Neurology disease analysis method and system based on deep learning, and medium

Abstract

The invention discloses a neurology disease analysis method and system based on deep learning, and a medium. The method comprises the following steps: S1, obtaining data and executing preprocessing; s2, constructing a time sequence convolutional neural network model; s3, performing global search optimization on the structure parameter set by adopting an improved fireworks algorithm, and generating a time sequence convolutional neural network model of structure optimization; s4, an improved water wave optimization algorithm is adopted to adjust the disturbance direction and amplitude according to the polarization vector, and optimal hyper-parameter configuration is obtained; s5, training the model; and S6, inputting data, loading the data to an edge computing device to execute a reasoning task, and presenting a prediction result. According to the invention, synchronous improvement of high-precision prediction and model deployment efficiency of intelligent neural condition identification is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of neurological disease analysis, and in particular to a neurological disease analysis method, system and medium based on deep learning. Background Art

[0002] With the continuous deepening of the integration of artificial intelligence and medical health technologies, the intelligent analysis of neurological diseases based on electroencephalogram (EEG) data has gradually become a research hotspot. As a non-invasive, low-cost, and high-time-resolution physiological signal, electroencephalogram has important value in the early detection and disease monitoring of neurological diseases such as epilepsy, Parkinson's disease, and Alzheimer's disease. Traditional electroencephalogram analysis relies on manual reading of images and feature engineering, which is not only inefficient, but also requires extremely high professional skills. The analysis results are subjective, lacking standardization, intelligence, and scalability.

[0003] To improve the diagnosis efficiency, in recent years, researchers have begun to introduce deep learning models into the electroencephalogram analysis process, and use convolutional neural networks (CNNs), recurrent neural networks (RNNs) and their variants to automatically extract features and classify and model electroencephalogram data. Some methods attempt to combine convolutional modules with temporal modeling capabilities to achieve tasks such as disease stage division or seizure prediction. However, existing methods generally have problems such as model structure design relying on manual experience, weak generalization ability, and unstable training process. At the structural configuration level, most studies use fixed-layer numbers and fixed convolutional kernel structures, lacking an automatic structure search mechanism and being difficult to dynamically optimize the network design according to data characteristics. At the same time, at the training level, most hyperparameter configurations use static settings or manual grid search, which is inefficient and non-adaptive. The model training is prone to getting stuck in local optima and having weak generalization ability.

[0004] In addition, for medical signals such as electroencephalogram with strong noise, non-stationarity, and large individual differences, existing models often adopt a uniform perturbation strategy in terms of structure and parameters, ignoring the uneven influence of different dimensions in optimization. Under the perturbation control without directional guidance, phenomena such as gradient oscillation, perturbation redundancy, and ineffective update are likely to occur in the optimization process. At the same time, existing optimization methods mainly rely on classical swarm intelligence algorithms such as genetic algorithms and particle swarm optimization, lacking improvement strategies suitable for high-dimensional spaces, non-convex objectives, and constrained structure optimization scenarios, and it is difficult to balance search efficiency and structural feasibility.

[0005] At the practical application level, due to the general requirement of the medical scenario for the system to have real-time response capabilities, the existing model deployment process lacks consideration of the compatibility with the edge computing environment. Most deep models have large computational requirements and redundant parameters, making it difficult to directly deploy them on hospital terminals or portable devices, resulting in high model analysis latency and large energy consumption, which limits their actual implementation in clinical practice. For application scenarios such as neurology that have high requirements for continuous state monitoring and trend prediction, the existing solutions still have significant deficiencies in aspects such as model structure adjustment ability, training parameter adaptability, and deployment portability.

[0006] In summary, the current intelligent analysis system for neurology conditions mainly faces the following problems: First, there is a lack of a collaborative mechanism that can automatically optimize the structure and parameter configuration to achieve the joint evolution of structure adjustment and training parameter tuning; second, there is a lack of a control mechanism for the directionality of the perturbation strategy, and it is impossible to dynamically guide according to the influence differences of different hyperparameter dimensions; third, there is a lack of a lightweight deployment process for edge computing, and it is difficult to perform efficient inference and rapid application after the model training is completed.

[0007] Therefore, how to provide a neurology condition analysis method, system, and medium based on deep learning is an urgent problem for those skilled in the art to solve. Summary of the Invention

[0008] An object of the present invention is to propose an intelligent analysis method and system for neurology conditions that combines an improved fireworks algorithm and polarization water wave optimization. Based on the deep learning model construction mechanism, the present invention automatically extracts features and analyzes the condition of the electroencephalogram time series signal, and realizes the co-evolution of the network structure and hyperparameters by introducing a structure search optimization and polarization parameter tuning mechanism. The present invention details the complete process from electroencephalogram data preprocessing, structure parameter search, adaptive perturbation generation, polarization vector guidance, model training optimization to edge deployment inference, and has the advantages of high optimization efficiency, stable model performance, fast response speed, and adaptation to edge computing, and can significantly improve the intelligent level and clinical application value of the neurology auxiliary diagnosis system.

[0009] According to the deep learning-based neurology condition analysis method of the embodiments of the present invention, it includes the following steps: S1. Obtain the electroencephalogram time series signal data of neurology patients and perform preprocessing; S2. Construct a time series convolutional neural network model including a set of structure parameters; S3. Use the improved fireworks algorithm to globally search and optimize the set of structural parameters. Generate structural parameter combinations based on the dimensions of the structural parameters and conduct training. Construct a structural legality verification module to eliminate the structural parameter combinations that do not meet the network configuration rules. Construct a multi-objective fitness function to score all the structural parameter combinations. Introduce an inverse fitness perturbation mechanism to select the structural parameter combination with the highest fitness and generate a temporally convolutional neural network model with optimized structure. S4. Set the hyperparameter set, attach polarization vectors to each group of hyperparameters. Use the improved water wave optimization algorithm to propagate, reflect, and refract the hyperparameter solutions in each round of training. Adjust the perturbation direction and amplitude according to the polarization vectors, and conduct evaluations based on the multi-objective fitness function. Trigger the hyperparameter perturbation self-repair mechanism when the fitness decreases after perturbation, and iteratively obtain the optimal hyperparameter configuration. S5. Under the control of the optimal structural parameter combination and the optimal hyperparameter configuration, train the temporally convolutional neural network model with optimized structure to obtain the final analysis model. S6. Input the preprocessed electroencephalogram temporal signal data into the final analysis model to obtain the disease stage classification result and the disease trend prediction result. Generate a deployment version and load it onto the edge computing device to execute the inference task, and present the disease stage classification result and the disease trend prediction result.

[0010] Optionally, in step S1, the electroencephalogram temporal signal data is the continuous voltage time series obtained from the acquisition channels, and the continuous voltage time series is preprocessed to generate standardized electroencephalogram data samples.

[0011] Optionally, in step S2, the structural parameters of the temporally convolutional neural network model include the convolutional kernel size, the number of network layers, and the dilation coefficient, the number of channels in each layer, the residual connection method, and the activation function type.

[0012] Optionally, S3 specifically includes: S31. Set the structural parameter search space, and combine each group of structural parameters into a structural vector. S32. Take the structural vectors in the structural parameter search space as the initial explosion centers, calculate the fitness values for each initial explosion center according to the prediction performance of the corresponding temporally convolutional neural network model on the validation set, and dynamically set the adaptive explosion radius based on the fitness values. S33. Based on each initial explosion center and the corresponding explosion radius, generate new structural parameter combinations in the neighborhood, and perform structural legality verification on the generated structural parameter combinations, only retaining the structural parameter combinations that meet the neural network construction rules. S34. Configure the structural parameter combinations that meet the neural network construction rules into the temporal convolutional neural network model respectively, train on the training set, record the accuracy, number of training rounds, and resource occupancy information on the validation set, calculate the comprehensive fitness score, and measure the overall performance of this structure; S35. Introduce the inverse fitness perturbation strategy into the explosion centers with lower fitness rankings, and continuously update the explosion center set during the iteration process; S36. Repeat steps S32 to S35 until the set maximum search rounds or the global fitness convergence criterion is reached, and finally determine the structural parameter combination with the highest fitness as the structural configuration of the temporal convolutional neural network model to generate the structure optimization model.

[0013] Optionally, the inverse fitness perturbation strategy in S35 includes: during each round of search iteration of the structural parameter combination, determine the part of the structural parameter combinations in the current explosion center set with fitness values lower than the median as the low-fitness explosion center set. For each structural parameter combination in the low-fitness explosion center set, increase the number of spark generations, expand the perturbation range within the search neighborhood, dynamically increase the explosion radius value, and generate a new round of structural parameter combinations in the search space according to the expanded explosion radius value, and include them in the next round of explosion center candidate sets.

[0014] Optionally, S4 specifically includes: S41. Set the hyperparameter set based on the structure optimization model, and the hyperparameter set includes the learning rate , regularization factor , the loss function weight coefficient for the disease stage classification task and the loss function weight coefficient for the disease trend prediction task , and define the hyperparameter set ; S42. Construct a polarization vector for each group of hyperparameters, where each component corresponds to the perturbation direction preference and perturbation amplitude adjustment factor of a hyperparameter dimension respectively; S43. Perform propagation perturbation on each hyperparameter set , and add the corresponding perturbation amounts to each component in the hyperparameter set in turn to obtain the hyperparameter set generated after perturbation; S44. Perform boundary detection on the hyperparameter set generated after perturbation. If a certain dimension exceeds the preset value range, reflect it back to the legal interval in a mirror image manner. When the fitness value improves compared with the previous round, trigger the refraction operation and extend the perturbation path along the forward direction; S45. When the fitness value of the current round decreases compared to the previous round, calculate the difference in the learning rate between this round and the previous round, introduce a first-order penalty term, and perform numerical correction on the fitness function. S46. After each round of iteration, update the components of each dimension in the polarization vector. S47. The components of each dimension in the updated polarization vector control the perturbation amount of the corresponding hyperparameter set in the next round of propagation process. After iterative update, the optimal hyperparameter configuration is obtained.

[0015] According to the embodiment of the present invention, a neurology disease analysis system based on deep learning includes: A data processing module, configured to collect neurology electroencephalogram time-series signal data, and perform filtering, denoising, normalization, and window slicing processing to generate preprocessed electroencephalogram time-series signal data. A structure optimization module, configured to construct a structure parameter space, perform structure combination search, legality screening, and fitness evaluation using an improved fireworks algorithm, and output a structure-optimized temporal convolutional neural network model. A perturbation enhancement module, configured to introduce inverse fitness perturbation for low-fitness structures, expand the search range, and enhance the local exploration ability in the structure optimization stage. A polarization parameter adjustment module, configured to set hyperparameters and allocate a polarization vector, guide the perturbation direction in propagation, reflection, and refraction, and dynamically update the polarization components in combination with the fitness gradient. A joint modeling module, which constructs a model based on the optimal structure parameter combination and the optimal hyperparameters, jointly completes the disease stage classification and trend prediction tasks, and generates a final analysis model. A lightweight deployment module, configured to perform pruning and compression on the model to generate a version that can be deployed on edge devices, and realize on-site real-time inference and auxiliary diagnosis. An operation control module, configured to perform model inference, output disease analysis results, support integration with the terminal system and clinical rapid feedback.

[0016] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor can execute a neurology disease analysis method based on deep learning.

[0017] The beneficial effects of the present invention are: (1) By introducing an improved fireworks algorithm to adaptively search for the neural network structure parameters, combining structure dimension-aware explosion radius setting, legality screening mechanism, and multi-objective fitness evaluation, the present invention realizes the automatic construction and fine optimization of the network structure, significantly reduces the dependence on artificial experience in structure design, and improves the structure matching degree and generalization ability of the model for complex electroencephalogram data.

[0018] (2) In the training stage of the present invention, an improved water wave optimization algorithm is adopted to adjust hyperparameters, combined with a polarization vector guiding mechanism, to achieve dynamic control of the perturbation direction and perturbation intensity. Through the coupling of the update and propagation - reflection - refraction mechanism of polarization components, the training process can focus on the fine optimization of highly sensitive dimensions, enhancing the convergence stability and robustness of parameter tuning and avoiding the training instability problems brought by traditional fixed perturbation strategies.

[0019] (3) After the model is constructed, through lightweight processing and edge deployment solutions, the final model is compressed into a real-time inference version adapted to on-site terminals and can be directly deployed to run on medical devices or edge computing nodes. This design realizes local inference feedback for the disease stage classification and trend prediction tasks, with advantages such as fast response speed, flexible deployment, and strong clinical adaptability, improving the practical applicability and usage value of the neurology disease analysis system. Description of the Drawings

[0020] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings: Figure 1 is the flowchart of the neurology disease analysis method based on deep learning proposed by the present invention; Figure 2 is the flowchart of the improved fireworks algorithm for the neurology disease analysis method based on deep learning proposed by the present invention; Figure 3 is the flowchart of hyperparameter adjustment of the polarization water wave optimization algorithm for the neurology disease analysis method based on deep learning proposed by the present invention. Detailed Embodiments

[0021] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0022] Refer to Figures 1-3 , the neurology disease analysis method based on deep learning includes the following steps: S1. Obtain the electroencephalogram time series signal data of neurology patients and perform preprocessing; S2. Construct a time series convolutional neural network model including a set of structural parameters. The model structure includes the convolutional kernel size, the number of network layers and the dilation coefficient, the number of channels in each layer, the residual connection method, and the type of activation function; S3. Use the improved fireworks algorithm to globally search and optimize the set of structural parameters. Set the adaptive explosion radius based on the dimensions of the structural parameters to generate combinations of structural parameters. Train each combination of structural parameters, construct a structural legality verification module to eliminate combinations of structural parameters that do not meet the network configuration rules, construct a multi-objective fitness function, score all combinations of structural parameters, introduce an inverse fitness perturbation mechanism, select the combination of structural parameters with the highest fitness, and generate a temporally convolutional neural network model with optimized structure; S4. Set the hyperparameter set based on the temporally convolutional neural network model with optimized structure. Attach polarization vectors to each set of hyperparameters. Use the improved water wave optimization algorithm to propagate, reflect, and refract the hyperparameter solutions in each round of training. Adjust the perturbation direction and amplitude according to the polarization vectors, evaluate based on the multi-objective fitness function, and trigger the hyperparameter perturbation self-repair mechanism when the fitness decreases after perturbation, and iteratively obtain the optimal hyperparameter configuration; S5. Under the control of the optimal combination of structural parameters and the optimal hyperparameter configuration, train the temporally convolutional neural network model with optimized structure to obtain the final analysis model. Specifically, after obtaining the optimal combination of structural parameters (including the convolution kernel size, number of network layers, number of channels, activation function type, etc.) optimized by the improved fireworks algorithm and the optimal hyperparameter configuration (including the learning rate, regularization factor, and weight coefficients of the multi-task loss function) searched by the polarization water wave optimization algorithm, input the pre-processed electroencephalogram temporal signal data into the temporally convolutional neural network model with this structural configuration, perform multiple rounds of training according to the set hyperparameters, jointly complete the task of classifying the disease stage and predicting the disease trend, and finally obtain an analysis model that performs optimally under multi-objective evaluations such as accuracy, resource occupancy, and training stability, providing a model basis for subsequent deployment and inference; S6. Input the pre-processed electroencephalogram temporal signal data into the final analysis model to obtain the disease stage classification result and the disease trend prediction result, generate a deployment version, load it into the edge computing device to perform the inference task, and present the disease stage classification result and the disease trend prediction result. Specifically, after obtaining the final analysis model, input the pre-processed electroencephalogram temporal signal data into the analysis model, perform forward inference under the optimal combination of structural parameters and the optimal hyperparameter configuration, and output the disease stage classification result and the disease trend prediction result simultaneously. Subsequently, perform lightweight processing on the model, including pruning, quantization, and compression, generate a deployment version adapted to the edge computing device, and load it into medical terminals, embedded devices, or other edge nodes to achieve local real-time inference of electroencephalogram data and presentation of prediction results, supporting rapid clinical judgment and auxiliary decision-making.

[0023] The present invention realizes the efficient and automatic modeling of the neurological disease analysis task by constructing a complete closed-loop process from electroencephalogram (EEG) data preprocessing, structure optimization, hyperparameter adjustment to edge deployment inference. Compared with the existing method that relies on manual setting of the structure and parameter tuning, the present invention introduces swarm intelligence optimization and dynamic perturbation mechanisms, effectively improving the model performance, training stability and clinical adaptability. At the same time, it supports lightweight deployment, can quickly complete the identification of disease stages and trend prediction on edge terminals, and has higher practicality and application value.

[0024] In this embodiment, the EEG time-series signal data in step S1 is a continuous voltage time series obtained from the acquisition channels. The continuous voltage time series is preprocessed to generate standardized EEG data samples. During preprocessing, first, a 0.5–40 Hz band-pass filter is applied to the original EEG time-series signal to remove low-frequency drift and high-frequency noise. Subsequently, artifact rejection is performed to strip interference signals. Then, the Z-score normalization method is used to standardize the signal amplitudes of each channel to ensure consistent feature distributions among different samples. Finally, the continuous signal is sliced into time windows according to a fixed duration and overlap rate to form structured short-time signal segments for input into the neural network model for subsequent analysis. In addition to the continuous voltage time series, the preprocessed EEG time-series signal data also includes: channel spatial structure information, the time-series segments after time window slicing, the clean signal after filtering and artifact rejection, the normalization result, and the disease stage labels and trend labels used in the training stage.

[0025] In this embodiment, the structural parameters of the temporal convolutional neural network model in step S2 include the convolutional kernel size, the number of network layers, and the dilation coefficient, the number of channels in each layer, the residual connection method, and the type of activation function.

[0026] In this embodiment, S3 specifically includes: S31. Set the search space for structural parameters and combine each set of structural parameters into a structural vector; S32. Use the structural vectors in the search space for structural parameters as the initial explosion centers, and calculate the fitness value for each initial explosion center according to the prediction performance of the corresponding temporal convolutional neural network model on the validation set. The fitness value is mainly based on the validation set accuracy, and at the same time takes into account the training time consumption and model complexity. The calculation formula is as follows: ; Where, is the fitness value of the th initial explosion center, is the validation set accuracy, with a value range of 0~1, is the normalized value of the training time or number of epochs, is the normalized value of the number of model parameters or computational complexity; the coefficient 0.1 represents a mild penalty for training efficiency and model complexity; dynamically set the adaptive explosion radius based on the fitness value, where the explosion radius is calculated as follows: ; where, is the explosion radius control coefficient, is the maximum fitness value, is the fitness value of the th initial explosion center, is the fitness value of the th initial explosion center, is the total number of initial explosion centers participating in the explosion radius setting in this round of fireworks algorithm, and the explosion radius is automatically adjusted according to the fitness difference; S33. Based on each initial explosion center and the corresponding explosion radius, generate new combinations of structural parameters in the neighborhood, and perform structural legality verification on the generated combinations of structural parameters. Only retain the combinations of structural parameters that meet the neural network construction rules. The specific process of structural legality verification includes: checking whether parameters such as the convolution kernel size, stride, dilation coefficient, etc. are compatible with the input size, ensuring that the number of channels between layers matches, the residual connection has dimensional consistency, the type of activation function is compatible with the context module, and the overall topological structure of the network is reasonable and acyclic, and the input and output are closed. For the structural combinations that do not meet the above rules, the system will automatically eliminate them, and only retain the combinations of structural parameters of the legal network structures that can be successfully constructed and trained; S34. Configure the combinations of structural parameters that meet the neural network construction rules into the temporal convolutional neural network model respectively, train on the training set, record the accuracy, number of training rounds, and resource occupancy information on the validation set, and calculate the comprehensive fitness score. The calculation method of the comprehensive fitness score is the same as the fitness value calculation method in S32, which measures the overall performance of this structure; S35. Introduce the inverse fitness perturbation strategy in the explosion centers with relatively low fitness rankings, and continuously update the set of explosion centers during the iteration process; S36. Repeat steps S32 to S35 until the set maximum search rounds or the global fitness convergence criterion is reached, and finally determine the combination of structural parameters with the highest fitness as the structural configuration of the temporal convolutional neural network model to generate a structure-optimized model.

[0027] In the structural parameter optimization stage, the present invention adopts an improved fireworks algorithm, and introduces a dimension-aware explosion radius, an adaptive perturbation distribution, a multi-objective fitness evaluation, and a legality screening mechanism to construct an efficient structure search process suitable for neural network modeling tasks. Compared with traditional random search or fixed-structure network methods, this solution can significantly improve the efficiency and feasibility of structure search, and enhance the local exploration ability of the structure space through an inverse fitness perturbation strategy, effectively avoiding falling into local optima.

[0028] In this embodiment, the inverse fitness perturbation strategy in S35 includes: in each round of structural parameter combination search iteration, determining the part of the structural parameter combinations with fitness values lower than the median in the current explosion center set as the low-fitness explosion center set. For each structural parameter combination in the low-fitness explosion center set, increase the number of sparks generated, expand the perturbation range within the search neighborhood, dynamically increase the explosion radius value, and generate a new round of structural parameter combinations in the search space according to the expanded explosion radius value, and incorporate them into the next round of explosion center candidate sets.

[0029] The inverse fitness perturbation mechanism proposed by the present invention improves the coverage ability of the inefficient solution region in the optimization process by actively identifying the structural solutions with low fitness scores and expanding their perturbation ranges. Compared with the traditional uniform perturbation strategy, this method can expand the search diversity in the early stage, accelerate the convergence to the high-quality solution region in the later stage, balance global exploration and local convergence, and further improve the accuracy and robustness of structural parameter optimization.

[0030] In this embodiment, S4 specifically includes: S41. Set a hyperparameter set based on the structure optimization model. The hyperparameter set includes the learning rate , the regularization factor , the loss function weight coefficient for the disease stage classification task and the loss function weight coefficient for the disease trend prediction task , and define the hyperparameter set ; S42. Construct a polarization vector for each group of hyperparameters, where each component corresponds to the perturbation direction preference and the perturbation amplitude adjustment factor of a hyperparameter dimension respectively; S43. Perform propagation perturbation on each hyperparameter set , add each component in the hyperparameter set to the corresponding perturbation amount in turn to obtain the hyperparameter set generated by perturbation, and the formula for calculating the perturbation amount is: ; where is the global perturbation coefficient, is the -dimensional component in the polarization vector, is the -dimensional sub-goal value of the hyperparameter at the -th iteration round, is the -dimensional sub-goal value of the hyperparameter at the -th iteration round, is a non-zero positive real constant to prevent division-by-zero error; this perturbation formula combines the polarization vector and the change trend of the sub-goal value, endowing each hyperparameter perturbation with directionality and intensity regulation. Its practical significance lies in enhancing the convergence efficiency and stability of hyperparameter optimization in complex electroencephalogram modeling tasks through a guiding and self-adaptive perturbation mechanism around the theme of "neurological condition analysis", making the model more accurate and robust in disease stage classification and trend prediction, and thus ensuring the reliability and real-time response ability of the final analysis model in clinical deployment; S44. Perform boundary detection on the hyperparameter set generated by perturbation, that is, judge whether its value exceeds the pre-set legal value range (such as learning rate ∈ [1e - 5, 1e - 1], regularization factor ∈ [0, 0.1], etc.). If a certain dimension exceeds the preset value range, reflect it back to the legal interval in a mirror image manner. When the fitness value improves compared to the previous round, trigger the refraction operation and extend the perturbation path along the forward direction; S45. When the fitness value of the current iteration round decreases compared to the previous round, calculate the learning rate difference between this round and the previous round, introduce a first-order penalty term, and numerically correct the fitness function; S46. After each iteration ends, update the components of each dimension in the polarization vector. The update formula is as follows: ; where: is the -dimensional component in the polarization vector at the -th iteration, is the -dimensional component in the polarization vector at the -th iteration, is the polarization enhancement adjustment coefficient, is the hyperbolic tangent function, is the fitness function with respect to the -dimensional hyperparameter; S47. The components of each dimension in the updated polarization vector control the perturbation amount of the corresponding hyperparameter set in the next propagation process. Through iterative update, the optimal hyperparameter configuration is obtained.

[0031] In the process of hyperparameter adjustment, the present invention introduces a polarization water wave optimization mechanism, guides the perturbation direction through polarization vectors, and realizes dynamic regulation by combining propagation, reflection, refraction operations and fitness gradient feedback, significantly improving the training stability and parameter adjustment efficiency. Compared with traditional static or manual parameter adjustment methods, this method has stronger adaptability and intelligence, and shows higher optimization quality and convergence speed especially when facing multi-objective loss weight configuration and non-convex search space, and has good scalability.

[0032] A neurology disease analysis system based on deep learning according to an embodiment of the present invention includes: A data processing module, configured to collect time-series electroencephalogram signal data of neurology, and perform filtering, denoising, normalization and window slicing processing to generate preprocessed time-series electroencephalogram signal data; A structure optimization module, configured to construct a structure parameter space, perform structure combination search, legality screening and fitness evaluation by using an improved fireworks algorithm, and output a time-series convolutional neural network model with optimized structure; A perturbation enhancement module, configured to introduce inverse fitness perturbation for low-fitness structures, expand the search range, and enhance the local exploration ability in the structure optimization stage; A polarization parameter adjustment module, configured to set hyperparameters and allocate polarization vectors, guide the perturbation direction in propagation, reflection, and refraction, and dynamically update the polarization components in combination with the fitness gradient; A joint modeling module, constructs a model based on the optimal structure parameter combination and the optimal hyperparameters, jointly completes the disease stage classification and trend prediction tasks, and generates a final analysis model; A lightweight deployment module, configured to perform pruning and compression on the model, generate a version deployable on edge devices, and realize on-site real-time inference and auxiliary diagnosis; An operation control module, configured to perform model inference, output disease analysis results, support integration with the terminal system and provide rapid clinical feedback.

[0033] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor is enabled to execute a neurology disease analysis method based on deep learning.

[0034] Example 1: To verify the feasibility of the present invention in implementation, the present invention is applied to the EEG monitoring system upgrade project in the neurology outpatient department of a certain top-three hospital to realize dynamic disease monitoring and prediction of epilepsy high-risk patients. In this project, the hospital hopes to deploy an intelligent analysis system to realize real-time processing of patient electroencephalogram data, automatic judgment of disease stages, and future trend prediction, so as to assist doctors in formulating more scientific treatment plans and alleviating the shortage of experts and the diagnostic burden.

[0035] In the actual deployment process, the system accesses the hospital's electroencephalogram (EEG) acquisition system to obtain multi-channel EEG time-series data of approximately 75 outpatients or patients under observation daily. The sampling frequency is 256 Hz, and the acquisition duration for each patient is approximately 30 minutes. The data processing module first performs 0.5–40 Hz band-pass filtering, artifact removal, Z-score normalization, and time window segmentation operations on the original data, converting the continuous data into a short-time segment format suitable for model processing. In the structural optimization stage, an improved fireworks algorithm is used to automatically search for combinations of network layers, convolutional kernel sizes, number of channels, and activation function configurations in the parameter space, avoiding the low efficiency and overfitting risks caused by manual repeated debugging of the structure in the past.

[0036] To improve the training efficiency and prediction stability, the system introduces a polarized water wave optimization algorithm to dynamically adjust the learning rate, regularization factor, and dual-task loss weight coefficient. By constructing a polarization vector for each dimension of the hyperparameter, the algorithm can accurately control the perturbation direction and intensity in each round of perturbation, and realize automatic parameter tuning by combining the propagation-reflection-refraction mechanism and fitness gradient feedback. The average time taken for each structure search in the entire optimization process is less than 2 hours, and the number of parameter tuning rounds is controlled within 60 times, improving the efficiency by more than 80% compared with the traditional grid search method.

[0037] After the model training is completed, lightweight processing of the deployment version is carried out, including pruning rate of 0.4, 8-bit quantization, and distillation compression, etc. The model size is compressed from the original 47 MB to 12.3 MB, and it can be directly deployed on the embedded diagnostic terminals already configured in the hospital, realizing the prediction analysis of the disease stage and trend of an EEG segment within 0.5 seconds on a single device. One month after the system was launched, the total number of patient samples accessed reached 2,164. By comparing and verifying with the results of manual judgment by the neurology department director, the accuracy rate of disease stage identification is 93.4%, and the correct rate of trend prediction is 89.7%, which is significantly higher than that of the static convolutional model used before deployment.

[0038] In addition, in actual use, the average feedback time of doctors for this system is 0.38 seconds per round of inference, greatly improving the clinical response speed. Especially in the early intervention warning of epileptic seizures, the model successfully detected 153 potential high-risk fluctuation signals in advance, and 89 of them were confirmed by doctors to have clinical significance, assisting in adjusting the treatment strategy in advance.

[0039] The system is launched in some cooperative medical units in Beijing, Tianjin, and Jiangsu. The deployment in the hospital does not rely on GPU and can run on low-power NPU terminals. The daily analyzed data volume is between 3,500 and 4,200 pieces. The system has high operating stability, with a daily average failure rate <0.2% and low maintenance workload. The actual use feedback shows that this solution can effectively reduce the workload of doctors, improve the diagnostic efficiency, and significantly enhance the processing ability of complex nerve signals.

[0040] Table 1: Comparison data table of the method of the present invention and existing methods in the scenario of neurological disease analysis

[0041] Through comparative experiments, it can be seen that the Baseline-CNN method, as a traditional static convolutional network, has relatively limited performance in the tasks of disease stage recognition and trend prediction, only achieving accuracies of 84.2% and 74.6% respectively. Moreover, due to its fixed structure and parameter tuning relying on manual work, the overall performance improvement space is limited, and the inference delay is as high as 1.12 seconds, which does not meet the requirements of edge real-time analysis. In contrast, the CNN model with grid search parameter tuning has a slight improvement in accuracy and trend prediction, reaching 88.1% and 82.3% respectively. However, due to the fixed structure, the search process takes 5.5 hours, with low efficiency, and the model compression is insufficient, and the deployment volume is still 42MB, and the on-site inference speed has not been significantly improved.

[0042] The system proposed by the present invention combines an improved fireworks algorithm and a polarized water wave optimization strategy, which not only has the ability of dynamic optimization at the structural design level, but also can achieve adaptive perturbation control based on the fitness gradient during the parameter tuning process. Under the same task test set, the disease stage recognition accuracy of the model of the present invention reaches 93.4%, and the trend prediction accuracy reaches 89.7%, which is significantly better than the control group, verifying the modeling effect of the structure-parameter collaborative optimization mechanism on complex electroencephalogram data. At the same time, due to considering the deployment constraints during the optimization process, the model can be compressed to 12.3MB after training, which is only 1 / 4 of the traditional model, realizing efficient deployment on edge terminal devices. The inference time is significantly shortened to 0.38 seconds, which can meet the high response requirements in outpatient clinics, mobile terminals and edge-side scenarios.

[0043] In summary, the present invention not only significantly improves the model accuracy and prediction ability, but also greatly optimizes the system deployment efficiency and runtime performance, taking into account both model quality and application feasibility, and has broad promotion value.

[0044] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution of the present invention and its inventive concept, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A method for analyzing the condition of neurology based on deep learning, characterized in that, The steps are as follows: S1. Obtain the electroencephalogram (EEG) time series signal data of neurology patients and perform preprocessing; S2. Construct a time series convolutional neural network model including a set of structural parameters; S3. Use an improved fireworks algorithm to globally search and optimize the set of structural parameters, generate structural parameter combinations based on the structural parameter dimensions and conduct training, construct a structural legality verification module to eliminate the structural parameter combinations that do not meet the network configuration rules, construct a multi-objective fitness function, score all the structural parameter combinations, introduce an inverse fitness perturbation mechanism, select the structural parameter combination with the highest fitness, and generate a structurally optimized time series convolutional neural network model; S4. Set a set of hyperparameters, attach polarization vectors to each group of hyperparameters, use an improved water wave optimization algorithm to propagate, reflect, and refract the hyperparameter solutions in each round of training, adjust the perturbation direction and amplitude according to the polarization vectors, evaluate based on the multi-objective fitness function, trigger the hyperparameter perturbation self-repair mechanism when the fitness decreases after perturbation, and iteratively obtain the optimal hyperparameter configuration; S5. Under the control of the optimal structural parameter combination and the optimal hyperparameter configuration, train the structurally optimized time series convolutional neural network model to obtain the final analysis model; S6. Input the preprocessed EEG time series signal data into the final analysis model to obtain the disease stage classification result and the disease trend prediction result, generate a deployment version, load it onto the edge computing device to execute the inference task, and present the disease stage classification result and the disease trend prediction result.

2. The method for analyzing neurological diseases based on deep learning according to claim 1, wherein In step S1, the EEG time series signal data is a continuous voltage time series obtained from the acquisition channel, and the continuous voltage time series is preprocessed to generate a standardized EEG data sample.

3. The method for analyzing neurological diseases based on deep learning according to claim 2, wherein In step S2, the structural parameters of the time series convolutional neural network model include the convolution kernel size, the number of network layers, and the dilation coefficient, the number of channels in each layer, the residual connection method, and the activation function type.

4. The method for analyzing neurological diseases based on deep learning according to claim 3, wherein The specific steps of S3 are as follows: S31. Set the structural parameter search space and combine each group of structural parameters into a structural vector; S32. Take the structural vectors in the structural parameter search space as the initial explosion centers, calculate the fitness value for each initial explosion center according to the prediction performance of the corresponding time series convolutional neural network model on the validation set, and dynamically set the adaptive explosion radius based on the fitness value; S33. Based on each initial explosion center and the corresponding explosion radius, generate new structural parameter combinations in the neighborhood, perform structural legality verification on the generated structural parameter combinations, and only retain the structural parameter combinations that meet the neural network construction rules; S34. Configure the structural parameter combinations that meet the neural network construction rules into the time series convolutional neural network model respectively, train on the training set, record the accuracy, the number of training rounds, and the resource occupancy information on the validation set, and calculate the comprehensive fitness score to measure the overall performance of this structure; S35. Introduce an inverse fitness perturbation strategy in the explosion centers with relatively low fitness rankings and continuously update the set of explosion centers during the iteration process; S36. Repeat steps S32 to S35 until the set maximum search round or the global fitness convergence criterion is reached, and finally determine the structural parameter combination with the highest fitness as the structural configuration of the temporal convolutional neural network model, and generate a structure optimization model.

5. The method for analyzing neurological diseases based on deep learning according to claim 4, wherein The inverse fitness perturbation strategy in S35 includes: in each round of structural parameter combination search iteration, determine the partial structural parameter combinations with fitness values lower than the median in the current explosion center set as the low-fitness explosion center set. For each structural parameter combination in the low-fitness explosion center set, increase the number of spark generations, expand the perturbation range within the search neighborhood, dynamically increase the explosion radius value, and generate a new round of structural parameter combinations in the search space according to the expanded explosion radius value, and include them in the next round of explosion center candidate sets.

6. The method for analyzing neurological diseases based on deep learning according to claim 5, wherein, S4 specifically includes: S41. Set a hyperparameter set based on the structure optimization model. The hyperparameter set includes a learning rate , a regularization factor , a weight coefficient of the loss function for the disease stage classification task , and a weight coefficient of the loss function for the disease trend prediction task , and define the hyperparameter set ; S42. Construct a polarization vector for each set of hyperparameters , where each component corresponds to the perturbation direction preference and perturbation amplitude adjustment factor of a hyperparameter dimension respectively; S43. For each set of hyperparameters perform propagation perturbation, and add the corresponding perturbation amounts to each component in the set of hyperparameters in turn to obtain a set of hyperparameters generated by perturbation; ​ S44. Perform boundary detection on the hyperparameter set generated by perturbation. If a certain dimension exceeds the preset value range, reflect it back to the legal interval in a mirror image manner. When the fitness value improves compared with the previous round, trigger the refraction operation and extend the perturbation path along the forward direction. S45. When the fitness value of the current round is lower than that of the previous round, calculate the learning rate difference between this round and the previous round, introduce a first-order penalty term, and perform numerical correction on the fitness function. S46. After each round of iteration, update the components of each dimension in the polarization vector. S47. The components of each dimension in the updated polarization vector control the perturbation amount of the corresponding hyperparameter set in the next round of propagation process. Through iterative update, the optimal hyperparameter configuration is obtained.

7. A neurological disease analysis system based on deep learning, applied to the neurological disease analysis method based on deep learning according to any one of claims 1 to 6, characterized in that, It includes: A data processing module, which is used to collect the electroencephalogram time series signal data of neurology, and perform filtering, denoising, normalization and window slicing processing to generate preprocessed electroencephalogram time series signal data. A structure optimization module, which is used to construct a structural parameter space, perform structural combination search, legality screening and fitness evaluation using an improved fireworks algorithm, and output a structure-optimized temporal convolutional neural network model. A perturbation enhancement module, which is used to introduce inverse fitness perturbation for low-fitness structures, expand the search range, and enhance the local exploration ability in the structure optimization stage. A polarization parameter adjustment module, which is used to set hyperparameters and allocate polarization vectors, guide the perturbation direction in propagation, reflection and refraction, and dynamically update the polarization components in combination with the fitness gradient. A joint modeling module, which constructs a model based on the optimal structural parameter combination and the optimal hyperparameters, jointly completes the task of disease stage classification and trend prediction, and generates a final analysis model. A lightweight deployment module, which is used to perform pruning and compression on the model to generate a version that can be deployed on edge devices, and realize on-site real-time inference and auxiliary diagnosis. An operation control module, which is used to perform model inference, output the disease analysis result, support integration with the terminal system and provide rapid clinical feedback.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor can execute the deep learning-based neurology disease analysis method according to any one of claims 1 to 6.

Images