Health data enhancement and anomaly detection method based on generative adversarial network (GAN)

Abstract

The present invention discloses a health data enhancement and anomaly detection method based on a generative adversarial network (GAN). By setting a generator module, a discriminator module, a federated learning protocol module, and a physiological signal generation module, and by adjusting multimodal features, hierarchical attention, and thresholds, the data system is enhanced. By training the system, the problems of scarcity, heterogeneity, and high false alarm rate of medical data are solved, which is conducive to improving the accuracy of medical diagnosis, thereby improving the level of medical care and bringing benefits to patients.

Description

Technical Field

[0001] The present invention relates to a health data enhancement and anomaly detection method based on a generative adversarial network (GAN). Background Art

[0002] The rapid development of medical artificial intelligence (AI) relies on high-quality and diverse medical data. However, acquiring medical data faces multiple challenges. According to a 2023 World Health Organization report, the world generates over 2.5 zettabytes of medical data annually, but only 0.7% of this data is effectively labeled and used for AI model training. In the field of rare diseases, such as amyotrophic lateral sclerosis, the average number of valid samples accumulated by a single medical institution is less than 50. Furthermore, privacy regulations for medical data strictly restrict the sharing and circulation of data, leading to the widespread phenomenon of "data silos." While traditional data augmentation techniques (such as rotation, translation, and noise injection) can partially alleviate the problem of insufficient data, the resulting images often lack anatomical plausibility, such as heart position deviation or vascular structure fracture. Therefore, a new approach to health data augmentation and anomaly detection is needed. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a data enhancement system that supports privacy protection and anatomical constraints, as well as a real-time anomaly detection method of multimodal fusion, to solve the problems of scarcity, heterogeneity and high false alarm rate of medical data, and to provide a practical and widely used health data enhancement and anomaly detection method based on generative adversarial network (GAN).

[0004] In order to solve the above problems, the present invention adopts the following technical solutions:

[0005] A health data enhancement method based on a generative adversarial network is implemented by a health data enhancement system based on a generative adversarial network, the system comprising the following components: a generator module, a discriminator module, a federated learning protocol module, and a physiological signal generation module;

[0006] The generator module includes a parallel image generation branch and a U-Net segmentation branch;

[0007] The image generation branch consists of a 5-layer deconvolution network, each layer is: deconvolution layer, batch normalization layer, LeakyReLU activation function, outputting a synthetic medical image with a resolution of 512×512;

[0008] The U-Net segmentation branch includes a 4-layer encoder and a 4-layer decoder, which outputs the segmentation mask of the target organ; the loss function of the segmentation branch is the Dice coefficient loss, with a weight of 30%;

[0009] The federated learning protocol module supports cross-institutional collaborative training. Participants use differentially private stochastic gradient descent for local training, with a gradient clipping threshold C = 0.5 and a Gaussian noise scale σ = 1.2. Global model aggregation is performed every 10 rounds of communication, using a weighted average, with weights allocated based on the amount of data from each participant. The privacy budget calculation satisfies ε = 0.8, complying with HIPAA privacy protection standards.

[0010] The physiological signal generation module uses the timing-constrained WGAN-GP to generate ECG signals. The constraints include: QRS wave width 80-120ms, heart rate 60-100 BPM;

[0011] The temporal continuity loss function of the physiological signal generation module is defined as the L2 norm difference of adjacent signal points: ,

[0012] in, is the smoothing loss value, which reflects the smoothness of the entire sequence data; The number of elements in the sequence minus 1, that is, the length of the sequence minus 1; Indicates the first elements; Indicates the +1 element; Calculates the L2 norm of the difference between two adjacent elements.

[0013] Preferably, the loss function of the generator is: ;in, It is the Wasserstein distance adversarial loss, with a gradient penalty coefficient of λ=10; The Dice coefficient loss of the segmentation branch constrains the positions of key organs such as the heart and lungs; For the federated aggregate loss, the weighted average of the model parameters of each participant is calculated;

[0014] The communication data encryption of the federated learning protocol module adopts a homomorphic encryption algorithm with a key length of 2048 bits.

[0015] Preferably, the physiological signal generation module further includes: a QRS wave width detection algorithm and abnormal signal filtering; the QRS wave width detection algorithm is based on dynamic time warping matching standard QRS template, with an error tolerance of ±5ms;

[0016] Abnormal signal filtering: If the heart rate or QRS width of the generated signal exceeds the preset range, it will be automatically discarded and regenerated.

[0017] Preferably, a health data enhancement method based on a generative adversarial network comprises the following steps:

[0018] 1) Data preparation and preprocessing;

[0019] This includes data collection and cleaning, collecting medical data, removing noise, and setting standardized data formats;

[0020] This includes data labeling, key area labeling by doctors, and privacy data desensitization.

[0021] 2) Select enhancement methods, using traditional methods and deep generation methods;

[0022] 3) Implement enhanced processing;

[0023] 4) Evaluate generated data: perform quality assessment and privacy protection assessment;

[0024] 5) Deploy augmented data.

[0025] A health data anomaly detection method based on a generative adversarial network includes the following steps:

[0026] 1) Perform multimodal feature extraction; specifically, for image feature extraction, a pre-trained ResNet-50 model was used, with the fully connected layer removed and replaced with an adaptive average pooling layer to output a 256-dimensional feature vector. For temporal signal feature extraction, a 1D-CNN network was used to extract local features, and sinusoidal position encoding was superimposed to inject temporal information. For text feature extraction, a BioBERT model was used to perform word segmentation and entity recognition on medical record text, outputting a 768-dimensional clinical semantic embedding.

[0027] 2) Perform hierarchical attention fusion:

[0028] Specifically: First-level attention: using image features as query and time series signal features as key and value, calculate the image-to-signal attention features: ;

[0029] in, is the query vector obtained by linear transformation of image features; is the key vector obtained by linear transformation of signal features; is the key vector Dimensions; Apply the Softmax function to the similarity score matrix obtained above to convert it into a probability distribution form; is the value vector obtained by linear transformation of signal features;

[0030] Second-level attention: Use the first-level output as the query and the text features as the key and value to calculate the final fusion features: ;

[0031] in, is the image-to-signal attention feature; To fuse features into attention features of text; This is the layer normalization operation, which normalizes the input feature vector in the layer dimension so that the input data of each layer has a similar distribution;

[0032] 3) Perform dynamic threshold adjustment and anomaly determination. Specifically, update the threshold based on the exponentially weighted moving average, set the smoothing factor β = 0.9, and calculate it as follows: ;

[0033] in, is the time step The mean estimate of ; is the time step The mean estimate of ; is the time step Observed values of is the time step The variance estimate of ; is the time step The variance estimate of ; is the current observation value and the current mean estimate square of the difference; is the time step The estimated standard deviation of

[0034] The abnormality judgment condition: if the abnormal probability score is , triggering an early warning signal, where Indicates the abnormal probability of the detected object; Representative time The mean of normal data under ; Corresponding time The standard deviation of normal data under Greater than the mean plus 3 times the standard deviation , it is judged as abnormal.

[0035] Preferably, the layer structure of the 1D-CNN network is: Conv1D(64)-ReLU-MaxPool1D(2)-Conv1D(128)-ReLU-MaxPool1D(2)-Flatten;

[0036] The position encoding formula is: ;in, represents position encoding, d=256 is the feature dimension, is the temporal position, is the dimension index in the positional encoding vector, counting starting from 0.

[0037] Preferably, the following optimizations are performed when deployed to edge devices, specifically including: model lightweighting and real-time interface optimization;

[0038] The model is lightweighted by using TensorRT to perform 8-bit integer quantization on the model, reducing memory usage by 75%; convolutional layers and batch normalization layers are merged to reduce inference latency;

[0039] The real-time interface is optimized to support the HL7 FHIR standard, receiving DICOM images and physiological signals in JSON format through a RESTful API; the output results are encapsulated as HL7 warning messages and pushed to the hospital information system.

[0040] Preferably, the update triggering conditions for the dynamic threshold adjustment are: time triggering, data volume triggering and emergency triggering; specifically:

[0041] Time trigger: The threshold is forced to be updated every 24 hours;

[0042] Data volume trigger: The threshold is updated after receiving 100 new samples cumulatively;

[0043] Emergency trigger: If the abnormal probability exceeds the current threshold for 5 consecutive times, the threshold will be recalculated immediately.

[0044] Preferably, a health data anomaly detection method based on a generative adversarial network comprises the following steps:

[0045] 1) Perform data preparation and feature analysis;

[0046] Perform data division and divide the training set, validation set, and test set by time or patient ID;

[0047] Perform data extraction. For image data, use pre-trained models to extract deep features. For time series data, extract time domain and frequency domain features. For text data, use BERT to extract semantic embeddings of medical record text.

[0048] 2) Select a detection model;

[0049] For image anomaly detection: the autoencoder + reconstruction error threshold method is used. The specific steps are as follows:

[0050] A. Only normal images are used for training AE;

[0051] B. Calculate the reconstruction error of the test sample;

[0052] C. Set the dynamic threshold: the mean of the normal sample error + 3 × standard deviation;

[0053] For time series anomaly detection, the LSTM-VAE hybrid model is used. The specific steps are as follows:

[0054] A. The encoder compresses the ECG signal into a latent vector;

[0055] B. The decoder reconstructs the signal and calculates the reconstruction error;

[0056] C. Jointly optimize ELBO and anomaly score;

[0057] For abnormal table data, the isolation forest method is used. The specific steps are as follows:

[0058] A. Randomly select features and split values to build a tree structure;

[0059] B. Calculate the path length of the sample;

[0060] C. Determine the abnormality level based on the path length sorting;

[0061] 3) Perform model training and tuning;

[0062] 4) Conduct model evaluation, specifically: indicator selection and interpretability analysis;

[0063] 5) Deploy and monitor; deploy on edge devices, convert the model into TFLite format, and deploy it to smart wearable devices; conduct continuous learning and regularly fine-tune the model with new data to prevent distribution drift.

[0064] The beneficial effects of the present invention are:

[0065] 1. Generate synthetic data that conforms to medical priors while protecting privacy through anatomically constrained federated generative adversarial networks.

[0066] 2. A hierarchical attention mechanism is used to fuse image, time series signal and text features, and combined with dynamic thresholds to achieve high-precision anomaly detection.

[0067] 3. The system's accuracy in pneumonia classification tasks has been improved to 92.3%, the false alarm rate for ICU sepsis warnings has been reduced to 1.1%, and it supports real-time deployment of edge devices.

[0068] 4. Generate synthetic data that is consistent with medical priors.

[0069] 5. Achieve cross-hospital collaborative federated training with a privacy attack defense success rate of ≥97%.

[0070] 6. By integrating image, time series signal and text features, the anomaly detection false alarm rate is ≤1.5%.

[0071] 7. Support real-time inference of edge devices with latency ≤ 100ms. DETAILED DESCRIPTION

[0072] A health data enhancement method based on a generative adversarial network is implemented by a health data enhancement system based on a generative adversarial network, the system comprising the following components: a generator module, a discriminator module, a federated learning protocol module, and a physiological signal generation module;

[0073] The generator module includes a parallel image generation branch and a U-Net segmentation branch;

[0074] The image generation branch consists of a 5-layer deconvolution network, each layer is: deconvolution layer, batch normalization layer, LeakyReLU activation function, outputting a synthetic medical image with a resolution of 512×512;

[0075] Layer 1: Input noise vector ( Represents a variable, represents a 100-dimensional real vector space), the deconvolution kernel is 4×4, the stride is 1, the padding is 1, and the output channels are 512;

[0076] Layers 2-4: Deconvolution kernel 4×4, stride 2, output channels 256, 128, and 64, respectively;

[0077] Layer 5: Deconvolution kernel 4×4, stride 2, output channel 1 (single-channel medical imaging), activation function is Tanh.

[0078] The U-Net segmentation branch includes a 4-layer encoder and a 4-layer decoder, which outputs the segmentation mask of the target organ; the loss function of the segmentation branch is the Dice coefficient loss, with a weight of 30%;

[0079] The segmentation targets include key organs such as the heart, lungs, and liver, and the loss function is expanded to multi-organ Dice loss:

[0080] ;

[0081] in is the number of organ categories, is the true segmentation mask, is the prediction mask.

[0082] Dynamic adjustment of loss function weights during implementation: Dynamic adjustment of loss weights according to the training phase, specifically: 1. Adaptive weight allocation;

[0083] Initial stage (first 50 rounds): the adversarial loss weight is increased to 0.8 to accelerate the convergence of the generator;

[0084] Mid-term stage (50-150 rounds): The anatomical constraint loss weight is increased to 0.4 to strengthen the rationality of the anatomy;

[0085] Late stage (after 150 rounds): The federated aggregate loss weight is increased to 0.2 to improve model generalization;

[0086] Adjustment formula: ; where t is the training round.

[0087] Second, perform gradient penalty optimization. The gradient penalty coefficient λ of Wasserstein adversarial loss changes with the training round:

[0088] ;in, is the total number of training rounds;

[0089] The federated learning protocol module supports cross-institutional collaborative training. Participants use differentially private stochastic gradient descent for local training, with a gradient clipping threshold C = 0.5 and a Gaussian noise scale σ = 1.2. Global model aggregation is performed every 10 rounds of communication, using a weighted average, with weights allocated based on the amount of data from each participant. The privacy budget calculation satisfies ε = 0.8, complying with HIPAA privacy protection standards.

[0090] Enhancements to the federated learning protocol during implementation:

[0091] Dynamic participant selection: Dynamically adjust participant weights based on data quality (such as annotation consistency and signal-to-noise ratio);

[0092] Gradient compression: Use Top-K gradient sparsification (retain the top 10% gradient values), reducing communication overhead by 40%;

[0093] Heterogeneous model support: allows the generator structures of each participant to differ (such as the number of layers and channels), and aligns the feature space through knowledge distillation.

[0094] The physiological signal generation module uses the timing-constrained WGAN-GP to generate ECG signals. The constraints include: QRS wave width 80-120ms, heart rate 60-100 BPM;

[0095] The temporal continuity loss function of the physiological signal generation module is defined as the L2 norm difference of adjacent signal points: ;in, is the smoothing loss value, which reflects the smoothness of the entire sequence data; The number of elements in the sequence minus 1, that is, the length of the sequence minus 1; Indicates the first elements; Indicates the +1 element; Calculates the L2 norm of the difference between two adjacent elements.

[0096] Furthermore, the loss function of the generator is: ;in, It is Wasserstein adversarial loss, with gradient penalty coefficient λ=10; for Dice loss; Average federal losses.

[0097] Furthermore, the federated learning protocol module uses a homomorphic encryption algorithm for communication data encryption with a key length of 2048 bits. It also has multi-algorithm compatibility: supporting Paillier, RSA-OAEP, and LWE homomorphic encryption algorithms, with optional key lengths (2048 / 4096 bits).

[0098] Furthermore, the physiological signal generation module further includes: QRS wave width detection algorithm and abnormal signal filtering; the QRS wave width detection algorithm is based on dynamic time regularization matching the standard QRS template, with an error tolerance of ±5ms; abnormal signal filtering: if the heart rate or QRS width of the generated signal exceeds the preset range, it is automatically discarded and regenerated; if the generated signal exceeds the tolerance range, the reinforcement learning mechanism is triggered and the generator is re-optimized.

[0099] Furthermore, the physiological signal generation module supports multimodal signals: including ECG, EEG, and PPG. The constraints are shown in Table 1 below;

[0100]

[0101] During the implementation process, QRS wave detection is enhanced, specifically:

[0102] Multi-template matching: 10 standard QRS waveform templates (normal, left bundle branch block, right bundle branch block, etc.) are pre-stored, and dynamic time warping (DTW) is used to calculate similarity;

[0103] Adaptive Tolerance: Adjusts error tolerance based on patient history data: ;

[0104] Where Tolerance is the tolerance value, 5ms is a fixed time threshold, and AvgQRSWidth represents the average width of the QRS complex;

[0105] Real-time correction: If the generated signal exceeds the limit three times in a row, fine-tuning of the generator is triggered, and the learning rate is reduced to 1 / 10 of the initial value.

[0106] At the same time, the abnormal signal filtering is expanded: a multi-level filtering mechanism is adopted; as shown in Table 2 below;

[0107] Signal quality scoring: Based on the weighted scoring model, the formula is: ; Only signals with a score ≥ 0.8 are retained;

[0108] Among them, Score is the comprehensive score; QRS_Acc is the characteristic indicator of the QRS complex. A higher QRS_Acc means more reliable detection of the QRS complex. HR_Stability: a heart rate stability indicator. This indicator measures the degree of fluctuation of the heart rate over a certain period of time. For example, it is obtained by calculating the standard deviation of the heart rate. ST_Slope is the ST segment slope indicator. The ST segment is the section from the end of the QRS complex to the starting point of the T wave in the electrocardiogram. The change in its slope can reflect physiological conditions such as the blood supply to the myocardium.

[0109] Furthermore, a health data enhancement method based on generative adversarial networks includes the following steps: 1) data preparation and preprocessing;

[0110] This includes data collection and cleaning, collecting medical data, removing noise, and setting standardized data formats;

[0111] This includes data labeling, key area labeling by doctors, and privacy data desensitization.

[0112] 2) Select enhancement methods, using traditional methods and deep generation methods;

[0113] 3) Implement enhanced processing;

[0114] 4) Evaluate the generated data and conduct quality assessment and privacy protection assessment;

[0115] 5) Deploy augmented data.

[0116] A health data anomaly detection method based on a generative adversarial network includes the following steps:

[0117] 1) Perform multimodal feature extraction; specifically, for image feature extraction, a pre-trained ResNet-50 model was used, with the fully connected layer removed and replaced with an adaptive average pooling layer to output a 256-dimensional feature vector. For temporal signal feature extraction, a 1D-CNN network was used to extract local features, and sinusoidal position encoding was superimposed to inject temporal information. For text feature extraction, a BioBERT model was used to perform word segmentation and entity recognition on medical record text, outputting a 768-dimensional clinical semantic embedding.

[0118] Image features: Supports 3D medical images (such as CT and MRI volume data) and uses 3D-ResNet to extract spatial-temporal features;

[0119] Time series signal: Adaptive sampling: Dynamically adjust the sampling rate according to the signal frequency (such as ECG 250Hz, EEG 512Hz);

[0120] Text features: Entity relationship graph construction: Identify "drug-symptom-disease" associations and generate graph neural network (GNN) embeddings.

[0121] 2) Perform hierarchical attention fusion:

[0122] Specifically: First-level attention: using image features as query and time series signal features as key and value, calculate the image-to-signal attention features: ; is the query vector obtained by linear transformation of image features; is the key vector obtained by linear transformation of signal features; is the key vector Dimensions; Apply the Softmax function to the similarity score matrix obtained above to convert it into a probability distribution form; is the value vector obtained by linear transformation of signal features.

[0123] Second-level attention: Use the first-level output as the query and the text features as the key and value to calculate the final fusion features: ;

[0124] in is the image-to-signal attention feature; To fuse features into attention features of text; This is a layer normalization operation that normalizes the input feature vector in the layer dimension so that the input data of each layer has a similar distribution.

[0125] Enhanced content of hierarchical attention during implementation:

[0126] 1. Each level of attention is expanded to 8 heads to enhance feature interaction capabilities;

[0127] 2. Introduce cross-attention between modalities (e.g. text → image), the formula is:

[0128] ;

[0129] in, is the query vector obtained by linear transformation of text features; is the key vector obtained by linear transformation of image features; is the value vector obtained by linear transformation of image features; is the key vector The dimension of CrossAttn is the cross-modal attention calculation function.

[0130] 3. Each level of attention output is superimposed on the original input to avoid gradient disappearance.

[0131] 3) Perform dynamic threshold adjustment and anomaly determination. Specifically, update the threshold based on the exponentially weighted moving average, set the smoothing factor β = 0.9, and calculate it as follows: ;

[0132] in, is the time step The mean estimate of ; is the time step The mean estimate of ; is the time step Observed values of is the time step The variance estimate of ; is the time step The variance estimate of ; is the current observation value and the current mean estimate square of the difference; is the time step The estimated standard deviation of .

[0133] The abnormality judgment condition: if the abnormal probability score is Triggering early warning signals, including Indicates the abnormal probability of the detected object;

[0134] Representative time The mean of normal data under ; Corresponding time The standard deviation of normal data under Greater than the mean plus 3 times the standard deviation , it is judged as abnormal.

[0135] During the implementation, the smoothing factor β is adjusted according to the abnormal probability distribution. ;in, Representing time series In from −100 to The variance within these 100 time steps; Representing time series The maximum variance over the entire history.

[0136] At the same time, combined with the patient's historical data (such as underlying diseases, age), the threshold offset Δ is adjusted individually:

[0137] ;in, Indicates time threshold value; Represents the time step The mean estimate of ; Represents time-step-based The estimated standard deviation of Calculated; Indicates patient-specific adjustments.

[0138] Furthermore, the layer structure of the 1D-CNN network is: Conv1D(64)-ReLU-MaxPool1D(2)-Conv1D(128)-ReLU-MaxPool1D(2)-Flatten;

[0139] The position encoding formula is: ;

[0140] in, represents position encoding, d=256 is the feature dimension, is the temporal position, is the dimension index in the positional encoding vector, counting starting from 0.

[0141] Furthermore, the following optimizations are performed when deployed to edge devices, including: model lightweighting and real-time interface optimization,

[0142] The model is lightweighted by using TensorRT to perform 8-bit integer quantization on the model, reducing memory usage by 75%; convolutional layers and batch normalization layers are merged to reduce inference latency;

[0143] During implementation, multi-level optimization of the model lightweighting is performed, as shown in Table 3 below;

[0144] At the same time, a layer fusion strategy is implemented: adjacent Conv1D and BatchNorm layers are merged to reduce the number of memory accesses; activation functions (such as ReLU) are integrated into the computational graph of the previous layer to save instruction cycles.

[0145] Perform expansion and optimization of the real-time structure and support multiple protocols, as shown in Table 4 below;

[0146] Perform extended optimization of the real-time structure and adopt a fault-tolerant mechanism, specifically:

[0147] Data check: CRC32 checksum verifies data integrity and discards data packets with an error rate greater than 1%;

[0148] Retransmission mechanism: If transmission fails three times in a row, it switches to the backup communication channel (e.g. Bluetooth → Wi-Fi);

[0149] For example, it can be deployed on smart home gateways and integrated with smart watches (PPG), weight scales (impedance data), and environmental sensors (temperature and humidity).

[0150] The real-time interface is optimized to support the HL7 FHIR standard, receiving DICOM images and physiological signals in JSON format through a RESTful API; the output results are encapsulated as HL7 warning messages and pushed to the hospital information system.

[0151] Furthermore, the update trigger conditions for the dynamic threshold adjustment are: time trigger, data volume trigger and emergency trigger; specifically:

[0152] Time trigger: The threshold is forced to be updated every 24 hours;

[0153] Data volume trigger: The threshold is updated after receiving 100 new samples cumulatively;

[0154] Emergency trigger: If the abnormal probability exceeds the current threshold for 5 consecutive times, the threshold will be recalculated immediately.

[0155] The calculation of the threshold needs to take into account offset, temperature compensation, and motion state compensation.

[0156] A health data anomaly detection method based on a generative adversarial network includes the following steps:

[0157] 1) Perform data preparation and feature analysis;

[0158] Perform data division and divide the training set, validation set, and test set by time or patient ID;

[0159] Perform data extraction. For image data, use pre-trained models (such as ResNet-50) to extract deep features. For time series data, extract time domain (mean, variance) and frequency domain (FFT) features. For text data, use BERT to extract semantic embeddings of medical record text.

[0160] 2) Select a detection model;

[0161] For image anomaly detection: the autoencoder (AE) + reconstruction error threshold method is used. The specific steps are as follows:

[0162] A. Only normal images are used for training AE;

[0163] B. Calculate the reconstruction error (MSE or SSIM) of the test sample;

[0164] C. Set the dynamic threshold: the mean of the normal sample error + 3 × standard deviation;

[0165] For time series anomaly detection, LSTM-VAE (Variational Autoencoder) is used. The specific steps are as follows:

[0166] A. The encoder compresses the ECG signal into a latent vector;

[0167] B. The decoder reconstructs the signal and calculates the reconstruction error;

[0168] C. Jointly optimize ELBO (evidence lower bound) and anomaly score;

[0169] For abnormal table data, the isolation forest method is used. The specific steps are as follows:

[0170] A. Randomly select features and split values to build a tree structure;

[0171] B. Calculate the path length of the sample (the path of the abnormal sample is shorter);

[0172] C. Determine the abnormality level based on the path length sorting;

[0173] 3) Perform model training and tuning;

[0174] 4) Conduct model evaluation, specifically: indicator selection and interpretability analysis;

[0175] 5) Deploy and monitor; deploy on edge devices, convert the model into TFLite format, and deploy it to smart wearable devices; conduct continuous learning and regularly fine-tune the model with new data to prevent distribution drift.

[0176] The above are only specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that are not conceived through creative work should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection defined in the claims.

Claims

1. A health data enhancement method based on a generative adversarial network, which is implemented by a health data enhancement system based on a generative adversarial network, characterized in that: The system includes the following components: generator module, discriminator module, federated learning protocol module and physiological signal generation module; The generator module includes a parallel image generation branch and a U-Net segmentation branch; The image generation branch consists of a 5-layer deconvolution network, each layer is: deconvolution layer, batch normalization layer, LeakyReLU activation function, outputting a synthetic medical image with a resolution of 512×512; The U-Net segmentation branch includes a 4-layer encoder and a 4-layer decoder, which outputs the segmentation mask of the target organ; the loss function of the segmentation branch is the Dice coefficient loss, with a weight of 30%; The federated learning protocol module supports cross-institutional collaborative training. Participants use differentially private stochastic gradient descent for local training, with a gradient clipping threshold C = 0.5 and a Gaussian noise scale σ = 1.

2. Global model aggregation is performed every 10 rounds of communication, using a weighted average, with weights allocated based on the amount of data from each participant. The privacy budget calculation satisfies ε = 0.8, complying with HIPAA privacy protection standards. The physiological signal generation module uses the timing-constrained WGAN-GP to generate ECG signals. The constraints include: QRS wave width 80-120ms, heart rate 60-100 BPM; The temporal continuity loss function of the physiological signal generation module is defined as the L2 norm difference of adjacent signal points: ;in, is the smoothing loss value, which reflects the smoothness of the entire sequence data; The number of elements in the sequence minus 1, that is, the length of the sequence minus 1; Indicates the first elements; Indicates the +1 element; To calculate the L2 norm of the difference between two adjacent elements; The loss function of the generator is: ;in, It is the Wasserstein distance adversarial loss, and the gradient penalty coefficient λ=10; The Dice coefficient loss of the segmentation branch constrains the positions of key organs such as the heart and lungs; The federated aggregation loss is calculated by weighted averaging the model parameters of each participant. The communication data encryption of the federated learning protocol module adopts a homomorphic encryption algorithm with a key length of 2048 bits.

2. The health data enhancement method based on generative adversarial networks according to claim 1, characterized in that: The physiological signal generation module further includes: a QRS wave width detection algorithm and abnormal signal filtering; the QRS wave width detection algorithm is based on dynamic time warping matching standard QRS template with an error tolerance of ±5ms; Abnormal signal filtering: If the heart rate or QRS width of the generated signal exceeds the preset range, it will be automatically discarded and regenerated.

3. The health data enhancement method based on generative adversarial networks according to claim 1, characterized in that: The steps include: 1) Data preparation and preprocessing; This includes data collection and cleaning, collecting medical data, removing noise, and setting standardized data formats; This includes data labeling, key area labeling by doctors, and privacy data desensitization. 2) Select enhancement methods, using traditional methods and deep generation methods; 3) Implement enhanced processing; 4) Evaluate generated data: perform quality assessment and privacy protection assessment; 5) Deploy augmented data.

Images