Epileptic seizure auxiliary prediction and cognitive function evaluation system and method based on multi-modal data fusion and intelligent adaptive optimization

Abstract

The invention discloses an epileptic seizure auxiliary prediction and cognitive function evaluation system and method based on multi-modal data fusion and intelligent adaptive optimization. All the modules work cooperatively, and intelligent optimization from multi-dimensional electroencephalogram signal, motion data and physiological data collection, intelligent preprocessing and feature deep extraction and fusion to epileptic seizure prediction and cognitive function evaluation is achieved. Through the combination of neural architecture search and self-supervised comparative learning, the network structure can be automatically optimized on the basis of multi-modal data, and the ability of learning and identifying an epileptic seizure mode and a cognitive state is effectively improved. The system realizes high-accuracy epileptic seizure prediction and cognitive state evaluation through adaptive optimization, and remarkably reduces the missing report rate and the false report rate. The system has the characteristics of high efficiency, intelligence and accuracy, can provide timely early warning for epilepsy patients, provides scientific cognitive function intervention suggestions for doctors, and has wide application prospects.

Description

Technical Field

[0001] The present invention belongs to the fields of artificial intelligence and medical health, and particularly relates to a system and method for auxiliary prediction of epileptic seizures and cognitive function assessment based on multimodal data fusion and intelligent adaptive optimization. Background Art

[0002] Epilepsy is a common neurological disease, characterized by suddenness and recurrence. Patients usually cannot predict the timing of epileptic seizures in advance, which greatly affects their daily lives. Traditional methods for predicting epileptic seizures mostly rely on doctors' experience and simple electroencephalogram (EEG) signal processing techniques. However, these methods have certain limitations. For example, they cannot effectively remove noise, the real-time analysis accuracy of signals is not high, and the prediction time is usually short. Although there are some EEG-based epileptic seizure prediction systems, existing technologies often lack personalized analysis and adaptive optimization capabilities, resulting in low prediction accuracy and difficulty in dealing with complex and multi-dimensional patient data. In addition, cognitive function assessment also plays an important role in clinical diagnosis, especially in the early identification of neurodegenerative diseases such as dementia and Alzheimer's disease. However, existing cognitive function assessment methods mostly rely on subjective judgments such as interviews and questionnaires, lack objective quantitative indicators, and cannot track the changes in patients' cognitive states in real time and dynamically. Although some systems based on EEG and behavior monitoring have been applied to cognitive function assessment, the algorithm accuracy, real-time performance, and personalization of these methods are still insufficient, and they cannot provide efficient and accurate assessments. With the rapid development of artificial intelligence and deep learning technologies, especially the application of advanced algorithms such as neural architecture search (NAS) and self-supervised learning, the potential of intelligent medical systems in epileptic seizure prediction and cognitive function assessment has gradually emerged. However, existing systems still face problems such as insufficient multimodal data fusion, high computational resource consumption, and poor model adaptability. Therefore, there is an urgent need for an intelligent system that can combine multiple physiological data and improve the prediction and assessment accuracy through adaptive optimization algorithms to meet the needs of precision medicine and personalized diagnosis. Summary of the Invention

[0003] The object of the present invention is to provide a system and method for epilepsy seizure assisted prediction and cognitive function evaluation based on multimodal data fusion and intelligent adaptive optimization, aiming to solve the problems of low data processing efficiency, poor prediction accuracy, lack of personalized analysis, and insufficient processing ability for complex and multi-dimensional data in traditional epilepsy seizure prediction and cognitive function evaluation systems. Traditional epilepsy seizure prediction methods mostly rely on expert experience or simple electroencephalogram analysis, and cannot effectively solve problems such as noise interference, complex pattern recognition, and individual differences. Cognitive function evaluation usually relies on subjective judgment, lacks real-time and accurate quantitative analysis, and fails to fully utilize the multimodal physiological data of patients. To solve these problems, the present invention conducts fusion analysis on epilepsy seizure prediction and cognitive function evaluation to form a complementarity, enhancing the accuracy, real-time performance, and personalized level in their respective intelligent analysis processes.

[0004] The technical solution of the present invention is: A system for epilepsy seizure assisted prediction and cognitive function evaluation based on multimodal data fusion and intelligent adaptive optimization, which is composed of multiple modules working in cooperation, including a multimodal physiological data acquisition and perception module, a dynamic data standardization and denoising preprocessing module, a multi-dimensional feature extraction and deep fusion module, an intelligent network architecture search and self-supervised learning optimization module, an adaptive multi-task optimization and learning module, and a real-time prediction and multi-dimensional health evaluation module connected in sequence. Through intelligent data acquisition and analysis, it provides accurate results for epilepsy seizure prediction and cognitive function evaluation;

[0005] The multimodal physiological data acquisition and perception module real-time collects the multimodal physiological data of patients through a variety of perception devices; specifically, an electroencephalogram device is used to record the brain waves of patients, obtaining time series electroencephalogram signals, including the amplitude and frequency changes of different brain waves; accelerometers and gyroscope wearable devices are used to monitor the movement data of patients, recording gait, posture, and movement patterns to reflect the movement characteristics and cognitive function status of patients; at the same time, this module also uses monitoring devices for physiological parameters such as heart rate and blood pressure to collect the physiological data of patients to assist in analyzing the correlation between epilepsy seizures and cognitive function changes; through the comprehensive collection of these multimodal data, this module provides rich inputs for subsequent data processing and analysis;

[0006] The dynamic data normalization and denoising preprocessing module normalizes and denoises the raw data obtained from the perception data acquisition module to ensure the accuracy and consistency of subsequent analysis; the electroencephalogram (EEG) signals are passed through a digital filter to remove high-frequency and low-frequency noise, ensuring that the signal-to-noise ratio of the data reaches a high level; for motion data and physiological data, normalization techniques are used to unify the data to the same dimension; for signals with missing or poor quality, interpolation techniques are used to complete or correct them; this module dynamically adjusts the processing flow according to the real-time collected data and environmental changes through an adaptive learning algorithm to ensure that the preprocessing flow can effectively handle different types of input data;

[0007] The multi-dimensional feature extraction and deep fusion module extracts effective information from the multi-modal data obtained from the perception data acquisition module using a variety of feature extraction techniques; for EEG signals, the power spectral density and waveform change features related to epileptic seizures are extracted; for motion data, dynamic time warping and acceleration feature analysis methods are used to extract features related to gait, posture, movement speed, and cognitive function; at the same time, by analyzing physiological data such as heart rate and blood pressure, physiological features affecting cognitive state and epileptic seizures are extracted; feature fusion effectively combines EEG signals, motion data, and physiological data through weighted average technology to generate a unified, multi-dimensional feature representation, providing comprehensive input for subsequent intelligent analysis;

[0008] The intelligent network architecture search and self-supervised learning optimization module optimizes the feature data by combining two advanced intelligent algorithms: neural architecture search and self-supervised contrast learning; during the neural architecture search process, the system automatically selects the most suitable neural network architecture for constructing the epileptic seizure prediction and cognitive function evaluation models; through the neural architecture search algorithm, the system can automatically adjust the depth, number of layers, and number of nodes of the neural network to find the architecture most suitable for the input data; at the same time, the self-supervised contrast learning method is used to learn the potential relationships between data based on unlabeled data;

[0009] The adaptive multi-task optimization and learning module uses multi-task learning technology to optimize the learning processes of epileptic seizure prediction and cognitive function evaluation; the system simultaneously performs epileptic seizure prediction and cognitive function evaluation tasks in the same network through a shared neural network model; through the shared parameters and feature information, the two tasks can promote each other, thereby improving the prediction accuracy of each task; using adaptive optimization technology, the weight distribution and resource allocation between tasks are automatically adjusted according to the complexity of the tasks and the non-linear characteristics of the data;

[0010] The real-time prediction and multi-dimensional health assessment module performs real-time data prediction and evaluation on the optimized model; for seizure prediction, the system uses the trained model to predict whether a patient will have a seizure based on the collected EEG signals and physiological data; for cognitive function assessment, the system evaluates the patient's cognitive state according to the patient's EEG signals, motion data, and physiological data; the prediction and evaluation module generates a health status assessment report for the patient based on the real-time data.

[0011] A method for an epilepsy seizure-assisted prediction and cognitive function assessment system based on multi-modal data fusion and intelligent adaptive optimization of the present invention includes the following steps:

[0012] S1. Data collection: Through the multi-modal physiological data collection and perception module, use an electroencephalogram device, an accelerometer, a gyroscope, and a physiological monitoring device to collect the patient's electroencephalogram signals, motion posture data, and physiological parameter data such as heart rate and blood pressure in real time; the collected data is encrypted and transmitted and stored in the data storage unit of the system in real time to ensure the security and integrity of the data, providing a solid foundation for subsequent data processing and analysis;

[0013] S2. Data preprocessing: Through the dynamic data standardization and denoising preprocessing module, process the raw data obtained from the perception data collection module, use a digital filtering algorithm to remove the noise in the electroencephalogram signals; perform standardization processing on the motion and physiological data to eliminate the dimension difference; for signals with poor quality or missing data, use the interpolation method to complete them to ensure the accuracy of subsequent analysis;

[0014] S3. Feature extraction and fusion: Through the multi-dimensional feature extraction and deep fusion module, based on the multi-modal data obtained from the perception data collection module, extract the relevant features for epilepsy seizure and cognitive function assessment; for electroencephalogram signals, extract the relevant features for epilepsy seizure; for motion data, use the dynamic time warping and acceleration feature extraction methods to extract gait and posture features; combine physiological data to extract the physiological features that affect cognitive function and epilepsy seizure; all the extracted features generate a unified multi-dimensional feature vector through deep fusion technology, providing an efficient input for subsequent intelligent analysis;

[0015] S4. Feature optimization: After the feature extraction and fusion are completed, use the intelligent network architecture search and self-supervised learning optimization module to optimize the extracted feature data; use the neural architecture search algorithm to automatically select the most suitable neural network architecture, optimize the parameters of the network depth, number of layers, and number of nodes to improve the accuracy and stability of the model; use the self-supervised contrast learning algorithm to learn the potential relationships between data on unlabeled data and enhance the feature learning ability of the model;

[0016] S5. Multi-task Learning and Adaptive Optimization: Through the adaptive multi-task optimization and learning module, the system simultaneously performs epilepsy seizure prediction and cognitive function assessment tasks in the same neural network model; by sharing network parameters and feature information, the two tasks can promote each other, thereby improving the prediction accuracy; through the adaptive optimization technology, the weights and computing resource allocation between tasks are automatically adjusted.

[0017] S6. Prediction and Health Assessment: Based on the optimized model, the system performs epilepsy seizure prediction and cognitive function assessment through real-time data; for epilepsy seizure prediction, the system uses the electroencephalogram (EEG) signals and physiological data collected in real-time to predict whether the patient will have an epilepsy seizure and generate corresponding warnings; for cognitive function assessment, the system evaluates the cognitive state of the patient according to EEG, motion data, and physiological data; the prediction results and assessment information will be integrated to generate a personalized health report, providing effective decision-making support for doctors.

[0018] S1 specifically includes:

[0019] Real-time collection includes EEG signals, gait, posture, acceleration motion data, heart rate, and blood pressure physiological parameters, providing accurate inputs for subsequent intelligent data processing and analysis.

[0020] EEG Signal Collection: Use EEG equipment to collect time-series EEG signals, mainly focusing on the amplitude and frequency changes of the Delta, Theta, Alpha, and Beta wave frequency bands; these signals contain key features related to epilepsy seizures.

[0021] Motion Data and Posture Data Collection: The collection of motion data is mainly completed through accelerometers and gyroscope wearable devices; by monitoring the patient's gait, posture, and acceleration motion behavior in real-time through these devices, additional physiological clues can be provided for analyzing the patient's cognitive function state; accelerometers are used to measure the patient's motion speed and acceleration; gyroscopes are used to record the patient's posture and direction changes.

[0022] Physiological Data Collection: Heart rate and blood pressure are important physiological indicators reflecting the human health state, which can provide auxiliary data for epilepsy seizure and cognitive function assessment; a heart rate monitoring device measures the patient's heart rate in real-time, with the unit of bpm; a blood pressure monitoring device monitors the patient's systolic and diastolic blood pressures in real-time, with the unit of mmHg.

[0023] Data Transmission: All collected data will be uploaded to the storage unit of the system in a secure and encrypted manner, and the data transmission method supports wired or wireless transmission. Wired transmission uses standard data interfaces for high-stability transmission; wireless transmission uses WiFi, Bluetooth, and Zigbee wireless communication technologies.

[0024] Data Format:

[0025] EEG signals: Time series data in the format of [timestamp, EEG signal value];

[0026] Motion data: Time series data in the format of [timestamp, acceleration value, velocity value, posture angle]

[0027] Physiological data: Time series data in the format of [timestamp, heart rate value, blood pressure value];

[0028] System storage unit: After all the collected physiological data are transmitted to the system, they will be stored in a dedicated storage unit.

[0029] Furthermore, S2 specifically includes:

[0030] Signal denoising: Eliminate the irrelevant noise introduced by various factors during the acquisition process; for different types of EEG signals, motion data, and heart rate data sensor data, the denoising methods will be different;

[0031] Denoising of EEG signals: EEG signals contain power grid noise (high-frequency noise) and baseline drift (low-frequency noise); to remove these noises, digital filtering technology (band-pass filter) is used;

[0032] Band-pass filter formula: The signal after the EEG signal is processed by the band-pass filter is X filt , by designing appropriate cut-off frequencies f c and order n, remove the noise in specific frequency bands and retain the EEG signals useful for seizure prediction; then the transfer function of the band-pass filter is expressed as:

[0033]

[0034] where: H(f) is the transfer function of the filter, f is the signal frequency, f c is the cut-off frequency of the filter, and n is the order of the filter;

[0035] Denoising of motion data and physiological data: The noise of motion data and physiological data appears as high-frequency vibrations and sensor errors; using the moving average method, perform moving average processing on the acceleration signal of motion data to eliminate short-term fluctuations; Moving average denoising formula:

[0036]

[0037] where, y(t) is the original signal, and M is the size of the moving window; by selecting an appropriate window size M, remove high-frequency noise and smooth the motion signal;

[0038] Missing value imputation: Due to equipment failures, sensor data loss, and signal interruptions, there may be missing values in the data; imputing missing values is crucial for ensuring the accuracy of subsequent analyses; for time series data, linear interpolation is used to impute missing values. For the missing data between time points t 0 and t 1 , linear interpolation is performed using the following formula:

[0039]

[0040] where x(t) is the imputed data, x(t o ) and x(t 1 ) are known adjacent data points, t 0 and t 1 are the corresponding timestamps, and t is the timestamp of the missing data;

[0041] Data normalization and standardization: Different types of electroencephalogram signals, motion data, heart rate data, and physiological data have different dimensions and value ranges. To compare and fuse different data on the same scale, data normalization and standardization are required;

[0042] Data normalization: Map all electroencephalogram signals, motion data, and heart rate data to a unified interval; the input data is X = [x 1 , x 2 ,..., x N , and the normalization formula is:

[0043]

[0044] where min(X) and max(X) are the minimum and maximum values of the dataset X respectively, and x i is the normalized data; by this method, all input data is compressed into the interval [0, 1], facilitating subsequent fusion analysis;

[0045] Data standardization: Transform electroencephalogram signals, motion data, and heart rate data into a distribution with a mean of 0 and a standard deviation of 1; for the dataset X = [x 1 , x 2 ,..., x N , the standardization formula is:

[0046]

[0047] where μ(X) and σ(X) are the mean and standard deviation of the dataset X respectively, and x i is the standardized data;

[0048] Adaptive learning algorithm: The denoising, missing value imputation, and normalization steps in the data preprocessing process need to be adjusted according to the changes in real-time data and the environment. Therefore, the system adopts an adaptive learning algorithm to dynamically adjust the preprocessing process. This algorithm automatically selects appropriate preprocessing methods based on the characteristics of the input data. The noise level of the current data is measured by the signal-to-noise ratio (SNR), and the algorithm adjusts the filter parameters according to the change in the current SNR. SNR calculation formula:

[0049]

[0050] When the SNR is low, the algorithm increases the order of the filter for stronger denoising. When the SNR is high, the algorithm selects a lower filtering intensity.

[0051] Furthermore, S3 includes:

[0052] Electroencephalogram (EEG) signal feature extraction: EEG signals are the core data source for seizure prediction. By extracting the features of EEG signals, abnormal activity patterns related to seizures can be identified. Common EEG analysis methods include time-frequency domain analysis.

[0053] Time-frequency domain analysis combines the advantages of the time domain and the frequency domain and can extract the frequency characteristics of the signal changing over time. Wavelet transform is used to obtain the resolution in both the time domain and the frequency domain simultaneously. For a signal x(t), the result of wavelet transform is:

[0054]

[0055] where, W φ (a, b) is the result of wavelet transform, ψ is the mother wavelet function, and Φ * is the complex conjugate of ψ; a is the scale factor, and b is the translation factor. Through wavelet transform, the characteristics of the detailed part and the low-frequency part of the signal can be extracted to identify the time and frequency characteristics of seizures.

[0056] Motion data feature extraction: Seizures can affect the patient's motor ability. Therefore, the analysis of motion data can help evaluate the relationship between the patient's cognitive function status and seizures.

[0057] Dynamic time warping is an algorithm for measuring the similarity between two time series, used to process asynchronous motion data. There are two time series A = {a 1 , a 2 ,..., a N} and B = {b 1 , b 2 ,..., b N}. The distance of dynamic time warping is calculated recursively to find the minimum matching path.

[0058]

[0059] Among them, dist(a i , b j ) is the distance between two points, and D(i, j) represents the minimum distance from a i to b j ; min{} represents the operation of taking the minimum value. Dynamic time warping can effectively align different gait and posture data, so as to extract the motion patterns related to epileptic seizures;

[0060] Acceleration feature extraction: The data provided by the acceleration sensor is used to extract gait and posture features. The acceleration features include mean acceleration, standard deviation acceleration, and maximum acceleration. For the acceleration data at each moment, calculate the total acceleration A(t) of the acceleration:

[0061]

[0062] Among them, a x (t) 2 , a y (t) 2 and a z (t) 2 are the components of the acceleration in the X, Y, and Z directions respectively; by extracting the total acceleration features in different time periods, identify the correlation with cognitive function and epileptic seizures;

[0063] Physiological data feature extraction: Changes in epileptic seizures and cognitive function are usually accompanied by changes in physiological parameters.

[0064] Heart rate feature: Heart rate is an important physiological index for evaluating the neurological status of patients; heart rate features include mean heart rate and heart rate variability; heart rate variability is an index to measure the irregularity of the inter-beat interval, and the commonly used calculation method is the standard deviation:

[0065]

[0066] Among them, RRi is the time interval between two consecutive heartbeats, RR is the mean of these time intervals, and N is the number of recorded heartbeats;

[0067] Blood pressure feature: Blood pressure is another important index reflecting cognitive status and epileptic seizures; the features in blood pressure data include the fluctuations of systolic blood pressure and diastolic blood pressure; by calculating the mean and variance statistics of the blood pressure signal, extract the correlation features of blood pressure fluctuations with cognitive function changes and epileptic seizures;

[0068] Feature fusion: All the extracted features will be fused into a unified, multi-dimensional feature vector for subsequent intelligent analysis and prediction; the feature fusion method is the weighted average method;

[0069] f fused = w EEG · f EEG + w Motion · f Motion + w Physio · f Physio

[0070] where f fused : the fused feature vector; it is a comprehensive feature vector obtained by weighted averaging of the input data; f EEG : the feature vector of the electroencephalogram signal; f Motion : the feature vector of the motion data; f Physio : the feature vector of the physiological data; w EEG : the weight coefficient of the electroencephalogram (EEG) feature; w Motion : the weight coefficient of the motion data feature; w Physio : the weight coefficient of the physiological data feature;

[0071] Furthermore, S4 specifically includes:

[0072] Neural Architecture Search (NAS) is an automated neural network architecture optimization method that automatically selects the optimal architecture by exploring different network structures to adapt to specific tasks; in the present invention, NAS is used to optimize the depth, number of layers, number of nodes and other hyperparameters of the neural network, so as to improve the prediction ability of the model and ensure high precision and high stability of epileptic seizure prediction and cognitive function evaluation;

[0073] Input data

[0074] In the first step of feature optimization, the system obtains various physiological signals from the multimodal data acquisition and perception module, and these signals include electroencephalogram (EEG) signals, motion data, heart rate and blood pressure; after appropriate preprocessing, the data is converted into feature vectors suitable for model processing; and provided to the NAS algorithm for network architecture optimization; the effective fusion and processing of these multimodal data are the key to improving the system performance;

[0075] Initial architecture generation

[0076] Use the NAS algorithm to generate an initial neural network architecture and select a model structure suitable for processing multimodal physiological data; in this invention, the long short-term memory network (LSTM) is selected as the initial neural network architecture;

[0077] Hyperparameter setting: Determine the initial hyperparameters, including the number of layers of the LSTM, the number of units in each layer, the learning rate, and the batch size; the selection of these hyperparameters provides a basis for subsequent optimization;

[0078] Number of units in each layer

[0079] The number of units in each layer determines the memory capacity of the LSTM in each layer; for seizure prediction, EEG signals usually contain various frequency and amplitude variations, so a higher number of units is required to capture these complex temporal features; while in cognitive assessment, due to the multi-modality of the tasks, the features of a single modality are relatively simple and the number of units is relatively small;

[0080] The number of layers of the LSTM is L, and the number of units in each layer is N u , and the output of the LSTM is h l , then the operation process of the LSTM network can be expressed as:

[0081] h l = LSTM l (h l-1 , X) for l = 1, 2,..., L

[0082] Among them, h l : This is the hidden state of the first layer; h l-1 : This is the hidden state of the (l - 1)-th layer; X: This is the input data; LSTM: This is the core operation of the long short-term memory network; L: This is the number of layers of the LSTM network;

[0083] Selection and optimization of the learning rate

[0084] The learning rate is an important hyperparameter that determines the training speed and convergence speed of the model; in the seizure prediction task, since the signal contains tiny abnormal variations, a smaller learning rate helps to improve the stability of the model and avoid skipping local optimal solutions; while in cognitive function assessment, due to the complexity of multi-modal data, a suitable learning rate selection accelerates model training and avoids overfitting at the same time;

[0085] In the gradient descent algorithm, the learning rate α updates the model parameters through the following formula:

[0086]

[0087] Among them, θ t : This is the value of the model parameter at the t-th iteration; θ t+1 : This is the updated value of the model parameter at the (t + 1)-th iteration; α: This is the learning rate, a very crucial hyperparameter that controls the step size of parameter updates in each iteration; This is the gradient of the loss function, the derivative with respect to the model parameter θ t ; The gradient represents the rate of change of the loss function at the current parameter value, indicating the direction and magnitude of the parameter adjustment: represents the direction in which the loss function changes fastest for each parameter; represents that the loss function is a measure of the error between the model prediction output and the true result;

[0088] Selection of Batch Size

[0089] The batch size controls the number of samples used for each weight update; in seizure prediction, due to the high-frequency fluctuations of EEG signals, a smaller batch size helps capture more detailed information; while in cognitive function assessment, due to the fusion of multi-modal data, a larger batch size helps improve the model's learning ability of the correlations between different data modalities; in batch gradient descent, the batch size B determines the number of samples used for each parameter update; the update formula for each time is as follows:

[0090]

[0091] where, x i is the i-th training sample, y i is its corresponding label, θ 0 : These are the parameters of the initial model: B is the batch size; Ψ(θ 0 x i , y i ) is the loss function, which measures the error between the predicted value θ 0 x i of the model and the label y i ;

[0092] Objective Function

[0093] One of the core tasks of NAS is to optimize the neural network architecture by minimizing the loss function; in the feature optimization module, the objective function measures the performance of the model through the error between the predicted value of the network and the true label, thereby guiding the optimization of the network architecture;

[0094] The loss function is defined as follows:

[0095]

[0096] where: f(x i , A) is the output of the model under the input x i and the architecture A, A is the current network architecture, the configuration of the number of layers and units of the LSTM; y i is the actual label: the occurrence or non-occurrence of seizures, the score of the cognitive state; L is the loss function, and N is the number of samples;

[0097] The goal of this loss function is to minimize the difference between the predicted value and the actual label, so as to optimize the network architecture; in seizure prediction, the goal is to minimize the error between the seizure probability predicted by the network and the true seizure event; in cognitive function assessment, the goal is to minimize the error between the cognitive score predicted by the model and the actual assessment result;

[0098] Optimization Process

[0099] The NAS algorithm automatically adjusts various hyperparameters of the network architecture through multiple rounds of iterative search, including the depth of the neural network, the number of layers, and the number of nodes in each layer. During each round of the search process, the network architecture is trained and evaluated, and the architecture is adjusted based on the value of the feedback loss function.

[0100] NAS determines the number of layers of the LSTM network according to the task complexity. For complex epilepsy prediction tasks, an LSTM network with more layers and units is required to better capture long-term dependencies. The number of nodes in each layer is optimized according to the task requirements. More nodes help capture more information while increasing the computational overhead.

[0101] Training and evaluation: After each adjustment, the network is trained on a specific training set, and whether the architecture needs to be further optimized is determined according to the evaluation results of the loss function. Through multiple rounds of iterative search, NAS gradually optimizes the network architecture and finally determines the neural network architecture most suitable for epilepsy seizure prediction and cognitive function evaluation tasks.

[0102] Application of self-supervised contrastive learning

[0103] In the epilepsy seizure prediction and cognitive function evaluation system, the workflow of self-supervised contrastive learning in the feature optimization module includes the following key steps:

[0104] Data input

[0105] At the front end of the system, the data acquisition module collects multimodal physiological signals from multiple physiological data sources. After these signals are preprocessed, they enter the feature extraction stage. For each type of data, the system extracts a set of high-dimensional feature vectors to represent the data features at different time points and different physiological states. The data from different sensors will be further fused to form a comprehensive high-dimensional feature vector. This high-dimensional feature vector will be used as input and processed by the self-supervised contrastive learning module to optimize the representation of the feature space.

[0106] Positive and negative sample pair generation

[0107] In self-supervised contrastive learning, the generation of positive and negative sample pairs is one of the core tasks. In the implementation of the present invention, the system generates "positive sample pairs" and "negative sample pairs" by transforming the original feature vectors.

[0108] Positive sample pairs: Sample pairs from similar situations, indicating samples that should be close in the feature space; EEG signal data of the same patient before and after epilepsy seizures, heart rate data of the same patient within the same time period.

[0109] Negative sample pairs: Sample pairs from different scenarios, indicating samples that should be far apart in the feature space; EEG signal data of different patients, or motion data during seizure and non-seizure periods;

[0110] Contrast loss function

[0111] To optimize the sample distribution in the feature space, the system uses a contrast loss function to train the model; the core objective of the contrast loss function is to maximize the similarity between positive sample pairs while minimizing the similarity between negative sample pairs, thereby enhancing the model's ability to distinguish different categories of samples in the feature space;

[0112] The contrast loss function is usually defined as:

[0113]

[0114] where sin(f(x i ), f(x j )) represents the similarity between samples x i and x j in the feature space, usually measured using cosine similarity: is the exponential function, represents the sum over all samples, N is the total number of samples; f(x i ) represents the feature representation of sample x i ;

[0115] The objective of this loss function is to make similar samples closer in the feature space and pull the distance between different samples farther through contrastive learning; in this way, the model can learn more discriminative feature representations;

[0116] Optimization of the feature space and model learning

[0117] The model automatically adjusts the distance relationship between samples by minimizing the contrast loss function; in each training iteration, the similarity of positive sample pairs is maximized while the similarity of negative sample pairs is minimized; this process continuously drives the model to optimize the sample distribution in the feature space, making similar samples close to each other while different samples remain at a greater distance;

[0118] This optimization helps the model capture potential patterns in the data. Especially in epilepsy seizure prediction and cognitive function assessment tasks, self-supervised contrastive learning can more sensitively identify subtle changes in EEG signals before and after epilepsy seizures and capture the complex relationship between motion data and cognitive function;

[0119] Through the above optimizations, the system can more accurately predict epileptic seizures and evaluate changes in cognitive function status, providing more valuable support for clinical decision-making and helping doctors and patients make more scientific and accurate decisions.

[0120] Furthermore, S5 specifically includes:

[0121] In the framework of multi-task learning, design a shared neural network architecture, which is divided into the following parts:

[0122] Shared underlying feature extraction layer: This layer extracts features from multi-modal data, EEG signals, motion data, heart rate, and blood pressure to generate shared low-level features; these low-level features provide the same basic data input for the two tasks of epileptic seizure prediction and cognitive function assessment;

[0123] Task-specific layers: After the shared underlying feature extraction layer, the network branches into two, respectively dealing with the tasks of epileptic seizure prediction and cognitive function assessment; these task-specific layers learn specific high-level features according to task requirements; the output of each branch corresponds to different task goals;

[0124] The role of the adaptive optimization algorithm in the multi-task learning framework is to automatically adjust the weights and resource allocation between tasks; since the epileptic seizure prediction task and the cognitive function assessment task have different dependencies on different features and the task priorities also change at any time, an adaptive mechanism is needed to balance the learning process of each task;

[0125] The learning process of each task is affected by different task difficulties, data quality, and interference between tasks; therefore, the adaptive optimization algorithm dynamically adjusts the weights between tasks according to the loss function of each task and the feedback of the model; uses a weighted loss function to achieve weight adjustment between tasks; for the epileptic seizure prediction task and the cognitive function assessment task, the total loss function is defined by the following formula:

[0126]

[0127] Where: is the total loss function; is the loss function of the epileptic seizure prediction task; is the loss function of the cognitive function assessment task; λ 1 and λ 2 are the weight parameters of the epileptic seizure prediction task and the cognitive function assessment task respectively;

[0128] Adaptive learning rate adjustment

[0129] The adaptive optimization algorithm also includes a learning rate adjustment mechanism to ensure that the model can effectively converge during training; by using the adaptive optimization algorithm, the learning rate of each task is dynamically adjusted according to the gradient update history of each task; the update rule is usually as follows:

[0130]

[0131] Where: θ t represents the model parameters; η t is the learning rate at the current moment, and η t will be adjusted according to the gradient of each task; m t and v t are the moving average and variance of the gradients calculated at time t; ∈ is a constant to prevent division by zero errors;

[0132] This adaptive learning rate adjustment mechanism ensures that when dealing with multiple tasks, the learning progress of different tasks remains balanced, avoiding the situation where a certain task converges prematurely or the learning progress of other tasks is slow;

[0133] The key to multi-task learning is to promote the cooperation between the two tasks during training through shared underlying features, and ultimately improve the overall performance; in the present invention, the two tasks of seizure prediction and cognitive function assessment are co-optimized by sharing the feature representation of the neural network and sharing parameters; specifically:

[0134] Shared feature representation: The seizure prediction task and the cognitive function assessment task share the underlying feature extraction layer. At this layer, the two tasks extract common features based on the same input data; this sharing not only reduces the complexity of the model, but also enhances the generalization ability of the model through shared feature representation;

[0135] Inter-task information flow: During the optimization process, the seizure prediction task and the cognitive function assessment task interact with each other through the shared underlying features and task-specific layers, promoting the flow and sharing of information; by introducing a joint training mechanism in the neural network, the cognitive function assessment task helps the seizure prediction task obtain richer context information by optimizing the task-specific layer, and vice versa;

[0136] The multi-task learning and adaptive optimization module of the present invention efficiently integrates the seizure prediction and cognitive function assessment tasks; through the dual effects of sharing the neural network architecture and the adaptive optimization algorithm, the system can optimize multiple tasks simultaneously with limited computing resources, improving the overall performance;

[0137] Mutual promotion between tasks: The tasks of seizure prediction and cognitive function assessment share the underlying features and network parameters of the neural network. This information sharing not only reduces the computational overhead but also enhances the robustness of each task. The seizure prediction task helps to reveal the physiological changes of the patient, which in turn helps the cognitive function assessment task to accurately judge the patient's cognitive state.

[0138] Efficient resource allocation: The adaptive optimization algorithm can dynamically adjust the weights between tasks to ensure that the system can flexibly allocate computational resources in a complex environment. By learning the relationships between tasks, the optimization algorithm can adjust the resource allocation in real time according to the performance of the tasks, ensuring the efficiency and accuracy of the model training process.

[0139] Enhanced system adaptability: In complex physiological signal data, the relationship between seizure prediction and cognitive function assessment is not simply linear, which requires the system to have a certain degree of adaptability to process these complex interaction information. Through adaptive optimization, the system can adjust the model according to the changes in data and the requirements of tasks, thereby improving the adaptability to different environments and individual differences.

[0140] 8. The method according to claim 2, wherein S6 specifically includes:

[0141] The core goal of seizure prediction is to calculate a probability of seizure occurrence; time series analysis and attention mechanism are introduced to improve the prediction accuracy; the long short-term memory network LSTM is used to fully consider the temporal sequence and dependence of data.

[0142] A multi-layer LSTM model is used to predict seizures; the input of the model includes multi-modal data: EEG signals, motion data, heart rate, blood pressure, and the seizure probability in the future is predicted based on the state at the past moment.

[0143] The input features are:

[0144] X t =(X EEG (t), X motion (t), X physiological (t))

[0145] Where: X EEG (t) represents the EEG signal at the t-th moment; X motion (t) represents the motion data at the t-th moment; X physiological (t) represents the physiological data at the t-th moment; the output of the LSTM model is the probability P epilepsy (t + 1) of seizure occurrence, that is, the probability of seizure occurrence at the (t + 1)-th moment:

[0146] P epilepsy(t + 1) = σ(W epilepsy ·H t + b epilepsy )

[0147] Where: H t represents the hidden state of the LSTM at the t-th moment; W epilepsy and b epilepsy are the weight matrix and the bias term respectively; σ is the Sigmoid activation function, and the output value is in the range of [0, 1], representing the occurrence probability of epileptic seizures;

[0148] Prediction Threshold Setting and Early Warning

[0149]

[0150] Once the model gives the occurrence probability P epilepsy (t + 1) of epileptic seizures, a threshold T is set. When P epilepsy (t + 1) > T, the system sends out an early warning signal, and the system issues a "high risk" early warning, recommending that the patient take emergency preventive measures;

[0151] Cognitive Function Assessment

[0152] For cognitive function assessment, the improvement goal is to generate a quantitative cognitive score through comprehensive analysis of the patient's electroencephalogram signals, motion data, and physiological data, so as to more accurately evaluate the patient's cognitive health status;

[0153] Similarly, using LSTM or other neural network structures, taking multi-modal data as input, and generating a cognitive score through the model; let the output of the cognitive function assessment model be C cognition (t), that is, the cognitive score of the patient:

[0154] C cognition (t) = f(X EEG (t), X motion (t), X physiological (t), θ)

[0155] Where, C cognition (t) represents the cognitive score of the patient at the t-th moment, usually a continuous value, indicating the patient's current cognitive state; θ represents the parameters of the cognitive function assessment model;

[0156] Based on the cognitive score C cognition (t), the system generates the following assessment results:

[0157] Mild Cognitive Impairment: When the cognitive score is in the range of [0, 0.3], it indicates that the patient has mild cognitive impairment;

[0158] Moderate cognitive impairment: When the cognitive score is within the range of [0.3, 0.7], it indicates that the patient's cognitive function is moderately impaired and further treatment and evaluation are required;

[0159] Normal cognitive state: When the cognitive score is within the range of [0.7, 1.0], it indicates that the patient's cognitive function is normal and no intervention is needed for the time being;

[0160] Through the above prediction and evaluation module, the system can more accurately predict epileptic seizures based on real-time data and provide personalized cognitive function evaluation according to the patient's physiological state; these improvements not only enhance the prediction accuracy of the system, but also improve its intelligence level, and can provide more effective decision-making support for doctors in practical applications.

[0161] Compared with the prior art, the beneficial effects of the present invention are:

[0162] 1. Improve prediction accuracy and real-time performance: By combining multi-modal data fusion technology and intelligent adaptive optimization algorithm, the present invention significantly improves the accuracy and real-time performance of epileptic seizure prediction and cognitive function evaluation. Traditional methods usually rely on a single data source or feature extraction method and cannot fully capture multi-dimensional physiological information, resulting in low prediction accuracy. Through the multi-modal data fusion of electroencephalogram, motion data and physiological parameters, the present invention can comprehensively analyze the patient's physical state and cognitive function changes to ensure that the prediction results are more accurate and timely.

[0163] 2. Achieve automated and personalized evaluation: Through the adaptive optimization algorithm, the present invention dynamically adjusts the model parameters according to the real-time data and physiological changes of each patient, realizing highly personalized epileptic seizure prediction and cognitive function evaluation. This adaptive learning mechanism based on individual characteristics breaks through the limitations of traditional evaluation methods, avoids the defects of relying on expert subjective judgment, and improves the objectivity and accuracy of evaluation.

[0164] 3. Comprehensive fusion and in-depth analysis of multi-modal data: Traditional epileptic seizure prediction and cognitive function evaluation systems often rely only on a single data source, such as electroencephalogram or physiological parameters, and ignore the potential value of other relevant data. Through the multi-modal data fusion method of the present invention, the system can comprehensively analyze various physiological signals such as electroencephalogram, motion data (such as gait, posture), heart rate and blood pressure, and extract richer and more comprehensive feature information, thereby improving the multi-dimensional analysis ability of seizure prediction and cognitive evaluation.

[0165] 4. Optimization and Adaptive Capability of Intelligent Algorithms: The present invention utilizes advanced Neural Architecture Search (NAS) and self-supervised contrastive learning algorithms to intelligently process and optimize multimodal data. Traditional methods usually rely on fixed algorithms and models, making it difficult to adapt to the physiological changes and data environments of different patients. However, through the adaptive optimization of intelligent algorithms, the present invention can adjust algorithm parameters according to real-time data and dynamically select the optimal network architecture through neural architecture search, improving the learning efficiency and processing ability of the model, enabling the system to flexibly handle various complex situations.

[0166] 5. Support for Large-scale and Real-time Data Processing and Analysis: The present invention introduces edge computing and distributed processing technologies, enabling efficient processing of large-scale and complex multimodal data. Through the distributed intelligent quantization analysis module and the adaptive intelligent prediction module, the system can real-time process a large amount of data from multiple data sources and can perform efficient real-time analysis and prediction according to requirements. Compared with traditional technologies, the present invention has stronger data processing capabilities and can achieve real-time and accurate seizure prediction and cognitive function assessment in practical applications.

[0167] 6. Simplification of the Clinical Diagnosis Process and Auxiliary Decision-making: The personalized health assessment report generated by the present invention is not only a prediction of epileptic seizures but also includes a detailed assessment of cognitive function status. Through the synchronous analysis of epileptic seizures and cognitive function, the system can provide a comprehensive health status assessment report for clinicians, helping doctors formulate personalized treatment plans. In addition, combined with the adaptive ability of intelligent algorithms, the system can real-time adjust the prediction results and assessment suggestions, thus providing forward-looking auxiliary decision-making support for doctors.

[0168] 7. The relevance of seizure auxiliary prediction and cognitive function assessment is reflected in physiological signal sources, data processing methods, algorithm models, and practical application scenarios. Seizure prediction can directly reflect the neurological health status of patients, while cognitive function assessment can supplement the assessment of the long-term impact of seizures on the brain (such as impaired memory, attention, etc.). Through multimodal data fusion and adaptive optimization technologies, the present invention can dynamically evaluate cognitive function changes on the basis of real-time seizure prediction, providing an efficient solution for precision medicine and personalized health management. The combination of the two can not only enhance the functional diversity of the system but also bring significant performance improvements in task collaboration and resource utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0169] Figure 1 : Overall flowchart of the present invention;

[0170] Figure 2 : Flowchart of the long short-term memory (LSTM) algorithm introduced by the present invention;

[0171] Figure 3: Flowchart of the adaptive optimization algorithm introduced in the present invention;

[0172] Figure 4 : Experimental simulation diagram of self-supervised contrast learning introduced in the present invention. (a) Diagram of the change in inter-class and intra-class distances; (b) Change in task loss; (c) Diagram of the accuracy of epilepsy prediction and cognitive assessment. Detailed implementation manners

[0173] As Figures 1 - 3 shown, the present invention provides a system and method for epilepsy seizure prediction and cognitive function assessment based on multi-modal data fusion and intelligent adaptive optimization. The system aims to perform epilepsy seizure prediction and cognitive function assessment through the comprehensive acquisition of multi-modal physiological data, intelligent data preprocessing and analysis, and advanced deep learning algorithms. By collecting and extracting features from multiple physiological signals, the present invention provides an efficient and accurate solution, providing a scientific basis for clinical diagnosis.

[0174] In the intelligent algorithm module, the present invention combines neural architecture search (NAS) and self-supervised contrast learning. For the tasks of epilepsy seizure prediction and cognitive function assessment, it automatically optimizes the neural network structure and enhances the model's ability to recognize complex patterns through unsupervised learning. Neural architecture search can automatically select the most suitable neural network architecture according to different input data, optimize the depth, number of layers, and number of nodes of the model, ensuring that the system can achieve efficient and accurate prediction on large-scale datasets; self-supervised contrast learning, through a self-learning mechanism, uses unlabeled data to strengthen the learning of potential relationships between data, improving the system's ability to handle abnormal patterns and complex situations.

[0175] To further improve the performance of the system, the present invention also introduces a multi-task learning and optimization module. This module uses multi-task learning technology to simultaneously optimize the tasks of epilepsy seizure prediction and cognitive function assessment through a shared neural network model. The shared feature information and optimization process can improve the prediction accuracy of each task, while ensuring that the system achieves balance and optimal performance between different tasks. Finally, the prediction and evaluation module generates real-time epilepsy seizure prediction results and cognitive function assessment reports according to the optimized model, providing personalized decision support for clinicians.

[0176] Multi-modal Physiological Data Acquisition and Sensing Module: This module collects the multi-modal physiological data of patients in real time through a variety of sensing devices. Specifically, an electroencephalogram (EEG) device is used to record the brain waves of patients, obtaining time-series EEG signals, and focusing on the amplitude and frequency changes of different brain waves. The amplitude range of the EEG signal is 50μV to 200μV, and the frequency range is 0.5Hz to 40Hz. During epileptic seizures, the frequency of the brain waves may show sudden changes, and the frequency may reach the range of 5Hz to 10Hz. At the same time, accelerometers and gyroscope wearable devices are used to monitor the movement data of patients, recording gait, posture, and movement patterns. The gait speed range of the movement data is 0.5m / s to 2.5m / s, and the accuracy range of the posture change data collected by the sensor is ±5 degrees. In addition, the system also uses physiological parameter monitoring devices such as heart rate and blood pressure to collect the physiological data of patients to assist in analyzing the correlation between epileptic seizures and cognitive function changes. The heart rate range is 60bpm to 100bpm, and the blood pressure range is 90 / 60mmHg to 130 / 85mmHg.

[0177] The comprehensive acquisition of these multi-modal data provides rich input for subsequent data processing and analysis, and provides comprehensive data support for epileptic seizure prediction and cognitive function assessment.

[0178] Dynamic Data Standardization and Denoising Preprocessing Module: The main function of this module is to standardize and denoise the raw data obtained from the sensing data acquisition module to ensure the accuracy and consistency of subsequent analysis. The EEG signal removes high-frequency and low-frequency noise through a digital filter, and the frequency range of the noise is 0Hz to 0.5Hz and 40Hz to 100Hz, ensuring that the signal-to-noise ratio of the data reaches a high level. Standardization techniques are used for the movement data and physiological data to unify the data to the same dimension, making the comparison between various types of data more fair and reasonable. After standardization, the unit of the movement data is m / s, and the units of blood pressure and heart rate after standardization are mmHg and bpm, making the data comparable.

[0179] For signals with missing or poor quality, interpolation techniques are used to complete or correct them to improve the data quality and integrity. This module uses an adaptive learning algorithm to dynamically adjust the data processing flow according to the real-time collected data and environmental changes to ensure that the preprocessing can effectively handle different types of input data.

[0180] Multi-dimensional Feature Extraction and Deep Fusion Module: Based on the multi-modal data obtained from the Perceptual Data Acquisition Module, this module uses a variety of feature extraction techniques to extract effective information from it. For electroencephalogram (EEG) signals, time-domain, frequency-domain, or time-frequency-domain analysis methods are used to extract the power spectral density and waveform change characteristics of brain waves related to epileptic seizures. The frequency range of the power spectral density of brain waves is from 0.5 Hz to 40 Hz, and the fluctuation amplitude of the power spectral density during epileptic seizures may reach 0.2 μV 2 to 3.0 μV 2 . For motion data, methods such as Dynamic Time Warping (DTW) and acceleration feature analysis are used to extract features related to cognitive function, such as gait, posture, and movement speed. The standardized feature value range of gait is from 0 to 1, reflecting the fluency and consistency of the patient's gait.

[0181] Through these feature extractions, the multi-dimensional feature vectors generated by the system will be transmitted to the Data Fusion Module for further integration. Feature fusion effectively combines EEG signals, motion data, and physiological data through weighted average technology to generate a unified, multi-dimensional feature representation, providing comprehensive input for subsequent intelligent analysis.

[0182] Intelligent Network Architecture Search and Self-supervised Learning Optimization Module: This module combines two advanced intelligent algorithms, Neural Architecture Search (NAS) and self-supervised contrast learning, to optimize the feature data. The depth of the neural network architecture is from 3 to 12 layers, and the number of nodes in each layer ranges from 64 to 1024. The most suitable network structure is automatically selected through neural architecture search to improve the performance of the epileptic seizure prediction and cognitive function assessment models.

[0183] The self-supervised contrast learning method conducts self-learning through unlabeled data. The contrast loss function designed for the self-supervised task ranges from 0 to 1. By maximizing the inter-class difference and minimizing the intra-class difference, the feature learning ability of the model is enhanced, and the network's ability to recognize complex patterns is improved.

[0184] Adaptive Multi-task Optimization and Learning Module: This module uses multi-task learning technology to optimize the learning processes of two tasks, epileptic seizure prediction and cognitive function assessment. The accuracy rate of epileptic seizure prediction ranges from 80% to 95%, and the accuracy rate of cognitive function assessment ranges from 75% to 90%. The system simultaneously conducts the learning of these two tasks in the same network through a shared neural network model. Through the shared parameters and feature information, the two tasks can promote each other and improve the prediction accuracy of each task. Through adaptive optimization technology, the system can automatically adjust the weight allocation and resource allocation between tasks according to the complexity of the tasks and the non-linear characteristics of the data, ensuring the optimal learning effect among different tasks.

[0185] Real-time Prediction and Multidimensional Health Assessment Module: This module uses an optimized model to predict and evaluate real-time data. For seizure prediction, the prediction accuracy rate of EEG signals ranges from 80% to 95%, and the prediction ranges for heart rate and blood pressure are 60 bpm to 100 bpm (heart rate) and 90 / 60 mmHg to 130 / 85 mmHg (blood pressure). For cognitive function assessment, the system combines the patient's EEG signals, motion data, and physiological data to evaluate the patient's cognitive state. Cognitive function assessment: The assessment accuracy rate is 75% to 90%. Report generation time: The generation time is 1 second to 5 seconds. The prediction and evaluation module generates a personalized health status assessment report based on real-time data and provides the prediction results of seizures and the assessment of cognitive state.

[0186] Specific Numerical Example Analysis:

[0187] 1. Data Acquisition (S1) Numerical Acquisition Collect the following types of data:

[0188] As shown in Table 1-3, for EEG signals: The sampling frequency of the EEG device used is 500 Hz, that is, 500 data points are collected per second, the duration is 10 seconds, and the total number of collected data points is 5000.

[0189] Table 1 EEG Signal Acquisition

[0190] Timestamp 0.000 0.002 0.004 ... 9.998 EEG signal value 0.0012 0.0034 -0.0021 ... 0.0053

[0191] Table 2 Motion Data (Acceleration)

[0192] Timestamp 0.000 0.010 0.020 ... 9.990 Acceleration X 0.12 0.13 0.11 ... 0.14 Acceleration Y -0.04 -0.05 -0.03 ... -0.06 Acceleration Z 0.10 0.11 0.09 ... 0.12

[0193] The sampling frequency of the accelerometer is 100 Hz, the acquisition time is 10 seconds, and the total number of data points is 1000.

[0194] Table 3 Physiological Data (Heart Rate, Blood Pressure)

[0195] Timestamp 0.000 1.000 2.000 ... 9.000 Heart rate (bpm) 72 74 76 ... 71 Blood pressure (mmHg) 120 / 80 118 / 78 122 / 82 ... 119 / 79

[0196] The heart rate monitoring device samples once per second for 10 seconds, with a total of 10 data points.

[0197] 2. Data Preprocessing (S2) Numerical Example

[0198] 2.1 Signal Denoising

[0199] As shown in Table 4-, for EEG signals, a band-pass filter is used for denoising, and the frequency band of the filter is 0.5 - 50 Hz to remove unnecessary low-frequency and high-frequency noise.

[0200] Table 4 Partial Data of Original EEG Signals

[0201] Timestamp 0.000 0.002 0.004 ... 9.998 EEG signal value 0.0012 0.0034 -0.0021 ... 0.0053

[0202] After band-pass filtering, the signal changes to: timestamp, EEG denoised signal value, as shown in Table 5:

[0203] Table 5 EEG denoised signal values

[0204] Timestamp 0.000 0.002 0.004 ... 9.998 EEG signal value 0.0007 0.0031 -0.0005 ... 0.0048

[0205] 2.2 Missing value imputation

[0206] If there are missing values in the EEG signal at certain timestamps, the missing values can be imputed by linear interpolation. For example, if the EEG value at the 3rd timestamp is missing, we can interpolate based on the previous and next values.

[0207] By linear interpolation, the EEG value at the 3rd timestamp can be interpolated to 0.002.

[0208] 2.3 Data normalization and standardization

[0209] The original value range of the EEG signal is [-0.05, 0.05]. We use Min-Max normalization to map the data to the [0, 1] interval:

[0210] Original signal: [-0.05, 0.05] Normalized signal: [0, 1]

[0211] 3. Feature extraction and fusion (S3) numerical example

[0212] 3.1 EEG signal feature extraction

[0213] Extract time-domain features, frequency-domain features, and time-frequency domain features from the EEG signal.

[0214] Time-domain features: Calculate the mean, standard deviation, maximum value, etc. of the EEG signal. The time-domain features of the EEG signal are as follows:

[0215] Mean: 0.001

[0216] Standard deviation: 0.005 Maximum value: 0.03 Minimum value: -0.03

[0217] Frequency-domain features: Calculate the power spectral density. The power spectral density within a certain frequency band is:

[0218] Delta wave (0.5 - 4 Hz): 0.02 Theta wave (4 - 8 Hz): 0.05

[0219] Alpha wave (8 - 13 Hz): 0.15 Beta wave (13 - 30 Hz): 0.10

[0220] These frequency-domain features can help identify the precursors of epileptic seizures.

[0221] 3.2 Motion data feature extraction

[0222] By extracting features from the acceleration data, such as mean, standard deviation, maximum value, etc., the following features are obtained:

[0223] Mean acceleration on the X-axis: 0.13

[0224] Standard deviation of acceleration on the Y-axis: 0.05 Maximum acceleration on the Z-axis: 0.14

[0225] 3.3 Physiological data feature extraction

[0226] The following features are extracted from the heart rate and blood pressure data: Mean heart rate: 73 bpm

[0227] Mean blood pressure (systolic, diastolic): 121 / 80 mmHg

[0228] 4. Numerical example of feature optimization (S4)

[0229] 4.1 Optimization by Neural Architecture Search (NAS)

[0230] Use NAS to optimize the hyperparameters of the LSTM model. The finally selected architecture is:

[0231] Number of LSTM layers: 3 layers

[0232] Number of LSTM units in each layer: 128

[0233] Optimal learning rate: 0.001

[0234] Batch size: 64

[0235] 4.2 Self-supervised contrastive learning

[0236] Use self-supervised contrastive learning to maximize the similarity between similar sample pairs. The loss value after contrastive learning optimization is 0.15, indicating that the model has a good learning of the feature representation of the samples.

[0237] 5. Multi-task learning and adaptive optimization (S5)

[0238] 5.1 Model architecture and shared features

[0239] In this module, a multi-task learning framework is used, which includes two main tasks: seizure prediction task and cognitive function assessment task; these two tasks are co-optimized through the underlying feature extraction layer that shares the neural network architecture. The adopted neural network structure is a hybrid architecture that combines a convolutional neural network (CNN) and a long short-term memory (LSTM) network.

[0240] Shared underlying feature extraction layer: This layer processes multimodal data (such as EEG signals, motion data, heart rate, blood pressure, etc.) and extracts low-level features of these data. The input data dimension is: sampled once per second, continuous data for 5 minutes, containing 4 features (EEG signals, motion data, heart rate, blood pressure), that is, each time step contains 4 input features. There are 60 time steps per minute, and there are a total of 300 time steps for 5 minutes of data. The input matrix dimension of the 4 features is (300, 4).

[0241] 5.2 Loss Function and Weight Adjustment for Multi-Task Learning

[0242] The goal of multi-task learning is to simultaneously improve the performance of seizure prediction and cognitive function assessment by optimizing a comprehensive loss function. The loss function we adopt is as follows:

[0243] Loss function for seizure prediction: Cross-entropy loss function, used for binary classification tasks, with the goal of minimizing the error between the prediction and the actual label.

[0244] Loss function for cognitive function assessment: Mean Squared Error (MSE), with the goal of predicting the difference between the cognitive score and the true cognitive score. λ1 = 1.0 and λ2 = 0.5, and the weights are adjusted through an adaptive optimization algorithm in the later stage.

[0245] 5.3 Adaptive Optimization and Learning Rate Adjustment

[0246] Learning rate adjustment: Use an adaptive optimization algorithm, such as Adam or RMSprop, to automatically adjust the learning rate for each task. The initial learning rate is 0.001. After several rounds of training, the gradient of the seizure prediction task becomes larger, while the gradient of the cognitive function assessment task is smaller. Then the optimization algorithm will dynamically adjust the learning rate according to the gradient of each task. For example, for the seizure prediction task, the learning rate may increase to 0.002, while the learning rate of the cognitive function assessment task may decrease to 0.0005.

[0247] Training process:

[0248] Initial stage: λ1 = 1.0, λ2 = 0.5

[0249] Mid-stage: According to the performance of the loss function, adjust to λ1 = 1.2, λ2 = 0.8. At this time, the loss of the seizure prediction task is larger, so its weight needs to be increased.

[0250] Final stage: When the performance of the two tasks is close, λ1 and λ2 are adjusted to balanced values, for example, λ1 = 1.0, λ2 = 1.0.

[0251] 5.4 Training and Evaluation

[0252] After multiple training cycles, the changes in loss and accuracy on the validation set are shown in Table 6 below:

[0253] Table 6 Changes in Loss and Accuracy on the Validation Set

[0254]

[0255] As can be seen from the above table, after training, the accuracy of seizure prediction gradually increases, and at the same time, the mean squared error (MSE) of the cognitive score also gradually decreases, indicating that the multi-task learning framework has successfully optimized the performance of both tasks.

[0256] 6. Prediction and Evaluation (S6)

[0257] Judgment Criteria:

[0258] 1. Probability Division of Seizures (Threshold Set at 0.5)

[0259] If the probability of a seizure is 0.5, assume this is a critical value for distinguishing whether a seizure has occurred. At this critical point, changes in physiological and motion data can be judged.

[0260] Electroencephalogram (EEG) Changes

[0261] Normal EEG (Probability < 0.5): The electroencephalogram shows normal frequency and amplitude, normal a waves and β waves, and normal fluctuations in θ waves and δ waves. There is no obvious high-frequency, sharp wave, or spike activity.

[0262] Seizure (Probability ≥ 0.5): The electroencephalogram shows abnormal discharge activities in frequency, such as high-frequency discharges (sharp waves or spikes), frequent θ waves and δ waves, and even local or whole-brain synchronous abnormal fluctuations.

[0263] Motion Data (Accelerometer, Gyroscope)

[0264] Normal Motion Characteristics (Probability < 0.5): Smooth gait, consistent motion pattern, and stable posture. There are no significant incoordinations or involuntary movements.

[0265] Motion Changes after Seizure (Probability ≥ 0.5): During or after a seizure, there are uncoordinated gaits, frequent falls, or sudden muscle twitches and loss of motor control. The motion data will show abnormal motion patterns.

[0266] Physiological Data (Heart Rate, Blood Pressure, etc.)

[0267] Normal Physiological Data (Probability < 0.5): Stable heart rate and blood pressure, without obvious fluctuations. The heart rate is within the normal range (60 - 100 beats per minute), and the blood pressure is maintained at a normal level.

[0268] Physiological changes during epileptic seizures (probability ≥ 0.5): increased heart rate (≥ 100 bpm), elevated blood pressure, accompanied by the occurrence of orthostatic hypotension or arrhythmia. During seizures, sympathetic nerve excitation causes these changes.

[0269] 2. Classification of cognitive function changes

[0270] According to the cognitive score (in the range of 0 to 1.0), cognitive function can be divided into three categories, and different thresholds are set through electroencephalogram, movement data, and physiological data.

[0271] Mild cognitive impairment (cognitive score [0, 0.3]):

[0272] Electroencephalogram (EEG): There is a mild increase in theta waves, a decrease in alpha and beta waves. In particular, the amplitude of alpha waves may become smaller. The amplitude of theta waves (4 - 7 Hz) increases, showing enhanced symmetry in the electroencephalogram, representing mild cognitive function decline.

[0273] Movement data:

[0274] Patients with mild cognitive impairment may show a slightly slower gait and slightly poorer motor coordination. The stability and movement pattern of the gait may decline, but not to the extent of obvious instability or falls.

[0275] Judgment: The posture and steps are relatively stable, but there may be a slight motor retardation.

[0276] Physiological data:

[0277] The heart rate and blood pressure may fluctuate slightly, but usually within the normal range.

[0278] Slight autonomic nervous system disorders are observed, such as orthostatic hypotension, but it does not seriously affect normal physiological activities.

[0279] Moderate cognitive impairment (cognitive score [0.3, 0.7]):

[0280] Electroencephalogram (EEG): There is an obvious increase in theta and delta waves, and a significant decrease in alpha and beta waves. There will be more electroencephalogram asymmetry or local brain wave abnormalities, indicating moderate impairment of cognitive function.

[0281] Judgment: Frequent theta and delta wave activities indicate a heavy cognitive load in the patient, and there may be a decrease in attention to the environment.

[0282] Movement data:

[0283] The gait becomes slower, and there may be obvious gait instability, motor retardation, and increased reaction time. The patient may have difficulty maintaining long - term walking or completing complex motor tasks.

[0284] Judgment: The patient's movement pattern may become more uncoordinated, possibly accompanied by repetitive movements while walking or difficulty in maintaining balance.

[0285] Physiological data:

[0286] Fluctuations in heart rate and blood pressure may be more obvious, especially during postural changes, presenting symptoms such as orthostatic hypotension and elevated heart rate.

[0287] Sometimes it may be accompanied by arrhythmia or large blood pressure fluctuations.

[0288] Normal cognitive state (cognitive score [0.7, 1.0]):

[0289] Electroencephalogram (EEG):

[0290] Judgment: The amplitude and frequency of the electroencephalogram are normal, without obvious abnormal waveforms.

[0291] Movement data:

[0292] Judgment: Posture is stable, movement pattern is normal, without obvious fall risk or movement incoordination.

[0293] Physiological data: Heart rate and blood pressure are within the normal range, without significant fluctuations.

[0294] 6.1 Seizure prediction

[0295] In the seizure prediction task, the LSTM model is adopted, which can process time series data and consider the influence of historical data on current prediction. The inputs of the model include the following multimodal data: EEG signals, movement data, heart rate, blood pressure, etc., and the data dimension is (300, 4).

[0296] Model output: The output of the LSTM is the seizure probability at each time step, obtained through the Sigmoid activation function. The prediction results of the trained LSTM model on the test set are shown in Table 7 below:

[0297] Table 7 Seizure probability and actual seizure situation

[0298] 10 20 30 40 50 15 12 60 75 10 No seizure No seizure Seizure Seizure No seizure

[0299] According to the prediction probability, a threshold τ = 0.5 is set, and an early warning is triggered when the onset probability > 0.5.

[0300] 6.2 Cognitive function assessment

[0301] For cognitive function assessment, the model generates a cognitive score by analyzing the patient's multimodal data. The cognitive scores output by the model are shown in Table 8 below:

[0302] Table 8 Cognitive score assessment results at each time step

[0303] 10 20 30 40 50 85 45 25 75 60 Normal cognition Moderate cognitive impairment Mild cognitive impairment Normal cognition Moderate cognitive impairment

[0304] The cognitive score ranges from [0, 1] and is interpreted according to the following rules:

[0305] 0 ≤ cognitive score ≤ 0.30: Mild cognitive impairment

[0306] 0.30 < cognitive score ≤ 0.70: Moderate cognitive impairment 0.70 < cognitive score ≤ 1.00: Normal cognition

[0307] 6.3 Comprehensive assessment and system feedback

[0308] Seizure warning: When the system predicts that the seizure probability > 0.5, the system will trigger a "high-risk" warning and recommend that the patient take emergency preventive measures.

[0309] Cognitive status feedback: Based on the patient's cognitive score, the system can provide feedback: If the cognitive score indicates mild or moderate cognitive impairment, it is recommended to further examine or treat.

[0310] As Figure 4 shown in (a), the intra-class and inter-class distance changes: The red line represents the intra-class distance gradually decreasing, and the green line represents the inter-class distance gradually increasing. This indicates that as the model is trained, the distribution of features in space is gradually optimized, the samples between classes are more separated, and the samples within classes are more concentrated. The decrease in intra-class distance and the increase in inter-class distance are important indicators to measure the success of self-supervised learning.

[0311] As Figure 4 shown in (b), the loss changes of the seizure prediction and cognitive assessment tasks: The blue line shows that the loss of the seizure prediction task gradually decreases with training, and the red line shows that the loss of the cognitive assessment task also gradually decreases. The continuous decrease in the loss value indicates that the model converges well during the learning process and reaches a high level of optimization.

[0312] As Figure 4 shown in (c), the accuracy improvement of seizure prediction and cognitive assessment: The purple line shows that the seizure prediction accuracy gradually increases from the initial 80% to 95%, and the orange line shows that the cognitive assessment accuracy increases from the initial 75% to 90%. The training process significantly improves the prediction performance of the two core tasks. It shows that under multi-modal data fusion and self-supervised learning optimization, the model's performance for different tasks has been significantly improved.

[0313] Through the deep integration of a systematic data processing process and intelligent algorithms, the present invention achieves precise feature extraction and quantitative analysis for epilepsy seizure prediction and cognitive function assessment, thereby providing efficient and objective support for clinical decision-making. By innovatively combining neural architecture search (NAS) with self-supervised contrastive learning algorithms, the system can provide high-precision prediction and assessment in a complex multi-modal data environment, significantly improving the prediction accuracy of epilepsy seizures and the comprehensiveness of cognitive function assessment.

[0314] The intelligent algorithm module of the present invention adopts a variety of programming environments and deep learning frameworks (such as TensorPlow, PyTorch, etc.), and is jointly deployed with efficient optimization algorithms (such as neural architecture search) to achieve efficient feature extraction, data fusion, and prediction analysis. These algorithm modules can run on high-performance computing platforms, support parallel computing and large-scale data processing, and ensure accurate and efficient results in a complex multi-modal data environment. Core algorithms, such as data preprocessing, feature extraction, intelligent optimization, and prediction, can be implemented through function or library calls in programming languages such as MATLAB and Python, and have the ability to adaptively adjust parameters to adapt to different clinical data inputs and application scenarios. By introducing advanced technologies such as self-supervised learning and neural architecture search, the present invention can improve the learning ability of the model without a large amount of labeled data, and at the same time automatically optimize the network structure, thereby improving prediction accuracy and efficiency. The system can dynamically adjust the model structure according to the input data and task requirements to ensure more accurate epilepsy seizure prediction and cognitive function assessment. In addition, the algorithm module has high flexibility and scalability, and supports running on common operating systems (such as Windows, Linux, etc.), ensuring the wide applicability of the technical solution on different hardware platforms.

[0315] To verify the effectiveness and advantages of the present invention, a large number of non-algorithm experimental verifications have also been carried out on the present invention, and tests are combined with actual patient data. The results show that the system of the present invention can significantly improve the accuracy of epilepsy seizure prediction, greatly extend the prediction lead time, and effectively reduce the false negative rate and false positive rate. At the same time, the accuracy of cognitive function assessment has also been improved from the traditional 60% to over 85%, which can accurately identify changes in cognitive status and warn of the trend of cognitive decline. Through the present invention, the system can effectively handle complex multi-modal physiological data in the process of epilepsy seizure prediction and cognitive function assessment, realize the prediction of patients' epilepsy seizures, the real-time assessment of cognitive status, and the dynamic analysis of related health conditions, improving the accuracy, real-time performance, and personalization level of medical diagnosis, and providing a new technical path for intelligent medicine, personalized health management, and precision medicine.

Claims

1. An epileptic seizure auxiliary prediction and cognitive function assessment system based on multimodal data fusion and intelligent adaptive optimization, characterized by: The system consists of multiple modules that work together, including a multimodal physiological data acquisition and perception module, a dynamic data standardization and denoising preprocessing module, a multidimensional feature extraction and deep fusion module, an intelligent network architecture search and self-supervised learning optimization module, an adaptive multi-task optimization and learning module, and a real-time prediction and multidimensional health assessment module. Through intelligent data collection and analysis, it provides accurate results for epileptic seizure prediction and cognitive function assessment. The multimodal physiological data acquisition and perception module collects the patient's multimodal physiological data in real time through a variety of sensing devices; specifically, it uses electroencephalogram equipment to record the patient's brain waves and obtain time series EEG signals, including the amplitude and frequency changes of different brain waves; it uses accelerometers and gyroscopes to monitor the patient's motion data, record gait, posture and movement patterns, to reflect the patient's movement characteristics and cognitive function status; at the same time, the module also uses heart rate and blood pressure physiological parameter monitoring equipment to collect the patient's physiological data, in order to assist in analyzing the correlation between epileptic seizures and cognitive function changes; Through the comprehensive collection of these multimodal data, this module provides rich input for subsequent data processing and analysis; The dynamic data standardization and denoising preprocessing module is used to standardize and denoise the raw data obtained from the sensory data acquisition module to ensure the accuracy and consistency of subsequent analysis; the EEG signal will remove high-frequency and low-frequency noise through a digital filter to ensure that the signal-to-noise ratio of the data reaches a high level; for motion data and physiological data, standardization technology is used to unify the data to the same dimension; for missing and poor-quality signals, interpolation technology is used to complete or correct them; this module uses an adaptive learning algorithm to dynamically adjust the processing flow according to real-time collected data and environmental changes to ensure that the preprocessing process can effectively cope with different types of input data; The multi-dimensional feature extraction and deep fusion module uses a variety of feature extraction techniques to extract effective information from the multi-modal data obtained from the perception data acquisition module. For EEG signals, the power spectrum density and waveform change characteristics of EEG related to epileptic seizures are extracted. For motion data, dynamic time warping and acceleration feature analysis methods are used to extract features related to gait, posture, movement speed and cognitive function. At the same time, through the analysis of heart rate and blood pressure physiological data, physiological characteristics that affect cognitive state and epileptic seizures are extracted. feature Fusion effectively combines EEG signals, motion data, and physiological data through weighted averaging technology to generate a unified, multi-dimensional feature representation, thereby providing comprehensive input for subsequent intelligent analysis; The intelligent network architecture search and self-supervised learning optimization module optimizes feature data by combining two advanced intelligent algorithms: neural architecture search and self-supervised comparative learning. During the neural architecture search process, the system automatically selects the most suitable neural network architecture for epilepsy seizure prediction and cognitive function assessment model construction. Through the neural architecture search algorithm, the system can automatically adjust the depth, number of layers and number of nodes of the neural network to find the most suitable architecture for the input data. At the same time, the self-supervised comparative learning method is used to learn the potential relationship between data based on unlabeled data. The adaptive multi-task optimization and learning module uses multi-task learning technology to optimize the learning process of the two tasks of epilepsy prediction and cognitive function assessment. The system uses a shared neural network model to simultaneously perform epilepsy prediction and cognitive function assessment tasks in the same network. Through shared parameters and feature information, the two tasks can promote each other, thereby improving the prediction accuracy of each task. Adaptive optimization technology is used to automatically adjust the weight distribution and resource allocation between tasks according to the complexity of the tasks and the nonlinear characteristics of the data. The real-time prediction and multi-dimensional health assessment module predicts and evaluates the optimized model in real time. For epileptic seizure prediction, the system uses the trained model to predict whether the patient will have an epileptic seizure based on the collected EEG signals and physiological data. For cognitive function assessment, the system evaluates the patient's cognitive status based on the patient's EEG signals, movement data and physiological data; the prediction and evaluation module generates a health status assessment report for the patient based on real-time data.

2. A method for epileptic seizure auxiliary prediction and cognitive function assessment system based on multimodal data fusion and intelligent adaptive optimization, characterized in that: The following steps are involved: S1. Data collection: Through the multimodal physiological data collection and perception module, the EEG device, accelerometer, gyroscope and physiological monitoring equipment are used to collect the patient's EEG signals, movement posture data, and heart rate and blood pressure physiological parameter data in real time; The collected data is stored in the system's data storage unit in real time through encrypted transmission, ensuring the security and integrity of the data, providing a solid foundation for subsequent data processing and analysis; S2. Data preprocessing: The raw data obtained from the sensory data acquisition module is processed through the dynamic data standardization and denoising preprocessing module, and the noise in the EEG signal is removed by using a digital filtering algorithm; the motion and physiological data are standardized to eliminate dimensional differences; for signals of poor quality or missing, interpolation is used to complete them to ensure the accuracy of subsequent analysis; S3. Feature extraction and fusion: Through the multi-dimensional feature extraction and deep fusion module, based on the multi-modal data obtained from the perception data acquisition module, the relevant features of epileptic seizures and cognitive function assessment are extracted; for EEG signals, the relevant features of epileptic seizures are extracted; for motion data, gait and posture features are extracted using dynamic time warping and acceleration feature extraction methods; physiological data are combined to extract physiological features that affect cognitive function and epileptic seizures; all extracted features are generated into a unified multi-dimensional feature vector through deep fusion technology, providing efficient input for subsequent intelligent analysis; S4. Feature optimization: After feature extraction and fusion, the extracted feature data is optimized using the intelligent network architecture search and self-supervised learning optimization module. The neural architecture search algorithm is used to automatically select the most suitable neural network architecture, optimize the network depth, number of layers, and number of nodes, and improve the accuracy and stability of the model. The self-supervised comparative learning algorithm is used to learn the potential relationship between data on unlabeled data to enhance the feature learning ability of the model. S5. Multi-task learning and adaptive optimization: Through the adaptive multi-task optimization and learning module, the system simultaneously performs epileptic seizure prediction and cognitive function assessment tasks in the same neural network model; by sharing network parameters and feature information, the two tasks can promote each other, thereby improving prediction accuracy; through adaptive optimization technology, the weights and computing resource allocation between tasks are automatically adjusted; S6. Prediction and health assessment: Based on the optimized model, the system uses real-time data to predict epileptic seizures and evaluate cognitive functions. For epileptic seizure prediction, the system uses real-time collected EEG signals and physiological data to predict whether the patient will have an epileptic seizure and generate corresponding warnings. For cognitive function assessment, the system evaluates the patient's cognitive status based on EEG, motor data and physiological data; the predicted results and evaluation information will be integrated to generate personalized health reports to provide effective decision-making support for doctors.

3. The method according to claim 2, characterized in that S1 specifically includes: Real-time collection of EEG signals, gait, posture, acceleration motion data, heart rate, and blood pressure physiological parameters, providing accurate input for subsequent intelligent data processing and analysis; EEG signal acquisition: Use EEG equipment to collect time series EEG signals, focusing on the amplitude and frequency changes of Delta, Theta, Alpha, and Beta waves; these signals contain key features related to epileptic seizures; Motion data and posture data collection: Motion data collection is mainly completed through wearable devices such as accelerometers and gyroscopes. These devices monitor the patient's gait, posture, and acceleration movement behavior in real time, which can provide additional physiological clues for analyzing the patient's cognitive function status. Accelerometers are used to measure the patient's movement speed and acceleration; gyroscopes are used to record the patient's posture and direction changes. Physiological data collection: Heart rate and blood pressure are important physiological indicators that reflect the health status of the human body, and can provide auxiliary data for epileptic seizure and cognitive function assessment; heart rate monitoring equipment measures the patient's heart rate in real time in bpm; blood pressure monitoring equipment monitors the patient's systolic and diastolic blood pressure in real time in mmHg; Data transmission: All collected data will be uploaded to the system's storage unit in a secure and encrypted manner. Data transmission supports wired or wireless transmission. Wired transmission uses a standard data interface for high-stability transmission; wireless transmission uses Wi-Fi, Bluetooth, and Zigbee wireless communication technologies. Data format: EEG signal: time series data, the format is [timestamp, EEG signal value]; Motion data: time series data in the format of [timestamp, acceleration value, speed value, posture angle] Physiological data: time series data in the format of [timestamp, heart rate value, blood pressure value]; System storage unit: All collected physiological data will be stored in a dedicated storage unit after being transmitted to the system.

4. The method according to claim 2, characterized in that: S2 specifically includes: Signal denoising: Eliminate irrelevant noise introduced by various factors during the acquisition process; the denoising methods will be different for different types of EEG signals, motion data, and heart rate data sensor data; Denoising of EEG signals: EEG signals contain high frequency noise from power grid noise and low frequency noise from baseline drift; in order to remove these noises, digital filtering technique band pass filter is used; Bandpass filter formula: The signal after the EEG signal is processed by the bandpass filter is x filtt , by designing an appropriate cutoff frequency f c And order n, remove the noise in a specific frequency band and retain the EEG signal that is useful for predicting epileptic seizures; the transfer function of the bandpass filter is expressed as: Where: H(f) is the filter transfer function, f is the signal frequency, and f c is the cutoff frequency of the filter, n is the order of the filter; Denoising of motion data and physiological data: The noise of motion data and physiological data is manifested as high-frequency vibration and sensor error. The sliding average method is used to perform sliding average processing on the acceleration signal of the motion data to eliminate short-term fluctuations. The sliding average denoising formula is: Where y(t) is the original signal and M is the size of the sliding window. By selecting an appropriate window size M, high-frequency noise can be removed and the motion signal can be smoothed. Missing value patching: Due to equipment failure, sensor data loss, and signal interruption, there will be missing values ​​in the data; patching of missing values ​​is crucial to ensure the accuracy of subsequent analysis; for time series data, linear interpolation is used to patch missing values. For missing data between time points t0 and t1, linear interpolation is performed using the following formula: Among them, x(t) is the supplementary data, x(t0) and x(t1) are known adjacent data points, t0 and t1 are the corresponding timestamps, and t is the timestamp of the missing data; Data normalization and standardization: Different types of EEG signals, motion data, heart rate data, and physiological data have different dimensions and value ranges. In order to compare and integrate different data at the same scale, data normalization and standardization are required; Data normalization: Map all EEG signals, motion data, and heart rate data to a unified interval; the input data is X = [x1, x2, ..., x N ], the normalization formula is: Among them, min(X) and max(X) are the minimum and maximum values ​​of the data set X, respectively. i ′ is the normalized data; through this method, all input data are compressed to the interval [0, 1], which is convenient for subsequent fusion analysis; Data standardization: transform EEG signals, motion data, and heart rate data into a distribution with a mean of 0 and a standard deviation of 1; for the data set X = [x1, x2, ..., x N ], the standardized formula is: Among them, μ(X) and σ(X) are the mean and standard deviation of the data set X, respectively, and x′ i It is the standardized data; Adaptive learning algorithm: The denoising, missing value patching, and normalization steps in the data preprocessing process need to be adjusted according to the changes in real-time data and environment; therefore, the system uses an adaptive learning algorithm to dynamically adjust the preprocessing process; the algorithm automatically selects the appropriate preprocessing method based on the characteristics of the input data; the noise level of the current data is measured by the signal-to-noise ratio, and the algorithm adjusts the filter parameters according to the changes in the current signal-to-noise ratio; the signal-to-noise ratio calculation formula is: When the SNR is low, the algorithm increases the order of the filter to perform stronger denoising; when the SNR is high, the algorithm chooses a lower filter strength.

5. The method according to claim 2, characterized in that S3 include: Electroencephalogram signal feature extraction: Electroencephalogram signals are the core data source for epileptic seizure prediction. By extracting the features of electroencephalogram signals, abnormal activity patterns related to epileptic seizures can be identified. Common electroencephalogram analysis methods include time-frequency domain analysis. Time-frequency domain analysis combines the advantages of time domain and frequency domain and can extract the frequency characteristics of signals that change over time; Wavelet transform is used to obtain the resolution of both time domain and frequency domain. For a signal x(t), the result of wavelet transform is: Among them, W φ (a, b) are the results of wavelet transform, ψ is the mother wavelet function, Φ * is the complex conjugate of ψ; a is the scale factor, and b is the translation factor; through wavelet transform, the features of the details and low-frequency parts in the signal can be extracted, and the time and frequency characteristics of epileptic seizures can be identified; Motion data feature extraction: Epilepsy attacks can affect patients’ motor abilities, so the analysis of motion data can help assess the relationship between patients’ cognitive function and epilepsy attacks. Dynamic time warping is an algorithm that measures the similarity between two time series and is used to process asynchronous motion data. There are two time series A = {a1, a2, ..., a N } and B = {b1, b2, ..., b N }, the distance of dynamic time warping is recursively calculated to obtain the minimum matching path; Among them, dist(a i , b j ) is the distance between two points, D(i, j) represents the distance from a i to b j The minimum distance; min{} represents the minimum value operation. Dynamic time warping can effectively align different gait and posture data to extract movement patterns related to epileptic seizures. Acceleration feature extraction: The data provided by the acceleration sensor is used to extract gait and posture features. Acceleration features include mean acceleration, standard deviation acceleration, and maximum acceleration. For the acceleration data at each moment, the total acceleration A(t) is calculated: Among them, a x (t) 2 , a y (t) 2 and a z (t) 2 They are the components of acceleration in the X, Y, and Z directions respectively; by extracting the total acceleration features in different time periods, the correlation with cognitive function and epileptic seizures is identified; Physiological data feature extraction,Epileptic seizures and changes in cognitive function are often accompanied by changes in physiological parameters, Heart rate characteristics: Heart rate is an important physiological indicator for evaluating the patient's nervous system status; heart rate characteristics include average heart rate and heart rate variability; heart rate variability is an indicator of the irregularity of heartbeat intervals, and the commonly used calculation method is standard deviation: Among them, RR i is the time interval between two consecutive heartbeats, RR is the mean of these time intervals, and N is the number of recorded heartbeats; Blood pressure characteristics: Blood pressure is another important indicator that reflects cognitive status and epileptic seizures. The characteristics of blood pressure data include fluctuations in systolic and diastolic blood pressure. By calculating the mean and variance statistics of the blood pressure signal, the correlation characteristics of blood pressure fluctuations on cognitive function changes and epileptic seizures are extracted. Feature fusion: All extracted features will be fused into a unified, multi-dimensional feature vector for subsequent intelligent analysis and prediction; the feature fusion method is the weighted average method; f fused =w EEG ·f EEG +w Motion ·f Motion +w Physio ·f Physio Among them, f fused : The fused feature vector; it is a comprehensive feature vector obtained by weighted averaging of the input data; f EEG : feature vector of EEG signal; f Motion : feature vector of motion data; f Physio : feature vector of physiological data; w EEG : Weight coefficient of electroencephalogram (EEG) feature; w Motion : Weight coefficient of motion data feature; w Physio : Weight coefficient of physiological data features.

6. The method according to claim 2, characterized in that S4 specifically includes: Neural Architecture Search (NAS) is used to optimize the depth, number of layers, number of nodes, and other hyperparameters of the neural network, thereby improving the model's predictive ability and ensuring high accuracy and stability in epileptic seizure prediction and cognitive function assessment; Input Data In the first step of feature optimization, the system obtains multiple physiological signals from the multimodal data acquisition and perception module, including EEG signals, motion data, heart rate and blood pressure. After appropriate preprocessing, the data is converted into feature vectors suitable for model processing and provided to the NAS algorithm for network architecture optimization. The effective fusion and processing of these multimodal data is the key to improving system performance. Initial architecture generation Generate a preliminary neural network architecture using the NAS algorithm and select a model structure suitable for processing multimodal physiological data; in this invention, a long short-term memory network LSTM is selected as the preliminary neural network architecture; Hyperparameter setting: Determine the initial hyperparameters, including the number of LSTM layers, the number of units per layer, the learning rate, and the batch size; the selection of these hyperparameters provides the basis for subsequent optimization; Number of units per floor The number of units in each layer determines the memory capacity of each LSTM layer. For epileptic seizure prediction, EEG signals usually contain multiple frequency and amplitude changes, so a higher number of units is required to capture these complex temporal features. In cognitive assessment, due to the multimodality of the task, the features of a single modality are relatively simple and the number of units is relatively small. The number of LSTM layers is L, and the number of units in each layer is N u , the output of LSTM is h l , then the operation process of the LSTM network can be expressed as: h l =LSTM l (h l-1 ,X)for l=1,2,...,L Among them, h l : This is the hidden state of layer l; h l -1: This is the hidden state of layer l-1; X: This is the input data; LSTM: This is the core operation of the long short-term memory network; L: This is the number of layers of the LSTM network; Learning rate selection and optimization The learning rate is an important hyperparameter that determines the speed of model training and convergence. In the task of epileptic seizure prediction, since the signal contains small abnormal changes, a smaller learning rate helps to improve the stability of the model and avoid skipping the local optimal solution. In cognitive function assessment, due to the complexity of multimodal data, the appropriate learning rate selection accelerates model training while avoiding overfitting. In the gradient descent algorithm, the learning rate α updates the model parameters by the following formula: Among them, θ t : This is the value of the model's parameters at the tth iteration; θ t+1 : This is the updated value of the model parameters at the t+1th iteration; α: This is the learning rate, a very critical hyperparameter that controls the step size of the parameter update in each iteration; This is the gradient of the loss function with respect to the model parameters θ t The derivative of ; the gradient represents the rate of change of the loss function at the current parameter value, indicating the direction and magnitude of the parameter adjustment: Indicates the direction in which the loss function changes fastest for each parameter; The loss function is a measure of the error between the model's predicted output and the true result. Batch size selection The batch size controls the number of samples used each time the weights are updated. In epileptic seizure prediction, due to the high-frequency fluctuations of EEG signals, a smaller batch size helps capture more detailed information. In cognitive function assessment, due to the fusion of multimodal data, a larger batch size helps improve the model's ability to learn the associations between different data modalities. In batch gradient descent, the batch size B determines the number of samples used each time the parameters are updated. The update formula for each time is: Among them, x i is the i-th training sample, y i is its corresponding label, θ0: This is the parameter of the initial model; B is the batch size; Ψ(θ0x i ,y i ) loss function, which measures the model prediction value θ0x i With label y i The error between Objective Function One of the core tasks of NAS is to optimize the neural network architecture by minimizing the loss function. In the feature optimization module, the objective function is to measure the performance of the model by the error between the network's predicted value and the true label, thereby guiding the optimization of the network architecture. The loss function is defined as follows: Where: f(x i , A) is the input x i The output of the model under architecture A, A is the current network architecture, the number of LSTM layers and units; y i is the actual label: whether an epileptic seizure occurs or not, the score of cognitive state; L is the loss function, and N is the number of samples; The goal of this loss function is to minimize the difference between the predicted value and the actual label to optimize the network architecture; in epileptic seizure prediction, the goal is to minimize the error between the seizure probability predicted by the network and the actual seizure event; in cognitive function assessment, the goal is to minimize the error between the cognitive score predicted by the model and the actual assessment result; Optimization process The NAS algorithm automatically adjusts various hyperparameters of the network architecture through multiple rounds of iterative searches, including the depth of the neural network, the number of layers, and the number of nodes in each layer. During each round of search, the network architecture is trained and evaluated, and the architecture is adjusted based on the value of the feedback loss function. NAS determines the number of layers of the LSTM network based on the complexity of the task. For complex epilepsy prediction tasks, LSTM networks with more layers and units are required to better capture long-term dependencies. The number of nodes in each layer will be optimized according to task requirements. More nodes help capture more information, but also increase computational overhead. Training and evaluation: After each adjustment, the network is trained on a specific training set and the loss function is evaluated to determine whether the architecture needs further optimization. Through multiple rounds of iterative search, NAS gradually optimizes the network architecture and ultimately determines the neural network architecture that is most suitable for epileptic seizure prediction and cognitive function assessment tasks. Applications of Self-Supervised Contrastive Learning In the epileptic seizure prediction and cognitive function assessment system, the self-supervised contrastive learning workflow in the feature optimization module includes the following key steps: Data Entry At the front end of the system, the data acquisition module will collect multimodal physiological signals from multiple physiological data sources; after preprocessing, these signals enter the feature extraction stage; for each data type, the system will extract a set of high-dimensional feature vectors to represent the data features at different time points and under different physiological states; these data from different sensors will be further fused to form a comprehensive high-dimensional feature vector; this high-dimensional feature vector will be used as input to enter the self-supervised contrastive learning module for processing to optimize the representation of the feature space; Positive and negative sample pair generation In self-supervised contrastive learning, the generation of positive and negative sample pairs is one of the core tasks; in the implementation of the present invention, the system generates "positive sample pairs" and "negative sample pairs" by transforming the original feature vectors; Positive sample pairs: sample pairs from similar situations, indicating samples that should be close in the feature space; EEG signal data of the same patient before and after an epileptic seizure, heart rate data of the same patient in the same time period; Negative sample pairs: sample pairs from different situations, representing samples that should be kept away in the feature space; EEG signal data from different patients, or movement data during epileptic seizures and non-seizures; Contrastive loss function In order to optimize the sample distribution in the feature space, the system uses the contrastive loss function to train the model. The core goal of the contrastive loss function is to maximize the similarity between positive sample pairs and minimize the similarity between negative sample pairs, thereby improving the model's ability to distinguish samples of different categories in the feature space. The contrastive loss function is usually defined as: where sin(f(x i ), f(x j )) represents sample x i and x j The similarity in feature space is usually measured using cosine similarity: exp() is the exponential function, represents the sum of all samples, N is the total number of samples; f(x i ) represents the sample x i The feature representation of .

7. The method according to claim 2, characterized in that: S5 specifically includes: In the framework of multi-task learning, a shared neural network architecture is designed, which is divided into the following parts: Shared underlying feature extraction layer: This layer generates shared low-level features by extracting features from multimodal data, such as EEG signals, motion data, heart rate, and blood pressure. These low-level features provide the same basic data input for the two tasks of epilepsy seizure prediction and cognitive function assessment. Task-specific layers: After sharing the underlying feature extraction layer, the network splits into two branches, handling epileptic seizure prediction and cognitive function assessment tasks respectively; these task-specific layers learn specific high-level features based on task requirements; the output of each branch corresponds to a different task goal; The role of the adaptive optimization algorithm in the multi-task learning framework is to automatically adjust the weights and resource allocation between tasks. Since the epileptic seizure prediction task and the cognitive function assessment task have different degrees of dependence on different features, the priority of the task changes at any time, so an adaptive mechanism is needed to balance the learning process of each task. The learning process of each task is affected by different task difficulties, data quality, and interference between tasks; therefore, the adaptive optimization algorithm dynamically adjusts the weights between tasks according to the loss function of each task and the feedback of the model; the weighted loss function is used to achieve the weight adjustment between tasks; for the epileptic seizure prediction task and the cognitive function assessment task, the total loss function is defined by the following formula: in: is the total loss function; is the loss function for the epileptic seizure prediction task; is the loss function of the cognitive function assessment task; λ1 and λ2 are the weight parameters of the epileptic seizure prediction task and the cognitive function assessment task, respectively; Adaptive learning rate adjustment The adaptive optimization algorithm also includes a learning rate adjustment mechanism to ensure that the model can effectively converge during training. By using the adaptive optimization algorithm, the learning rate of each task is dynamically adjusted according to the gradient update history of each task. The update rule is usually as follows: Where: θ t represents the model parameters; η t is the learning rate at the current moment, η t It will be adjusted according to the gradient of each task; m t and v t is the moving average and variance of the gradient calculated at time t; ∈ is a constant to prevent division by zero errors.

8. The method according to claim 2, characterized in that: S6 specifically includes: The core goal of epileptic seizure prediction is to calculate the probability of an epileptic seizure; introduce time series analysis and attention mechanism to improve the accuracy of prediction; and fully consider the temporal sequence and dependency of data through the long short-term memory network LSTM; A multi-layer LSTM model is used to predict epileptic seizures. The model input includes multimodal data: EEG signals, motion data, heart rate, blood pressure, and predicts the probability of future epileptic seizures based on the state at the past moment. The input features are: X t =(X EEG (t),X motion (t),X physiological (t)) Where: X EEG (t) represents the EEG signal at time t; X motion (t) represents the motion data at the tth moment; X physiological (t) represents the physiological data at the tth moment; the output of the LSTM model is the probability P of an epileptic seizure occurring epilepsy (t+1), that is, the probability of an epileptic seizure occurring at time t+1: P epilepsy (t+1)=σ(W epilepsy ·H t +b epilepsy ) Where: H t represents the hidden state of LSTM at the tth moment; W epilepsy and b epilepsy are the weight matrix and bias term respectively; σ is the Sigmoid activation function, and the output value is in the range of [0, 1], which indicates the probability of epileptic seizure; Prediction threshold setting and warning Once the model gives the probability of an epileptic seizure occurring, P epilepsy (t+1), set a threshold τ, when P epilepsy When (t+1)>τ, the system sends out an early warning signal, and the system issues a "high risk" warning, suggesting that the patient take emergency preventive measures; Cognitive function assessment For cognitive function assessment, the goal of improvement is to generate a quantitative cognitive score through comprehensive analysis of the patient's EEG signals, movement data, and physiological data in order to more accurately assess the patient's cognitive health status; Similarly, LSTM or other neural network structures are used to take multimodal data as input and generate cognitive scores through the model; let the output of the cognitive function evaluation model be C cognition (t), the patient's cognitive score: C cognition (t)=f(X EEG (t),X motion (t),X physiological (t),θ) Among them, C cognition (t) represents the patient's cognitive score at time t, which is usually a continuous value, indicating the patient's current cognitive state; θ represents the cognitive function assessment model parameter; Based on cognitive score C cognition (t), the system generates the following evaluation results: Mild cognitive impairment: When the cognitive score is in the range of [0, 0.3], it indicates that the patient has mild cognitive impairment; Moderate cognitive impairment: When the cognitive score is in the range of [0.3, 0.7], it means that the patient's cognitive function is moderately impaired and further treatment and evaluation are required; Normal cognitive status: When the cognitive score is within the range of [0.7, 1.0], it means that the patient's cognitive function is normal and no intervention is needed temporarily; Through the above prediction and evaluation modules, the system can more accurately predict epileptic seizures based on real-time data and provide personalized cognitive function assessment according to the patient's physiological state; these improvements not only enhance the system's predictive accuracy, but also improve its intelligence, and can provide doctors with more effective decision support in practical applications.

Images