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

Abstract

This invention discloses a method, system, and medium for analyzing neurological conditions based on deep learning. The method comprises the following steps: S1, acquiring data and performing preprocessing; S2, constructing a time-series convolutional neural network model; S3, using an improved fireworks algorithm to perform global search optimization on a set of structural parameters to generate a structurally optimized time-series convolutional neural network model; S4, using an improved water wave optimization algorithm to adjust the perturbation direction and amplitude based on the polarization vector to obtain the optimal hyperparameter configuration; S5, training the model; and S6, inputting data, loading it to an edge computing device to perform inference tasks, and presenting prediction results. This invention achieves high-precision prediction for intelligent identification of neurological conditions and simultaneously improves model deployment efficiency.

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 deepening integration of artificial intelligence and healthcare technologies, intelligent analysis of neurological diseases based on electroencephalogram (EEG) data has become a research hotspot. As a non-invasive, low-cost, and high-temporal-resolution physiological signal, EEG plays a vital role in the early detection and monitoring of neurological diseases such as epilepsy, Parkinson's disease, and Alzheimer's disease. Traditional EEG analysis relies on manual image reading and feature engineering, which is not only inefficient but also requires extremely high professional skills, resulting in highly subjective analysis results and a lack of standardization, intelligence, and scalability.

[0003] To improve diagnostic efficiency, researchers have begun to introduce deep learning models into the EEG analysis process in recent years, using convolutional neural networks (CNNs), recurrent neural networks (RNNs), and their variants to automatically extract features and perform classification modeling on EEG data. Some methods attempt to combine convolutional modules with time series modeling capabilities to achieve tasks such as disease stage classification or epileptic seizure prediction. However, existing methods generally suffer from problems such as reliance on manual experience in model structure design, weak generalization capabilities, and unstable training processes. At the structural configuration level, most studies use a fixed number of layers and a fixed convolution kernel structure, lacking an automatic structural search mechanism, making it difficult to dynamically optimize the network design based on data characteristics. At the same time, at the training level, hyperparameter configuration mostly uses static settings or manual grid searches, which are inefficient and non-adaptive. Model training is prone to falling into local optimality and poor generalization capabilities.

[0004] Furthermore, for EEGs, a medical signal characterized by strong noise, non-stationarity, and significant individual variability, existing models often employ uniform perturbation strategies in both structure and parameters, ignoring the uneven impact of different dimensions on optimization. Without directional guidance in perturbation control, the optimization process is prone to gradient oscillation, redundant perturbations, and ineffective updates. Furthermore, existing optimization methods primarily rely on classic swarm intelligence algorithms such as genetic algorithms and particle swarms, lacking improved strategies suitable for optimizing high-dimensional spaces, non-convex objectives, and constrained structures, making it difficult to balance search efficiency and structural feasibility.

[0005] In practical applications, because medical scenarios generally require systems to have real-time response capabilities, the existing model deployment process lacks compatibility considerations for edge computing environments. Most deep models have large computational loads and redundant parameters, making them difficult to deploy directly on hospital terminals or portable devices. This results in high model analysis latency and high energy consumption, limiting their actual implementation in clinical practice. For application scenarios such as neurology, which require high continuous state monitoring and trend prediction, existing solutions still have significant deficiencies in terms of model structure adjustment capabilities, training parameter adaptability, and deployment portability.

[0006] In summary, the current intelligent analysis system for neurological conditions faces the following main 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 structural adjustment and training parameter adjustment; second, there is a lack of a control mechanism for the direction of perturbation strategies, and it is impossible to dynamically guide the impact differences of different hyperparameter dimensions; third, there is a lack of a lightweight deployment process for edge computing, which makes it difficult to efficiently infer and quickly apply the model after training.

[0007] Therefore, how to provide a neurological disease analysis method, system and medium based on deep learning is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] One purpose of the present invention is to propose a method and system for intelligent analysis of neurological conditions that combines an improved fireworks algorithm with polarization water wave optimization. Based on a deep learning model construction mechanism, the present invention automatically extracts features and analyzes the EEG time series signals, and realizes the co-evolution of network structure and hyperparameters by introducing structural search optimization and polarization parameter adjustment mechanisms. The present invention describes in detail the complete process from EEG data preprocessing, structural parameter search, adaptive perturbation generation, polarization vector guidance, model training optimization to edge deployment reasoning. It has the advantages of high optimization efficiency, stable model performance, fast response speed, and adaptation to edge computing, and can significantly improve the intelligence level and clinical application value of the neurology auxiliary diagnosis system.

[0009] A neurological condition analysis method based on deep learning according to an embodiment of the present invention includes the following steps:

[0010] S1. Obtain EEG time series signal data of patients in the neurology department and perform preprocessing;

[0011] S2. Construct a temporal convolutional neural network model containing a set of structural parameters;

[0012] S3. Use the improved fireworks algorithm to perform global search optimization on the structural parameter set, generate and train structural parameter combinations based on the structural parameter dimensions, build a structural legitimacy verification module to eliminate structural parameter combinations that do not meet the network configuration rules, construct a multi-objective fitness function, score all structural parameter combinations, introduce an inverse fitness perturbation mechanism, select the structural parameter combination with the highest fitness, and generate a temporal convolutional neural network model for structural optimization;

[0013] S4. Set a set of hyperparameters, attach a polarization vector to each set of hyperparameters, and use an improved water wave optimization algorithm to propagate, reflect, and refraction hyperparameter solutions in each round of training. Adjust the perturbation direction and amplitude based on the polarization vector and evaluate based on a multi-objective fitness function. When the fitness drops after the perturbation, trigger the hyperparameter perturbation self-repair mechanism, and iterate to obtain the optimal hyperparameter configuration.

[0014] S5. Under the control of the optimal structural parameter combination and the optimal hyperparameter configuration, the structure-optimized temporal convolutional neural network model is trained to obtain the final analysis model;

[0015] S6. Input the preprocessed EEG time series signal data into the final analysis model to obtain the disease stage classification results and disease trend prediction results, generate a deployment version, load it to the edge computing device to perform the inference task, and present the disease stage classification results and disease trend prediction results.

[0016] Optionally, the EEG time series signal data in step S1 is a continuous voltage time series obtained by an acquisition channel, and the continuous voltage time series is preprocessed to generate standardized EEG data samples.

[0017] Optionally, the structural parameters of the temporal convolutional neural network model in step S2 include the convolution kernel size, the number of network layers and the expansion coefficient, the number of channels per layer, the residual connection method and the activation function type.

[0018] Optionally, the S3 specifically includes:

[0019] S31, setting a structural parameter search space, and combining each set of structural parameters into a structural vector;

[0020] S32, using the structure vector in the structure parameter search space as the initial explosion center, calculating 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 setting the adaptive explosion radius based on the fitness value;

[0021] S33, based on each initial explosion center and the corresponding explosion radius, generating new structural parameter combinations in the neighborhood, performing a structural validity check on the generated structural parameter combinations, and retaining only the structural parameter combinations that meet the neural network construction rules;

[0022] S34. Assign the structural parameter combinations that meet the neural network construction rules to the temporal convolutional neural network model respectively, train it on the training set, record the accuracy, number of training rounds and resource usage information on the validation set, calculate the comprehensive fitness score, and measure the overall performance of the structure;

[0023] S35. Introduce an inverse fitness perturbation strategy in the explosion centers with lower fitness rankings, and continuously update the explosion center set during the iteration process;

[0024] S36. Repeat steps S32 to S35 until the set maximum number of search rounds or the global fitness convergence standard 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 a structural optimization model.

[0025] Optionally, the inverse fitness perturbation strategy in S35 includes: in each round of structural parameter combination search iteration process, determining some structural parameter combinations in the current explosion center set whose fitness values are lower than the median as the low fitness explosion center set, and for each structural parameter combination in the low fitness explosion center set, increasing the number of sparks generated, expanding the perturbation range within the search neighborhood, and dynamically increasing the explosion radius value; based on the expanded explosion radius value, generating a new round of structural parameter combinations in the search space and incorporating them into the next round of explosion center candidate set.

[0026] Optionally, the S4 specifically includes:

[0027] S41, setting a hyperparameter set based on the structural optimization model, the hyperparameter set including the learning rate , regularization factor , the loss function weight coefficient of the disease stage classification task And the loss function weight coefficient of the disease trend prediction task , define the set of hyperparameters ;

[0028] S42. Construct a polarization vector for each set of hyperparameters , where each component corresponds to a perturbation direction preference and perturbation amplitude adjustment factor of a hyperparameter dimension;

[0029] S43. For each hyperparameter set Perform propagation perturbations and set the hyperparameters Each component in the equation is added with the corresponding disturbance Get the set of hyperparameters generated by perturbation;

[0030] S44: Perform boundary detection on the perturbed hyperparameter set. If a dimension exceeds the preset value range, it is reflected back to the legal range in a mirror manner. When the fitness value is improved compared to the previous round, a refraction operation is triggered to extend the perturbation path in the forward direction.

[0031] S45. When the fitness value of the current round decreases compared with the previous round, the learning rate difference between the current round and the previous round is calculated, a first-order penalty term is introduced, and the fitness function is numerically corrected;

[0032] S46. After each round of iteration, update the component of each dimension in the polarization vector;

[0033] S47. The component of each dimension in the updated polarization vector controls the perturbation amount of the corresponding hyperparameter set in the next round of propagation, and the optimal hyperparameter configuration is obtained through iterative updating.

[0034] A neurological disease analysis system based on deep learning according to an embodiment of the present invention includes:

[0035] The data processing module is used to collect neurological EEG time series signal data and perform filtering, denoising, normalization and window slicing processing to generate pre-processed EEG time series signal data;

[0036] The structural optimization module is used to construct the structural parameter space, use the improved fireworks algorithm to search for structural combinations, screen for legality, and evaluate fitness, and output a temporal convolutional neural network model for structural optimization;

[0037] The perturbation enhancement module is used to introduce inverse fitness perturbations for low-fitness structures, expand the search range, and enhance the local exploration capability during the structure optimization phase;

[0038] Polarization parameter adjustment module, used to set hyperparameters and assign polarization vectors, guide the perturbation direction during propagation, reflection, and refraction, and dynamically update the polarization component based on the fitness gradient;

[0039] The joint modeling module builds a model based on the optimal combination of structural parameters and optimal hyperparameters, jointly completing the tasks of disease stage classification and trend prediction to generate the final analysis model;

[0040] A lightweight deployment module that prunes and compresses the model to generate a deployable version for edge devices, enabling real-time inference and assisted diagnosis on-site.

[0041] The operation control module is used to execute model reasoning, output disease analysis results, and support integration with terminal systems and rapid clinical feedback.

[0042] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the processor is enabled to perform a neurological disease analysis method based on deep learning.

[0043] The beneficial effects of the present invention are:

[0044] (1) The present invention introduces an improved fireworks algorithm to adaptively search for neural network structural parameters. Combined with the structural dimension-aware explosion radius setting, the legitimacy screening mechanism, and the multi-objective fitness evaluation, the present invention realizes the automatic construction and fine optimization of the network structure, significantly reduces the dependence of structural design on manual experience, and improves the structural matching degree and generalization ability of the model for complex EEG data.

[0045] (2) During the training phase, the present invention uses an improved water wave optimization algorithm for hyperparameter adjustment. Combined with the polarization vector guidance mechanism, this method achieves dynamic control of the perturbation direction and intensity. By coupling the polarization component update with the propagation-reflection-refraction mechanism, the training process can focus on the fine-grained optimization of highly sensitive dimensions, enhancing the convergence stability and robustness of the parameter adjustment and avoiding the training instability problem caused by traditional fixed perturbation strategies.

[0046] (3) After the model is built, the present invention compresses the final model into a real-time inference version adapted to the on-site terminal through lightweight processing and edge deployment solutions. This version can be directly deployed on medical equipment or edge computing nodes. This design implements local inference feedback for disease stage classification and trend prediction tasks, and has the advantages of fast response speed, flexible deployment, and strong clinical adaptability, which improves the feasibility and practical value of the neurology disease analysis system in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The accompanying 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 of the present invention. In the accompanying drawings:

[0048] Figure 1 This is a flowchart of the neurological disease analysis method based on deep learning proposed by the present invention;

[0049] Figure 2 This is a flowchart of the neurological disease analysis method based on deep learning proposed by the present invention using the improved fireworks algorithm;

[0050] Figure 3 This is a flowchart of the hyperparameter adjustment of the polarized water wave optimization algorithm for the deep learning-based neurological disease analysis method proposed in the present invention. DETAILED DESCRIPTION

[0051] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0052] refer to Figure 1-3 , a neurological disease analysis method based on deep learning includes the following steps:

[0053] S1. Obtain EEG time series signal data of patients in the neurology department and perform preprocessing;

[0054] S2. Build a temporal convolutional neural network model containing a set of structural parameters, including the convolution kernel size, the number of network layers and expansion coefficients, the number of channels per layer, the residual connection method, and the activation function type;

[0055] S3. Use the improved fireworks algorithm to perform global search optimization on the structural parameter set, set the adaptive explosion radius based on the structural parameter dimension, generate structural parameter combinations, train each set of structural parameter combinations, build a structural legitimacy verification module to eliminate structural parameter combinations that do not meet the network configuration rules, build a multi-objective fitness function, score all structural parameter combinations, introduce an inverse fitness perturbation mechanism, select the structural parameter combination with the highest fitness, and generate a temporal convolutional neural network model for structural optimization;

[0056] S4. A time-series convolutional neural network model based on structural optimization sets a set of hyperparameters. A polarization vector is added to each set of hyperparameters. An improved water wave optimization algorithm is used to propagate, reflect, and refraction hyperparameter solutions in each round of training. The perturbation direction and amplitude are adjusted according to the polarization vector. Evaluation is performed based on a multi-objective fitness function. When the fitness decreases after the perturbation, a hyperparameter perturbation self-repair mechanism is triggered to iteratively obtain the optimal hyperparameter configuration.

[0057] S5. Under the control of the optimal structural parameter combination and the optimal hyperparameter configuration, the structurally optimized time-series convolutional neural network model is trained to obtain the final analysis model. Specifically, after obtaining the optimal structural parameter combination (including 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 learning rate, regularization factor, and weight coefficient of the multi-task loss function) searched by the polarized water wave optimization algorithm, the standardized EEG time-series signal data is input into the time-series convolutional neural network model with this structural configuration, and multiple rounds of training are performed according to the set hyperparameters to jointly complete the disease stage classification task and the disease trend prediction task, and finally obtain the analysis model that performs best under multiple objective evaluations such as accuracy, resource usage, and training stability, providing a model foundation for subsequent deployment and reasoning;

[0058] S6. Input the pre-processed EEG time series signal data into the final analysis model, obtain the disease stage classification results and disease trend prediction results, generate a deployment version, load it into the edge computing device to perform the reasoning task, and present the disease stage classification results and disease trend prediction results. Specifically including: after obtaining the final analysis model, input the pre-processed EEG time series signal data into the analysis model, perform forward reasoning under the optimal structural parameter combination and optimal hyperparameter configuration, and output the disease stage classification results and disease trend prediction results at the same time. Subsequently, the model is lightweight processed, including pruning, quantization and compression, to generate a deployment version adapted to the edge computing device, and loaded into the medical terminal, embedded device or other edge node to achieve local real-time reasoning and prediction result presentation of EEG data, and support rapid clinical judgment and auxiliary decision-making.

[0059] The present invention achieves efficient automated modeling of neurological disease analysis tasks by constructing a complete closed-loop process from EEG data preprocessing, structure optimization, hyperparameter adjustment to edge deployment reasoning. Compared with the existing method that relies on manual setting of structure and parameter adjustment, the present invention introduces swarm intelligence optimization and dynamic perturbation mechanism, which effectively improves model performance, training stability and clinical adaptability. At the same time, it supports lightweight deployment and can quickly complete disease stage identification and trend prediction at the edge terminal, with higher feasibility and practical application value.

[0060] In this embodiment, the EEG time series signal data in step S1 is a continuous voltage time series obtained by the acquisition channel. The continuous voltage time series is preprocessed to generate standardized EEG data samples. During preprocessing, the original EEG time series signal is first subjected to a 0.5-40Hz bandpass filter to remove low-frequency drift and high-frequency noise, and then artifact removal is performed to strip off the interference signal. The Z-score normalization method is then used to standardize the signal amplitude of each channel to ensure that the feature distribution between different samples is consistent. Finally, the continuous signal is time-window sliced 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, time series segments after time window slicing, clean signals after filtering and artifact removal, standardization results, and disease stage labels and trend labels used in the training phase.

[0061] In this embodiment, the structural parameters of the temporal convolutional neural network model in step S2 include the convolution kernel size, the number of network layers and the expansion coefficient, the number of channels per layer, the residual connection method and the activation function type.

[0062] In this embodiment, S3 specifically includes:

[0063] S31, setting a structural parameter search space, and combining each set of structural parameters into a structural vector;

[0064] S32. The structure vector in the structure parameter search space is used as the initial explosion center. For each initial explosion center, the fitness value is calculated based on the prediction performance of the corresponding time series convolutional neural network model on the validation set. The fitness value is mainly based on the accuracy of the validation set, while taking into account the training time and model complexity. The calculation formula is as follows:

[0065] ;

[0066] in, For the The fitness value of the initial explosion center, is the accuracy of the validation set, ranging from 0 to 1, is the normalized value of the training time or number of rounds, is the normalized value of the model parameters or computational complexity; a coefficient of 0.1 indicates a slight penalty on training efficiency and model complexity;

[0067] Dynamically set the adaptive explosion radius based on the fitness value, where the explosion radius is calculated as follows:

[0068] ;

[0069] in, is the explosion radius control coefficient, is the maximum fitness value, For the The fitness value of the initial explosion center, For the The fitness value of the initial explosion center, The total number of initial explosion centers that participate in the explosion radius in this round of fireworks algorithm. The explosion radius is automatically adjusted with the difference in fitness.

[0070] S33. Based on each initial explosion center and the corresponding explosion radius, a new structural parameter combination is generated in the neighborhood, and a structural legitimacy check is performed on the generated structural parameter combination, and only the structural parameter combination that meets the neural network construction rules is retained. The structural legitimacy check process specifically includes: checking whether parameters such as convolution kernel size, stride, expansion coefficient, etc. are compatible with the input size, ensuring that the number of channels between layers matches, the residual connection has dimensional consistency, the activation function type is compatible with the context module, and the overall network topology is reasonable and acyclic, with closed input and output. For structural combinations that do not meet the above rules, the system will automatically eliminate them and only retain structural parameter combinations of legal network structures that can be successfully constructed and trained;

[0071] S34. Assign the structural parameter combinations that meet the neural network construction rules to the temporal convolutional neural network model respectively, train on the training set, record the accuracy, number of training rounds and resource usage information on the validation set, and calculate the comprehensive fitness score. The comprehensive fitness score is calculated in the same way as the fitness value calculation method in S32 to measure the overall performance of the structure;

[0072] S35. Introduce an inverse fitness perturbation strategy in the explosion centers with lower fitness rankings, and continuously update the explosion center set during the iteration process;

[0073] S36. Repeat steps S32 to S35 until the set maximum number of search rounds or the global fitness convergence standard 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 a structural optimization model.

[0074] This paper utilizes an improved fireworks algorithm during the structural parameter optimization phase, and incorporates a dimensionally aware explosion radius, adaptive perturbation distribution, multi-objective fitness assessment, and a legitimacy screening mechanism to create an efficient structural search process tailored to neural network modeling tasks. Compared to traditional random search or fixed-structure network approaches, this approach significantly improves structural search efficiency and feasibility. Furthermore, through an inverse fitness perturbation strategy, it enhances local exploration of the structural space, effectively avoiding regression into local optima.

[0075] In this embodiment, the inverse fitness perturbation strategy in S35 includes: in each round of structural parameter combination search iteration process, determining some structural parameter combinations in the current explosion center set whose fitness values are lower than the median as the low fitness explosion center set, and for each structural parameter combination in the low fitness explosion center set, increasing the number of sparks generated, expanding the perturbation range within the search neighborhood, and dynamically increasing the explosion radius value. Based on the expanded explosion radius value, a new round of structural parameter combinations is generated in the search space and included in the next round of explosion center candidate set.

[0076] The inverse fitness perturbation mechanism proposed in this paper improves coverage of inefficient solution regions during the optimization process by proactively identifying structural solutions with low fitness scores and expanding their perturbation range. Compared to traditional uniform perturbation strategies, this method can expand search diversity in the early stages and accelerate convergence to high-quality solution regions in the later stages, balancing global exploration with local convergence, further improving the accuracy and robustness of structural parameter optimization.

[0077] In this embodiment, the S4 specifically includes:

[0078] S41, setting a hyperparameter set based on the structural optimization model, the hyperparameter set including the learning rate , regularization factor , the loss function weight coefficient of the disease stage classification task And the loss function weight coefficient of the disease trend prediction task , define the set of hyperparameters ;

[0079] S42. Construct a polarization vector for each set of hyperparameters , where each component corresponds to a perturbation direction preference and perturbation amplitude adjustment factor of a hyperparameter dimension;

[0080] S43. For each hyperparameter set Perform propagation perturbations and set the hyperparameters Each component in the equation is added with the corresponding disturbance Get the set of hyperparameters generated by perturbation and calculate the perturbation amount The formula is:

[0081] ;

[0082] in, is the global perturbation coefficient, is the first polarization vector The weight of the dimension, For the The hyperparameter of dimension is in the iteration The sub-goal value under the round, For the The hyperparameter of dimension is in the iteration The sub-goal value of the round, To prevent division by zero errors, a non-zero positive real constant is used. This perturbation formula combines the polarization vector with the trend of the sub-objective value to give each hyperparameter perturbation directionality and intensity control. Its practical significance lies in improving the convergence efficiency and stability of hyperparameter optimization in complex EEG modeling tasks through a guided and adaptive perturbation mechanism, focusing on the theme of "neurological disease analysis." This makes the model more accurate and robust in disease stage classification and trend prediction, thereby ensuring the reliability and real-time responsiveness of the final analysis model in clinical deployment.

[0083] S44. Perform boundary detection on the hyperparameter set generated by the perturbation, that is, determine 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 dimension exceeds the preset value range, it is reflected back to the legal range in a mirror manner. When the fitness value is improved compared with the previous round, the refraction operation is triggered to extend the perturbation path in the forward direction;

[0084] S45. When the fitness value of the current round decreases compared with the previous round, the learning rate difference between the current round and the previous round is calculated, a first-order penalty term is introduced, and the fitness function is numerically corrected;

[0085] S46. After each round of iteration, update the component of each dimension in the polarization vector. The update formula is as follows:

[0086] ;

[0087] in: For the In the iteration, the polarization vector The weight of the dimension, For the In the iteration, the polarization vector The weight of the dimension, is the polarization enhancement adjustment coefficient, is the hyperbolic tangent function, is the fitness function About The partial derivatives of the hyperparameters with respect to dimensionality;

[0088] S47. The component of each dimension in the updated polarization vector controls the perturbation amount of the corresponding hyperparameter set in the next round of propagation, and the optimal hyperparameter configuration is obtained through iterative updating.

[0089] This method introduces a polarized water wave optimization mechanism into the hyperparameter adjustment process. By guiding the perturbation direction through the polarization vector, this method combines propagation, reflection, and refraction operations with fitness gradient feedback to achieve dynamic control, significantly improving training stability and parameter adjustment efficiency. Compared with traditional static or manual parameter adjustment methods, this method is more adaptable and intelligent, demonstrating higher optimization quality and convergence speed when faced with multi-objective loss weight configurations and non-convex search spaces, and possesses good scalability.

[0090] A neurological disease analysis system based on deep learning according to an embodiment of the present invention includes:

[0091] The data processing module is used to collect neurological EEG time series signal data and perform filtering, denoising, normalization and window slicing processing to generate pre-processed EEG time series signal data;

[0092] The structural optimization module is used to construct the structural parameter space, use the improved fireworks algorithm to search for structural combinations, screen for legality, and evaluate fitness, and output a temporal convolutional neural network model for structural optimization;

[0093] The perturbation enhancement module is used to introduce inverse fitness perturbations for low-fitness structures, expand the search range, and enhance the local exploration capability during the structure optimization phase;

[0094] Polarization parameter adjustment module, used to set hyperparameters and assign polarization vectors, guide the perturbation direction during propagation, reflection, and refraction, and dynamically update the polarization component based on the fitness gradient;

[0095] The joint modeling module builds a model based on the optimal combination of structural parameters and optimal hyperparameters, jointly completing the tasks of disease stage classification and trend prediction to generate the final analysis model;

[0096] A lightweight deployment module that prunes and compresses the model to generate a deployable version for edge devices, enabling real-time inference and assisted diagnosis on-site.

[0097] The operation control module is used to execute model reasoning, output disease analysis results, and support integration with terminal systems and rapid clinical feedback.

[0098] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the processor is enabled to perform a neurological disease analysis method based on deep learning.

[0099] Example 1:

[0100] To verify the feasibility of this invention, it was applied to a project to upgrade the EEG monitoring system in the neurology outpatient department of a tertiary hospital. This project aims to enable dynamic condition monitoring and prediction for patients at high risk of epilepsy. The hospital aims to deploy an intelligent analysis system that processes patient EEG data in real time, automatically determines the stage of the disease, and predicts future trends. This system will assist physicians in developing more scientific treatment plans, alleviating the shortage of specialists and the burden of diagnosis.

[0101] During actual deployment, the system accesses the hospital's EEG acquisition system to obtain multi-channel EEG time-series data from approximately 75 outpatients or observation patients daily, with a sampling frequency of 256 Hz and a collection time of approximately 30 minutes per patient. The data processing module first performs 0.5–40 Hz bandpass filtering, artifact removal, Z-score normalization, and time window segmentation on the raw data, converting the continuous data into a short-term segment format suitable for model processing. During the structural optimization phase, an improved fireworks algorithm is used to automatically search for combinations of network layers, convolution kernel size, number of channels, and activation function configurations in the parameter space, avoiding the low efficiency and overfitting risks caused by repeated manual structural debugging in the past.

[0102] To improve 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 coefficients. By constructing a polarization vector for each hyperparameter dimension, the algorithm can precisely control the direction and intensity of the perturbation in each round, combining the propagation-reflection-refraction mechanism with fitness gradient feedback to achieve automatic parameter adjustment. The entire optimization process takes less than two hours per structure search on average, with parameter adjustment rounds limited to 60 times, improving efficiency by over 80% compared to traditional grid search methods.

[0103] After model training, the deployment version was lightweighted, including a pruning rate of 0.4, 8-bit quantization, and distillation compression. This reduced the model size from the original 47MB to 12.3MB, enabling direct deployment on the hospital's pre-configured embedded diagnostic terminals. This enabled a single device to complete disease stage and trend prediction analysis of EEG segments within 0.5 seconds. One month after the system went live, a total of 2,164 patient samples were received. Comparisons with manual judgment by the director of neurology demonstrated a 93.4% accuracy rate for disease stage identification and an 89.7% accuracy rate for trend prediction, significantly higher than the static convolutional model used before deployment.

[0104] Furthermore, in actual use, the average physician feedback time for this system was 0.38 seconds per reasoning round, significantly improving clinical response speed. In particular, in early intervention warnings for epileptic seizures, the model successfully detected 153 potential high-risk fluctuation signals, 89 of which were confirmed by physicians to be clinically significant, enabling early adjustments to treatment strategies.

[0105] The system has been deployed in selected partner medical institutions in Beijing, Tianjin, and Jiangsu. Deployment in hospitals does not rely on GPUs, but instead runs on low-power NPU terminals. The system analyzes 3,500–4,200 data items daily, boasting high system stability, an average daily failure rate of less than 0.2%, and minimal maintenance. User feedback indicates that this solution effectively reduces physician workload, improves diagnostic efficiency, and significantly enhances the ability to process complex neural signals.

[0106] Table 1: Comparative data between the method of the present invention and the existing method in the neurology disease analysis scenario

[0107]

[0108] Comparative experiments show that the Baseline-CNN approach, a traditional static convolutional network, performs relatively poorly in disease stage identification and trend prediction tasks, achieving only 84.2% and 74.6% accuracy, respectively. Due to its fixed structure and manual parameter tuning, overall performance improvement is limited, with inference latency reaching as high as 1.12 seconds, failing to meet the demands of real-time edge analysis. In contrast, a CNN model using grid search parameter tuning achieved slight improvements in accuracy and trend prediction, reaching 88.1% and 82.3%, respectively. However, due to its fixed structure, the search process took 5.5 hours, resulting in low efficiency. Furthermore, the model was insufficiently compressed, resulting in a deployed size of 42MB, and no significant improvement in inference speed on the field.

[0109] The system proposed in the present invention combines the improved fireworks algorithm with the polarized water wave optimization strategy. It not only has dynamic optimization capabilities at the structural design level, but also can achieve adaptive perturbation control based on the fitness gradient during the parameter adjustment process. Under the same task test set, the model of the present invention has an accuracy rate of 93.4% in identifying the stage of illness and an accuracy rate of 89.7% in trend prediction, which are significantly better than the control group, verifying the modeling effect of the structure-parameter collaborative optimization mechanism on complex EEG data. At the same time, since deployment constraints are taken into account 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.

[0110] To sum up, 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 wide promotion value.

[0111] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A neurological disease analysis method based on deep learning, characterized by: The steps include: S1. Obtain EEG time series signal data of patients in the neurology department and perform preprocessing; S2. Construct a temporal convolutional neural network model containing a set of structural parameters; S3. Use the improved fireworks algorithm to perform global search optimization on the structural parameter set, generate and train structural parameter combinations based on the structural parameter dimensions, build a structural legitimacy verification module to eliminate structural parameter combinations that do not meet the network configuration rules, construct a multi-objective fitness function, score all structural parameter combinations, introduce an inverse fitness perturbation mechanism, select the structural parameter combination with the highest fitness, and generate a temporal convolutional neural network model for structural optimization; S4. Set a set of hyperparameters, attach a polarization vector to each set of hyperparameters, and use an improved water wave optimization algorithm to propagate, reflect, and refraction hyperparameter solutions in each round of training. Adjust the perturbation direction and amplitude based on the polarization vector and evaluate based on a multi-objective fitness function. When the fitness drops after the perturbation, trigger the hyperparameter perturbation self-repair mechanism, and iterate to obtain the optimal hyperparameter configuration. S5. Under the control of the optimal structural parameter combination and the optimal hyperparameter configuration, the structure-optimized temporal convolutional neural network model is trained 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 results and disease trend prediction results, generate a deployment version, load it to the edge computing device to perform the inference task, and present the disease stage classification results and disease trend prediction results; The S3 specifically includes: S31, setting a structural parameter search space, and combining each set of structural parameters into a structural vector; S32, using the structure vector in the structure parameter search space as the initial explosion center, calculating 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 setting the adaptive explosion radius based on the fitness value; S33, based on each initial explosion center and the corresponding explosion radius, generating new structural parameter combinations in the neighborhood, performing a structural validity check on the generated structural parameter combinations, and retaining only the structural parameter combinations that meet the neural network construction rules; S34. Assign the structural parameter combinations that meet the neural network construction rules to the temporal convolutional neural network model respectively, train it on the training set, record the accuracy, number of training rounds and resource usage information on the validation set, calculate the comprehensive fitness score, and measure the overall performance of the structure; S35. Introduce an inverse fitness perturbation strategy in the explosion centers with lower fitness rankings, and continuously update the explosion center set during the iteration process; S36, repeating steps S32 to S35 until the set maximum number of search rounds or the global fitness convergence standard is reached, and finally determining the structural parameter combination with the highest fitness as the structural configuration of the temporal convolutional neural network model to generate a structural optimization model; The inverse fitness perturbation strategy in S35 includes: in each round of structural parameter combination search iteration process, determining some structural parameter combinations with fitness values lower than the median in the current explosion center set as the low fitness explosion center set, increasing the number of sparks generated for each structural parameter combination in the low fitness explosion center set, expanding the perturbation range within the search neighborhood, dynamically increasing the explosion radius value, and generating a new round of structural parameter combinations in the search space based on the expanded explosion radius value, and incorporating them into the next round of explosion center candidate set.

2. The neurological disease analysis method based on deep learning according to claim 1, characterized in that: The EEG time series signal data in step S1 is a continuous voltage time series obtained by the acquisition channel, and the continuous voltage time series is preprocessed to generate standardized EEG data samples.

3. The neurological disease analysis method based on deep learning according to claim 2, characterized in that: The structural parameters of the temporal convolutional neural network model in step S2 include the convolution kernel size, the number of network layers and the expansion coefficient, the number of channels per layer, the residual connection method and the activation function type.

4. The neurological disease analysis method based on deep learning according to claim 3, characterized in that: The S4 specifically includes: S41, setting a hyperparameter set based on the structural optimization model, the hyperparameter set including the learning rate , regularization factor , the loss function weight coefficient of the disease stage classification task And the loss function weight coefficient of the disease trend prediction task , define the set of hyperparameters ; S42. Construct a polarization vector for each set of hyperparameters , where each component corresponds to a perturbation direction preference and perturbation amplitude adjustment factor of a hyperparameter dimension; S43. For each hyperparameter set Perform propagation perturbations and set the hyperparameters Each component in the equation is added with the corresponding disturbance Get the set of hyperparameters generated by perturbation; S44: Perform boundary detection on the perturbed hyperparameter set. If a dimension exceeds the preset value range, it is reflected back to the legal range in a mirror manner. When the fitness value is improved compared to the previous round, a refraction operation is triggered to extend the perturbation path in the forward direction. S45. When the fitness value of the current round decreases compared with the previous round, the learning rate difference between the current round and the previous round is calculated, a first-order penalty term is introduced, and the fitness function is numerically corrected; S46. After each round of iteration, update the component of each dimension in the polarization vector; S47. The component of each dimension in the updated polarization vector controls the perturbation amount of the corresponding hyperparameter set in the next round of propagation, and the optimal hyperparameter configuration is obtained through iterative updating.

5. 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 4, characterized in that: include: The data processing module is used to collect neurological EEG time series signal data and perform filtering, denoising, normalization and window slicing processing to generate pre-processed EEG time series signal data; The structural optimization module is used to construct the structural parameter space, use the improved fireworks algorithm to search for structural combinations, screen for legality, and evaluate fitness, and output a temporal convolutional neural network model for structural optimization; The perturbation enhancement module is used to introduce inverse fitness perturbations for low-fitness structures, expand the search range, and enhance the local exploration capability during the structure optimization phase; Polarization parameter adjustment module, used to set hyperparameters and assign polarization vectors, guide the perturbation direction during propagation, reflection, and refraction, and dynamically update the polarization component based on the fitness gradient; The joint modeling module builds a model based on the optimal combination of structural parameters and optimal hyperparameters, jointly completing the tasks of disease stage classification and trend prediction to generate the final analysis model; A lightweight deployment module that prunes and compresses the model to generate a deployable version for edge devices, enabling real-time inference and assisted diagnosis on-site. The operation control module is used to execute model reasoning, output disease analysis results, and support integration with terminal systems and rapid clinical feedback.

6. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, enables the processor to execute the neurological disease analysis method based on deep learning according to any one of claims 1 to 4.

Images