Methods for improving small neural network training models for ECG data
By training small electrocardiogram data neural network models on edge computing devices, the network delay problem caused by large cloud platform models is solved, real-time and accuracy are improved, and data consistency and model updates are ensured in medical and health scenarios. .
Patent Information
- Application Number
- CN202111285450.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-11-02
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2041-11-02
AI Technical Summary
The existing large-scale ECG data neural network model based on cloud platforms leads to inconsistent data and results due to transmission loss and network delay, especially in the medical field, real-time requirements are high, and real-time data loses significance.
Use customized edge computing equipment to train small neural network models, collect ECG hardware equipment, perform neural network calculations and aggregation model operations, form thresholds, optimize data models based on terminal equipment, and update the model in collaboration with the cloud, and use volumes Integrated neural network, iterative neural network or deep neural network for calculation.
It effectively avoids data and results in the second-level network delay, improves real-time, enhances the accuracy and universality of the model, and updates the small neural network model across time zones and regions.
Smart Images

Figure CN114005543B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer technology, and in particular to a method for improving a small neural network training model for electrocardiogram data. Background Art
[0002] Neural network learning based on medical big data is an effective way to train large neural network models and assist in data analysis. However, the ECG data neural network training model in the prior art has the following shortcomings: the existing large network model based on the cloud platform has a network delay of more than seconds due to transmission loss. The huge network delay caused by the communication between the terminal device and the cloud is mainly caused by the long communication link. The reason for the need for this communication link is that the final judgment of the conditions and algorithms needs to be run on the cloud, which leads to inconsistent data and results due to the second-level network delay during use. Especially in the ECG data related set, the real-time requirement is of great value to the final judgment, and it will play an important auxiliary role in preventing and predicting the occurrence of diseases. The current mainstream technology uses the cloud for basic calculations, and the huge network delay caused by it will cause a lot of real-time data to lose its meaning. Therefore, it is necessary to propose a method for improving the small neural network training model for ECG data to solve the above problems. Summary of the invention
[0003] The present invention provides a method for improving a small neural network training model for electrocardiogram data, so as to solve the problem that large neural network models in the prior art all use the cloud for basic calculations, resulting in huge network delays that cause a lot of real-time data to lose their meaning.
[0004] The present invention provides a method for improving a small neural network of electrocardiogram data based on a customized edge computing device training model, comprising:
[0005] The edge computing device receives the ECG data of the first cycle, wherein the ECG data is collected by an ECG hardware device, and the ECG data includes original ECG waveform data, ECG additional information, ECG rhythm information, and PQRST waveform data;
[0006] The edge computing device performs neural network calculation on the electrocardiogram data of the first cycle to obtain first neural network calculation data;
[0007] The edge computing device performs a convergence model operation on the first neural network computing data to form a threshold M1;
[0008] The edge computing device receives the ECG data of the second cycle;
[0009] The edge computing device performs neural network calculation on the electrocardiogram data of the second cycle to obtain second neural network calculation data;
[0010] The edge computing device performs a convergence model operation on the second neural network computing data to form a threshold M2;
[0011] By analogy, the edge computing device receives ECG data of multiple time periods of at least 3 days, and obtains thresholds M1, M2...Mn, M1+M2+...+Mn=1, thereby forming a small neural network training model based on the ECG data of individual users, and continuously optimizes the data model for individual users. The small neural network training model for ECG data relies on the data interaction of terminal devices to continuously improve the accuracy of the model. At the same time, in the process of collaboration with the cloud, the iterative model has greater versatility: data across time zones and physical regions are updated in the cloud based on the ECG data cleaning algorithm of aggregation, aggregation, clustering and weighting, and the small neural network training model for ECG data is updated.
[0012] Furthermore, the electrocardiogram rhythm information includes average ventricular heart rate, average RR interval, the difference between the longest RR interval and the shortest RR interval, standard deviation of the RR interval, extrasystole information, extrasystole type data, extrasystole morphology data, F wave type data of atrial flutter and detection results of F wave of atrial fibrillation, and detection results of asynchronous P wave.
[0013] Furthermore, the neural network calculation is any one of a convolutional neural network, an iterative neural network, and a deep neural network.
[0014] Furthermore, the ECG hardware device is a 1-lead, 2-lead, 4-lead or 12-lead basic ECG acquisition device.
[0015] Furthermore, the edge computing device is equipped with an ARM Cortex-A72 architecture, a six-core 64-bit processor, a main frequency of 1.8GHz, an integrated quad-core Mali-T860 GPU, supports H.265 HEVC and VP9, H.264 encoding, 4K HDR, and supports up to 4K hard decoding.
[0016] Furthermore, the edge computing device has an onboard AI embedded neural network processor NPU with a peak computing power of 5.6 Tops, a typical computing power of 2.8 Tops, and an efficiency-to-energy ratio of 9.3 Tops / W.
[0017] Furthermore, the edge computing device adopts the AI-specific MPE matrix engine and APiM architecture, integrates storage and computing into one local parallel AI operation, and upgrades the network preload at one time.
[0018] Furthermore, the edge computing device supports three network training model instances based on VGG: GNet1, GNet18 and GNetfc.
[0019] The beneficial effects of the present invention are as follows: The present invention provides a method for improving a small neural network training model for ECG data, which can effectively avoid the inconsistency of data and results caused by second-level network delays during use, especially in the ECG data related set, where the real-time requirement is of high value for the final judgment, and plays an important auxiliary role in preventing and predicting the occurrence of diseases. The small model training of the ECG data neural network model contributes to the popularization and application of edge computing devices in this medical and health scenario. The ECG data neural network model relies on the data interaction of terminal devices to continuously improve the accuracy of the model, and in the process of collaboration with the cloud, it helps to make the iterative model more universal: data across time zones and physical regions can be effectively updated in the cloud based on ECG data cleaning algorithms such as aggregation, aggregation, clustering and weighting. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] In order to more clearly illustrate the technical solution of the present invention, the drawings required for use in the embodiments are briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0021] Figure 1 A flowchart of a method for improving a small neural network training model for ECG data provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0022] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. The technical solutions provided by the embodiments of the present invention are described in detail below in conjunction with the drawings.
[0023] See also Figure 1 The present invention provides a method for improving a small neural network training model for electrocardiogram data, comprising:
[0024] In step S101, the edge computing device receives the ECG data of the first cycle, wherein the ECG data is collected by an ECG hardware device, and the ECG data includes original ECG waveform data, ECG additional information, ECG rhythm information, and representative PQRST waveform data.
[0025] In this embodiment, a small ECG neural network model with computing power and operator adaptation can be trained based on the edge computing device. The edge computing device can be equipped with ARM's new Cortex-A72 architecture, a six-core 64-bit high-performance processor with a main frequency of up to 1.8GHz, an integrated quad-core Mali-T860 GPU, support for H.265 HEVC and VP9, H.264 encoding, 4K HDR, and has powerful hardware decoding capabilities, which can support up to 4K hard decoding. The onboard AI embedded neural network processor NPU has a peak computing power of up to 5.6Tops, a typical computing power of 2.8 Tops, and an efficiency-to-energy ratio of up to 9.3 Tops / W. It has super computing power while maintaining extremely low energy consumption, making it extremely advantageous in the field of edge computing of terminal devices. It adopts the AI-specific MPE matrix engine and APiM (AI processing in Memory) architecture, local parallel AI operations that integrate storage and computing, and one-time network pre-loading. It does not require instructions or buses, and does not require external DDR cache. The processing speed far exceeds that of processors with other traditional architectures, while also greatly reducing processing energy consumption.
[0026] It supports three network training model instances based on VGG: GNet1, GNet18 and GNetfc. More network instances will be added in the future, making it easy to test a large number of deep learning applications on the device.
[0027] In this embodiment, the electrocardiogram rhythm information includes the average ventricular heart rate, the average RR interval, the difference between the longest RR interval and the shortest RR interval, the standard deviation of the RR interval, extrasystole information, extrasystole type data, extrasystole morphology data, F wave type data of atrial flutter and the detection results of F wave of atrial fibrillation, and the detection results of asynchronous P wave.
[0028] Step S102: The edge computing device performs neural network calculation on the electrocardiogram data of the first cycle to obtain first neural network calculation data.
[0029] In this embodiment, the neural network calculation is any one of a convolutional neural network, an iterative neural network, and a deep neural network. The model calculation of the neural network calculation can use standard network model calculation methods such as Fourier series and Gaussian function to perform calculations after big data collection. The calculation type data includes data collected from the standard module ECG module for medical device certification. The supported ECG module types are single lead, dual lead, 4 leads, and 12 leads. The data type collected by the ECG module is used as the input data of the neural network model, specifically including: original electrocardiogram waveform data, additional effective information of the electrocardiogram (the effective information here is customized according to the actual judgment standard), data extraction representing the PQRST waveform, etc.
[0030] Step S103: The edge computing device performs a convergence model operation on the first neural network computing data to form a threshold M1.
[0031] Specifically, the threshold of Mn here is the weighted average threshold after the actual data is collected. In combination with each different data collection comparison direction of this network model, there will be different standard thresholds for real-time comparison with the collected and calculated threshold Mn, so as to continuously optimize and clarify the threshold range for individuals. For example, for heartbeat data, assuming that the normal value of heartbeat data is 80-120 times / minute, but the actual situation is due to individual differences, it cannot be judged that the data is problematic from a single measurement of 75-115 times / minute. By applying this small model and edge computing capabilities, long-term collection can be carried out, such as 3 minutes each time, more than 200 times in 7 days, and other valid data can be used to collaboratively determine what the actual threshold value for the individual is, thereby effectively increasing the individual user's continuous understanding and tracking of the effectiveness of their actual data.
[0032] Step S104: The edge computing device receives the ECG data of the second cycle.
[0033] Step S105: The edge computing device performs neural network calculation on the electrocardiogram data of the second cycle to obtain second neural network calculation data.
[0034] Step S106: The edge computing device performs a convergence model operation on the second neural network computing data to form a threshold M2.
[0035] Step S107, and so on, the edge computing device receives ECG data of multiple time periods of at least 3 days, and obtains thresholds M1, M2...Mn, M1+M2+...+Mn=1 (weighted process), thereby forming a small neural network training model based on the ECG data of individual users, and continuously optimizing the data model for individual users. The small neural network training model for ECG data relies on the data interaction of terminal devices (such as edge computing devices and ECG hardware equipment collection) to continuously improve the accuracy of the model, and at the same time, in the process of collaboration with the cloud, the iterative model has greater versatility: data across time zones and physical regions are based on aggregation, aggregation, clustering and weighted ECG data cleaning algorithms in the cloud to update the small neural network training model for ECG data.
[0036] Specifically, for the weighted average method of M1+...+Mn=1, M is the threshold calculated in a unit cycle. There will be N similar thresholds. In the case of a period of up to 7 days, some different ECG data will be collected for a month, and N will be larger. Finally, it is necessary to clarify the specific value of the data in a cycle in the process of valid display. Therefore, the threshold of each stage will be weighted and averaged with 1 as the unit, and 1 can be understood as 100%. For example, the first 5 seconds of each heartbeat data collection are extremely inaccurate, so the weight can be set to less than 1%. This is a weighted method for a cycle. The measurement data at 8 am and 6 pm every day are the most important. It may be that within the time span of 7 days, the weight of this part of the data will be set to 40%-40%. This is a weighted method that spans a single cycle but cooperates with a small cycle, and so on. The main function of weighting is to remove the significant impact of invalid data in the overall trend and expand the importance of important time period data in the overall data.
[0037] Specifically, the above process is based on individual data collection and calculation, so what comes out in the end is an individualized trained small model.
[0038] Specifically, the above-mentioned terminal device refers to a terminal device connected to an edge computing device, and the specific form can be a watch, a treadmill, a docking information system, etc.
[0039] The training mechanism of the small neural network training model for electrical data is as follows: Through a customized neural network training system and deep learning model, a relevant miniaturized neural network model is trained on a training cluster.
[0040] Furthermore, the present invention also has a distributed training and distributed distribution system for ECG data neural network models, which can realize a directional distribution model: based on the user ID description and actual usage data evaluation, a directional distribution mechanism is performed through a matching algorithm. Fixed-point distribution mechanism: regional users have their own ECG model basis. By analyzing and sorting out the effectiveness of the model across a single user, the regional model of the entire region is continuously corrected to continuously correct the user experience and accuracy of regional users. Based on general-purpose edge computing devices, relevant corrections to the neural network model by coordinating relevant medical big data across a single platform, across physical areas, and across time areas will effectively improve the accuracy of the relevant neural network model. Based on the training of small neural network models of edge computing devices, the cloud platform is coordinated to train efficient small neural network models for exclusive areas and form a model distribution closed loop in the distributed system.
[0041] Specific subscription and publishing methods:
[0042] 1. Subscribe to the small neural network model adapted for edge computing devices through the SubscribeECGNetworkModel() [Subscribe ECG neural network model] interface
[0043] Subscription method example:
[0044] Try()
[0045] {
[0046] Example.SybscribeECTNetWorkModel(Now, this.networkDodelData, Normal)
[0047] }Catch{}
[0048] 2. Perform cloud-based sampling and analysis of edge device metadata through the RawDataUpload() interface. The sampling method is: 1.
[0049] Subscription method example:
[0050] Try()
[0051] {
[0052] Example.RawDataUpload(1430, this.rawData, week)
[0053] }Catch{}
[0054] 3. Customize the neural network model of the edge device through the Backup&Restore() [Local neural network data recovery and backup] interface for periodic maintenance. The specific maintenance method is based on the current basic definition that when it is automated, no user participation is required, only subscription is required. Model training and update, the backup cycle is weekly. The basic strategy is to divide the current data and replace the updated data, and it can be updated regularly according to user needs.
[0055] Subscription method example:
[0056] Try()
[0057] {
[0058] Example.Backup&Restore(Backup, this.wholeData)
[0059] }Catch{}
[0060] 4. Customize the relevant neural network model for specific individual users through SubscribeECGNetworkModelPerPerson() [Subscribe to the miniaturized neural network model based on individual users]
[0061] Subscription method example:
[0062] Try()
[0063] {
[0064] Example.SubscribeECGNetworkModelPerPerson(Now, this.personalData,month)
[0065] }Catch{}.
[0066] Due to transmission loss, the network delay of large network models on traditional cloud platforms will be more than seconds, which is a big problem in medical scenarios. The ECG data neural network model based on edge computing can ensure timely response, and the transmission delay is below milliseconds. The huge network delay caused by the communication between the terminal device and the cloud is mainly caused by the long communication link. The reason for the need for this communication link is that the final judgment of conditions and algorithms needs to be run on the cloud. Based on the edge computing solution, small models can be run on the terminal side, thereby completely releasing the link and bandwidth requirements for instant communication with the cloud, thereby reducing latency. Small neural network models can run normally in AIOT terminal devices with limited hardware configuration, and the current network model does not have this capability. The formation of a clear mechanism training method and system based on AIOT general configuration can clearly solve this problem. The neural network model is a model that can be continuously improved based on big data optimization and retraining. The current distributed distribution mechanism in the medical field, especially based on the continuous improvement of ECG data, is not sound. The method of improving the small neural network training model of ECG data based on edge computing and cloud collaborative automation can efficiently provide a synchronous system and method for training and distributing new models and form a closed-loop system for small model release and model training. ECG data network models have always lacked group-related big data and neural network model training based on this big data. The method of improving the small neural network training model of ECG data based on edge computing and cloud collaborative automation can effectively solve the collection and integration of cross-individual data.
[0067] The basic parameters of edge computing devices are as follows:
[0068] Parameters: The CPU uses dual large cores (Cortex-A72) + four small cores (Cortex-A53), with a maximum main frequency of 1.8GHz;
[0069] The GPU uses quad-core ARM Mali-T860; supports OpenGL ES 1.1 / 2.0 / 3.0, OpenVG1.1, OpenCL, Directx11; the NPU uses SPR2801S, adopts the unique AI architecture of MPE and APiM; the typical computing power is 2.8Tops; DDR uses 2GB / 4GB dual-channel LP DDR4; the memory uses 8GB -128GB high-speed eMMC, TF card slot.
[0070] Hardware features: The network uses RJ45 interface Gigabit Ethernet; onboard WIFI / BT module, supports 2.4GHz / 5GHz dual-band WiFi, 802.11a / b / g / n / ac protocol; supports Bluetooth 4.1 (supports BLE); Mini PCIe seat (used to expand 3G / 4G module, used with Micro SIM card slot); multimedia decoding, supports 4K VP9 and 4K 10bitsH265 / H264 video decoding, up to 60fps 1080P multi-format video decoding (VC-1, MPEG-1 / 2 / 4, VP8); 1080P video encoding, supports H.264, VP8 format; video post-processor: de-interlacing, denoising, edge / detail / color optimization; the interface uses dual ISP pixel processing capability up to 13MPix / s, supports dual-channel camera data input at the same time; supports USB3.0 HOST, and Type-C interface; 1-channel ADC, SPI / GPIO, controllable LED × 2, I²C interface × 1, gravity sensor × 1 (expandable); support SD card; support RTC real-time clock; support timed power on and off; serial port uses RS232 × 1, RS485 × 1, debug serial port × 1, and 2-way TTL port on board; infrared uses 1-way infrared receiver head and supports infrared remote control function; power supply uses DC 12V-2A (DC5.5 × 2.1mm), supports external connection (power interface × 1); the mainboard can use POE+ (802.3 AT, output power 30W) Ethernet power supply.
[0071] System / Software: The system supports Android\Linux\Ubuntu systems; the supported frameworks support PyTorch and Caffe frameworks, and will support TensorFlow in the future; the development tool PLAI model training tool (currently supports VGG-based GNet1, GNet18 and GNetfc network models).
[0072] The embodiment of the present invention also provides a storage medium, and the embodiment of the present invention also provides a storage medium, wherein a computer program is stored in the storage medium, and when the computer program is executed by a processor, some or all of the steps in each embodiment of the method for improving the small neural network training model of electrocardiogram data provided by the present invention are implemented. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM).
[0073] Those skilled in the art can clearly understand that the technology in the embodiments of the present invention can be implemented by means of software plus a necessary general hardware platform. Based on this understanding, the technical solution in the embodiments of the present invention can be essentially or partly contributed to the prior art in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present invention or certain parts of the embodiments.
[0074] In this specification, the same or similar parts between the various embodiments can be referred to each other. In particular, for the device embodiment for improving the small neural network training model for electrocardiogram data, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the description in the method embodiment.
[0075] The above-described embodiments of the present invention do not limit the protection scope of the present invention.
Claims
1. A method for improving a small neural network training model for electrocardiogram data, characterized in that: include: The edge computing device receives the electrocardiogram data of the first cycle, wherein the electrocardiogram data is collected by an ECG hardware device, and the electrocardiogram data includes original electrocardiogram waveform data, electrocardiogram additional information, electrocardiogram rhythm information, and representative PQRST waveform data, wherein the electrocardiogram rhythm information includes average ventricular heart rate, average RR interval, difference between the longest RR interval and the shortest RR interval, standard deviation of RR interval, extrasystole information, extrasystole type data, extrasystole morphology data, F wave type data of atrial flutter and detection results of F wave of atrial fibrillation, and detection results of asynchronous P wave; The edge computing device performs neural network calculation on the electrocardiogram data of the first cycle to obtain first neural network calculation data; The edge computing device performs a convergence model operation on the first neural network calculation data to form a threshold M1, wherein the convergence model operation is a weighted average of the neural network calculation data for a time period; The edge computing device receives the ECG data of the second cycle; The edge computing device performs neural network calculation on the electrocardiogram data of the second cycle to obtain second neural network calculation data; The edge computing device performs a convergence model operation on the second neural network computing data to form a threshold M2; By analogy, the edge computing device receives ECG data of multiple time periods of at least 3 days, and obtains thresholds M1, M2...Mn, M1+M2+...+Mn=1, thereby forming a small neural network training model based on the ECG data of individual users, and continuously optimizes the data model for individual users. The small neural network training model for ECG data relies on the data interaction of terminal devices to continuously improve the accuracy of the model, and at the same time, in the process of collaboration with the cloud, the iterative model has greater versatility; data across time zones and physical regions are updated in the cloud based on the ECG data cleaning algorithm of aggregation, aggregation, clustering and weighting, and the small neural network training model for ECG data is updated.
2. The method according to claim 1, characterized in that The neural network calculation is any one of a convolutional neural network, an iterative neural network, and a deep neural network.
3. The method according to claim 1, characterized in that The ECG hardware device is a 1-lead, 2-lead, 4-lead or 12-lead basic ECG acquisition device.
4. The method according to claim 1, characterized in that The edge computing device is equipped with an ARM Cortex-A72 architecture, a six-core 64-bit processor, a main frequency of 1.8GHz, an integrated quad-core Mali-T860 GPU, supports H.265 HEVC and VP9, H.264 encoding, 4K HDR, and supports up to 4K hard decoding.
5. The method according to claim 4, characterized in that The edge computing device has an onboard AI embedded neural network processor NPU with a peak computing power of 5.6 Tops, a typical computing power of 2.8 Tops, and an efficiency power consumption ratio of 9.3 Tops / W.
6. The method according to claim 5, characterized in that The edge computing device adopts the AI-specific MPE matrix engine and APiM architecture, integrates storage and computing, and integrates local parallel AI operations, with one-time network preloading upgrade.
7. The method according to claim 5, characterized in that The edge computing device supports three network training model instances based on VGG: GNet1, GNet18 and GNetfc.
