseases according to claim 1, characterized in that, Based on tremor intensity characteristics, gait symmetry characteristics, autonomic rhythm characteristics, motor function scores, and patterns of neurological dysfunction, a neurological functional state atlas was constructed. The contribution weight of each characteristic to disease risk was calculated using a graph attention mechanism, including: Tremor intensity features, gait symmetry features, autonomic rhythm features, motor function scores, and abnormal neurological function patterns are used as node features of the graph neural network, and initial edge connections are constructed based on the physiological correlations between the node features. A graph attention mechanism is used to model the spatial dependence of node features, and to calculate the motor function association weight between tremor intensity features and gait symmetry features, the autonomic nervous regulation weight between autonomic rhythm features and abnormal neural function patterns, and the comprehensive coordination weight between motor function scores and other node features. The neighborhood node information is aggregated by a multi-layer graph convolutional network, the embedding representation of each node feature is iteratively optimized, and the graph embedding vector reflecting the overall state of neural function is output. Based on the graph embedding vector, the dynamic contribution weights of each node feature to disease risk are generated through the attention score calculation module. The weights of tremor intensity features and gait symmetry features are used for motor function assessment, the weights of autonomic rhythm features and abnormal neurological function patterns are used for autonomic function assessment, and the weights of motor function scores are used for comprehensive behavioral function assessment.

3. The method for detecting and analyzing neurological diseases according to claim 1, characterized in that, The system acquires users' historical monitoring data, uses neural differential equations to learn the continuous-time evolution of neural functional states within a neural functional state atlas based on this data, predicts disease progression trends within a preset future timeframe, and generates prediction results, including: Historical monitoring data is integrated into a neural functional state atlas to form multiple historical embedding vectors. These vectors are then arranged in a time series to construct a temporal evolution sequence of neural functional states. The time evolution sequence is calculated using neural differential equations. The gradual change process of neural functional state is described by a learnable continuous-time dynamic function, which is expressed by formula (1), as follows: ； In formula (1), For neural differential equations, It is a learnable continuous-time dynamic function. for The neural functional state vector at time t, for The weight parameters, For continuous time variables; A gating mechanism is introduced into the neural differential equation framework to dynamically adjust the rate of state change at different time scales in order to capture the short-term fluctuations and long-term degenerative characteristics of neural function. The neural differential equation model is obtained and expressed by formula (2), which is as follows: ； In formula (2), This is the gate vector corresponding to the gate control mechanism. For the residual term, This is element-wise multiplication; Based on the current neural functional state map and the trained neural differential equation model, multi-step state prediction is performed within a preset future time period, and the neural functional state prediction trajectory containing temporal continuity is output. The predicted trajectory of neurological function status is probabilistically processed to calculate the disease risk probability distribution at each time point and identify key time points where the risk exceeds a preset threshold. The predicted trajectory of neurological function status is compared and analyzed with the user's individual baseline data to generate personalized prediction results of disease progression trends. The prediction results include risk level change curves and early warning information for key time nodes.

4. The method for detecting and analyzing neurological diseases according to claim 1, characterized in that, At least one of the following is collected within a preset future time period: a signal from a wristband, a signal from a blood pressure monitor, and motion behavior images captured by a mobile phone camera. This is denoted as feedback correction information. The prediction result is then corrected based on the feedback correction information using an adaptive Kalman filter, including: The feedback correction information is input into the feature extraction module. Based on the extracted feedback features, at least one of the following is updated: tremor intensity feature, gait symmetry feature, autonomic rhythm feature, and motor function score, to generate the current neurological function state observation vector. A state-space model is constructed, the predicted trajectory output by the neural differential equation is used as the state transition prior, and the neural functional state observation vector is used as the measurement value to generate the state estimation result. An adaptive Kalman filter algorithm is used to generate confidence weights for the state estimation results by dynamically adjusting the process noise covariance matrix and the observation noise covariance matrix. In addition, when the deviation between the observed and predicted values ​​exceeds a dynamic threshold, the recalibration process of the prediction model is triggered; The state estimation result is corrected based on the confidence weight; The corrected state estimation results are fed back into the neural differential equation model, the parameters of the dynamic function are adjusted to adapt to the user's individualized neural function change patterns, and the corrected prediction results are output.

5. The method for detecting and analyzing neurological diseases according to claim 1, characterized in that, The method further includes: Determine whether the parameters in the feedback correction information exceed their corresponding normal fluctuation range; if so, trigger the re-acquisition mode. Alternatively, the deviation between the corrected prediction result and the initial prediction result can be calculated, and a re-acquisition mode can be triggered when the deviation exceeds a preset deviation threshold. The re-acquisition mode is configured as follows: Re-collect user wearable device monitoring data and behavioral test data; The wristband-acquired signal was re-decomposed into multiple frequency bands, and wavelet transform was used to separate the tremor component, gait component and autonomic nerve fluctuation component. The tremor component was re-extracted using a long short-term memory network to extract tremor intensity features, the gait component was re-extracted using a temporal convolutional network to extract gait symmetry features, and the autonomic nervous system fluctuation component was re-extracted using a Gaussian process regression algorithm to extract autonomic nervous system rhythm features. We re-extracted eye-tracking trajectory features and motion coordination features from the motion behavior images using a deep learning model, and calculated the motor function score by combining the results of standardized motion tests. Heart rate variability analysis was performed again on the blood pressure monitor and wristband signals to calculate the diurnal rhythm fluctuation characteristics, and the abnormal patterns of neurological function were identified by density clustering algorithm. Based on tremor intensity characteristics, gait symmetry characteristics, autonomic rhythm characteristics, motor function scores, and abnormal neurological function patterns, a neurological function state atlas was reconstructed, and the contribution weight of each characteristic to disease risk was calculated through graph attention mechanism. Based on historical test data, neural differential equations are re-applied to learn the continuous time evolution of neural functional states in the neural functional state atlas, predict the disease progression trend in the preset future time period, and generate updated prediction results.

6. The method for detecting and analyzing neurological diseases according to claim 1, characterized in that, The wristband collects signals including acceleration signals and gyroscope signals. These signals are decomposed into multiple frequency bands, and wavelet transform is used to separate tremor components, gait components, and autonomic nervous system fluctuation components, including: The acceleration signal and gyroscope signal are decomposed into multiple scales using wavelet basis functions with time-frequency localization characteristics. The original signal is separated into a first frequency band signal containing tremor characteristics, a second frequency band signal containing gait cycle characteristics, and a third frequency band signal reflecting autonomic nervous system regulation according to a preset frequency band. An adaptive decomposition algorithm based on signal energy distribution is applied to the first frequency band signal, the second frequency band signal, and the third frequency band signal. The wavelet decomposition level is dynamically adjusted according to the signal-to-noise ratio of each frequency band signal to obtain the decomposed first sub-frequency band signal, the second sub-frequency band signal, and the third sub-frequency band signal. Motion artifact suppression processing is then applied to each decomposed sub-frequency band signal. The first sub-band signal, after motion artifact suppression, is merged into a time-domain continuous flutter component using a phase consistency reconstruction algorithm. The second sub-band signal, after motion artifact suppression processing, is used to perform motion cycle phase calibration using a dynamic time warping algorithm to obtain gait components. The third sub-band signal, after motion artifact suppression, is sequentially subjected to ensemble empirical mode decomposition to remove baseline drift, maximum likelihood estimation to recover physiological rhythm components, and 0.01-0.1Hz bandpass filtering to extract sympathetic-parasympathetic neural modulation features, thus obtaining the autonomic nervous wave component.

7. The method for detecting and analyzing neurological diseases according to claim 1, characterized in that, Tremor intensity features were extracted from the tremor component using a long short-term memory network, including: Construct a bidirectional long short-term memory network model with an attention mechanism; Time series analysis of the tremor components yields the amplitude change rate and temporal fluctuation characteristics of the tremor components. Based on the amplitude change rate and temporal fluctuation characteristics, a tremor intensity feature characterizing the severity of the tremor is generated; Gait symmetry features are extracted from gait components using a temporal convolutional network, including: Gait components are input into a multi-scale dilated temporal convolutional network, and gait periodic features at different time scales are extracted through multiple dilated convolutional layers set in parallel. Calculate the cross-correlation function of the left and right limb movement signals in the time and frequency domains based on the gait periodic characteristics; The phase difference is obtained by using the Hilbert transform, and the symmetric feature vector representing the gait asymmetry is output. The autonomic nervous system wave components were extracted using a Gaussian process regression algorithm to extract autonomic rhythm features, including: The autonomic nervous system fluctuation component is used as input, and its nonlinear change trend is fitted by a Gaussian process regression model. Furthermore, the autocorrelation characteristics of the autonomic nervous system wave components are modeled using the squared exponential kernel function; The output includes the autonomic rhythm characteristics of the low-frequency oscillation power ratio, the high-frequency oscillation power ratio, and the ratio of the two.

8. The method for detecting and analyzing neurological diseases according to claim 1, characterized in that, Deep learning models were used to extract eye-tracking and motion coordination features from motion behavior images, and motor function scores were calculated by combining these features with standardized motion test results. The motion behavior images are encoded using a three-dimensional convolutional neural network to extract spatiotemporal sequence features of eye movements through a spatiotemporal convolutional layer, and an eye movement trajectory feature vector is output. A topological graph model of limb joint motion data is constructed using graph convolutional networks, and a motion coordination feature matrix is ​​generated based on the spatial connection relationship of joints. The eye-track feature vector and the motion coordination feature matrix are fused together, and the time-frequency features of electromyography signals collected in the standardized motion test are also incorporated. The cross-modal features are weighted through a multilayer perceptron to output a motor function score.

9. The method for detecting and analyzing neurological diseases according to claim 1, characterized in that, Heart rate variability analysis was performed on signals collected by blood pressure monitors and wristbands to calculate diurnal rhythm fluctuation characteristics. Density clustering algorithms were used to identify patterns of abnormal neurological function, including: Dynamic blood pressure trend extraction is performed on the signals collected by the blood pressure monitor to obtain the periodic change curve of systolic / diastolic blood pressure; The RR interval sequence analysis was performed on the signal collected by the wristband, and the day-night ratio of low-frequency power to high-frequency power was extracted based on frequency domain transformation; The periodic variation curve and the day-night ratio are input into a density clustering model based on DBSCAN. Through adaptive neighborhood radius parameter optimization, abnormal clusters representing sympathetic-parasympathetic imbalance are divided in the feature space of the density clustering model. The data points corresponding to the periodic variation curves and the day-night ratios that fall into the abnormal clusters are marked as abnormal neural function patterns and the abnormal probability is output.

10. A neurological disease detection and analysis system, characterized in that, The system applicable to the method of any one of claims 1 to 9 comprises: Wearable devices, including at least one of wristbands, glasses, helmets, and headphones; Behavior data collection devices, including at least one of mobile phones, tablets, and cameras; The signal processing unit is communicatively connected to the wearable device and the behavior acquisition device. The signal processing unit acquires user wearable device monitoring data and behavior test data; performs multi-band decomposition on the wristband-acquired signals, using wavelet transform to separate tremor components, gait components, and autonomic nervous system fluctuation components; extracts tremor intensity features from the tremor component using a long short-term memory network, extracts gait symmetry features from the gait component using a temporal convolutional network, and extracts autonomic nervous system rhythm features from the autonomic nervous system fluctuation components using a Gaussian process regression algorithm; extracts eye movement trajectory features and movement coordination features from motion behavior images using a deep learning model, and calculates a motor function score based on standardized movement test results; performs heart rate variability analysis on the blood pressure monitor and wristband-acquired signals, calculates diurnal rhythm fluctuation features, and identifies abnormal neural function patterns using a density clustering algorithm. The disease analysis unit is used to construct a neurological function status map based on tremor intensity characteristics, gait symmetry characteristics, autonomic rhythm characteristics, motor function scores, and neurological function abnormality patterns. It calculates the contribution weight of each feature to disease risk using a graph attention mechanism. It acquires the user's historical monitoring data and uses neural differential equations to learn the continuous-time evolution of neurological function status within the neurological function status map based on historical data, predicting disease progression trends within a preset future time period and generating prediction results. Within the preset future time period, it collects at least one of the following: signals from a wristband, a blood pressure monitor, and motion behavior images captured by a mobile phone camera, recording these as feedback correction information. Based on this feedback correction information, it corrects the prediction results using an adaptive Kalman filter. Finally, it outputs a neurological disease risk assessment report based on the prediction results.

Images