Multifunctional neurology intelligent diagnosis system based on 5G and Internet of Things

Abstract

The invention relates to the technical field of medical data processing, in particular to a multifunctional neurology intelligent diagnosis system based on 5G and Internet of Things. The system comprises a remote neurology data transmission module used for obtaining neurology data through a 5G vehicle-mounted and indoor base station; the edge data processing module is used for performing edge fusion on the neurology data to obtain neurology fusion data, and performing feature extraction on the neurology fusion data by using a preset edge feature extraction model to obtain neurology feature data; the intelligent diagnosis reasoning module is used for performing feature reasoning on the neurology feature data by using a preset deep feature analysis model to obtain feature reasoning data; and the terminal network regulation and display module is used for carrying out broadband distribution on the feature reasoning data so as to carry out data transmission operation and data visualization operation. According to the method, efficient collaboration between the edge and the cloud can be realized, and an interpretable auxiliary service is provided.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data processing, and particularly to a multi-functional intelligent diagnosis system for neurology based on 5G and the Internet of Things. Background Art

[0002] With the intensification of the aging population trend, the incidence of neurology-related diseases (such as epilepsy, Parkinson's disease, stroke, etc.) has been continuously rising, posing higher requirements for real-time, intelligent, and remote diagnosis and treatment of the medical system. At present, the clinical diagnosis of neurology mainly relies on the manual interpretation of multimodal physiological signals such as electroencephalogram (EEG), rheoencephalogram, and electrocardiogram by doctors, which has problems such as diagnosis relying on experience, low efficiency, and difficult standardization of results. In addition, traditional medical systems mostly use wired or low-speed wireless networks for data collection and transmission, making it difficult to meet the requirements of high-timeliness and high-bandwidth application scenarios such as emergency care. Summary of the Invention

[0003] In order to solve the above technical problems, the present invention proposes a multi-functional intelligent diagnosis system for neurology based on 5G and the Internet of Things to solve at least one of the above technical problems.

[0004] The present application provides a multi-functional intelligent diagnosis system for neurology based on 5G and the Internet of Things, including: A remote neurology data transmission module for collecting data from nerve sensors through 5G vehicle-mounted and indoor base stations to obtain neurology data; An edge data processing module for performing edge fusion on the neurology data to obtain fused neurology data, and using a preset edge feature extraction model to extract features from the fused neurology data to obtain neurology feature data, which is sent to the cloud platform; An intelligent diagnosis and reasoning module for controlling the cloud platform to obtain the neurology feature data, and using a preset deep feature analysis model to perform feature reasoning on the neurology feature data to obtain feature reasoning data; A terminal network regulation and display module for performing broadband allocation on the feature reasoning data to obtain broadband allocation data; performing data transmission operations on the feature reasoning data according to the broadband allocation data to send it to the edge terminal for data visualization operations.

[0005] In the present invention, by utilizing the high bandwidth and low latency characteristics of 5G vehicle-mounted and indoor base stations, the remote neurology data transmission module can collect multi-channel physiological signals in real time and synchronize them to the edge node, greatly improving the response speed to sudden neurological events in emergency and ICU scenarios. The edge data processing module adopts a multi-modal fusion and lightweight feature extraction model to complete time-frequency feature preprocessing and compressed transmission locally, which not only ensures the richness of the input data of the diagnostic model but also effectively reduces the network bandwidth occupancy and the computing pressure on the cloud. The intelligent diagnostic inference module deploys a deep feature parsing model on the cloud platform. Through the dual strategies of advanced semantic reasoning and knowledge graph path search, it not only improves the accuracy and interpretability of the diagnostic results but also can dynamically adjust the inference process according to real-time features, enhancing the system's scene processing ability. The terminal network regulation and display module dynamically allocates 5G bandwidth based on the urgency of the feature inference data and the priority of the network slice, and pushes the visualization results to the edge terminal, providing intuitive and timely decision-making support for medical staff.

[0006] Preferably, the data collection includes: Through the 5G vehicle-mounted base station and the indoor base station, based on the PTP protocol, clock synchronization is performed on the nerve sensor nodes, and electroencephalogram, hemodynamic and vital sign data are collected to obtain the original neurology data.

[0007] In the present invention, high-precision clock synchronization of the nerve sensor nodes is achieved through the combination of 5G vehicle-mounted and indoor base stations with the PTP protocol, which can ensure the timing alignment and low-latency transmission of multi-modal data such as electroencephalogram, hemodynamics, and vital signs. This can significantly improve the accuracy and reliability of fusion and analysis, effectively avoid data misalignment or information loss caused by clock drift, and meet the strict requirements for high-timeliness and high-precision data collection in emergency and real-time monitoring scenarios.

[0008] Preferably, the feature extraction includes: Perform window slicing on the neurology fusion data to obtain window slice data; Perform parallel feature operator calculations on the window slice data to obtain parallel feature data; Identify the parallel feature data according to a preset edge feature extraction model to obtain feature extraction method data; Extract features from the neurology fusion data according to the feature extraction method data to obtain neurology feature data.

[0009] In the present invention, based on data characteristics or multi-resolution window slicing, continuous multi-modal signals are divided into independent segments with temporal and frequency characteristics, providing a structured input for subsequent parallel feature operators. By parallelly executing multiple operators such as wavelet decomposition, STFT, and autocorrelation, not only the short-time transients and long-time trends of the signals are covered, but also time-frequency and self-similarity features from multiple angles can be extracted in parallel, greatly enriching the data representation. A preset lightweight edge feature extraction model intelligently identifies the optimal feature extraction method according to the parallel feature data, realizing dynamic weight allocation for the outputs of different windows and different operators. Customized feature extraction is performed on the original fused data in this way to obtain neurological feature data with high distinctiveness and strong diagnostic relevance. This process not only takes into account the constraints of real-time performance and computing resources, but also ensures the feature quality and diagnostic accuracy through multi-strategy parallelism and intelligent selection, providing a reliable basis for cloud deep reasoning and clinical decision-making.

[0010] Preferably, the window slicing includes: Slicing the neurological fused data according to preset window slicing parameter data to obtain primary window slicing data; Calculating the instantaneous change rate of the primary window slicing data to obtain instantaneous change rate data; Updating the window length of the preset window slicing parameter data according to the instantaneous change rate data to obtain parameter update data; Generating parallel windows according to the parameter update data to obtain parallel window data; Slicing the neurological fused data according to the parallel window data to obtain secondary window slicing data; Calculating the boundary mean difference of the secondary window slicing data to obtain boundary mean difference data; Performing boundary smoothing according to the boundary mean difference data to obtain window slicing data.

[0011] In the present invention, primary slicing based on preset parameters provides a basic comparison framework for the system; subsequently, the instantaneous change rate is calculated in real time and the window length is dynamically updated, enabling the window to be quickly shortened to capture short-time key events when the signal mutates, and extended to cover more context information during the stable phase; multiple sets of parallel windows are generated through the parameter update data, taking into account both short-time high accuracy and long-time global perspectives; subsequently, the parallel windows are sliced twice, enriching the diversity of temporal features; through boundary mean difference evaluation and smoothing processing, the mutation artifacts at the slice boundaries are eliminated, improving the stability and accuracy of feature extraction. This process not only improves the detection sensitivity of abnormal events, but also effectively reduces the noise interference and the risk of feature loss, providing a high-quality, continuous and multi-scale information-rich data basis for subsequent intelligent diagnosis.

[0012] Preferably, the parallel window generation includes: Update data according to parameters to perform multi-level window parameter mapping, and obtain multi-level window mapping data; Generate preliminary slice indexes based on the multi-level window mapping data to obtain preliminary slice index data; Measure alignment errors based on the preliminary slice index data to obtain alignment error data; Perform jitter compensation based on the alignment error data to obtain parallel window data.

[0013] In the present invention, multi-level window parameter mapping is performed by updating data according to parameters, which can simultaneously generate window configurations at multiple time resolutions, taking into account high-frequency anomaly detection and low-frequency trend modeling, and effectively improving the temporal diversity of data coverage; through the parallel generation of preliminary slice indexes and alignment error measurement, it is possible to identify the overlapping deviations in time positions between windows of different resolutions, with high accuracy; the further introduced jitter compensation mechanism can automatically fine-tune the boundary positions of each window, minimizing the feature distortion caused by information misalignment or truncation at the window junction. Compared with traditional fixed window slicing, this method can dynamically adapt to local changes in data and achieve cross-scale collaboration, effectively improving the temporal consistency and multi-scale robustness of feature extraction, and is applicable to sensitive perception scenarios for sudden changes and trend evolution in neural physiological signals.

[0014] Preferably, the steps for constructing the preset edge feature extraction model include: Obtain historical parallel feature data and corresponding historical feature extraction method data; Perform global average pooling on the historical parallel feature data to obtain parallel feature channel data; Perform lightweight multi-layer perceptron calculation on the parallel feature channel data to obtain channel attention data; Perform one-dimensional convolutional neural calculation based on the channel attention data to obtain one-dimensional convolutional neural data; Perform one-layer long short-term memory network calculation on the one-dimensional convolutional neural data to obtain one-layer long short-term memory network data; Perform full connection dimensionality reduction layer processing on the one-layer long short-term memory network data to obtain a primary edge feature extraction model; Perform offline iterative training on the primary edge feature extraction model according to the historical feature extraction method data to obtain a secondary edge feature extraction model; Perform lightweight pruning on the secondary edge feature extraction model to obtain an edge feature extraction model.

[0015] In the present invention, by utilizing historical parallel feature data and corresponding feature extraction method data, the model can perform global average pooling and channel attention calculation based on the actual operation situation, effectively identifying the importance differences of each channel in different scenarios; combining the tandem structure of a one-dimensional convolutional neural network and a long short-term memory network can not only quickly extract temporal and local patterns, but also capture long-range dependence features, enhancing the expression ability of the edge side for multi-modal signals; compressing high-dimensional temporal features into low-dimensional vectors through a fully connected dimensionality reduction layer takes into account both feature integrity and transmission efficiency; offline iterative training based on historical method data enables the model to continuously self-correct and optimize, adapting to different patients and device environments; the lightweight pruning step significantly reduces model parameters and computational complexity through dual sensitivity evaluation and structure rearrangement, while maintaining high accuracy. The overall process not only ensures the real-time and low-power online inference ability of the edge side, but also can maintain the accuracy and stability of the model in the long term through continuous learning and pruning maintenance, providing high-quality and sustainable evolving edge computing support for intelligent diagnosis in neurology.

[0016] Preferably, the lightweight pruning includes: Calculating the average activation intensity and gradient importance of the convolutional layer channels in the secondary edge feature extraction model to obtain the average activation intensity data and gradient importance data respectively; Performing channel sensitivity evaluation on the average activation intensity data and gradient importance data to obtain channel sensitivity data; Calculating the time sensitivity of the long short-term memory network layer channels in the secondary edge feature extraction model to obtain time sensitivity data; Performing similarity weight sorting on the secondary edge feature extraction model according to the channel sensitivity data and time sensitivity data to obtain channel sorting data; Calculating the intra-group average sensitivity according to the channel sorting data to obtain intra-group average sensitivity data; Performing cross-group pruning ratio allocation on the channel sorting data according to the intra-group average sensitivity data to obtain cross-group pruning ratio data; Pruning the secondary edge feature extraction model according to the cross-group pruning ratio data to obtain a pruned model; Rearranging the structure of the pruned model to obtain an edge feature extraction model.

[0017] In the present invention, the convolutional layer channels are evaluated by using dual metrics of average activation intensity and gradient importance, which can accurately identify the channels that contribute the most to the inference result and avoid the risk of misdeletion caused by traditional single pruning metrics. The time sensitivity analysis of the long short-term memory network layer is introduced to quantify the key points of different time steps in the time series model, ensuring the complete retention of important time information. Based on the channel and time sensitivity data, similarity weight sorting and intra-group average sensitivity calculation are performed to achieve dynamic allocation of the cross-group pruning ratio, taking into account the compression requirements of both high-sensitivity groups and low-sensitivity groups. By optimizing the memory layout of the pruned model through structure rearrangement and combining hardware-aware kernel fusion and quantization strategies, the final edge feature extraction model achieves lower inference latency and smaller power consumption on the target device. The innovation points of this process compared with traditional pruning methods lie in dual sensitivity evaluation, dynamic grouping and cross-group ratio allocation, as well as structure rearrangement and hardware co-optimization, significantly improving the practicality and reliability of the model in resource-constrained environments.

[0018] Preferably, the feature inference includes: Using a preset deep feature parsing model to identify the neurology feature data, and obtaining graph inference mode data; Performing path inference on a preset medical knowledge graph model according to the graph inference mode data, and obtaining candidate path data; Performing path score screening on the candidate path data to obtain feature inference data.

[0019] In the present invention, using a preset deep feature parsing model to identify the neurology feature data extracted from the edge can automatically extract clinical-related entity and relationship candidates from high-dimensional features, reducing the manual annotation cost and improving the coverage of preliminary inference. According to the generated graph inference mode data, path inference is performed in the structured medical knowledge graph, which can not only discover potential causal links hidden behind multi-modal data, but also integrate data-driven discovery with expert knowledge to enhance the credibility of the diagnosis result. Comprehensive scoring and screening of all candidate paths based on graph embedding similarity and model confidence ensure that the output feature inference data has both high confidence and interpretability. On the one hand, the inference intention is automatically generated by the deep model, reducing the complexity of knowledge graph queries; on the other hand, the optimal output of multiple candidate explanations is achieved through the path scoring mechanism, providing rich and reliable diagnostic basis for clinical decision-making.

[0020] Preferably, the construction steps of the deep feature parsing model include: Obtaining historical neurology feature data and historical graph inference mode data; Constructing feature points for the historical neurology fusion data to obtain feature point data; Generate tree nodes based on feature point data to obtain tree node data; Construct a tree model based on the tree node data to obtain a primary depth feature parsing model; Iteratively train the primary depth feature parsing model according to the historical graph reasoning mode data to obtain a secondary depth feature parsing model; Perform pruning on the secondary depth feature parsing model to obtain a depth feature parsing model.

[0021] In the present invention, by constructing feature points and generating tree nodes for historical neurology fusion data, a large amount of continuous time series information can be structured into key discriminant points and decision nodes, significantly improving the model's semantic understanding ability of complex physiological signals; constructing a primary tree model based on tree node data not only speeds up the model training speed but also improves the transparency of the reasoning process through a hierarchical decision path; using historical graph reasoning mode data to iteratively train the primary model realizes the organic integration of data-driven and knowledge-driven, enabling the model to be optimized in different disease scenarios; through lightweight pruning of the secondary model, redundant nodes and branches are removed, greatly reducing the model complexity and running delay while maintaining the reasoning accuracy. The technical feature of the overall process different from the traditional black box depth model lies in that on the one hand, the feature points and tree structure are used to strengthen the interpretability, and on the other hand, the adaptability of the model and the lightweight requirements of edge deployment are taken into account through iterative training and pruning optimization, providing high-performance and traceable depth parsing capabilities for intelligent neurology diagnosis.

[0022] Preferably, the feature point construction includes: Perform modal normalization on the historical neurology feature data to obtain modal normalization data; Calculate the saliency map according to the modal normalization data to obtain saliency map data; Perform multi-modal fusion saliency on the historical neurology fusion data according to the saliency map data to obtain saliency data; Perform autoencoder residual calculation on the saliency data to obtain residual training data; Perform residual threshold segmentation according to the residual training data to obtain residual threshold segmentation data; Calculate the mutual information according to the residual threshold segmentation data and the saliency data to obtain mutual information data; Select key point candidates according to the mutual information data to obtain key point candidate data; Perform density peak clustering on the key point candidate data to obtain feature point data.

[0023] In the present invention, modal normalization and saliency map calculation are performed on multi-channel fusion data, which can highlight the mutation regions in different physiological signals and ensure that the saliency data reflects real physiological events. Autoencoder residual calculation and residual threshold segmentation can identify potential abnormal points under unsupervised conditions, providing a second verification for key points. Then, mutual information calculation is used to fuse saliency and residual information, ensuring that the selected candidate points have high information content in both the signal mutation and anomaly dimensions. Density peak clustering is used to extract the most representative feature points from the candidate points, avoiding the redundancy and omission caused by traditional threshold hard segmentation. The technical feature of this process, which is different from single saliency or single residual methods, lies in multi-level and multi-dimensional fusion screening, which not only improves the accuracy of feature points but also enhances the robustness of the model to noise and individual differences, providing a high-quality and interpretable input basis for subsequent tree model construction and in-depth analysis.

[0024] The beneficial effects of the present invention are as follows: 5G vehicle-mounted and indoor base stations are coordinated to collect neural signals, and multi-channel high-precision clock synchronization is achieved based on the PTP protocol to ensure the time alignment and integrity of multi-modal data such as electroencephalogram and hemodynamics. The edge data processing module uses a fusion algorithm and a window slicing mechanism based on data characteristics to achieve data compression and key feature retention, and dynamically extracts multi-dimensional features through a lightweight edge feature extraction model, improving the processing speed and real-time performance. The cloud intelligent inference module introduces a map-driven deep feature analysis model, which not only improves the diagnostic accuracy but also enhances the interpretability of the inference process through path inference. The terminal network regulation module dynamically adjusts network resources using the feature inference results to achieve fast distribution of high-priority data, and at the same time provides clear and intuitive diagnostic support for doctors through a visualization interface. Compared with traditional solutions, this system has significant technical advantages in low-latency linkage, high-robust feature extraction, and interpretable intelligent inference, and is suitable for real-time intelligent diagnosis data assistance tasks in neurology first aid, remote monitoring, and edge scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Other features, objects, and advantages of the present application will become more apparent by reading the detailed description of the non-limiting embodiments with reference to the following drawings: Figure 1 Shows the structural flowchart of a multi-functional neurology intelligent diagnosis system based on 5G and the Internet of Things in one embodiment; Figure 2 Shows the processing flowchart of a neurology feature extraction sub-module in one embodiment; Figure 3 Shows the processing flowchart of a neurology feature inference sub-module in one embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0026] The technical method of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative work belong to the scope of protection of the present invention.

[0027] In addition, the accompanying drawings are only schematic diagrams of the present invention and are not necessarily drawn to scale. The functional entities can be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor methods and / or microcontroller methods.

[0028] It should be understood that although terms such as "first" and "second" may be used here to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, the first unit can be called the second unit, and similarly the second unit can be called the first unit. The term "and / or" used here includes any and all combinations of one or more of the listed associated items.

[0029] Please refer to Figures 1 to 3 , this application provides a multi-functional intelligent diagnosis system for neurology based on 5G and the Internet of Things, including: A remote neurology data transmission module 10, configured to collect data from a nerve sensor through a 5G vehicle-mounted and indoor base station to obtain neurology data; In one embodiment, the neural sensor is deployed through surface electrodes attached to the surface of the patient's scalp or the nerve-sensitive area of the limb, and is used to collect electroencephalogram (EEG) and electromyogram (EMG) signals in real time. Among them, the sampling frequency of the EEG channel is set to 512 Hz, and the sampling frequency of the EMG channel is set to 1000 Hz to meet the requirement of capturing neural signal responses with high time resolution. The sensor establishes a communication connection with the local data acquisition device through low-power Bluetooth (BLE) or narrowband Internet of Things protocol (NB-IoT). The local device may include a Raspberry Pi or other embedded edge computing units. In a mobile application scenario, the local device can be connected to a vehicle-mounted base station configured with a 5G network module, supporting two deployment modes: non-standalone networking (NSA) and standalone networking (SA); in a fixed scenario, such as in a neurology ward, an enhanced mobile broadband (eMBB) communication channel can be established through an indoor base station to achieve high-frequency neural data transmission under large bandwidth conditions. The collected neural data is preliminarily structured by the edge acquisition device and encapsulated into data packets in the JavaScript Object Notation (JSON) format, and uploaded to the edge server through the Message Queuing Telemetry Transport Protocol over TLS (MQTT over TLS) based on the TLS protocol to ensure the real-time and security of data transmission.

[0030] The edge data processing module 20 is used to perform edge fusion on the neurology data to obtain the neurology fusion data, and use a preset edge feature extraction model to extract features from the neurology fusion data to obtain the neurology feature data, which has been sent to the cloud platform; In one embodiment, the edge data processing module first synchronizes and aligns the data collected from multi-source neural sensors (such as electroencephalogram EEG, electromyogram EMG, and galvanic skin sensors) based on timestamp information, with the maximum allowable time deviation set to ±10 milliseconds to ensure inter-modal temporal consistency. Subsequently, a sliding window mechanism is introduced to achieve temporal fusion processing, where the sliding window length is set to 2 seconds and the step size is 0.5 seconds; within each window period, a frame of feature representation data is output after fusion. To remove abnormal pseudo-signals, the system performs outlier rejection, that is, for each physiological parameter, if its value exceeds the mean ±3 standard deviations, it is regarded as a pseudo-signal and ignored. In the data amplitude normalization stage, the Z-score normalization strategy is used to normalize the amplitudes of signals in different modalities, mapping the data in each modality to a standard normal distribution with zero mean and unit variance to ensure the consistency of feature input. The fused signal data is transformed into a time-frequency map input tensor through short-time Fourier transform (STFT), and then input into a preset edge feature extraction model, which adopts a lightweight convolutional neural network architecture (based on MobileNet) to automatically extract spatio-temporal texture features in the spectrogram. The tensor size of each time-frequency map input is 128×128, and the output after extraction is a feature vector with a dimension of 256. The feature vector is securely transmitted via the HTTPS protocol and serialized and packed in the Protocol Buffers (protobuf) format, and uploaded to the cloud platform by the edge node through a reverse proxy to achieve remote synchronous update of structured features.

[0031] The intelligent diagnosis and reasoning module 30 is used to control the cloud platform to obtain the neurological department feature data and perform feature reasoning on the neurological department feature data by using a preset deep feature analysis model to obtain feature reasoning data; In one embodiment, the service unit deployed in the cloud receives the neural feature vector data uploaded by the edge device through a RESTful API (RESTful API) based application programming interface and writes it into the feature buffer for model invocation. The deep feature parsing model is designed with a multi-channel Transformer structure, where each channel corresponds to different modality data inputs, such as electroencephalogram (EEG) channels, electromyogram (EMG) channels, etc., aiming to enhance the multi-modal fusion expression ability. The model input is a multi-channel feature vector sequence in the form of a time series, where each time window segment corresponds to a 256-dimensional feature vector, and the total length of the sequence is denoted as T. The main body of the model is composed of 4 layers of Transformer encoders stacked, and a set of fully connected inference output layers are connected at the end, which are used to generate the probability distribution results of multi-class labels. A position encoding mechanism is introduced inside the model to maintain the consistency of the time series structure; at the same time, a Dropout layer is set, and the dropout rate is set to 0.3, which is used to alleviate overfitting and improve the generalization ability. The model output is a multi-class probability distribution vector, indicating the confidence levels of different candidate classes. The system can select the top K (such as Top-3) class labels as candidate output results based on the confidence ranking. To enhance the interpretability of the model output, the system introduces a feature contribution evaluation method based on the SHAP mechanism. This mechanism realizes the interpretability modeling of the model inference results by calculating the marginal contribution values of each input feature (including time window segments and modality channels) to the model output, facilitating users to review or focus on the key feature regions.

[0032] The terminal network regulation and display module 40 is used to allocate broadband for the feature inference data to obtain broadband allocation data; and perform data transmission operations on the feature inference data according to the broadband allocation data to send it to the edge terminal for data visualization operations.

[0033] In one embodiment, the system introduces a Quality of Service (QoS) scheduling policy based on the Differentiated Services (DiffServ) mechanism to achieve priority differentiation management of different types of data streams. High-priority channel resources are allocated to the model inference result data with high real-time requirements, while low-priority bandwidth is allocated to low-frequency communication flows such as device control instructions. For different types of edge terminal devices, the system presets bandwidth allocation upper limit parameters. For example, the bandwidth upper limit of mobile terminals is set to not exceed 2 Mbps, and the bandwidth upper limit of fixed terminals is set to not exceed 5 Mbps. At the same time, combined with traffic shaping technology, such as setting maximum burst rate and average rate limits for each type of service flow; using the Token Bucket algorithm at the edge gateway to control the data inflow rate; if there is a sudden increase in bandwidth for a certain service category, the system first caches the excess part and releases the data pressure through methods such as delayed sending and intermittent release; throttle and smooth the bursty data streams to avoid congestion caused by individual data packets occupying the entire channel resources. The system uses a communication channel based on a full-duplex transmission protocol (such as WebSocket) to push content such as inference results, visualization heatmaps, and feature interpretability diagrams (such as feature weight maps generated based on the SHAP algorithm, such as obtaining the output prediction of the inference model for the target sample; enumerating feature subset combinations, estimating the marginal contribution of each feature to the output under different combinations; calculating the SHAP value of each feature as its contribution degree; constructing an interpretive data structure) to the front-end interface. The data structure includes structured feature result data (encoded in JSON format) and image visualization content (binary encoded in PNG format) to achieve multimodal push and decoding processing. The entire terminal data transmission link is optimized to ensure that the data transmission delay from the cloud to the terminal is controlled within 200 milliseconds to meet the quasi-real-time interaction requirements. The front-end display interface is built using a development framework based on an extensible component-based architecture, supporting multi-level interactive display and feature backtracking functions. The content presented on the interface includes an inference result list, time series feature charts, and an interactive heatmap area. Users can click on a specific time window area based on the control functions provided by the interface to trigger the backtracking read operation of the original neural data segment, thereby assisting in feature interpretation and visualization analysis.

[0034] Preferably, the data acquisition includes: Through 5G vehicle-mounted base stations and indoor base stations, clock synchronization of neural sensor nodes is performed based on the PTP protocol, and electroencephalogram, hemodynamic, and vital sign data are collected to obtain original neurology data.

[0035] In one embodiment, in a mobile environment, a 5G CPE device that supports the NSA / SA mode is used as the core mobile base station device, combined with a Global Positioning System (GPS) module for dynamic positioning broadcast, so as to achieve regional coverage of mobile terminals. Its access method includes the combination of Ethernet backhaul and the 5G NR (New Radio) air interface uplink channel. The system design ensures that the end-to-end communication latency (E2E Latency) is less than 10 milliseconds (ms). A 5G femtocell is deployed inside a fixed medical department (such as the neurology department) and connected to the hospital local area network (LAN). The femtocell supports the TDD (Time Division Duplex) frame structure configuration, for example, using the "DDDSU" structure to optimize the uplink data stream transmission efficiency. For medical data, a QoS (Quality of Service) policy is used for classification and marking, and the priority is set to 5QI = 5 or 7 to ensure transmission stability.

[0036] The system deploys the Precision Time Protocol (PTP) compliant with the IEEE 1588v2 standard as the core clock synchronization mechanism. A PTP master clock device is deployed in the local network to periodically broadcast the current reference time. Each neural sensor node acts as a PTP client and performs clock drift correction every 0.5 seconds. The synchronization process includes four-phase synchronization messages (Sync), Follow-up messages (Follow_Up), Delay Request (Delay_Req), and Delay Response (Delay_Resp). The system requires that the synchronization error does not exceed ±1 microsecond (μs). All sensor nodes enable the hardware timestamp function, preferably using a network card at the physical layer or media access control layer that supports the PTP protocol. The collected neurology data includes electroencephalogram (EEG), cerebral hemodynamics (such as blood oxygen saturation, deoxyhemoglobin concentration), and various vital sign signals, specifically including the following modalities, such as electroencephalogram data acquisition, using a surface electrode cap with 8 to 32 channels, and the electrode distribution follows the internationally common 10-20 electrode placement system; the sampling frequency is set to 512 Hz. The collected analog signals are digitized through an analog-to-digital converter (ADC) with a conversion accuracy of 16 bits (bit). The data is sent to the local edge gateway through a low-power communication method (BLE or NB-IoT) and then uplink transmitted through the 5G channel. Cerebral blood flow and blood oxygen signal acquisition, such as sampling using a near-infrared spectrometer (NIRS) or a transcranial Doppler (TCD) cerebral blood flow meter, with a frequency of 10 Hz. The output parameters include blood oxygen concentration ( and the concentration of deoxyhemoglobin (HbR). Signal calibration of ±5% is performed through the resting state baseline before collection. Collection of vital signs, such as electrocardiogram (ECG), photoplethysmogram (PPG) for blood pressure estimation, body temperature, and respiratory rate (through RIP respiratory inductive belt). All modal signals are uniformly collected through a microcontroller unit (such as nRF52840). The microcontroller supports the BLE Mesh communication protocol and can achieve multi-node self-organizing network communication. The collected data is uniformly encapsulated into data frames, and the high-precision timestamp generated by the PTP protocol is used as the first field of each frame. The timestamp format follows the ISO 8601 standard "Data storage and exchange format - Information exchange - Representation of dates and times", and the Coordinated Universal Time (UTC) representation method is adopted. Each data frame is equipped with an increasing frame number, set as frame_id = last_id + 1, where frame_id is the unique serial number of the current data frame, used to identify the relative position of the frame in the data stream, and last_id is the number value of the previous frame successfully transmitted and confirmed, to support packet loss retransmission and fault tolerance tracking. During the data transmission process, the edge gateway performs preliminary caching and verification operations. The data communication uses the MQTT protocol and performs lightweight encryption based on TLS. The topic structure of the MQTT message is set as / hospital / neuro / {device_type} / {device_id} / raw, where hospital / neuro indicates that the message originates from the neurology department scenario of the hospital, {device_type} represents the type of sensing device, {device_id} is the unique identifier of the device, and raw indicates that the current data is unprocessed original sampling data. The message service quality level (QoS) is set to 1 to ensure at least one successful transmission. Each data packet is attached with a checksum generated using the SHA-256 algorithm. The edge node performs integrity verification on the received data; if the verification fails, the corresponding sensor is notified to retransmit the data frame through the last will and testament message mechanism (LWT) of MQTT.

[0037] Preferably, the feature extraction includes: 21. Perform window slicing on the integrated neurology data to obtain window slice data; In one embodiment, the integrated neurology data is a multi-modal time series set, specifically including electroencephalogram (EEG) signals, blood flow signals (BF), heart rate signals (HR), and corresponding timestamp information. The electroencephalogram signals are represented in the form of multi-channel time series, denoted as , where is the number of channels, is the total number of time points. Apply the sliding window slicing strategy to process this type of time series data, and set the window length to 2 seconds (denoted as ), and the sliding step size to 0.5 seconds (denoted as ), to cover the entire acquisition period. For each type of modal signal, starting from 0 to iterate, and each time intercept a continuous data segment of length L as a slice, forming a sequence of window slices numbered . All modalities are sliced based on a unified time axis to ensure that the data contained under the same window number for each modality is temporally aligned. For easy identification and management, each window slice is attached with a window number and a start timestamp, and the subsequence data of each modal signal under this window is recorded separately, thus forming a window slice data unit with a standardized structure.

[0038] 22. Perform parallel feature operator calculations on the window slice data to obtain parallel feature data; In one embodiment, the edge system introduces a variety of statistical operators for parallel feature calculations on the neurophysiological signal data within each slice window to obtain rich numerical representations. The feature operators used cover time-domain statistics, frequency-domain power distribution, and information entropy-based metrics, including but not limited to the following typical forms, such as the mean, which reflects the central tendency of the overall amplitude of the signal; the standard deviation, which measures the range of signal fluctuations; skewness, which describes the degree of asymmetry of the distribution; kurtosis, which measures the sharpness of the signal; the main peak of the spectrum, which is the main frequency peak extracted by the fast Fourier transform; specific band power (BandPower(f1,f2)): extracting the power value of the signal in a specific frequency band, such as the α band (8–13 Hz) or the β band (13–30 Hz). The above feature extraction operations are processed with each window slice as the basic unit, and the system uses a parallel computing mechanism, such as thread pool scheduling or graphics processing unit (GPU) acceleration, to concurrently execute various statistical operators on all signal channels to improve the calculation efficiency. The feature output result is a multi-channel feature set indexed by the window number, where each type of feature is identified by the channel name + feature type. For example, the output features may include various combined information such as the mean of the EEG channel, the power of the β band, and the standard deviation of the heart rate signal. To avoid the problem of model overfitting caused by high-dimensional features, the system limits the maximum number of statistical features extracted for each type of original signal to no more than 20 dimensions, thereby controlling the overall dimension of the model input and improving the generalization ability and training stability of the model.

[0039] 23. Identify the parallel feature data according to a preset edge feature extraction model to obtain feature extraction method data; In one embodiment, a lightweight gradient boosting decision tree model (such as LightGBM) is selected to perform feature importance ranking; alternatively, a one-dimensional convolutional neural network (1D-CNN) combined with an attention mechanism is used to learn feature weights between different modalities or identify based on the feature extraction method labels corresponding to historical data, obtaining the corresponding feature extraction method data. The input of the model is the multi-modal parallel feature vector generated in the previous stage, including structured data of multiple modalities such as electroencephalogram frequency band energy features, heart rate variability features, and respiratory cycle features. The model outputs a set of weight values corresponding to each type of feature, denoted as the feature weight vector. Among them, each item represents the importance coefficient of a certain feature, and the value range is [0,1]. For example, the weight of the electroencephalogram alpha band power feature is 0.87, indicating its high importance in the current task, and then the corresponding feature extraction method is retrieved. The system presets an importance threshold, for example, set to 0.5, to distinguish high-importance and low-importance feature channels; among all features, those with corresponding weights greater than or equal to the threshold will be marked as extraction method features and used as the main reference channels for subsequent analysis.

[0040] 24. Perform feature extraction on the neurology fusion data according to the feature extraction method data to obtain neurology feature data: In one embodiment, the feature extraction is performed based on a retrospective window mechanism. The system retrospectively analyzes each item of the original signal window data according to the selected feature type and recalculates each type of feature value. The feature type may include time-domain features (such as heart rate variability indicators, amplitude mean), frequency-domain features (such as power of a specific frequency band), statistical features (such as kurtosis, skewness), and entropy features (such as signal entropy, sample entropy), etc. For example, for electroencephalogram features from the frequency domain, such as alpha band power, the system first performs a fast Fourier transform (FFT) on the electroencephalogram signal to transform the signal from the time domain to the frequency domain; then performs band-pass filtering on the target frequency band (such as 8 - 13 Hz); integrates the power spectral density within this frequency band to obtain the total power value of this frequency band as the extracted feature. After extraction, all retained features are integrated into a structured feature vector and bound to the corresponding window number (window_id) to form a standardized data structure. Each feature vector includes multiple clearly named feature items, such as electroencephalogram alpha wave power value, heart rate signal entropy value, etc.

[0041] Preferably, the window slicing includes: Slice the neurology fusion data according to the preset window slicing parameter data to obtain primary window slice data; In one embodiment, the system sets window slice parameter data, including a window length parameter (window_length_sec), which represents the time range covered by each slice, such as 2.0 seconds; and a sliding step parameter (stride_sec), which represents the time interval between adjacent slices, such as 0.5 seconds. According to the sampling rate of different modality signals, the time dimension is converted into the number of sampling points. The original signal sequence is sliced using a sliding window method. The start timestamp (represented in UTC time format) of each slice is recorded for subsequent time alignment and time series analysis operations. The output primary window slice data consists of multiple sliding window segments, and each segment corresponds to a data sequence of a specific duration and its start time.

[0042] Calculate the instantaneous change rate of the primary window slice data to obtain the instantaneous change rate data; In one embodiment, for each window, calculate the average first-order difference of the signals within the window: , is the instantaneous change rate data, is the primary window slice quantity data, is the index data of the primary window slice data, is the th sampling point signal value in the primary window, is the th sampling point signal value in the primary window.

[0043] Update the window length of the preset window slice parameter data according to the instantaneous change rate data to obtain parameter update data; In one embodiment, the length parameter of the current window is dynamically corrected based on the calculated instantaneous change rate IVR. The IVR is used to quantify the degree of signal change within the current window and can be implemented through methods such as sliding variance or gradient statistics, reflecting the fluctuation amplitude of the data within the time segment. The system completes window parameter update based on the following mapping rules. When the IVR is less than 0.01, it indicates that the data change is stable, and the system sets the current window length to 4 seconds to expand the time coverage range; when the IVR is between 0.01 and 0.05, it indicates that there is a moderate degree of fluctuation in the data, and the system adjusts the window length to 2 seconds to achieve a balanced response; when the IVR is not less than 0.05, it indicates that the data fluctuates violently, and the system shortens the window length to 1 second to enhance the instantaneous capture ability of sudden changes.

[0044] Generate parallel windows according to the parameter update data to obtain parallel window data; In one embodiment, the parallel window generation process includes: (1) The system maintains a global time pointer as the initial reference time for data window division; (2) Determine the length of the current window and the subsequent sliding step based on the IVR value of the data segment where the current time point is located; (3) Generate the current window data according to the updated parameters and advance the time pointer to the next sliding starting point; (4) Repeat the above process until the entire signal data is traversed. Although the window sequence generated in this way has inconsistent time spans and some windows overlap or jump, the overall still follows a unified time benchmark, thus forming a pseudo-asynchronous window sequence.

[0045] Slice the neurology fusion data according to the parallel window data to obtain secondary window slice data; In one embodiment, the system performs a multi-modal synchronous cropping operation on the neurology fusion data according to the parallel window time parameters generated above to form secondary slice data segments with unified start and end boundaries. The cropping is based on the start time point and end time point defined by each parallel window, and continuous signals are extracted in segments according to the dynamic window length. To ensure the temporal consistency between multi-modal data, the system requires that all signal modalities (including but not limited to electroencephalogram signal EEG, heart rate signal HR, etc.) be synchronously cropped based on a unified timestamp, that is, the start time and end time of all modalities are exactly the same, so as to ensure the horizontal alignment structure of the slice data. Each slice window contains the following three types of structured data content: window identification information, that is, the unique number of each slice window, which is convenient for data tracking and indexing; modal data structure, including the data arrays corresponding to each signal modality within the current window time period; window time length information, recording the duration corresponding to the current slice, with the unit of seconds (for example, 1.5 seconds).

[0046] Calculate the boundary mean difference of the secondary window slice data to obtain the boundary mean difference data; In one embodiment, for any two consecutive windows and , respectively extract the signal mean within the last fixed-length segment (such as 100 milliseconds) before the end of window , denoted as ; the signal mean within the first fixed-length segment (such as 100 milliseconds) after the start of window , denoted as . The absolute difference between the above two means is defined as the boundary mean difference index, denoted as . Extract the means of the last 100ms of each window and the first 100ms of the next window; calculate their absolute difference as the jump index.

[0047] Perform boundary smoothing on the boundary mean difference data to obtain the window slice data.

[0048] In one embodiment, when the mean difference of adjacent window slices in the overlapping region is greater than a set threshold, a weighted average method is used to fuse the overlapping data segments. Let the current slice number be , and its subsequent slice be . The smoothed output signal within the overlapping time interval t is denoted as , is the original signal value of the th window slice in the overlapping region, is the original signal value of the th window slice in the same interval, is the weighting coefficient, and its value range is from 0.5 to 0.7.

[0049] In one embodiment, a filter kernel is constructed for the splicing area, and the FIR filter is used to perform a convolution operation on the signals in this area, thereby achieving a smooth transition of the edges. The filter bandwidth parameter is set according to the modal characteristics to ensure that the edge fluctuations are smoothed without affecting the key frequency components.

[0050] Preferably, the parallel window generation includes: Performing multi-level window parameter mapping according to the parameter-updated data to obtain multi-level window mapping data; In one embodiment, the following three-level window mapping rules are set: when the instantaneous change rate IVR ∈ [0.00, 0.01], it indicates that the signal change is small, and the system allocates a larger window length (4.0 seconds) and a sliding step (1.0 second) for this time period; when the instantaneous change rate IVR ∈ (0.01, 0.05], it indicates that the signal has a medium degree of change, and the corresponding window length is 2.0 seconds and the sliding step is 0.5 second; when the instantaneous change rate IVR ∈ (0.05, 1.00], it indicates that the signal fluctuates violently, and a shorter window (1.0 second) and a finer step (0.25 second) are allocated. Through this mapping mechanism, the system can achieve local adaptive window segmentation in the time dimension according to the signal characteristics, taking into account both the operation efficiency and the signal expression accuracy.

[0051] Generating a preliminary slice index according to the multi-level window mapping data to obtain preliminary slice index data; In one embodiment, the index generation process starts from the starting position of the data (time point t = 0) and iterates. In each iteration, the system obtains the starting position and the ending position and is continuously iteratively generated on the premise of satisfying the condition of end ≤ total_length (the total length of the target signal data, usually in the unit of the number of sample points, equivalent to the upper limit of the number of time steps of the original signal sequence) until the entire signal sequence is covered. After each round of iteration, the time pointer advances once according to the preset stride. In addition, before performing the window division operation, the system needs to map all modalities to a unified timeline coordinate system, and preferably selects the modality with the highest sampling rate as the time reference to ensure that different signal modalities can be accurately aligned within the same time period, thereby achieving slice consistency. The generated preliminary slice index data includes the start position index, end position index, and corresponding window length of multiple window segments, and the structure information can be expressed as start index; end index; window length. For example, the system can generate the following window index sequence: the first window index is from start 0 to end 512; the second window starts at 256 and ends at 768.

[0052] Perform alignment error measurement based on the preliminary slice index data to obtain alignment error data; In one embodiment, before calculating the multi-modal alignment error, a certain modality is set as the reference modality, denoted as modality A, and the other modality is the target alignment modality, denoted as modality B. Define the window , ; Use the alignment error function: , is the alignment error between modality A and modality B, is the start point of the window of modality A, is the start point of the window of modality B, is the end point of the window of modality A, is the end point of the window of modality B, is the reference window length. This function gives equal weights to the start time offset and the end time offset in the calculation, and the result is mapped to the [0,1] interval, indicating the alignment degree of the window boundaries of the two modalities. The closer the value is to 0, the more accurate it is.

[0053] Perform jitter compensation based on the alignment error data to obtain parallel window data.

[0054] In one embodiment, adjust the slice boundary within the error tolerance range to achieve cross-modal soft alignment. Set the maximum error tolerance threshold: , if the error : Use linear interpolation or local window scaling compensation, , is the window length of modality B after compensation adjustment, is the window length of modality B before compensation, is to make the window of modality B approach the end point of A (average alignment), is the modality The window end position of is modal The window end position. If the error : directly discard the parallel window, or split it into independent modal processing. Alternatively, apply an interpolation method (such as cubic spline interpolation) to the original signal of modality B, and adjust its start and end indices while preserving the waveform continuity to form a new time segment.

[0055] Preferably, the steps for constructing the preset edge feature extraction model include: Obtain historical parallel feature data and corresponding historical feature extraction method data; In one embodiment, for the input data source, the historical parallel feature data includes , where each is the feature vector corresponding to the th time window, with a preset dimension (such as 128 dimensions), covering statistical quantities, spectral features, and information entropy features of multiple modal signals; the historical feature extraction method data for each window is such as , and each is a structured label recording the feature selection information corresponding to window . In the label structure, each feature is attached with a binary selection marker indicating whether it is determined to have high analysis value in this window (1 indicates retention, 0 indicates discard), and the feature labels determined to be of high value in the corresponding window.

[0056] Perform global average pooling on the historical parallel feature data to obtain parallel feature channel data; In one embodiment, the system performs global average pooling on the historical parallel feature data to integrate its time dimension information and only retains the statistical features at the channel level for analysis. The historical feature data has a three-dimensional structure corresponding to the sample batch, time series steps, and feature channel dimension respectively. For example, if the batch size is 32, the time length is 50 steps, and each step contains 128-dimensional features, its original data structure can be represented as a three-dimensional tensor with dimensions 32×50×128. In the global average pooling operation, the system takes the time dimension as the integration axis and performs an average operation on the time series values of each channel.

[0057] Perform lightweight multi-layer perceptron calculation on the parallel feature channel data to obtain channel attention data; In one embodiment, the structure includes two layers of fully connected neural networks. The first layer is a fully connected layer with an output dimension set to 64, and the activation function is the rectified linear unit (ReLU), which is used to enhance the non-linear expression ability of features. The second layer is a fully connected layer with an output dimension equal to the number of channels (e.g., 128), and the activation function is the Sigmoid function, which is used to map the output to the interval [0,1], representing the normalized importance weight of the channels. The input of the above structure is the parallel feature channel data obtained in the previous step, and the output is a channel attention weight vector consistent with the channel dimension, expressed as: , is the channel attention weight vector, is the attention weight of the -th channel, and the value range is [0,1]. The larger the value, the higher the importance of the channel in the current task. is the total number of channels, e.g., 128.

[0058] Perform one-dimensional convolutional neural computing based on the channel attention data to obtain one-dimensional convolutional neural data; In one embodiment, let the input feature sequence be representing the original feature vector of the -th channel; is the attention weight coefficient corresponding to the channel, and the value range is [0,1]. Through channel-level element multiplication for weighted fusion, an attention-enhanced feature vector is formed, and its calculation method is: ; The weighted feature vector sequence is used as the input and fed into a one-dimensional convolutional neural network (1D-CNN) for local feature extraction. The convolutional network is configured as follows: a one-dimensional convolutional layer (Conv1D); 32 filters (filters = 32); the convolutional kernel size is 3 (kernelsize = 3), indicating that each convolution operation covers 3 time steps; the stride is 1 (stride = 1); the boundary processing method uses the same padding strategy (padding='same'); the rectified linear unit (ReLU) is used as the activation calculation function.

[0059] Perform one layer of long short-term memory network computing on the one-dimensional convolutional neural data to obtain one layer of long short-term memory network data; In one embodiment, the used LSTM single-layer structure contains 64 memory units, and the network is configured to only output the hidden state of the last time step of the sequence, that is, the high-level temporal semantic representation at the end of the sequence. The output result is a tensor with a dimension of [B,64], where B represents the sample batch size and 64 represents the temporal global feature dimension corresponding to each sample.

[0060] Perform a fully-connected dimensionality reduction layer processing based on the data of a layer of long short-term memory network to obtain a primary edge feature extraction model; In one embodiment, the system further introduces a fully-connected dimensionality reduction structure based on the time series feature representation extracted by the previous-stage long short-term memory network to compress its output and generate a weight prediction vector that can be used for feature selection. The output of the long short-term memory network is a time series representation vector with a fixed length, for example, with a dimension of 64. Subsequently, the system performs a dimensionality reduction mapping operation on this time series vector through a layer of fully-connected network. The output dimension of this fully-connected layer is equal to the total number of feature channels defined in the system (denoted as feature_dim), and the Sigmoid activation function is used to normalize the output result to generate a continuous probability distribution with a value range between [0, 1]. The mapping output can be expressed as: , where is the feature channel selection probability vector corresponding to each input sample; is the probability that the th feature channel is selected; is the total dimension of the feature channels, that is, feature_dim.

[0061] Perform offline iterative training on the primary edge feature extraction model according to the historical feature extraction method data to obtain a secondary edge feature extraction model; In one embodiment, the training task is modeled using a multi-label binary classification model. Each feature channel is regarded as an independent binary classification target, that is, to judge whether the feature is important in the current task. The training loss function selects the multi-label cross-entropy loss function, which is defined as follows: , is the multi-label cross-entropy loss value, is the feature channel sequence term, is the feature channel quantity data, is the true label, 1 indicates that this channel is a key feature, and 0 indicates non-key, is the model prediction value, indicating the probability that the th feature channel is selected. The Adam optimizer is used for model training, and the learning rate ; the input of each batch of samples is [batch_size, time_steps, feature_dim], indicating that each batch contains batch_size samples, and each sample is a feature sequence of feature_dim dimensions with time_steps time steps. The corresponding label tensor is two-dimensional, and the number of training epochs is 30 - 100 epochs. The system monitors the loss of the validation set. If the loss does not decrease for several consecutive epochs, the training is stopped in advance to prevent overfitting.

[0062] Perform lightweight pruning according to the secondary edge feature extraction model to obtain an edge feature extraction model.

[0063] In one embodiment, the pruning process is based on the importance evaluation of model weights to sparsify and prune redundant structures. Specifically, the system first analyzes the parameter weights of the connections in each channel of the convolutional layer, identifies the channels that contribute less to the model performance, and performs a sparsification operation based on their absolute weight values. The L1 regularization method is used to impose a sparsity constraint on the training process, and combined with the pruning threshold determination rule, the channel connections with weight magnitudes lower than the preset threshold are removed. To further compress the model structure and control the computational complexity, the system sorts the channels according to the importance scores of the channel weights, and only retains the key channels with an accumulated contribution degree of up to 90%. The remaining channels and associated nodes will be removed to optimize the model volume. The pruned model needs to meet the edge operation constraint conditions, that is: the total number of parameters does not exceed 500,000 (total number of parameters ≤ 0.5M), and the inference latency does not exceed 10 milliseconds on a representative edge computing platform (such as Raspberry Pi 4 or a processor based on the ARM Cortex-A72 architecture). To ensure deployment compatibility, the optimized model can be exported in a lightweight deployment format, including TensorFlow Lite or ONNX format, and supports loading and running on mainstream edge hardware platforms such as Google Edge TPU and NVIDIA Jetson Nano.

[0064] Preferably, the lightweight pruning includes: Calculate the average activation intensity and gradient importance for the convolutional layer channels in the secondary edge feature extraction model, respectively, to obtain the average activation intensity data and gradient importance data; In one embodiment, the average response amplitude of each convolutional channel on the input sample set is statistically calculated. For the th convolutional channel, the absolute values of the feature map output values at each time step (or spatial position) of all input samples are summed, and its mean value is calculated to obtain the average activation intensity of this channel , , is the average activation intensity of the th convolutional channel, is the total number of input samples, is the number of time steps or spatial positions included in each sample, is the sample sequence term, is the time step (or position) sequence term, is the number of time steps (or spatial positions) of each sample, is the th sample at time step on the Feature map output of a channel.

[0065] For the th convolutional channel, its gradient importance is defined as the average absolute value of the partial derivative of the loss function with respect to the feature map output over the sample and time dimensions: ; where is the gradient importance score of the th channel, is the total number of input samples, is the number of time steps or spatial positions included in each sample, is the sample sequence term, is the time step (or position) sequence term, is the number of time steps (or spatial positions) of each sample, is the th sample at time step on the th channel's feature map output, is the loss function of the model, is the partial derivative symbol, is the partial derivative of the loss function with respect to the output of the th channel at the th sample, th time point (or spatial position).

[0066] Perform channel sensitivity evaluation on the average activation intensity data and gradient importance data to obtain channel sensitivity data; In one embodiment, the channel sensitivity calculation formula: , is the sensitivity score of the th channel, is the mean of the average activation intensity of the channel, is the th channel's average activation intensity, is the maximum activation intensity value among all channels, is the weight of the average gradient strength of the channel, is the th channel's average gradient intensity, is the maximum gradient intensity value among all channels, , , and the result is the sensitivity score , the lower the less important.

[0067] Perform time sensitivity calculation on the channels of the long short-term memory network layer in the secondary edge feature extraction model to obtain time sensitivity data; In one embodiment, the time sensitivity is used to measure the degree of change in the response of each hidden unit in the long short-term memory network (LSTM) to the model output over the entire time series range, reflecting the time selectivity and dynamic feature expression ability of the unit. The degree of difference in the influence of each hidden unit on the model output at different time steps is calculated, and the variance of the activation of the unit in the time dimension is calculated: , is the time variance sensitivity, is the variance calculation in the time dimension, is the th hidden unit activation value at time step The larger the value, the stronger the response fluctuation of the unit on the time axis. Alternatively, calculate the entropy of the activation of the unit in the time dimension , is the time activation entropy sensitivity, is the time step index, is the activation probability. , is the activation probability, is the th hidden unit activation value at time step

[0068] According to the channel sensitivity data and the time sensitivity data, the secondary edge feature extraction model is sorted by similarity weights to obtain the channel sorting data; In one embodiment, before merging and sorting all channels, the dimensions need to be unified. The convolutional sensitivity and the LSTM time sensitivity are dimensionless and then merged. The sorting basis is: , is the fusion sensitivity score of the th channel, is the weighting factor (the weight for balancing the spatial sensitivity and the time sensitivity, which is 0.5), is the dimensionless value of the convolutional channel sensitivity (dimensionless with the max or min-max standard), is the channel sensitivity data, is the dimensionless value of the STM time sensitivity (dimensionless with the max or min-max standard), is the time sensitivity data. The system is based on the above fusion sensitivity score Sort all channels in descending order to obtain a unified channel sorting result.

[0069] Calculate the average sensitivity within the group based on the channel sorting data to obtain the average sensitivity data within the group; In one embodiment, according to the structural characteristics of different modules in the model, all channels are divided into two subsets. The first group ( ): The convolutional channel set, including all feature channels from the one-dimensional convolutional network; the second group ( ): The LSTM channel set, including all channel outputs from the recurrent neural network (such as long short-term memory units). For each channel group, calculate the average sensitivity within the group: , is the average sensitivity score within the group, is the grouping in the number of channels, is all belonging to this grouping channel index, is the th channel fusion sensitivity score (obtained from the previous formula).

[0070] Allocate the cross-group pruning ratio to the channel sorting data according to the average sensitivity data within the group to obtain the cross-group pruning ratio data; In one embodiment, the system sets the overall model pruning target ratio TotalRatio (for example, 0.2 means that 20% of the model parameters are expected to be pruned), and performs cross-group allocation of the pruning ratio based on the average time sensitivity value of each channel group. The average sensitivity is used to measure the th channel group's importance in the time series modeling process. The larger the value, the more significant the change in its channel activation value between different time steps, and the stronger the time series expression ability. To preferentially retain the structure with strong time series expression ability during pruning, the system uses the following pruning ratio allocation function: , is the pruning ratio allocated to group , is the average sensitivity of group (the larger the value, the more important), is the inverse weighted sum of the importance of all groups, is the channel group index variable, is the th group average sensitivity, is the total pruning ratio (such as 0.2 means 20% pruning).

[0071] Prune the secondary edge feature extraction model according to the cross-group pruning ratio data to obtain the pruned model; In one embodiment, the system performs a structured pruning operation on each channel in the secondary edge feature extraction model based on the obtained channel sorting result and the cross-group pruning ratio data described above. Adhering to the strategy of ascending sensitivity pruning within a group, that is, preferentially removing channels with lower sensitivity scores from each channel group. For each channel group, the system determines the number of channels to be pruned in the following way: multiply the original number of channels in the group by the corresponding preset pruning ratio, and take the floor of the resulting value as the number of channels to be pruned. For example, if a convolutional layer contains 64 channels and the corresponding pruning ratio is 0.13, then this layer will retain 87% of the channels, that is, at most 8 channels can be pruned; For the convolutional layers in the convolutional network, the system adopts a structured channel pruning method, that is, performs unified channel-level sparsification processing on the entire convolutional kernel weight tensor according to the channel dimension, so as to achieve explicit channel pruning of the convolutional layer weights and computational graph optimization. This pruning method ensures the deployability and running efficiency of the model in the inference stage; For the sub-module containing the long short-term memory network (LSTM), the system realizes sparse structure optimization by pruning the hidden state dimension channels. Specifically, the system removes several hidden state channels according to the sorting result, and correspondingly performs dimension reconstruction and parameter remapping on the state transition matrices in the LSTM (including the hidden state transition matrix and the input state matrix) to ensure that the model structure after pruning is still operable and complete; after the above pruning operation is completed, the system packages the model structure and the corresponding weight parameters after structured sparse optimization and stores them in the form of an intermediate model state, providing a basis for subsequent fine-tuning training, online deployment or model reloading.

[0072] Perform a structural rearrangement on the pruned model to obtain an edge feature extraction model.

[0073] In one embodiment, for the pruned convolutional channels, the corresponding indices in the subsequent batch normalization layer (BN) also need to be removed synchronously to avoid invalid parameters from participating in the calculation. At the same time, the convolutional kernel parameters and bias terms associated with this channel are removed. The forward propagation function (forward logic) of the model is adjusted to ensure that the channel dimensions of the outputs of each layer after pruning are continuous and consistent, avoiding dimension misalignment or undefined tensor connection problems. This step can be completed by reconstructing the computational graph structure or rewriting the tensor operation chain. If the model contains long short-term memory networks (LSTM) or other recurrent neural structures and is implemented using a framework such as PyTorch, the internal weight tensors need to be re-initialized. The shape of the new weight tensors should match the number of remaining channels to maintain the consistency of the LSTM input and hidden states. All unused parameter nodes and intermediate layer outputs are removed to generate a pure model file. This model can be saved in a format compatible with edge deployment, such as .tflite (TensorFlowLite); .onnx (Open Neural Network Exchange). During the deployment process, the model can be combined with the TensorRT engine for inference acceleration. If the ONNX model format is used, quantization processing can be further performed on the model after structure rearrangement, converting the parameter precision from floating-point type (FP32) to integer type (INT8) to adapt to embedded devices such as Jetson Nano and Jetson Xavier, improving the running speed and resource utilization while ensuring the accuracy of the model.

[0074] Preferably, the feature inference includes: 31. Using a preset depth feature parsing model to identify the neurology feature data to obtain graph inference mode data; In one embodiment, the neurological feature data is a feature vector FFF containing a multi-modal structure, where the feature items include but are not limited to electroencephalogram signal power features, heart rate variability features, hemodynamic fluctuation features, etc. The specific form is a set of key-value pairs composed of a feature name and a corresponding value. This feature vector is input into the constructed deep feature parsing model. The model adopts a lightweight Transformer Encoder structure, or consists of a feed-forward neural network and a multi-head attention mechanism, and has the following processing flow: the input layer receives the feature vector; the first non-linear transformation layer raises the feature dimension to 128 dimensions; the multi-head attention module models the relationship between features (the default setting is 4 attention heads); the second transformation layer compresses the result into a 64-dimensional semantic space; the Softmax layer outputs a normalized semantic weight distribution. The model output is a graph semantic embedding vector, which is used to express the potential inference intention in the current neurological state. The output graph inference method data includes the query type, such as path inference; the target entity, such as epilepsy; the semantic association weight, indicating the relative association degree of several intermediate entities in the semantic graph, such as neuroinflammation (0.71), cerebral ischemia (0.22), hypoxia (0.07), etc.

[0075] 32. Perform path inference on the preset medical knowledge graph model according to the graph inference method data to obtain candidate path data; In one embodiment, the knowledge graph model structure is a triple graph , nodes represent medical entities (such as "epilepsy"), relationships represent causality, accompaniment, etc., is: a set of triples, and each element is in the form of (h, r, t), representing a directed semantic edge of head entity - relationship - tail entity; the system applies a graph embedding pre-training model to the above graph structure to map entities into a continuous vector space. The graph embedding pre-training model can be selected from: TransE (a low-dimensional embedding method based on vector translation) / RotatE (an embedding method based on complex space rotation relationship modeling) / GAT (a graph neural network based on the attention mechanism) / R-GCN (a graph convolutional network for multi-relational graphs), which is used to embed the entity vector space and support path search and nested inference. In the path inference stage, the system uses the target medical entity in the current task as the starting point for graph search, and combines the preset semantic guidance strategy to carry out multi-hop path mining operations, with the medical entity node corresponding to the inference target as the starting point of the path; the maximum number of hops of the path is set to 3 hops; at each hop, select the next entity that is consistent with the predefined correlation weight direction to enhance the semantic relevance of the path; use a heuristic path search method (such as Search or Beam Search is used to control the search efficiency; the maximum number of candidate paths generated is set to 10 to reduce the computational overhead and limit the output scale. Each candidate path is represented as a sequence of adjacent medical entity nodes and their directed relationship edges, such as "Entity A → Entity B → Entity C → Target Entity", indicating the step-by-step semantic association process of the path in the knowledge graph.

[0076] 33. Screen the candidate path data through path scoring to obtain the feature inference data.

[0077] In one embodiment, each candidate path has the following defined scoring function: , is the score of path , is the weight of the correlation index between the path and the input feature or inference intention, with a value of 0.5, is the correlation between the path and the input feature or inference intention, is the weight of the knowledge reliability index of the nodes or edges in the path, with a value of 0.3, is the reliability of the knowledge edges or nodes in the path, is the weight of the compactness index of the path structure, with a value of 0.2, is the path compactness (the fewer the number of hops, the higher the score). All candidate paths are sorted in descending order according to the score values, and the top K paths are selected as the inference output. The value of K is a system-configurable parameter and can be set to 3 (i.e., the Top-3 paths), for example.

[0078] Preferably, the construction steps of the deep feature parsing model include: Obtain historical neurology feature data and historical knowledge graph inference method data; In one embodiment, the historical data is composed in units of time windows, and each window serves as a training sample, including (1) neurology feature data, represented in the form of a structured vector, which integrates signal information of multiple modalities, including electroencephalogram (EEG), heart rate signal (HR), blood flow signal (BF), etc. After fusion, a unified feature vector with a dimension of 128 is formed for model input; (2) knowledge graph inference method data, as the supervision label, which describes the expected inference target and path guidance information of the current feature sample in the medical knowledge graph. This structure includes the target entity, referring to the semantic central entity node corresponding to the feature, such as "epilepsy"; the key concept set, which are the intermediate concepts or semantic bridging nodes that play an important role in the inference process, such as "oxidative stress" and "neuron inflammation".

[0079] Construct feature points from the historical neurology fusion data to obtain feature point data; In one embodiment, the system calculates a set of statistical feature indicators for each modal signal channel (such as electroencephalogram EEG, heart rate HR, skin conductance, etc.) in each input sample to represent the representative state of the modality within a specific time period. Taking the EEG modality as an example, the system can calculate the following typical statistical indicators based on 16 channels, such as the mean value, which represents the average amplitude of each channel; the maximum value, which represents the maximum value of the signal; the kurtosis, which represents the sharpness of the signal distribution; and the band-pass frequency band energy, which represents the power values of frequency bands such as the α band (8–13 Hz) and the β band (13–30 Hz). The system unifies and splices the above statistical features to form a feature point vector representation of each sample. Each sample retains its unique identification number during the feature point construction process and includes a structured feature set with semantic labels as keys and numerical values as values in the output, such as EEG mean value, β band energy, heart rate entropy value, etc.

[0080] Generate tree nodes based on the feature point data to obtain tree node data; In one embodiment, let the original sample data set be D, which contains multiple feature points. For each candidate feature , the information gain calculation formula is defined as follows: , is the information gain brought by feature , is the entropy of the original data set , is the th subset after splitting by feature , is the entropy of subset , is the proportion of the subset in the original data set. A binary splitting method is used to construct the splitting node. For continuous numerical features, the best splitting point is selected within its value range, and the samples are divided into two parts: less than or equal to the threshold and greater than the threshold; Construct a tree model based on the tree node data to obtain a primary deep feature parsing model; In one embodiment, the model is an integrated tree structure, trained and constructed based on structured input feature data. A Gradient Boosted Tree model such as XGBoost or LightGBM can be used, or a gcForest (deep forest model) can be adopted. This model is integrated by multiple weak learners (i.e., decision trees), and the overall Boosting mechanism is used for optimization to iteratively improve the prediction performance. During the construction process, the structural complexity of each tree is controlled. The maximum depth is generally set to 6 to 12 layers, and the number of leaf nodes does not exceed 64, ensuring that the model maintains reasonable computational overhead while having the ability of non-linear expression. In addition to directly outputting classification / prediction results, the model can also be used as a feature expression tool to generate structured semantic embedding vectors for input samples. The expression is: , is the embedding vector representation of the sample obtained through the th tree, is the index of the leaf node where the sample falls in this tree, and this path / leaf node is mapped to a certain vector representation (for downstream learning tasks). The model saving format is , exporting the tree structure in JSON or text format for visualization and pruning optimization.

[0081] Iteratively train the primary deep feature parsing model according to the historical graph inference method data to obtain the secondary deep feature parsing model; In one embodiment, the system calculates the logarithmic loss value between the corresponding predicted probability and the actual label for each training sample and each class label . The overall loss function expression is: , where is the loss function value, is the sample index (a total of N samples), is the class label index (a total of C classes), is the true label value of the sample in the class (0 or 1), is the predicted probability of the sample for the class The prediction probability. In terms of the optimization mechanism, the system uses the Gradient Boosting Decision Tree method (eXtreme Gradient Boosting, abbreviated as XGBoost) to iteratively train the deep feature parsing model. Its training process is as follows. In each round of training, a new classification tree (weak learner) is constructed based on the residual gradient of the prediction result of the previous round to correct the current error; the newly generated tree model in each round is weighted and combined with the historical model to gradually approach the optimal prediction boundary; if the loss function value on the validation set does not decrease significantly for several consecutive rounds (reaching the convergence condition); or the number of training rounds reaches the upper limit, which is set between 50 and 100 rounds.

[0082] Prune the secondary deep feature parsing model to obtain the deep feature parsing model.

[0083] In one embodiment, the deep feature parsing model is a decision model based on the gradient boosting structure (such as XGBoost or LightGBM). Its pruning process includes two stages: pre-pruning and post-pruning, which are used to limit the model structure complexity before training and remove redundant calculation nodes after training, respectively. Pre-pruning sets control parameters in advance during the model training stage to suppress the risk of model overfitting and the growth of complexity. The settings are as follows: the maximum tree depth is set to 8 layers; the minimum sample weight sum of leaf nodes is set to 5, which is used to limit the total minimum number of samples contained in each leaf node; the minimum split gain threshold is set to 0.1, and only when the gain brought by node splitting is greater than this value is it allowed to continue splitting. The above parameters control the complexity of the model structure and reduce unnecessary deep branches and overfitting paths at the source. After the model training is completed, importance analysis is performed on each non-leaf node to further remove redundant structures. The post-pruning process is as follows: statistically analyze the contribution degree of each split node to the model performance, and the evaluation method can be based on the following two types of indicators, such as split frequency, the number of times the node is used as a split in all training rounds; information gain, the reduction in the loss function brought by node splitting (such as the degree of improving accuracy or reducing error); if the average gain of a certain node is lower than the set threshold (for example, 0.005), then merge all its subordinate subtree nodes into a leaf node, thereby compressing the model structure. The model can automatically identify redundant or non-contributing features (such as features with a split frequency of zero) during the construction process and automatically remove them during the training process to avoid them entering the model path, thereby further reducing model redundancy and storage overhead.

[0084] Preferably, the construction of the feature points includes: Perform modal normalization on the historical neurology feature data to obtain the modal normalized data; In one embodiment, the historical neurology feature data is in the form of structured multi-modal vectors, containing multiple feature items from different physiological modalities. The system performs Z-score normalization on the feature subsets of each modality separately. The system performs automatic classification based on the prefix of the feature name. Features starting with "EEG_" belong to the electroencephalogram modality; features starting with "HR_" belong to the heart rate modality; features starting with "BF_" belong to the blood flow modality.

[0085] Calculate the saliency map based on the modality-normalized data to obtain saliency map data; In one embodiment, for the th feature dimension, its saliency score is defined as the proportion of the variance of this dimension in the total variance of all dimensions , is the variance of the th feature dimension, is the total variance of all feature dimensions. The larger this index, the more obvious the change of this feature dimension among different samples, and the higher its distinguishability in the feature space. The system can also use the KL divergence in information theory to quantify the saliency of each feature dimension. For each feature dimension , the system calculates its distribution among all samples and the standard uniform distribution The Kullback-Leibler divergence (denoted as ) between them, is the eigenvalue distribution, is the uniform distribution (significant deviation indicates importance). The larger this value, the more structured and uneven the distribution of this feature dimension in the data, and the higher the saliency.

[0086] Perform multi-modal fusion saliency on the historical neurology fusion data according to the saliency map data to obtain saliency data; In one embodiment, perform weighted summation of the features within each modality: , is the fusion eigenvalue of modality (representing the aggregation under the entire modality), is all the feature indices in modality , is the value of the th feature, is the weight of feature (obtained through saliency normalization), , is the saliency score of feature (calculated by variance / KL divergence as before), is the total saliency of all features in this modality.

[0087] Calculate the auto - encoder residuals for the significance data to obtain residual training data; In one embodiment, the significance data is a structured set of feature samples with a dimension of 128. The system constructs an auto - encoder (AE) structure to learn the principal component representation of the samples. The AE structure includes an encoder part that reduces the 128 - dimensional input to 32 dimensions through a fully - connected transformation; an activation function layer that uses the ReLU activation function to enhance the non - linear expression ability; a decoder part that raises the hidden representation back to the original input dimension to form a reconstructed output; and the model training uses the mean squared error (MSE) as the loss function to calculate the mean squared difference between the original sample and its reconstructed value. For any input sample x, its residual is defined as follows , where is the sample after being reconstructed by the AE model, is the vector squared norm. By calculating the residual values of each sample, the system can identify the input regions that the model fails to effectively restore, that is, the feature segments with high heterogeneity or potential high analytical value. The relevant information can be recorded as structured sample scoring data, such as a record item containing the sample number and the corresponding residual value, the sample number is , and the reconstruction residual is .

[0088] Perform residual threshold segmentation on the residual training data to obtain residual threshold segmentation data; In one embodiment, the system sorts the residual values corresponding to all samples (for example, the residuals between the model output and the true target) and calculates its key statistical quantiles, including the first quartile (Q1, 25% quantile) and the third quartile (Q3, 75% quantile). Accordingly, the residual values of all samples are divided into the following three interval levels. For example, the low - residual level indicates that the residual value is lower than the first quartile (<Q1), representing the area where the model fits well; the medium - residual level indicates that the residual value is between the first quartile and the third quartile (Q1 - Q3), representing the area where the error is at a medium level; the high - residual level indicates that the residual value is higher than the third quartile (>Q3), representing the area where the model fitting effect is poor. The system adds the residual level of each sample as annotation information to the corresponding data structure to form unified residual segmentation data. For example, for the data with the sample number W123, if its residual value is higher than Q3, it is correspondingly labeled as the high - residual level.

[0089] Calculate the mutual information based on the residual threshold segmentation data and the significance data to obtain mutual information data; In one embodiment, for each feature , calculate its mutual information value with the residual level ​ , which is used to characterize the recognition ability of this feature for different residual regions. , is the mutual information between the feature and the residual level . is the predefined residual level category, is the residual level label, and its value is low / mid / high. is all the values of the feature . is the th feature, is the joint probability that the feature value is and the residual is . is the marginal probability that the feature takes the value of . is the marginal probability that the residual level is .

[0090] Candidate key points are selected according to the mutual information data to obtain candidate key point data; In one embodiment, the system first calculates the mutual information value (MI) between all candidate feature items and the target semantic label to measure the explanatory ability of the feature for the target variable at the information level. Subsequently, all feature items are sorted in descending order according to the mutual information value, and the top K feature items with the largest mutual information are selected as the candidate key point set, where K is the set candidate quantity parameter, for example, K = 50. To avoid noise interference, a mutual information threshold (for example, MI≥0.3) can be set to filter the sorting result, and only the feature items with mutual information values exceeding this threshold are retained.

[0091] Density peak clustering processing is performed on the candidate key point data to obtain feature point data.

[0092] In one embodiment, the system extracts the vector representation of all candidate key points in the multi-modal fusion feature space. Each candidate point is set as a high-dimensional feature vector. The system uses any of the following methods to evaluate the similarity between any two feature points. When the feature representation is mainly continuous real values (such as statistics, spectral energy), the Euclidean distance is used as the similarity measurement criterion. That is, the distance between two feature vectors is obtained by taking the square root of the sum of the squares of the differences in each dimension. When the feature representation is mainly unitized or sparse direction vectors (such as high-dimensional normalized features), the cosine distance is used to calculate the cosine value difference of the angle between the two vectors. Thus, a similarity matrix between feature points is constructed. The system then calculates the local density value of each candidate key point. The specific operation is as follows. Taking this point as the center, within the given neighborhood search radius ( Count the number of neighboring points within a certain range ( ); if this number exceeds the set minimum sample point threshold (MinPts = 5), then this point is regarded as a core point, that is, its local density is relatively high and it has the potential to form a cluster. At the same time, the system records the minimum distance from each point to a point with higher density. Among all the key points, the system preferentially looks for candidate points with the following characteristics as the initial cluster centers, such as the local density value being significantly high (ranked among the top in all candidates); the distance to other points with higher density being relatively far (indicating that it is in the center position of the cluster); the system identifies the points with a high comprehensive density-distance score of this type as the density peak points of the cluster, and takes these as the cluster centers, and gradually incorporates its neighboring points through the neighborhood expansion strategy to form a cluster. After clustering, the system extracts representative feature points for each cluster. The system can also randomly or based on the structural uniformity strategy select several supplementary feature points from the boundary region of the cluster to enhance the feature coverage and boundary expression ability. By default, a fixed number of points (such as 10) can be additionally extracted from each cluster. These points have a relatively high spatial distribution diversity relative to the center point and can effectively reflect the internal distribution profile of the cluster. The central density point in each cluster is identified as the main representative of this type of structure and has the following characteristics: it is in the high-density region within the cluster and is the core of the formation of the cluster; it can effectively express the common structural patterns of other feature points within the cluster; it has a strong discrimination ability with the center points of other clusters in the feature space to avoid information redundancy and clustering overlap. The system forms the final feature point data from the central density points and the selected boundary representative point sets identified in all clusters.

[0093] Therefore, from any perspective, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended application documents rather than the above description. Therefore, it is intended to cover all changes falling within the meaning and scope of the equivalent elements of the application documents within the present invention.

[0094] The above description is only a specific implementation manner of the present invention, enabling those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A multi-functional intelligent diagnosis system for neurology based on 5G and the Internet of Things, characterized in that, Including: A remote neurology data transmission module for collecting data from nerve sensors through 5G vehicle-mounted and indoor base stations to obtain neurology data; An edge data processing module for performing edge fusion on the neurology data to obtain fused neurology data, and using a preset edge feature extraction model to extract features from the fused neurology data to obtain neurology feature data, which is sent to the cloud platform; An intelligent diagnosis and reasoning module for controlling the cloud platform to obtain the neurology feature data, and using a preset deep feature analysis model to perform feature reasoning on the neurology feature data to obtain feature reasoning data; A terminal network regulation and display module for performing broadband allocation on the feature reasoning data to obtain broadband allocation data; performing data transmission operations on the feature reasoning data according to the broadband allocation data to send it to the edge terminal for data visualization operations.

2. The system according to claim 1, characterized in that, The data collection includes: Through 5G vehicle-mounted base stations and indoor base stations, clock synchronization is performed on nerve sensor nodes based on the PTP protocol, and electroencephalogram, hemodynamic, and vital sign data are collected to obtain raw neurology data.

3. The system according to claim 1, wherein The feature extraction includes: Performing window slicing on the fused neurology data to obtain window slice data; Performing parallel feature operator calculations on the window slice data to obtain parallel feature data; Identifying the parallel feature data according to a preset edge feature extraction model to obtain feature extraction method data; Extracting features from the fused neurology data according to the feature extraction method data to obtain neurology feature data.

4. The system according to claim 3, characterized in that, The window slicing includes: Slicing the fused neurology data according to preset window slicing parameter data to obtain primary window slice data; Calculating the instantaneous change rate of the primary window slice data to obtain instantaneous change rate data; Updating the window length of the preset window slicing parameter data according to the instantaneous change rate data to obtain parameter update data; Generating parallel windows according to the parameter update data to obtain parallel window data; Slicing the fused neurology data according to the parallel window data to obtain secondary window slice data; Calculating the boundary mean difference of the secondary window slice data to obtain boundary mean difference data; Performing boundary smoothing according to the boundary mean difference data to obtain window slice data.

5. The system according to claim 4, wherein The parallel window generation includes: Performing multi-level window parameter mapping according to the parameter update data to obtain multi-level window mapping data; Generating preliminary slice indexes according to the multi-level window mapping data to obtain preliminary slice index data; Performing alignment error measurement according to the preliminary slice index data to obtain alignment error data; Performing jitter compensation according to the alignment error data to obtain parallel window data.

6. The system according to claim 1, wherein The construction steps of the preset edge feature extraction model include: Obtaining historical parallel feature data and corresponding historical feature extraction method data; Performing global average pooling on the historical parallel feature data to obtain parallel feature channel data; Performing lightweight multi-layer perceptron calculations on the parallel feature channel data to obtain channel attention data; Performing one-dimensional convolutional neural calculations according to the channel attention data to obtain one-dimensional convolutional neural data; Perform a layer of long short-term memory network calculation on one-dimensional convolutional neural data to obtain long short-term memory network data for one layer; Perform a fully connected dimensionality reduction layer process on the long short-term memory network data for one layer to obtain a primary edge feature extraction model; Perform offline iterative training on the primary edge feature extraction model according to historical feature extraction method data to obtain a secondary edge feature extraction model; Perform lightweight pruning on the secondary edge feature extraction model to obtain an edge feature extraction model.

7. The system according to claim 6, characterized in that, The lightweight pruning includes: Calculate the average activation intensity and gradient importance for the convolutional layer channels in the secondary edge feature extraction model to obtain average activation intensity data and gradient importance data respectively; Evaluate the channel sensitivity for the average activation intensity data and gradient importance data to obtain channel sensitivity data; Calculate the time sensitivity for the long short-term memory network layer channels in the secondary edge feature extraction model to obtain time sensitivity data; Perform similarity weight sorting on the secondary edge feature extraction model according to the channel sensitivity data and time sensitivity data to obtain channel sorting data; Calculate the intra-group average sensitivity according to the channel sorting data to obtain intra-group average sensitivity data; Allocate the cross-group pruning ratio to the channel sorting data according to the intra-group average sensitivity data to obtain cross-group pruning ratio data; Prune the secondary edge feature extraction model according to the cross-group pruning ratio data to obtain a pruned model; Rearrange the structure of the pruned model to obtain an edge feature extraction model.

8. The method according to claim 1, wherein The feature reasoning includes: Use a preset deep feature parsing model to identify neurology feature data to obtain atlas reasoning method data; Perform path reasoning on a preset medical knowledge atlas model according to the atlas reasoning method data to obtain candidate path data; Perform path score screening on the candidate path data to obtain feature reasoning data.

9. The system according to claim 1, characterized in that The construction steps of the deep feature parsing model include: Obtain historical neurology feature data and historical atlas reasoning method data; Construct feature points for the historical neurology fusion data to obtain feature point data; Generate tree nodes according to the feature point data to obtain tree node data; Construct a tree model according to the tree node data to obtain a primary deep feature parsing model; Perform iterative training on the primary deep feature parsing model according to the historical atlas reasoning method data to obtain a secondary deep feature parsing model; Perform pruning processing on the secondary deep feature parsing model to obtain a deep feature parsing model.

10. The system according to claim 9, wherein The feature point construction includes: Perform modal normalization on the historical neurology feature data to obtain modal normalization data; Calculate a saliency map according to the modal normalization data to obtain saliency map data; Perform multi-modal fusion saliency on the historical neurology fusion data according to the saliency map data to obtain saliency data; Calculate the autoencoder residual for the saliency data to obtain residual training data; Perform residual threshold segmentation according to the residual training data to obtain residual threshold segmentation data; Calculate the mutual information according to the residual threshold segmentation data and the saliency data to obtain mutual information data; Candidate key points are selected based on the mutual information data to obtain candidate key point data; Density peak clustering processing is performed on the candidate key point data to obtain feature point data.

Images