A method and system for epilepsy monitoring and early warning

Abstract

This invention discloses a method and system for epilepsy monitoring and early warning, relating to the field of medical signal processing. It solves the problems of high false alarms and false negatives, and low energy efficiency in existing monitoring technologies due to physiological artifact interference, individual differences, and continuous high power consumption. The key technical points are: first, achieving deep artifact removal of the signal through multi-angle collaborative filtering; second, constructing a personalized lightweight prediction model based on historical data similarity matching and transfer learning; and finally, employing a dynamic signal filtering and graded power adjustment mechanism. This achieves the technical effect of significantly improving the accuracy of early warning for patients with specific lesions while greatly reducing system redundancy and energy consumption.

Description

Technical Field

[0001] This invention relates to the field of medical signal processing, and more specifically, to a method and system for epilepsy monitoring and early warning. Background Technology

[0002] Epilepsy is a common chronic neurological disorder, affecting approximately 0.6% of the global population, with children being 10-15 times more likely to develop the condition than adults. The unpredictability and randomness of epileptic seizures are its most troubling characteristics, posing a serious threat to the physical and mental health of affected children and potentially leading to accidental injury or even death. Currently, video electroencephalography (EEG) monitoring, widely used in hospitals, is considered the "gold standard" for epilepsy diagnosis. It records brain activity and behavioral manifestations in real time, correlates specific electrical signals during seizures with clinical behavior, and establishes a universal baseline threshold model for risk assessment, enabling real-time monitoring and early warning of epilepsy. However, the following problems still exist:

[0003] In everyday environments, a patient's normal physiological activities can produce artifacts that overlay EEG signals, easily confusing them with epileptiform discharges and leading to false alarms and missed alarms. Furthermore, current technologies use fixed threshold models for epilepsy prediction, but for patients with other brain lesions (such as a history of head injury), their background EEG rhythms differ significantly from those of ordinary epilepsy patients. This results in inconsistent baseline values ​​for abnormal discharges causing seizures, leading to low accuracy in early warning. Moreover, current technologies primarily rely on real-time monitoring. During long-term monitoring, even for patients with low incidence rates, the system continues to record and process all EEG data, generating a large amount of repetitive and invalid information. This increases redundant computation and energy consumption, leading to higher system costs.

[0004] Therefore, how to research and design an epilepsy monitoring and early warning method and system that can overcome the above-mentioned defects is an urgent problem that we need to solve. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide an epilepsy monitoring and early warning method and system. This system achieves deep purification of EEG signals through a collaborative filtering network; constructs a personalized prediction model based on historical data similarity matching and weighted adjustment to solve the baseline value problem caused by individual differences; and achieves on-demand allocation of computing resources and reduces energy consumption through signal filtering and graded power adjustment. Finally, it triggers graded early warnings and generates nursing countermeasures based on real-time risk indices, realizing a leap from passive monitoring to proactive intervention and providing a reliable solution for family monitoring of children with epilepsy.

[0006] The above-mentioned technical objective of the present invention is achieved through the following technical solution:

[0007] Firstly, a method for epilepsy monitoring and early warning is provided, including the following steps:

[0008] S1: Acquire real-time multimodal signals including EEG, ECG, EMG and behavior, perform multi-angle collaborative filtering to remove physiological artifacts, and obtain pure EEG signals and multimodal auxiliary features;

[0009] S2: Based on the large-scale historical epilepsy patient data in the hospital database and the patient's own medical history data, feature extraction and similarity matching are performed to obtain a reference sample set that is most similar to the target patient's features;

[0010] S3: Based on the reference sample set and the target patient's own medical history data, transfer learning and model fine-tuning are performed on the basic model to obtain the epilepsy prediction model corresponding to the target patient.

[0011] S4: Input the pure EEG signal and the multimodal auxiliary features into the epilepsy prediction model, perform feature intensity threshold screening, adjust the calculation power according to the screening results, perform seizure risk probability calculation and symptom type analysis, and obtain real-time risk index and symptom type;

[0012] S5: Trigger graded early warning based on the real-time risk index and the symptom type, and generate targeted nursing and intervention suggestions in conjunction with the patient's medical history data.

[0013] Furthermore, the multi-angle collaborative filtering is implemented based on multi-modal data correlation decomposition:

[0014] Wavelet transform is used to decompose the multimodal signal from the time domain to the transform domain. The correlation between the EEG signal and signals such as ECG and EMG is calculated by the transfer entropy of symbolic processing, so as to accurately identify ECG artifacts and EMG artifacts.

[0015] Behavioral features are extracted using behavioral signals to help determine the time period and pattern of motion artifacts. Collaborative filtering is performed in the transform domain to adaptively attenuate the artifact portion.

[0016] The filtered EEG signal is separated using a blind source separation algorithm to obtain the purified EEG signal.

[0017] Furthermore, the similarity matching includes:

[0018] A multidimensional feature vector containing time-domain, frequency-domain, and nonlinear dynamic features is constructed based on the patient's own medical history data. The multidimensional feature vector is then compared with the multidimensional feature vectors of historical patients in the hospital database using cosine similarity calculation to form the reference sample set. The contribution of each sample is quantified based on the similarity measure, and the weight of the sample in the model fine-tuning is dynamically allocated.

[0019] Furthermore, the epilepsy prediction model is generated using a hierarchical adjustment strategy:

[0020] A general epilepsy prediction model pre-trained based on large-scale historical epilepsy patient data is obtained as the base model;

[0021] The parameters of the base model are loaded into the new model as initial values, and the base model is fine-tuned with weights using the reference sample set to obtain the epilepsy prediction model.

[0022] The fine-tuning process employs a hierarchical learning strategy, where the parameters at the bottom layer of the model used to extract general features are updated slightly with an update magnitude less than a first threshold, while the parameters at the top layer of the model used for specific decision-making are adjusted significantly based on the weighted sample set with an update magnitude greater than a second threshold, in order to quickly adapt to the physiological characteristics of the target patient; the first threshold is less than or equal to the second threshold.

[0023] Furthermore, the feature intensity threshold screening is based on the energy and rate of change of the multimodal auxiliary features, and is used to preliminarily determine whether the signal is worth conducting in-depth analysis.

[0024] Furthermore, the graded adjustment of calculation power adopts a dynamic power management strategy, which dynamically switches the power level based on the characteristic strength of the signal. When the characteristic strength is lower than the first threshold, it maintains low power consumption and only records data. When it is between the first and second thresholds, it enables medium power consumption to calculate the risk probability. When it is higher than the second threshold, it switches to full power for complete analysis and early warning.

[0025] Furthermore, the real-time risk index and symptom type are generated through multimodal information fusion:

[0026] The real-time risk index is dynamically weighted by integrating symptom activity index, short-term trend and model confidence.

[0027] The symptom type is obtained by matching the current multimodal auxiliary features with a patient-specific symptom template library.

[0028] Furthermore, the intervention recommendations include instructing family members to maintain observation and record during low-risk periods, to perform anti-asphyxiation positioning and remove hazardous materials during medium-risk periods, and to clear nasal and oral secretions, time the event, and prepare for emergency treatment during high-risk periods.

[0029] Furthermore, it also includes model self-updating and optimization: periodically or after each warning event, newly generated effective data is incorporated into the target patient's own medical history data, and steps S2 and S3 are re-executed to update the reference sample set and optimize the epilepsy prediction model.

[0030] In a second aspect, an epilepsy monitoring and early warning system is provided, the system being used to implement an epilepsy monitoring and early warning method as described in any one of the first aspects, comprising:

[0031] The collaborative filtering module, including a three-level filtering network, is used to perform multi-angle collaborative filtering based on real-time multimodal signals to remove physiological artifacts and obtain pure EEG signals and multimodal auxiliary features.

[0032] The sample matching module includes a data storage unit and a feature matching unit, which are used to extract features and perform similarity matching based on large-scale historical epilepsy patient data in the hospital database and the patient's own medical history data, to obtain a reference sample set that is most similar to the target patient's features.

[0033] The model generation module includes a data storage unit, a basic model unit, and a transfer adjustment unit. It is used to perform transfer learning and model fine-tuning on the basic model based on the reference sample set and the target patient's own medical history data to obtain an epilepsy prediction model tailored for the target patient.

[0034] The predictive analysis module includes a feature intensity filter, a hierarchical power network, and a risk prediction module. It is used to perform feature intensity threshold filtering based on the pure EEG signal and the multimodal auxiliary features, and to adjust the calculation power hierarchically to calculate the probability of attack risk and analyze the symptom type, so as to obtain the real-time risk index and symptom type.

[0035] The graded early warning module includes an alarm unit and a strategy matcher. It triggers graded early warnings based on the real-time risk index and the symptom type, and generates targeted nursing and intervention suggestions by combining patient medical history data.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] 1. This invention achieves deep purification of EEG signals through a collaborative filtering network; it constructs a personalized prediction model based on historical data similarity matching and weighted adjustment to solve the baseline value problem caused by individual differences; through signal filtering and graded power adjustment, it realizes on-demand allocation of computing resources, reduces energy consumption, and finally triggers graded early warning and generates nursing countermeasures based on real-time risk index, realizing the leap from passive monitoring to active intervention and providing a reliable solution for family monitoring of children with epilepsy.

[0038] 2. This invention decomposes multimodal signals into the transform domain using wavelet transform, and accurately identifies physiological artifacts such as electrocardiograms and electromyograms by combining transfer entropy calculation, and performs adaptive attenuation. Finally, it reconstructs a pure EEG signal with a high signal-to-noise ratio through blind source separation. This significantly reduces false alarms and false negatives caused by signal confusion, provides a reliable data foundation for subsequent analysis, and improves the robustness of the system in real-world environments.

[0039] 3. This invention selects the most similar cases from historical data using cosine similarity and dynamically allocates weights by quantifying sample contribution using Shapley value; combined with transfer learning and hierarchical fine-tuning strategies, the model retains the ability to extract general features at the bottom layer and quickly adapts to patient-specific patterns at the top layer; it breaks through the limitations of traditional fixed threshold models, improves the accuracy of early warning, and avoids overfitting problems caused by data scarcity.

[0040] 4. This invention monitors the energy changes of multimodal auxiliary features in real time through a low-power program. It triggers hierarchical calculations only when the feature intensity exceeds a dynamic threshold, thereby reducing the average power consumption and data storage of the system, significantly extending the device's battery life, and providing feasibility for long-term home monitoring. Attached Figure Description

[0041] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0042] Figure 1 This is a flowchart from Embodiment 1 of the present invention;

[0043] Figure 2 This is a schematic diagram of wavelet decomposition in Embodiment 1 of the present invention;

[0044] Figure 3 This is a schematic diagram of the transfer learning fine-tuning principle in Embodiment 1 of the present invention;

[0045] Figure 4 This is a structural diagram of the epilepsy prediction model in Embodiment 1 of the present invention;

[0046] Figure 5 This is a comparison chart of the accuracy of the epilepsy prediction model in Embodiment 1 of the present invention;

[0047] Figure 6 This is a graph of the loss function of the training set of the epilepsy prediction model in Embodiment 1 of the present invention;

[0048] Figure 7 This is a system block diagram in Embodiment 2 of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of this invention are only for explaining this invention and are not intended to limit this invention.

[0050] Example 1: A method for epilepsy monitoring and early warning, such as Figure 1 As shown, it includes the following steps:

[0051] S1: Acquire real-time multimodal signals including EEG, ECG, EMG and behavior, perform multi-angle collaborative filtering to remove physiological artifacts, and obtain pure EEG signals and multimodal auxiliary features;

[0052] S2: Based on the large-scale historical epilepsy patient data in the hospital database and the patient's own medical history data, feature extraction and similarity matching are performed to obtain a reference sample set that is most similar to the target patient's features;

[0053] S3: Based on the reference sample set and the target patient's own medical history data, transfer learning and model fine-tuning are performed on the basic model to obtain the epilepsy prediction model corresponding to the target patient.

[0054] S4: Input pure EEG signals and multimodal auxiliary features into the epilepsy prediction model, perform feature intensity threshold screening, adjust the calculation power according to the screening results, perform seizure risk probability calculation and symptom type analysis, and obtain real-time risk index and symptom type;

[0055] S5: Triggers graded early warnings based on real-time risk index and symptom type, and generates targeted nursing and intervention recommendations based on patient medical history data.

[0056] In step S1, wavelet transform and blind source separation techniques are used to perform collaborative analysis and targeted attenuation of multimodal signals, thereby achieving precise removal of physiological artifacts and obtaining a pure EEG signal with a high signal-to-noise ratio.

[0057] Data such as electroencephalogram (EEG), electrocardiogram (ECG), electromyography (EMG), and behavioral data are acquired through wearable devices such as headbands or patches, as shown in Table 1. Specifically, this includes multi-channel EEG signals, epilepsy-specific waves, surface electromyography signals, heart rate, heart rate variability, R-wave peak value, muscle activity energy, explosive electromyography signals, triaxial acceleration, angular velocity, amplitude of motion, and body position.

[0058] Table 1

[0059]

[0060] To ensure that the signals are aligned in time, the sampling clocks of all signals must be synchronized, or precise interpolation alignment must be performed using timestamps.

[0061] Artifacts are generated by the patient's normal physiological activities. However, existing filtering methods, such as amplitude limiting, median, and arithmetic mean filtering, mainly rely on simple rules and fixed parameters in the time or frequency domain. The algorithms are simple but have poor adaptability, making it impossible to remove artifacts and seriously affecting the accuracy of monitoring. Therefore, this invention adopts multi-angle collaborative filtering to attenuate physiological artifacts in a targeted manner and obtain a pure EEG signal with a high signal-to-noise ratio.

[0062] In some examples, multi-angle collaborative filtering is implemented based on correlation and behavioral analysis of transfer entropy, and targeted elimination is performed in the transform domain:

[0063] After synchronizing the signals, such as Figure 2 As shown, wavelet transform (such as Symlets wavelet) is used to decompose multimodal signals from the time domain to the transform domain.

[0064] Identification of ECG or EMG artifacts: In the transform domain, the transfer entropy between EEG signals and ECG / EMG data is calculated. Transfer entropy is an information theory-based metric that quantifies nonlinear causal relationships and directional information flow between signals; for example, calculating the transfer entropy from ECG signals to EEG signals can clearly assess the degree of interference of ECG activity on EEG signals.

[0065] A pre-defined transfer entropy threshold is used. When the transfer entropy of a component in the EEG signal exceeds this threshold with that of the ECG or EMG signal, the component is identified as a corresponding ECG or EMG artifact. This method is better able to capture complex nonlinear interferences and achieves higher recognition accuracy than existing linear correlation analysis.

[0066] Auxiliary analysis of behavioral signals: The system performs real-time analysis of behavioral data and extracts features to identify periods of significant movement in patients, such as turning over, walking, and chewing. The EEG signals corresponding to these periods are marked as low-confidence motion artifacts.

[0067] Collaborative filtering: In the transform domain, components identified as ECG or EMG artifacts undergo adaptive attenuation. This means that the weight of the component is reduced proportionally to its transfer entropy, rather than being set to zero, to avoid signal distortion. For low-confidence periods marked as motion artifacts, the system can adopt conservative strategies, such as reducing the weight of data from those periods in subsequent analysis or marking them for special processing later.

[0068] Blind source separation and clean signal reconstruction: After the targeted attenuation described above, the processed signal is subjected to blind source separation using the Fast Independent Component Analysis (FastICA) algorithm. This algorithm decomposes the multi-channel EEG signal into several statistically independent components, identifies and removes the remaining artifact components based on the spatiotemporal distribution characteristics of these components, and then linearly combines the remaining components to reconstruct a clean EEG signal with a high signal-to-noise ratio for subsequent epilepsy prediction analysis.

[0069] The artifact removal system of this invention is a multi-level, multi-angle intelligent filtering system. Its core lies in utilizing the complementarity of multimodal information to achieve accurate identification and effective suppression of physiological artifacts. It can significantly reduce false alarms and false negatives caused by physiological activity artifacts, providing a highly reliable input signal for subsequent epilepsy prediction, and is the foundation for the system to achieve high-accuracy early warning.

[0070] In step S2, multidimensional features of patients are extracted, and similar samples are matched from historical data using cosine similarity and Shapley value to quantify their contribution. Weights are dynamically assigned to construct a personalized reference benchmark.

[0071] In epilepsy monitoring, considering the individualization of abnormal discharge benchmark values ​​due to differences in brain lesions among patients, and the inability of existing technologies to adapt to this difference using a uniform threshold model, this invention selects a reference sample set most similar to the target patient from historical data to provide personalized benchmark references for subsequent analysis and improve the accuracy of early warning.

[0072] In some examples, the reference sample set of this invention is obtained through feature extraction and similarity matching:

[0073] Feature extraction source: Extract time-domain, frequency-domain, and nonlinear features such as approximate entropy, power spectral density, and Hjorth parameter from the target patient's own medical history data to form a multidimensional feature vector.

[0074] Similarity Measurement: The multidimensional feature vector of the target patient obtained above is compared with the feature vectors of a large number of historical epilepsy patients in the hospital database using cosine similarity calculation. Since cosine similarity focuses on the similarity of feature vectors in direction, and is not sensitive to the absolute value of the vectors, it is very suitable for comparing the similarities and differences in EEG patterns between different patients, rather than the difference in absolute values.

[0075] Efficient Nearest Neighbor Search: To address the computational efficiency issue under massive data, an efficient approximate nearest neighbor search library is adopted to quickly find the K most similar samples to the target patient's feature vector from the historical database, forming an initial reference sample set, where K is not less than 50.

[0076] Dynamic weight allocation: After initial matching, not all similar samples have the same reference value. This invention introduces the Shapley value method from cooperative game theory to accurately quantify the contribution of each similar sample:

[0077] By calculating the marginal contribution of each similar sample to the model's predictive accuracy when added to different subsets, its importance is fairly assessed; samples with greater contributions are assigned higher weights in subsequent model fine-tuning; ultimately, a weighted, personalized set of reference samples is obtained. This ensures that subsequent modeling can draw on a wide range of sources, prioritizing the most relevant and valuable historical experience for the target patients, rather than simply averaging samples.

[0078] This invention extracts unique multidimensional feature vectors from each patient, performs cosine similarity calculation and matching, and adjusts weights through contribution quantification, establishing a personalized analytical benchmark for each patient instead of using a fixed threshold. This allows the early warning model to adapt to the specificities of patients with different etiologies and physiological states, significantly improving the early warning accuracy for special populations.

[0079] In step S3, a transfer learning and hierarchical learning rate strategy is adopted to fine-tune the general model using a weighted sample set, so that the bottom layer of the model retains general knowledge while the top layer quickly adapts to patient-specific patterns.

[0080] Considering the specificity of patients with different etiologies and physiological states, existing technologies that train models from scratch or use general models are difficult to adapt to individual differences. Furthermore, using only patient data can lead to overfitting or non-convergence due to data scarcity. Therefore, this invention utilizes a pre-trained basic model for fine-tuning to adapt to the target patient, achieving efficient and accurate personalized modeling.

[0081] In some examples, individualized epilepsy prediction models are obtained through transfer learning and hierarchical fine-tuning of the base model.

[0082] Fine-tuning is a commonly used technique in transfer learning, and its principle is as follows: Figure 3 As shown, a source model is trained on the source dataset, and a new target model is created. All model designs and parameters of the source model except for the output layer are copied. An output layer with an output size equal to the number of categories in the target dataset is added to the target model, and the model parameters of this layer are randomly initialized. The target model is then trained on the target dataset.

[0083] The model fine-tuning of this invention is as follows: First, a pre-trained general base model is obtained: using large-scale historical epilepsy patient data from a hospital database, a general epilepsy prediction model is pre-trained as the base model. This model typically employs a deep learning architecture that can effectively extract spatiotemporal features, such as a hybrid model that combines a one-dimensional convolutional neural network (1D-CNN) for local feature extraction and a long short-term memory network (LSTM) for temporal dependency modeling.

[0084] This basic model has learned general feature patterns of pre-seizure EEG and multimodal signals abstracted from massive amounts of data, and has powerful feature representation capabilities; this avoids training the model from scratch for each new patient and solves the problems of poor initial model performance and high training resource consumption.

[0085] Then, a training set is generated based on the weighted reference sample set and patient data: the parameters of the base model are loaded into a new model instance as initial values. Subsequently, the weighted reference sample set finally generated in step S2 is merged with the target patient's own medical history data to form a personalized training set.

[0086] A hierarchical learning rate strategy is used for model fine-tuning:

[0087] Fine-tuning of low-level parameters using a low learning rate: The lower layers of the model (such as convolutional layers in a CNN) are primarily responsible for extracting basic, general features (such as EEG edges and waveforms). These parameters are fine-tuned using a low learning rate, allowing for only small updates, such as updates less than a first threshold. The aim is to maximize the preservation of the general feature representation capabilities learned from massive datasets when adapting to new patients, preventing feature forgetting. Low-level parameters (such as convolutional kernels) need to remain stable to avoid overfitting the target patient data. Small update increments limit parameter fluctuations; for example, the first threshold can be set to 8%, 10%, or 12% of the initial learning rate (0.0001~0.01). The optimal value is selected by validating the loss convergence speed using small batches of data.

[0088] High learning rates are used for fine-tuning top-level parameters: The top layer of the model (such as the fully connected classification layer) is mainly responsible for mapping high-level features to specific prediction results (such as real-time risk index and symptom type). A high learning rate is used to fine-tune these parameters, allowing for significant adjustments, such as large adjustments with update increments greater than a second threshold, where the first threshold is less than or equal to the second threshold. Top-level parameters (such as the fully connected layer) need to quickly adapt to the patient's physiological characteristics; larger update increments can accelerate convergence. For example, the second threshold can be set to 30%, 40%, or 50%.

[0089] The first and second thresholds can be determined based on cross-validation, such as dividing the weighted sample set into a training set (80%) and a validation set (20%). Different threshold combinations are tested within a preset range, and the validation set accuracy is recorded. The threshold combination with the highest validation accuracy is determined as the final threshold.

[0090] The layer structure of the epilepsy prediction model of this invention is as follows: Figure 4 As shown, the system includes an input layer that receives multimodal temporal data; a one-dimensional convolutional neural network (CNN) module that extracts meaningful local patterns from the multi-channel signals at each instant; a sequence flattening and transformation layer that rearranges and deforms the feature maps output by the CNN module to form a prescribed format; a long short-term memory network module that is used to understand the evolution of features over time and identify the gradual change patterns of features before an epileptic seizure; and a fully connected classification layer that maps the high-level spatiotemporal features learned by the LSTM module onto the final prediction output to complete the decision.

[0091] The accuracy of the epilepsy prediction model of this invention is compared to, for example... Figure 5 As shown, the accuracy trends of the proposed model and traditional models on the validation set are illustrated under different training cycles. The traditional CNN model starts with a low accuracy, converges slowly, and quickly reaches a performance bottleneck of approximately 88%, indicating limited learning ability. The general CNN-LSTM model, as the foundational model of this invention, outperforms the traditional CNN in both starting point and convergence speed, eventually stabilizing at approximately 93.3%, demonstrating the advantages of the hybrid architecture. The proposed model (personalized model) uses transfer learning, resulting in a very high starting point (90%), thanks to the knowledge gained from pre-training the general model. After fine-tuning, its accuracy steadily improves, eventually stabilizing at a higher level of approximately 96.3%, significantly outperforming the others.

[0092] After multiple training iterations, the loss function diagram of the epilepsy prediction model of this invention is shown below. Figure 6 As shown, the loss value decreases during training. The initial loss value does not start from an extremely high random value, indirectly reflecting that the model is fine-tuned based on a pre-trained general-purpose base model, rather than training from scratch. Therefore, the starting point is lower, improving training efficiency. The curve decreases smoothly without drastic fluctuations, proving that the training strategies such as hierarchical learning rates are effective and ensure the stability of the training process. The loss value eventually converges to a very low level, indicating that the model architecture can effectively extract key features from multimodal physiological data and accurately fit the complex patterns of epileptic seizures.

[0093] This invention generates a personalized epilepsy prediction model that combines high accuracy and high adaptability through transfer learning and hierarchical fine-tuning. It possesses powerful general knowledge and is deeply adapted to the individual characteristics of patients. It reduces the risk of overfitting to small-scale data, ensures the robustness of the model, and achieves accurate prediction.

[0094] In step S4, auxiliary features are monitored by a low-power sentinel program, and hierarchical computing power is triggered by a dynamic threshold to realize on-demand allocation of computing resources driven by feature intensity, which significantly reduces average energy consumption.

[0095] In real clinical settings, about half of epileptic seizures go undetected, while long-term monitoring generates a large amount of normal data that does not require manual verification. Existing technologies cannot optimize resource allocation by continuously processing all data. Therefore, this invention uses power grading calculation to allocate computing power on demand, significantly reducing the average energy consumption of the system and the power consumption of the equipment.

[0096] In some examples, power grading calculations are triggered by filtering based on feature strength determined by multimodal auxiliary features, specifically:

[0097] Preliminary signal screening based on auxiliary features: Instead of constantly running complex personalized prediction models, this invention uses a computationally inexpensive "sentinel" program that runs continuously in the background. This program monitors in real time the multimodal auxiliary features (such as heart rate values ​​and the energy of electromyography signals) extracted from signals such as electrocardiograms and electromyograms, and calculates their real-time energy intensity and rate of change.

[0098] Dynamic threshold judgment: The system presets dynamic thresholds. It will only be triggered when the combination of the energy and rate of change of these auxiliary features exceeds the preset threshold, and then it will be activated from the dormant state and a complete personalized prediction model will be loaded for in-depth analysis.

[0099] Hierarchical power adjustment based on model output: Once triggered, the loaded personalized prediction model performs in-depth analysis of the pure EEG signal and multimodal auxiliary features, and outputs a feature intensity normalization value between 0 and 1. This value reflects the probability of the current signal's role in epilepsy prediction.

[0100] Level 3 Power Management Mode: Based on this characteristic strength value, the system dynamically adjusts the operating frequency and voltage of computing cores such as CPU / GPU, i.e., enters a tiered computing mode.

[0101] Low-power mode (characteristic strength < 0.3): This indicates a very low level of signal abnormality, possibly due to isolated interference or false triggering. The system will save the data for this period but will not perform intensive calculations, quickly reverting to "sentinel" monitoring mode.

[0102] Medium computing power mode (0.3 ≤ feature strength ≤ 0.7): This indicates a suspicious anomaly in the signal, but with high uncertainty. The system will activate medium computing power, for example, by increasing the analysis frequency of the model and performing more intensive risk probability calculations to confirm the trend.

[0103] Full power mode (feature intensity > 0.7): This indicates that the signal exhibits highly epileptic seizure characteristics. The system will activate the maximum performance mode to perform a high-frequency, highly complex full analysis and prepare to initiate the early warning process.

[0104] In some examples, the hierarchical power management effect of the present invention is shown in Table 2. Table 2 compares the performance indicators of the non-hierarchical calculation mode and the hierarchical power calculation mode of the present invention based on multiple time dimensions, and simulates the energy consumption and data performance of the system at different runtimes.

[0105] Table 2

[0106]

[0107] This invention utilizes low-computational-cost auxiliary features as triggers to achieve "no calculation without events," greatly reducing the average power consumption of the system. The graded power adjustment strategy enables on-demand allocation of computing resources, resulting in extremely low system energy consumption during periods when the patient is not experiencing an attack or between attacks. This solves the problem of high energy consumption during long-term monitoring and is particularly beneficial for the long-term battery life of portable devices.

[0108] In some examples, the calculation of the real-time risk index is not a simple linear mapping, but a multi-factor dynamic weighted fusion process. It comprehensively considers the current symptom activity index, short-term trends, and model confidence, and its calculation formula is as follows:

[0109] ;

[0110] in, This is a real-time risk index; The symptom activity index represents the raw value that the model directly outputs after analyzing the signal at the current moment, reflecting the probability of an attack. The short-term trend indicates the slope and acceleration of the symptom activity index in recent times, which is used to determine whether the risk is in an upward, plateau, or downward channel. Model confidence represents the certainty of this prediction and can usually be measured by the entropy of the classification probability or the distribution of activation values ​​in the output layer. This is the trend weighting coefficient. This is the confidence weighting coefficient.

[0111] Symptom types are determined by comparing real-time signals with a personalized symptom template library:

[0112] By using a machine learning classifier to match the current multimodal auxiliary features with a patient-specific symptom template library, a specific symptom type description is output.

[0113] In step S5, based on the real-time risk index and specific symptom type output by the model, a graded early warning is triggered and personalized nursing countermeasures are generated, realizing the upgrade from alarm to clinical decision support.

[0114] Considering that different risk levels in epilepsy monitoring require different response measures, and that existing technologies using a uniform alarm method are prone to information overload or neglect, this invention generates graded early warnings and personalized intervention suggestions by quantifying real-time risk indices and matching symptom types, thereby achieving a leap from simple alarms to precise decision support and improving the effectiveness of monitoring.

[0115] First, an S-shaped function is used to map the real-time risk index to a real-time risk value ranging from 0 to 100.

[0116] When the risk value is less than 30, it is considered low risk. Provide gentle reminders to the caregiver, and continue to observe and record.

[0117] When the risk value is between 30 and 70, it is considered medium risk. Immediately send a flashing alarm to medical staff and caregivers, along with instructions on anti-asphyxiation positioning and removal of hazardous materials.

[0118] When the risk value is greater than 70, it is considered high risk. An alarm will be sounded immediately to medical staff and caregivers, and a plan will be generated to guide the clearing of secretions from the mouth and nose, time the event, and prepare for emergency treatment.

[0119] In some examples, model self-updating and optimization are also included: periodically or after each warning event, newly generated effective data are incorporated into the target patient's own medical history data, and steps S2 and S3 are re-executed to update the reference sample set and optimize the epilepsy prediction model.

[0120] The multi-level early warning mechanism of this invention ensures differentiated responses to different levels of risk, reducing unnecessary disruptions to daily life while ensuring sufficient attention is paid in emergency situations, thus optimizing the allocation of monitoring resources.

[0121] Example 2: An epilepsy monitoring and early warning system, used to implement an epilepsy monitoring and early warning method as described in Example 1, such as... Figure 7 As shown, it includes a collaborative filtering module, a sample matching module, a model generation module, a predictive analysis module, and a hierarchical early warning module, wherein:

[0122] The collaborative filtering module includes a three-level filtering network for multi-angle collaborative filtering based on real-time multimodal signals to remove physiological artifacts and obtain pure EEG signals and multimodal auxiliary features. The sample matching module includes a data storage unit and a feature matching unit for feature extraction and similarity matching based on large-scale historical epilepsy patient data in the hospital database and the patient's own medical history data to obtain a reference sample set most similar to the target patient's features. The model generation module includes a data storage unit, a basic model unit, and a transfer adjustment unit for transfer learning and model fine-tuning of the basic model based on the reference sample set and the target patient's own medical history data to obtain an epilepsy prediction model tailored to the target patient. The prediction analysis module includes a feature intensity filter, a graded power network, and a risk prediction module for feature intensity threshold filtering based on pure EEG signals and multimodal auxiliary features, graded adjustment of calculation power, calculation of seizure risk probability, and symptom type analysis to obtain a real-time risk index and symptom type. The graded early warning module includes an alarm unit and a strategy matcher for triggering graded early warnings based on the real-time risk index and symptom type, and generating targeted nursing and intervention suggestions based on the patient's medical history data.

[0123] Working Principle: This invention performs deep artifact removal on multimodal signals through multi-angle collaborative filtering to obtain high-quality input. Based on this, it generates personalized, lightweight prediction models for each patient through similarity matching and transfer learning, solving the problem of individual differences. Finally, through dynamic signal filtering and hierarchical power calculation, it achieves on-demand allocation of computing resources. This represents a leap from passive monitoring to precise early warning and proactive intervention, significantly improving early warning accuracy and greatly reducing system energy consumption.

[0124] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0125] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0126] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0127] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0128] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monitoring and early warning of epilepsy, characterized in that, Includes the following steps: S1: Acquire real-time multimodal signals including EEG, ECG, EMG and behavior, perform multi-angle collaborative filtering to remove physiological artifacts, and obtain pure EEG signals and multimodal auxiliary features; S2: Based on the large-scale historical epilepsy patient data in the hospital database and the patient's own medical history data, feature extraction and similarity matching are performed to obtain a reference sample set that is most similar to the target patient's features; S3: Based on the reference sample set and the target patient's own medical history data, transfer learning and model fine-tuning are performed on the basic model to obtain the epilepsy prediction model corresponding to the target patient. S4: Input the pure EEG signal and the multimodal auxiliary features into the epilepsy prediction model, perform feature intensity threshold screening, adjust the calculation power according to the screening results, perform seizure risk probability calculation and symptom type analysis, and obtain real-time risk index and symptom type; S5: Trigger graded early warning based on the real-time risk index and the symptom type, and generate targeted nursing and intervention suggestions in conjunction with the patient's medical history data.

2. The method for epilepsy monitoring and early warning according to claim 1, characterized in that, The multi-angle collaborative filtering is implemented based on multi-modal data correlation decomposition: Wavelet transform is used to decompose the multimodal signal from the time domain to the transform domain. The correlation between the EEG signal and signals such as ECG and EMG is calculated by the transfer entropy of symbolic processing, so as to accurately identify ECG artifacts and EMG artifacts. Behavioral features are extracted using behavioral signals to help determine the time period and pattern of motion artifacts. Collaborative filtering is performed in the transform domain to adaptively attenuate the artifact portion. The filtered EEG signal is separated using a blind source separation algorithm to obtain the purified EEG signal.

3. The method for epilepsy monitoring and early warning according to claim 1, characterized in that, The similarity matching includes: A multidimensional feature vector containing time-domain, frequency-domain, and nonlinear dynamic features is constructed based on the patient's own medical history data. The multidimensional feature vector is then compared with the multidimensional feature vectors of historical patients in the hospital database using cosine similarity calculation to form the reference sample set. The contribution of each sample is quantified based on the similarity measure, and the weight of the sample in the model fine-tuning is dynamically allocated.

4. The epilepsy monitoring and early warning method according to claim 3, characterized in that, The epilepsy prediction model was generated using a hierarchical adjustment strategy: A general epilepsy prediction model pre-trained based on large-scale historical epilepsy patient data is obtained as the base model; The parameters of the base model are loaded into the new model as initial values, and the base model is fine-tuned with weights using the reference sample set to obtain the epilepsy prediction model. The fine-tuning process employs a hierarchical learning strategy, where the parameters at the bottom layer of the model used to extract general features are updated slightly with an update magnitude less than a first threshold, while the parameters at the top layer of the model used for specific decision-making are adjusted significantly based on the weighted sample set with an update magnitude greater than a second threshold, in order to quickly adapt to the physiological characteristics of the target patient; the first threshold is less than or equal to the second threshold.

5. The method for epilepsy monitoring and early warning according to claim 1, characterized in that, The feature intensity threshold screening is based on the energy and rate of change of the multimodal auxiliary features, and is used to initially determine whether the signal is worth conducting in-depth analysis.

6. The method for epilepsy monitoring and early warning according to claim 1, characterized in that, The graded adjustment of calculation power adopts a dynamic power management strategy, which dynamically switches the power level based on the characteristic strength of the signal. When the characteristic strength is lower than the first threshold, it maintains low power consumption and only records data. When it is between the first and second thresholds, it enables medium power consumption to calculate the risk probability. When it is higher than the second threshold, it switches to full power for complete analysis and early warning.

7. The method for epilepsy monitoring and early warning according to claim 1, characterized in that, The real-time risk index and the symptom type are generated through multimodal information fusion: The real-time risk index is dynamically weighted by integrating symptom activity index, short-term trend and model confidence. The symptom type is obtained by matching the current multimodal auxiliary features with a patient-specific symptom template library.

8. The method for epilepsy monitoring and early warning according to claim 1, characterized in that, The intervention recommendations include instructing family members to keep an eye on the patient and record information when the risk is low; instructing them to perform anti-asphyxiation positioning and remove hazardous materials when the risk is medium; and instructing them to clear nasal and oral secretions, keep track of time, and prepare for emergency treatment when the risk is high.

9. The method for epilepsy monitoring and early warning according to claim 1, characterized in that, It also includes model self-updating and optimization: periodically or after each warning event, newly generated effective data is incorporated into the target patient's own medical history data, and steps S2 and S3 are re-executed to update the reference sample set and optimize the epilepsy prediction model.

10. An epilepsy monitoring and early warning system, characterized in that, The system is used to implement an epilepsy monitoring and early warning method as described in any one of claims 1-9, comprising: The collaborative filtering module, including a three-level filtering network, is used to perform multi-angle collaborative filtering based on real-time multimodal signals to remove physiological artifacts and obtain pure EEG signals and multimodal auxiliary features. The sample matching module includes a data storage unit and a feature matching unit, which are used to extract features and perform similarity matching based on large-scale historical epilepsy patient data in the hospital database and the patient's own medical history data, to obtain a reference sample set that is most similar to the target patient's features. The model generation module includes a data storage unit, a basic model unit, and a transfer adjustment unit. It is used to perform transfer learning and model fine-tuning on the basic model based on the reference sample set and the target patient's own medical history data to obtain an epilepsy prediction model tailored for the target patient. The predictive analysis module includes a feature intensity filter, a hierarchical power network, and a risk prediction module. It is used to perform feature intensity threshold filtering based on the pure EEG signal and the multimodal auxiliary features, and to adjust the calculation power hierarchically to calculate the probability of attack risk and analyze the symptom type, so as to obtain the real-time risk index and symptom type. The graded early warning module includes an alarm unit and a strategy matcher. It triggers graded early warnings based on the real-time risk index and the symptom type, and generates targeted nursing and intervention suggestions by combining patient medical history data.

Images