WO2022088442A1 - 健康监测设备 - Google Patents

健康监测设备 Download PDF

Info

Publication number
WO2022088442A1
WO2022088442A1 PCT/CN2020/137983 CN2020137983W WO2022088442A1 WO 2022088442 A1 WO2022088442 A1 WO 2022088442A1 CN 2020137983 W CN2020137983 W CN 2020137983W WO 2022088442 A1 WO2022088442 A1 WO 2022088442A1
Authority
WO
WIPO (PCT)
Prior art keywords
index
output
component
cardiac
health
Prior art date
Application number
PCT/CN2020/137983
Other languages
English (en)
French (fr)
Inventor
王景峰
黄凯
陈样新
张玉玲
郭思璐
宋日辉
Original Assignee
生物岛实验室
中山大学孙逸仙纪念医院
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 生物岛实验室, 中山大学孙逸仙纪念医院 filed Critical 生物岛实验室
Publication of WO2022088442A1 publication Critical patent/WO2022088442A1/zh

Links

Images

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/05—Detecting, measuring or recording for diagnosis by means of electric currents or magnetic fields; Measuring using microwaves or radio waves 
    • A61B5/053—Measuring electrical impedance or conductance of a portion of the body
    • A61B5/0531—Measuring skin impedance
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/68—Arrangements of detecting, measuring or recording means, e.g. sensors, in relation to patient
    • A61B5/6801—Arrangements of detecting, measuring or recording means, e.g. sensors, in relation to patient specially adapted to be attached to or worn on the body surface
    • A61B5/6802—Sensor mounted on worn items

Definitions

  • the present disclosure relates to the technical field of wearable devices, and in particular, to a health monitoring device.
  • Cardiovascular disease is a serious threat to human health due to its hidden and sudden characteristics, so early diagnosis and prevention are very important.
  • the 12-lead dynamic detector (Holter) commonly used in hospitals has large power consumption and volume, and is generally operated under the guidance of a doctor when wearing it, and cannot meet the needs of 3-7 days of continuous medical testing.
  • the current Holter's ECG monitoring and judgment needs to transmit data to the background server for calculation. This calculation lag is very obvious and cannot meet the needs of real-time calculation and all-weather real-time warning.
  • the inventors found that the use of ECG data alone is not a good predictor of heart failure.
  • embodiments of the present disclosure provide a health monitoring device.
  • the health monitoring equipment includes:
  • the wearable component is woven by mixing conductive material and insulating material, the conductive material forms a plurality of electrodes on the wearable component and a multimodal physiological signal acquisition circuit is formed between the plurality of electrodes, the Multimodal physiological signals include at least ECG signals and bioimpedance signals;
  • a processing component including a scheduler and a plurality of operators, wherein the scheduler is configured to receive the multimodal physiological signal, schedule the plurality of operators to determine health status information of the user, and based on the The health state information controls the output component to output early warning information, the arithmetic unit includes a programmable logic gate array, and a task computing unit of a binary neural network model is fixedly arranged in the arithmetic unit.
  • the weights and activation values of the hidden layer are binary data for determining health state information based on multimodal physiological signals;
  • a battery for supplying power to the wearable component, the processing component and the output component.
  • the scheduler includes a central processing unit, the arithmetic unit further includes a general-purpose graphics processor and an integrated chip customized based on a programmable logic gate array, and the scheduler and the arithmetic unit are deployed on a printed circuit board .
  • the processing component further includes:
  • a preprocessing unit configured to perform preprocessing on the multimodal physiological signal, the preprocessing including at least one of noise processing, baseline drift processing, heart beat segmentation and data resampling.
  • the health monitoring device further includes an input component, and the processing component is further configured to determine the body surface area of the user based on the height and weight information input by the user.
  • the multimodal physiological signal acquisition circuit is further used to acquire heart rate, pre-ejection period, ventricular ejection time, mean arterial pressure, left ventricular systolic time, cardiac contractility index, central venous pressure, pulmonary artery occlusion some or all of the pressure.
  • the processing component further includes a feature pre-extraction unit configured to perform one or more of the following:
  • peripheral vascular resistance based on said cardiac output, mean arterial pressure and central venous pressure
  • the peripheral vascular resistance index is determined based on the cardiac index, mean arterial pressure and central venous pressure.
  • the input of the binary neural network model includes ECG signal, bioimpedance signal, left ventricular function index, cardiac systolic function index, preload index and afterload index, wherein:
  • the left ventricular function index includes at least one of stroke volume, stroke index, cardiac output, and cardiac index;
  • the myocardial contractility index includes at least one of cardiac contractility index, left heart work index, left heart contraction time, and ejection fraction;
  • the preload index includes central venous pressure and/or left ventricular end-diastolic volume
  • the afterload index includes peripheral vascular resistance and/or peripheral vascular resistance index.
  • the binary neural network model is obtained by training the following methods:
  • the weights and activation values of the hidden layer of the real-number neural network model are real-number data
  • the training operation and the compression operation are alternately performed until the convergence conditions are met, and a binary neural network model is obtained, wherein, in the training operation, the weights in the model are updated through the training data; in the compression operation, the Size of branches and/or binarized compressed models.
  • the output component has a first output mode and a second output mode
  • the processing component is configured to:
  • the output device is controlled to output early warning information according to the second output mode.
  • the output device includes a three-color indicator light
  • the processing component is configured to:
  • the red indicator light is controlled to light up.
  • a plurality of electrodes are formed on the wearable component by using a conductive material, and a multimodal physiological signal acquisition circuit is formed between the plurality of electrodes, and the multimodal physiological signals are at least Including ECG signal and bio-impedance signal;
  • the scheduler is configured to receive multi-modal physiological signals, schedule a plurality of operators to determine the health status information of the user, and control the output component to output early warning information, and the operators include programmable logic gates
  • An array, a task computing unit of a binary neural network model is fixedly arranged in the operator, and the binary neural network model is used to determine the health state information based on the multimodal physiological signal;
  • the battery is used to send the wearable components, processing components and output The components are powered, so that the health monitoring device can be made smaller in size, has lower power consumption, is easy to wear for a long time, and can more accurately predict heart failure through multimodal physiological signals including bioimpedance signals.
  • FIG. 1 shows a block diagram of a health monitoring device according to an embodiment of the present disclosure
  • FIG. 2 shows a schematic diagram of a health monitoring device according to an embodiment of the present disclosure
  • FIG. 3 illustrates a block diagram of processing components according to an embodiment of the present disclosure
  • FIG. 4 shows a schematic diagram of a master device according to an embodiment of the present disclosure.
  • the average age of the aging population in China and the world is increasing, and the number of people is increasing year by year. More and more people pay attention to the detection of the physiological data of the aging population.
  • the generated monitoring physiological data continues to grow at an exponential growth rate.
  • cardiovascular disease has become the leading cause of death in our country.
  • myocardial infarction, arrhythmia, sudden cardiac death and other cardiovascular diseases have sudden onset, rapid changes, serious illness and high concealment, which seriously threaten human health.
  • the common clinical ECG and 24-hour ECG are difficult to capture this type of abnormal ECG signals in patients.
  • Some dynamic ECG detectors are equipped with real-time Bluetooth transmission system, the data can be read on the smartphone terminal, and the 4G network data can be remotely transmitted to the background to achieve the monitoring effect.
  • background staff to perform 24-hour detection, which requires huge manpower and material resources, and when the network is unstable or even unavailable, the ECG signal of the monitored person may be lost, and important abnormalities may be missed. signal, causing the early warning function to fail.
  • these ambulatory electrocardiographs have large power consumption and volume, and are generally operated under the guidance of a doctor when worn, and cannot meet the needs of continuous medical testing for 3-7 days.
  • a plurality of electrodes are formed on the wearable component by using a conductive material, and a multimodal physiological signal acquisition circuit is formed between the plurality of electrodes, and the multimodal physiological signals are at least Including ECG signal and bio-impedance signal;
  • the scheduler is configured to receive multi-modal physiological signals, schedule a plurality of operators to determine the health status information of the user, and control the output component to output early warning information, and the operators include programmable logic gates
  • An array, a task computing unit of a binary neural network model is fixedly arranged in the operator, and the binary neural network model is used to determine the health state information based on the multimodal physiological signal;
  • the battery is used to send the wearable components, processing components and output The components are powered, so that the health monitoring device can be made smaller in size, has lower power consumption, is easy to wear for a long time, and can more accurately predict heart failure through multimodal physiological signals including bioimpedance signals.
  • the health monitoring device can solve the problem of network instability by performing local analysis on multi-modal physiological signals through a small, portable processing component.
  • Users can detect multi-modal physiological signals by wearing wearable devices, and the signals can be given real-time early warning in the local machine through the algorithm implanted in the machine, which can help residents, patients, health and health managers to achieve early warning of sudden cardiovascular diseases and solve a large number of problems. health management issues.
  • the mobile terminal realizes intelligent medical treatment. It can detect the physiological signals of patients in real time in an all-round way, and capture the paroxysmal abnormal heart rhythm that cannot be recorded by conventional short-term detection. mistake.
  • FIG. 1 shows a block diagram of a health monitoring device 100 according to an embodiment of the present disclosure.
  • the health monitoring device 100 includes:
  • the wearable component 120 is woven by mixing conductive material and insulating material, the conductive material forms a plurality of electrodes on the wearable component and a multimodal physiological signal acquisition circuit is formed between the plurality of electrodes, so
  • the multimodal physiological signals at least include ECG signals and bioimpedance signals;
  • the processing component 130 includes a scheduler and a plurality of operators, wherein the scheduler is configured to receive the multimodal physiological signal, schedule the plurality of operators to determine the user's health status information, and based on the The health state information controls the output component to output early warning information, the arithmetic unit includes a programmable logic gate array, and the arithmetic unit is fixedly provided with a task computing unit of a binary neural network model, and the binary neural network model
  • the weights and activation values of the hidden layers of are binary data for determining health state information based on multimodal physiological signals; and
  • the battery 140 is used to supply power to the wearable component, the processing component and the output component.
  • a plurality of electrodes are formed on the wearable component by using a conductive material, and a multimodal physiological signal acquisition circuit is formed between the plurality of electrodes, and the multimodal physiological signals are at least Including ECG signal and bio-impedance signal;
  • the scheduler is configured to receive multi-modal physiological signals, schedule a plurality of operators to determine the health status information of the user, and control the output component to output early warning information, and the operators include programmable logic gates
  • An array, a task computing unit of a binary neural network model is fixedly arranged in the operator, and the binary neural network model is used to determine the health state information based on the multimodal physiological signal;
  • the battery is used to send the wearable components, processing components and output The components are powered, so that the health monitoring device can be made smaller in size, has lower power consumption, is easy to wear for a long time, and can more accurately predict heart failure through multimodal physiological signals including bioimpedance signals.
  • the output assembly 110, the processing assembly 130 and the battery 140 may be packaged as the main control device through a casing, and the casing may be provided with a fixing structure, such as a spring clip, a fixing belt, etc., for fixing the main control device on the main control device.
  • a fixing structure such as a spring clip, a fixing belt, etc.
  • the master device may be fastened to the user's pants.
  • FIG. 2 shows a schematic diagram of a health monitoring device according to an embodiment of the present disclosure.
  • the health monitoring device may include a wearable component 210 and a main control device 220 .
  • the wearable component 210 is woven by a mixture of conductive material and insulating material, for example, it may include a plurality of elastic corset loops and an elastic connecting belt connecting the plurality of elastic corset loops, the conductive material is in the A plurality of electrodes are formed on the wearable component and a multimodal physiological signal acquisition circuit is formed between the plurality of electrodes.
  • the conductive material may be an elastic conductive yarn, for example, may contain a copper-nickel material or a conductive sponge material.
  • the insulating material can be, for example, cotton, hemp, silk or various chemical fibers.
  • the wearable component 210 is formed by mixing the conductive material and the insulating material, and an electrode with good bioconductivity is formed on the inner surface of the wearable component 210, thereby obtaining electrophysiological signals on the body surface.
  • the textile has the functions of comfort, fixation, high resilience, ultra-light and breathability.
  • the plurality of elastic corsets may include, for example, three elastic corsets located on the user's neck, chest and abdomen when the health monitoring device is worn by the user, or more or less may be provided.
  • elastic waistband located near the user's chest is usually necessary because many standardized indicators require electrodes at this location to be collected. However, if these standardized indicators are not required, elastic corsets may not be provided near the user's chest.
  • the elastic body girdle since only the elastic body girdle is used, it has better air permeability compared with the overall clothing, and in the case of long-term use, for example, more than 24 hours, or in hot weather In summer, the wearable components of the embodiments of the present disclosure bring a more comfortable experience to use.
  • Devices such as Holter require the help of a doctor to smooth the patient's skin before the electrodes can be attached to the skin, but they can only last for about 24 hours and cannot be worn for a long time.
  • the user can even take off the wearable component halfway, and when wearing it again, no complicated preparation is required. It can automatically fit with the user's body, reducing the difficulty of operation.
  • the elastic connecting straps can be positioned on the front and back sides of the user's spine, and the elastic corset straps on the neck can be pulled from the front and rear sides to fix it.
  • the elastic connecting straps are not limited to the form shown in the figure, for example, the elastic connecting straps located on the ventral side and the back side can only be provided between the elastic girdle straps located near the neck and near the chest, while the elastic connecting straps near the chest and near the abdomen
  • the elastic connecting belts between the body girdle belts can be arranged on the left and right sides of the human body.
  • the length of the elastic corset can be adjusted, for example, a length adjustment structure can be provided on the back side, so as to further improve the adaptability of the wearable component and reduce the probability of the electrode falling off.
  • the elastic corset band on the user's chest has an asymmetric structure in the left-right direction. Since the position of the heart is on the left side, in order to arrange more acquisition electrodes around the position of the heart, the elastic corset band near the chest in the embodiment of the present disclosure has an asymmetric structure in the left-right direction, as shown in FIG. 2 , The area on the left is larger than the area on the right to accommodate more electrodes. While the right side does not require more acquisition electrodes, it can have a smaller area to maintain comfort.
  • electrodes with good bioconductivity are formed on the inner surface of the wearable component 210 through the mixed weaving of conductive materials, so as to obtain electrophysiological signals on the body surface.
  • these electrodes may be located in at least two of the following positions of the user: the fourth intercostal space on the right sternal border, the fourth intercostal space on the left sternal border, the intersection of the left midclavicular line and the fifth intercostal space, The midpoint of the line connecting the fourth intercostal space on the left sternal border and the intersection, the left anterior axillary line and the intersection at the same level, the left midaxillary line at the same level as the intersection, and the left posterior axillary line at the same level as the intersection ,
  • the paraspinal area is at the same level as the intersection point, the left and right anterior sides of the abdomen, the anterior sides of the left and right clavicle, near the neck, and near the central position of the back of the spine
  • more sensors may be set on the health monitoring device, for example, for detecting heart rate, pre-ejection period, ventricular ejection time, mean arterial pressure, left ventricular systolic time, cardiac contractility index, central vein Pressure, pulmonary artery occlusion pressure, some or all of the various sensors, the multimodal physiological signal acquisition circuit can transmit these data to the processing component, that is, in addition to ECG data and bioimpedance data, multimodal physiological signals
  • the acquisition circuit may also acquire some or all of heart rate, pre-ejection, ventricular ejection time, mean arterial pressure, left ventricular systolic time, cardiac contractility index, central venous pressure, and pulmonary artery occlusion pressure.
  • the processing component may further include a preprocessing unit configured to perform preprocessing on the multimodal physiological signal, the preprocessing including noise processing, baseline drift processing, beat segmentation and data resampling at least one of them.
  • Multi-modal physiological signals can be preprocessed, for example, through noise processing and baseline drift processing, the information in multi-modal physiological signals can be expressed more effectively; through heartbeat segmentation, the input data is made in heartbeat units, which simplifies the processing of the model It is difficult and easy to train; through data resampling, the size of the input data can be reduced, the amount of calculation can be reduced, and the processing efficiency can be improved.
  • the health monitoring device may further include an input component, and a user may use the input component to input necessary information, such as height and weight information.
  • the input component may be, for example, a keyboard, a touch screen, a stylus, a camera, a wireless signal receiving device, and the like.
  • the processing component may further include a feature pre-extraction unit, configured to preliminarily process the acquired data, and use the preliminarily processed result obtained as a pre-extracted feature, which is input to the binary value together with the ECG signal and the bioimpedance signal. in the neural network model.
  • a feature pre-extraction unit configured to preliminarily process the acquired data, and use the preliminarily processed result obtained as a pre-extracted feature, which is input to the binary value together with the ECG signal and the bioimpedance signal.
  • the feature pre-extraction may include, for example, one or more of the following.
  • BSA 0.024265 ⁇ height 0.3964 ⁇ weight 0.5378 .
  • TFIT can be determined on the first derivative of the bioimpedance dZ/dT, where the TFIT is the first time after the cardiac cycle begins.
  • the interval between a zero-crossing location and the location of the first minimum after the peak ventricular ejection rate (dZ/dT max ), then SVI and SV are calculated by the following equations:
  • TFIT cal the weight used to balance the relationship among TFIT cal , HR HR, and (systolic arterial pressure SAP-diastolic arterial pressure DAP) .
  • the ejection fraction EF based on the pre-ejection PEP and the ventricular ejection time VET.
  • the input of the binary neural network model may include ECG signals, bioimpedance signals, left ventricular drainage function indexes, cardiac systolic function indexes, preload indexes, and afterload indexes.
  • the left ventricular function index includes at least one of stroke volume SV, stroke index SVI, cardiac output CO, and cardiac index CI;
  • the myocardial contractility index includes cardiac contractility index CTI, left ventricular At least one of work index LCWi, left ventricular systolic time LVET, and ejection fraction EF;
  • the preload index includes central venous pressure CVP and/or left ventricular end-diastolic volume EDV;
  • the afterload index includes peripheral vascular resistance SVR and/or peripheral vascular resistance index SVRi.
  • the health status recognition model based on deep learning has more network layers to classify features, which can improve the accuracy.
  • the existing single commercial semiconductor platform is often difficult to balance computing power and power consumption.
  • the related algorithms become more and more complex, so the amount of computation required to complete a specific task is also increasing, and the power consumption during runtime also increases.
  • the embodiments of the present disclosure provide a customized processing component, which can effectively alleviate the above problems.
  • FIG. 3 shows a block diagram of a processing component 300 in accordance with an embodiment of the present disclosure.
  • the processing component 300 includes a scheduler 311 and a plurality of operators 321 , 322 and 323 , wherein the scheduler 311 is configured to receive multimodal physiological signals, and schedule the multiple operators to determine the user’s health status information, and control the output component 110 to output early warning information based on the health status information, the operator includes a programmable logic gate array (FPGA), and the operator is fixedly provided with a binary neural network model.
  • FPGA programmable logic gate array
  • a task computing unit wherein the weights and activation values of the hidden layer of the binary neural network model are binary data for determining health state information based on multimodal physiological signals.
  • the processing assembly 300 may be implemented as a printed circuit board (PCB).
  • the printed circuit board may include, for example, a fixed portion 310 and a reconfigurable portion 320 .
  • the fixed part 310 includes a scheduler 311 for allocating computing tasks to operators according to a scheduling policy;
  • the reconfigurable part 320 includes multiple operators, such as operators 321, 322, and 323, among the multiple operators Part or all of the task computing unit is fixedly provided with a binary neural network model.
  • the above layout can reduce the area of the original hardware circuit and reduce the volume of the processing components.
  • the reconfigurable portion 320 may, for example, contain hundreds of parallel cores, and the spatial redundancy of computing resources allows safety and non-safety critical applications to coexist, thereby providing an appropriate partitioning mechanism that can be used to achieve high reliability, secure authentication and re-certified multi-core computing processing systems.
  • the scheduler may be implemented by, for example, a central processing unit (CPU).
  • the scheduler is only used for task scheduling and does not participate in any operation. All operation tasks are executed by the operation unit, which can reduce system power consumption.
  • the fixed part 310 may further include functional modules such as a runtime environment 312, an operating system 313, and an input and output management 314, so as to ensure that the processing components realize necessary functions.
  • the RTE Resourcetime Environment
  • the RTE is to aggregate different devices, build an installable client driver loader, and act as a proxy between the user program and the actual implementation, so that different vendors can be called smoothly.
  • OpenCL implementation without any conflict.
  • These modules may be implemented in a software manner, may be implemented in a programmable hardware manner, or may be implemented in a combination of software and hardware, which is not limited in the present disclosure.
  • the operator may be an accelerator supporting PCIe.
  • the arithmetic unit may include one or more of the following: General-purpose computing on graphics processing units (GPGPU, General-purpose computing on graphics processing units), Field Programmable Gate Array (FPGA, Field Programmable Gate Array), Field Programmable Gate Array integrated chip.
  • FPGA appears as a semi-custom circuit in the field of special application circuits, which not only solves the shortcomings of full-custom circuits, but also overcomes the shortcomings of the limited number of gate circuits of the original programmable logic device.
  • One of its notable features is low power consumption.
  • GPGPU enables stream processing components to process non-graphics data due to the powerful parallel processing capabilities and programmable pipelines of modern graphics processing components.
  • SIMD single instruction stream multiple data stream
  • GPGPU greatly surpasses the traditional central processing unit in performance.
  • the feasibility of making the FPGA into a chip is verified by simulation technology, so that the FPGA can be made into a smaller integrated chip based on the field programmable logic gate array, which can further reduce the processing time. size, weight and power consumption of the device.
  • the FPGA in the operator is customized based on the trained neural network model.
  • the process of customizing the FPGA includes: constructing a neural network-based health state recognition model, using ECG data as training data to train and compress the health state recognition model, and customizing a programmable logic gate array based on the compressed health state model.
  • the step of compression can include pruning or parameter binarization, pruning reduces the association between neurons, and parameter binarization makes neurons easily implemented as hardware circuits, both of which greatly reduce the neural network. It also reduces the complexity of the deployed FPGA, which can improve computing efficiency and reduce power consumption.
  • the weight value of the binary neural network is -1 or 1. Therefore, the multiplication calculation of the weight value between neurons can be simplified as a bit operation, and the multiplication by -1 can be realized by complementing the code. operation, which greatly improves the efficiency of model calculation.
  • the binary neural network model can be obtained by training the following methods:
  • the real-number neural network model is trained and compressed using the training data to obtain a binary neural network model.
  • a real-number type neural network can be trained on a platform with higher computing power first, and the real-number type neural network can include, for example, an input layer, four hidden layers and an output layer.
  • the real-number type neural network can include, for example, an input layer, four hidden layers and an output layer.
  • the use of training data to train and compress the real-number neural network model to obtain a binary neural network model includes:
  • the training operation and the compression operation are alternately performed until the convergence conditions are met, and a binary neural network model is obtained, wherein, in the training operation, the weights in the model are updated through the training data; in the compression operation, the Size of branches and/or binarized compressed models.
  • acquiring training data includes:
  • preprocessing on the ECG data to obtain training samples, where the preprocessing includes at least one of noise processing, baseline drift processing, heart beat segmentation, feature pre-extraction, and data resampling;
  • the training samples are marked to obtain sample labels, and the sample labels represent different health states corresponding to the ECG data, and the training samples and the sample labels are determined as training data.
  • multi-modal physiological signals can be de-noised first.
  • the EMG interference noise is eliminated through a low-pass filter, and the baseline drift is corrected through an IIR zero-phase-shift digital filter.
  • the embodiments of the present invention do not limit the noise removal method.
  • the heartbeats of the heartbeat signals can be divided to obtain multiple heartbeat signals of each heartbeat signal, and the sampled data obtained by performing data resampling on the heartbeat signals can be used as training samples , the ECG signal to which the heartbeat signal belongs is a patient with abnormal ECG or a normal person as the sample label of the training sample, and the training sample and the sample label constitute the training data.
  • the real-number neural network model can be trained by using the training data to update the real-number weights of the real-number neural network model, and then the real-number neural network model is performed pruning, and binarizing the real weights to obtain a binary neural network model, and training the binary neural network model through the training data until convergence.
  • the binary neural network model is obtained through one compression and trained until convergence.
  • compression can be performed in multiple rounds. Each compression can prune and/or binarize some nodes, and continue training after compression. Then enter the next round of compression until both the model size and prediction accuracy meet the predetermined convergence conditions.
  • the FPGA since the FPGA is customized based on a specific neural network, it has smaller volume, lower power consumption and higher efficiency than general processing components.
  • the comprehensive index of the processing component according to the embodiment of the present disclosure can reach the following levels:
  • two TPS54386 dual 3A asynchronous converters may be used for the power supply interface of the processing component to ensure stable and sufficient dual power supply.
  • the processing component may further include a power supply management unit configured to adjust the number of powered arithmetic units based on a current task load. For example, when the current task load is not less than the first threshold, all operators may be enabled; when the current task load is less than the first threshold, the number of enabled budgeters may be determined according to the current task load. The number of enabled operators may be determined according to the ratio of the current task load to the first threshold, or an interval may be defined, and the number of enabled operators may be determined according to a predetermined correspondence.
  • the processor of the embodiment of the present disclosure can effectively reduce its power consumption.
  • the power supply management unit, the preprocessing unit, or the feature pre-extracting unit described in the embodiments of the present disclosure may be implemented in a software manner, or may be implemented in a programmable hardware manner.
  • the power management unit, the preprocessing unit or the feature pre-extracting unit, for example, may be set in the scheduler, and the names of these units or modules do not constitute a limitation on the unit or the module itself.
  • the output component may have at least a first output mode and a second output mode
  • the processing component is configured to:
  • the output device is controlled to output early warning information according to the second output mode.
  • FIG. 4 shows a schematic diagram of a master device 400 according to an embodiment of the present disclosure.
  • the main control device 400 may include, for example, a screen 410 , indicator lights 421 , 422 , and 423 , buttons 431 , 432 , and the like.
  • the types and numbers of the above input/output units are only exemplary, more or less input/output units may be provided, or different types of input/output units may be provided, or one of the above input/output units may be omitted. one or more.
  • the output device may have different output modes for prompting different types of health states.
  • the indicator light 421 or the screen 410 can be turned on, and prompt information can also be displayed on the screen 410; in the case of).
  • the first output mode may be that the indicator light 423 or the screen 410 is blinking
  • the second output mode may be that the indicator light 423 or the screen 410 is always on.
  • the first output mode may be a local prompt
  • the second output mode may be sending early warning information to a predetermined device to remind a predetermined contact.
  • early warning information can be sent to the predetermined device to remind the predetermined contact.
  • the output device includes three-color indicator lights.
  • indicator lights 421, 422, and 423 may be green indicator lights, yellow indicator lights, and red indicator lights, respectively, or may be integrated in the same indicator light.
  • the processing component is configured to:
  • the red indicator light is controlled to light up.
  • Some related algorithms process the ECG signal to obtain the patient's disease name, such as arrhythmia, atrial fibrillation, heart failure, sudden cardiac death, etc. If these contents are directly used for the prompts of health monitoring equipment, patients who lack the relevant knowledge cannot respond well. According to the technical solutions provided by the embodiments of the present disclosure, users can understand the monitoring results without having to master medical knowledge, and thus take appropriate countermeasures. For example, if you find that the red indicator light is on, you can call an ambulance immediately. If you find that the yellow indicator light is on, you can go to the hospital by yourself. If the green indicator light is on, you don't need to take any measures.
  • the health monitoring device can give an early warning of the user's health status in time; local analysis of multi-modal physiological signals through a small, portable processing component can well solve the problem of network problems.
  • Stability issue Users can detect multi-modal physiological signals by wearing wearable devices, and the signals can be given real-time early warning in the local machine through the algorithm implanted in the machine, which can help residents, patients, health and health managers to achieve early warning of sudden cardiovascular diseases and solve a large number of problems. health management issues.
  • the mobile terminal realizes intelligent medical treatment. It can detect the physiological signals of patients in real time in an all-round way, and capture the paroxysmal abnormal heart rhythm that cannot be recorded by conventional short-term detection. mistake.

Landscapes

  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Surgery (AREA)
  • Biophysics (AREA)
  • Pathology (AREA)
  • Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Medical Informatics (AREA)
  • Molecular Biology (AREA)
  • Physics & Mathematics (AREA)
  • Animal Behavior & Ethology (AREA)
  • General Health & Medical Sciences (AREA)
  • Public Health (AREA)
  • Veterinary Medicine (AREA)
  • Dermatology (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
  • Measuring And Recording Apparatus For Diagnosis (AREA)

Abstract

一种健康监测设备（100），包括输出组件（110）、可穿戴组件（120，210）、处理组件（130，300）以及电池（140）。其中，可穿戴组件（120，210）由导电材料和绝缘材料混合编织而成，导电材料在可穿戴组件（120，210）上形成多个电极以及在多个电极之间形成多模态生理信号采集电路；处理组件（130，300）包括调度器（311）和多个运算器（321，322，323），调度器（311）被配置为接收多模态生理信号，调度多个运算器（321，322，323）以确定使用者的健康状态信息，并控制输出组件（110）输出预警信息，运算器（321，322，323）包括可编程逻辑门阵列，运算器（321，322，323）内固化地设置有用于确定健康状态信息的二值神经网络模型的任务运算单元；电池（140），用于向可穿戴组件（120，210）、处理组件（130，300）以及输出组件（110）供电。该设备功耗低，体积小，便于长时间佩戴，通过包括生物阻抗信号的多模态生理信号可以更加准确地预测心力衰竭。

Description

健康监测设备
相关申请的交叉引用
本公开要求2020年10月26日提交的中国专利申请号为“CN 202011158916.1”的优先权，其全部内容作为整体并入本申请中。
技术领域
本公开涉及可穿戴设备技术领域，具体涉及一种健康监测设备。
背景技术
心血管疾病由于其隐蔽性和突发性等特点，严重威胁着人类的健康，因此早期诊断和预防是非常重要的。目前，医院普遍使用的12导联动态检测仪(Holter)功耗和体积较大，穿戴时一般在医生的指导下进行操作，且无法满足3-7天的医用连续检测需求。并且，目前的Holter的心电监测判断需要把数据传输到后台服务器计算，这种计算滞后非常明显，不能实现实时计算和全天候即时预警的需求。此外，本发明人发现，单独使用心电数据不能很好地预测心力衰竭的情况。
发明内容
为了解决相关技术中的问题，本公开实施例提供一种健康监测设备。
具体地，所述健康监测设备包括：
输出组件；
可穿戴组件，由导电材料和绝缘材料混合编织而成，所述导电材料在所述可穿戴组件上形成多个电极以及在所述多个电极之间形成多模态生理信号采集电路，所述多模态生理信号至少包括心电信号和生物阻抗信号；
处理组件，包括调度器和多个运算器，其中，所述调度器被配置为接收所述多模态生理信号，调度所述多个运算器以确定使用者的健康状态信息，并基于所述健康状态信息控制所述输出组件输出预警信息，所述运算器包括可编程逻辑门阵列，所述运算器内固化地设置有二值神经网络模型的任务运算单元，所述二值神经网络模型的隐藏层的权值和激活值为二值数据，用于 基于多模态生理信号确定健康状态信息；以及
电池，用于向所述可穿戴组件、处理组件以及输出组件供电。
根据本公开实施例，所述调度器包括中央处理器，所述运算器还包括通用图形处理器和基于可编程逻辑门阵列定制的集成芯片，所述调度器和运算器部署于印制电路板。
根据本公开实施例，所述处理组件还包括：
预处理单元，被配置为对所述多模态生理信号进行预处理，所述预处理包括噪声处理、基线漂移处理、心拍分割和数据重采样中的至少一种。
根据本公开实施例，所述健康监测设备还包括输入组件，所述处理组件还被配置为基于用户输入的身高体重信息，确定用户的体表面积。
根据本公开实施例，所述多模态生理信号采集电路还用于采集心率、射血前期、心室射血时间、平均动脉压、左心收缩时间、心收缩力指数、中心静脉压、肺动脉阻塞压中的部分或全部。
根据本公开实施例，处理组件还包括特征预提取单元，被配置为执行以下一种或多种：
基于所述生物阻抗信号确定每搏输出量和/或每搏指数；
基于所述射血前期和心室射血时间确定射血分数；
基于所述每搏输出量和射血分数确定左室舒张末期容积；
基于所述每搏输出量和心率确定心输出量；
基于所述心输出量和体表面积确定心脏指数；
基于所述心脏指数、平均动脉压和肺动脉阻塞压确定左心做功指数；
基于所述心输出量、平均动脉压和中心静脉压确定外周血管阻力；
基于所述心脏指数、平均动脉压和中心静脉压确定外周血管阻力指数。
根据本公开实施例，所述二值神经网络模型的输入包括心电信号、生物阻抗信号、左室排血功能指标、心机收缩功能指标、前负荷指标以及后负荷指标，其中：
所述左室排血功能指标包括每搏输出量、每搏指数、心输出量、心脏指数中的至少一种；
所述心肌收缩指标包括心收缩力指数、左心做功指数、左心收缩时间、射血分数中的至少一种；
所述前负荷指标包括中心静脉压和/或左室舒张末期容积；
所述后负荷指标包括外周血管阻力和/或外周血管阻力指数。
根据本公开实施例，所述二值神经网络模型通过以下方法训练得到：
构造实数型神经网络模型，所述实数型神经网络模型的隐藏层的权值和激活值为实数型数据；
交替执行训练操作和压缩操作，直至满足收敛条件，得到二值神经网络模型，其中，在所述训练操作中，通过所述训练数据更新模型中的权值；在所述压缩操作中，通过剪枝和/或二值化压缩模型的大小。
根据本公开实施例，所述输出组件具有第一输出模式和第二输出模式，所述处理组件被配置为：
在所述健康状态为第一类状态的情况下，控制所述输出装置按照第一输出模式输出预警信息；
在所述健康状态为第二类状态的情况下，控制所述输出装置按照第二输出模式输出预警信息。
根据本公开实施例，所述输出装置包括三色指示灯，所述处理组件被配置为：
在所述健康状态为正常状态的情况下，控制绿色指示灯点亮；
在所述健康状态为异常状态的情况下，控制黄色指示灯点亮；
在所述健康状态为紧急状态的情况下，控制红色指示灯点亮。
根据本公开实施例提供的技术方案，通过导电材料在所述可穿戴组件上形成多个电极以及在所述多个电极之间形成多模态生理信号采集电路，所述多模态生理信号至少包括心电信号和生物阻抗信号；调度器被配置为接收多模态生理信号，调度多个运算器以确定使用者的健康状态信息，并控制输出组件输出预警信息，运算器包括可编程逻辑门阵列，运算器内固化地设置有二值神经网络模型的任务运算单元，二值神经网络模型用于基于多模态生理信号确定健康状态信息；电池，用于向可穿戴组件、处理组件以及输出组件供电，从而使该健康监测设备的体积可以制作得更小，功耗更低，便于长时间佩戴，通过包括生物阻抗信号的多模态生理信号可以更加准确地预测心力衰竭。
应当理解的是，以上的一般描述和后文的细节描述仅是示例性和解释性 的，并不能限制本公开。
附图说明
结合附图，通过以下非限制性实施方式的详细描述，本公开的其它特征、目的和优点将变得更加明显。在附图中：
图1示出根据本公开实施例的健康监测设备的框图；
图2示出根据本公开实施例的健康监测设备的示意图；
图3示出根据本公开实施例的处理组件的框图；
图4示出根据本公开实施例的主控设备的示意图。
具体实施方式
下文中，将参考附图详细描述本公开的示例性实施例，以使本领域技术人员可容易地实现它们。此外，为了清楚起见，在附图中省略了与描述示例性实施例无关的部分。
在本公开中，应理解，诸如“包括”或“具有”等的术语旨在指示本说明书中所公开的特征、数字、步骤、行为、部件、部分或其组合的存在，并且不欲排除一个或多个其他特征、数字、步骤、行为、部件、部分或其组合存在或被添加的可能性。
另外还需要说明的是，在不冲突的情况下，本公开中的实施例及实施例中的特征可以相互组合。下面将参考附图并结合实施例来详细说明本公开。
目前，中国乃至全球老龄化平均年龄增高、人数逐年增加，对老龄人口生理数据检测受到越来越多人的重视，同时产生的监护生理数据以指数的增长速度持续增长。
近年来，心血管疾病已成为我国居民死亡的首要原因。其中心肌梗死、心律失常、心脏性猝死等心血管疾病发病突然、变化快、病情重、隐蔽性高，严重威胁着人类的健康。而临床上常见的普通心电图、24小时心电图很难捕捉到患者这种类型异常的心电信号。
发明人通过市场调研发现，包括Holter和植入式心电监测仪等动态心电检测仪在内，大多数的动态心电检测仪仅能实现数据记录的功能，均无本机植入算法来达到心血管事件发生时的本机预警功能，均需要终端临床医学专 家解读数据。有部分动态心电检测仪配置有实时蓝牙传输系统，数据可在智能手机端读取，并实现4G网络数据远程传输至后台以达到监控效果。但是，目前尚需要后台工作人员来24小时检测，需要耗费巨大的人力物力，并且在网络不稳定甚至不可用的情况下，被监测的人的心电信号可能发生丢失，从而可能错过重要的异常信号，导致预警作用失效。另一方面，该些动态心电检测仪功耗和体积较大，穿戴时一般在医生的指导下进行操作，且无法满足3-7天的医用连续检测需求。
根据本公开实施例提供的技术方案，通过导电材料在所述可穿戴组件上形成多个电极以及在所述多个电极之间形成多模态生理信号采集电路，所述多模态生理信号至少包括心电信号和生物阻抗信号；调度器被配置为接收多模态生理信号，调度多个运算器以确定使用者的健康状态信息，并控制输出组件输出预警信息，运算器包括可编程逻辑门阵列，运算器内固化地设置有二值神经网络模型的任务运算单元，二值神经网络模型用于基于多模态生理信号确定健康状态信息；电池，用于向可穿戴组件、处理组件以及输出组件供电，从而使该健康监测设备的体积可以制作得更小，功耗更低，便于长时间佩戴，通过包括生物阻抗信号的多模态生理信号可以更加准确地预测心力衰竭。
本公开实施例提供的健康监测设备通过小型、可以随身携带的处理组件对多模态生理信号进行本地分析可以很好的解决网络不稳定的问题。使用者通过佩戴可穿戴设备检测多模态生理信号，信号通过本机植入的算法在本机实时预警，可以帮助居民、患者、健康及卫生管理人员实现突发性心血管疾病预警、解决大量健康管理问题。移动终端实现智能医疗能够全方位实时检测患者生理信号，捕捉到常规短时检测记录不到的阵发性心律异常，同时还能够避免收到医生技术水平、熟练程度、疲劳程度等影响导致的诊断失误。
图1示出根据本公开实施例的健康监测设备100的框图。
如图1所示，该健康监测设备100包括：
输出组件110；
可穿戴组件120，由导电材料和绝缘材料混合编织而成，所述导电材料在所述可穿戴组件上形成多个电极以及在所述多个电极之间形成多模态生理信号采集电路，所述多模态生理信号至少包括心电信号和生物阻抗信号；
处理组件130，包括调度器和多个运算器，其中，所述调度器被配置为接收所述多模态生理信号，调度所述多个运算器以确定使用者的健康状态信息，并基于所述健康状态信息控制所述输出组件输出预警信息，所述运算器包括可编程逻辑门阵列，所述运算器内固化地设置有二值神经网络模型的任务运算单元，所述二值神经网络模型的隐藏层的权值和激活值为二值数据，用于基于多模态生理信号确定健康状态信息；以及
电池140，用于向所述可穿戴组件、处理组件以及输出组件供电。
根据本公开实施例提供的技术方案，通过导电材料在所述可穿戴组件上形成多个电极以及在所述多个电极之间形成多模态生理信号采集电路，所述多模态生理信号至少包括心电信号和生物阻抗信号；调度器被配置为接收多模态生理信号，调度多个运算器以确定使用者的健康状态信息，并控制输出组件输出预警信息，运算器包括可编程逻辑门阵列，运算器内固化地设置有二值神经网络模型的任务运算单元，二值神经网络模型用于基于多模态生理信号确定健康状态信息；电池，用于向可穿戴组件、处理组件以及输出组件供电，从而使该健康监测设备的体积可以制作得更小，功耗更低，便于长时间佩戴，通过包括生物阻抗信号的多模态生理信号可以更加准确地预测心力衰竭。
根据本公开实施例，输出组件110、处理组件130和电池140可以通过外壳封装为主控装置，该外壳上可以设置有固定结构，例如弹簧夹、固定带等，用于将主控装置固定于使用者衣物、可穿戴组件或使用者身体。例如，主控设备可以固定在使用者的裤子上。
图2示出根据本公开实施例的健康监测设备的示意图。
如图2所示，该健康监测设备可以包括可穿戴组件210和主控装置220。
根据本公开实施例，可穿戴组件210由导电材料和绝缘材料混合编织而成，例如可以包括多条弹性束身环带以及连接所述多条弹性束身环带的弹性连接带，所述导电材料在所述可穿戴组件上形成多个电极以及在所述多个电极之间形成多模态生理信号采集电路。
根据本公开实施例，该导电材料可以是弹性导电纱线，例如可以包含铜镍材料或导电海绵材料。该绝缘材料例如可以是棉、麻、蚕丝或各种化学纤维。导电材料与绝缘材料通过混合编制形成可穿戴组件210，在可穿戴组件 210的内表面形成具有良好生物导电性的电极，从而获取体表电生理信号。该纺织物具有舒适、固定、高回弹、超轻、透气等功能。
根据本公开实施例，多条弹性束身环带例如可以包括所述健康监测设备被使用者穿戴时处于使用者颈部、胸部和腹部的三条弹性束身环带，或者也可以设置更多或更少的弹性束身环带。其中，位于使用者胸部附近的弹性束身环带通常情况下是必要的，因为许多标准化的指标需要该位置的电极进行采集。但如果不需要这些标准化的指标，也可以不在使用者胸部附近设置弹性束身环带。
根据本公开实施例提供的技术方案，由于仅采用弹性束身环带，相对于制成整体衣物而言，具有更好的透气性，在长时间使用的情况下，例如24h以上，或者在炎热的夏季，本公开实施例的可穿戴组件为使用带来更加舒适的体验。
另外，多条弹性束身环带之间几乎不会产生相互干扰，可以更好地适应身体局部的情况，例如不同的胸围和腰围，使弹性束身环带与身体之间固定得更为稳定，弹性束身环带上的电极与身体之间不易脱落。
Holter等设备需要在医生的帮助下，将患者的皮肤打磨平滑才能将电极贴在皮肤上，但也只能维持24h左右，无法长期佩戴。本公开实施例的健康监测设备在长期佩戴的场景中，使用者甚至可以中途脱下该可穿戴组件，再次穿戴时，无需复杂的准备，由于具有较强的弹性，弹性束身环带在佩戴时可以自动与使用者的身体贴合，降低操作难度。
根据本公开实施例，弹性连接带可以位于使用者脊椎的前侧和后侧，可以从前后两侧拉动位于颈部的弹性束身环带，使其固定。弹性连接带不限于图中所示意的形式，例如位于腹侧和背侧的弹性连接带可以仅设置在位于颈部附近与胸部附近的弹性束身环带之间，而胸部附近与腹部附近的弹性束身环带之间的弹性连接带可以设置于人体的左右两侧。
根据本公开实施例，弹性束身环带的长度可以是可调节的，例如可以在背侧设置长度调节结构，从而可以进一步提高可穿戴组件的适应性，降低电极脱落的概率。
根据本公开实施例，所述处于使用者胸部的弹性束身环带在左右方向上具有非对称的结构。由于心脏位置在偏左的位置，为了在心脏位置周围可以 布置较多的采集电极，本公开实施例的胸部附近的弹性束身环带在左右方向上具有非对称的结构，如图2所示，左侧的面积比右侧的面积更大，以容纳更多的电极。而右侧不需要更多采集电极的情况下，可以具有较小的面积，以保持舒适性。
根据本公开实施例，通过导电材料的混合编织，在可穿戴组件210的内表面形成具有良好生物导电性的电极，从而获取体表电生理信号。健康监测设备被使用者穿戴时，这些电极例如可以处于使用者以下部位中的至少两种：胸骨右缘第四肋间，胸骨左缘第四肋间、左锁骨中线与第五肋间交点、胸骨左缘第四肋间与所述交点连线的中点、左腋前线与所述交点同一水平处、左腋中线与所述交点同一水平处、左腋后线与所述交点同一水平处、脊柱旁与所述交点同一水平处、腹部左右前侧、左右锁骨前侧、颈部附近、背部的脊椎中央位置附近。通过以上设置，可以方便地采集人体12导联的心电数据和生物阻抗数据，其中，四肢的电极被替换为锁骨和腹部附近的电极。
根据本公开实施例，该健康监测设备上还可以设置更多的传感器，例如用于检测心率、射血前期、心室射血时间、平均动脉压、左心收缩时间、心收缩力指数、中心静脉压、肺动脉阻塞压中的部分或全部的各种传感器，多模态生理信号采集电路可以将该些数据传输到处理组件，即，除了心电数据和生物阻抗数据之外，多模态生理信号采集电路还可以采集心率、射血前期、心室射血时间、平均动脉压、左心收缩时间、心收缩力指数、中心静脉压、肺动脉阻塞压中的部分或全部。
根据本公开实施例，所述处理组件还可以包括预处理单元，被配置为对所述多模态生理信号进行预处理，所述预处理包括噪声处理、基线漂移处理、心拍分割和数据重采样中的至少一种。
根据本公开实施例，采集的原始数据直接输入到健康状态识别模型难以取得较好的识别效果。可以对多模态生理信号进行预处理，例如通过噪声处理和基线漂移处理使多模态生理信号中的信息得到更加有效地表达；通过心拍分割，使输入数据以心拍为单位，简化模型的处理难度，易于训练；通过数据重采样，可以减小输入数据的大小，降低计算量，提高处理效率。
根据本公开实施例，该健康监测设备还可以包括输入组件，用户可使用输入组件输入必要的信息，例如身高体重信息。该输入组件例如可以是键盘、 触摸屏、手写笔、摄像头、无线信号接收装置等。
根据本公开实施例，处理组件还可以包括特征预提取单元，用于初步处理已获得的数据，将得到的初步处理结果作为预提取的特征，与心电信号和生物阻抗信号共同输入到二值神经网络模型中。通过特征预提取，可以利用已有的经验，提高模型预测的准确度。
根据本公开实施例，特征预提取例如可以包括以下一种或多种。
(1)基于身高体重信息确定用户的体表面积BSA，例如可以通过下式确定，BSA＝0.024265×身高 0.3964×体重 0.5378。
(2)基于所述生物阻抗信号Z确定每搏输出量SV和/或每搏指数SVI，例如，可以在生物阻抗的一阶导数dZ/dT上确定TFIT，该TFIT为心动周期开始后的第一个过零点位置与心室射血速率峰值(dZ/dT max)之后的第一个极小值位置之间的间隔，然后通过如下公式计算SVI和SV：
SVI＝k×((dZ/dT max)/(Zmax-Zmin))×W(TFIT cal)      (在校准阶段)
SV＝SVI×((dZ/dT max)/(dZ/dT max) cal×TFIT cal/TFIT) 1/3×BSA   (在校准之后)
其中，k是常数，下标cal表示在校准阶段测量的参数，W(TFIT cal)是用于平衡TFIT cal、心率HR、(收缩期动脉压SAP-舒张期动脉压DAP)三者关系的权重。
(3)基于所述射血前期PEP和心室射血时间VET确定射血分数EF，例如可以通过下式确定射血分数，EF＝0.84-(0.64×PEP)/VET。
(4)基于所述每搏输出量SV和射血分数EF确定左室舒张末期容积EDV，EDV＝SV/EF。
(5)基于所述每搏输出量SV和心率HR确定心输出量CO，CO＝SV×HR/1000。
(6)基于所述心输出量CO和体表面积BSA确定心脏指数CI，CI＝CO/BSA。
(7)基于所述心脏指数CI、平均动脉压MAP和肺动脉阻塞压PAOP确定左心做功指数LCWi，例如可以通过下式确定，LCWi＝0.0144×CI×(MAP-PAOP)。
(8)基于所述心输出量CO、平均动脉压MAP和中心静脉压CVP确定外周血管阻力SVR，SVR＝80×(MAP-CVP)/CO。
(9)基于所述心脏指数CI、平均动脉压MAP和中心静脉压CVP确定外周血管阻力指数SVRi，SVR＝80×(MAP-CVP)/CI。
(10)基于心电数据确定例如P波长度，PQ间长度，QRS波长度等。
根据本公开实施例，二值神经网络模型的输入可以包括心电信号、生物阻抗信号、左室排血功能指标、心机收缩功能指标、前负荷指标以及后负荷指标。其中，所述左室排血功能指标包括每搏输出量SV、每搏指数SVI、心输出量CO、心脏指数CI中的至少一种；所述心肌收缩指标包括心收缩力指数CTI、左心做功指数LCWi、左心收缩时间LVET、射血分数EF中的至少一种；所述前负荷指标包括中心静脉压CVP和/或左室舒张末期容积EDV；所述后负荷指标包括外周血管阻力SVR和/或外周血管阻力指数SVRi。通过上述指标的引入，可以有效提高神经网络模型的预测准确率。
本发明人发现，用于心电信号处理的传统算法的准确率不是很高，必须依靠大量的手动特征提取，具有很强的主观性，获得的特征也没有层次可言，需要由相关领域的专家来进行人工鉴别，对相关专业知识要求较高。基于深度学习的健康状态识别模型拥有更多网络层数来进行特征的分类，能够提高准确率。然而，随着芯片设计规模的与日俱增，其功能日趋复杂，功能、架构、设计思路都存在较大差异，现有的单一商用半导体平台在算力和功耗上往往难以平衡。一方面，随着功能和能力的演进，相关算法变得越来越复杂，因此完成特定任务所需的计算量也越来越大，从而运行时所产生的功耗也随之增大。另一方面，由于对可穿戴设备的便携性要求，很难部署复杂的智能算法。在载荷和能耗有限的环境中，也很难放置体积大，质量大，高能耗的高性能图形处理器(GPU)板卡来处理复杂的应用。因此，往往由于载荷、功耗、体积的限制，大大约束了可穿戴设备处理器的处理能力。
本公开实施例提供了一种定制化的处理组件，可以有效缓解上述问题。
图3示出根据本公开实施例的处理组件300的框图。
如图3所示，该处理组件300包括调度器311和多个运算器321、322、323，其中，调度器311被配置为接收多模态生理信号，调度多个运算器以确定使用者的健康状态信息，并基于所述健康状态信息控制所述输出组件110输出预警信息，所述运算器包括可编程逻辑门阵列(FPGA)，所述运算器内固化地设置有二值神经网络模型的任务运算单元，所述二值神经网络模型的 隐藏层的权值和激活值为二值数据，用于基于多模态生理信号确定健康状态信息。
通过将隐藏层的权值和激活值限制为二值数据，并将神经网络模型固化于如FPGA等运算器中，由调度器进行调度，可以极大地降低计算量，减少设备的复杂度和功耗，能够适应便携设备的需求。
根据本公开实施例，该处理组件300可以实现为印制电路板(PCB)。该印制电路板例如可以包括固定部分310和可重新配置部分320。其中，固定部分310包括调度器311，用于根据调度策略将计算任务分配到运算器；可重新配置部分320包括多个运算器，例如运算器321、322、323，所述多个运算器中的部分或全部固化地设置有二值神经网络模型的任务运算单元。上述布局可以减少原始硬件电路的面积，减小处理组件的体积。可重新配置部分320例如可以包含数百个并行内核，计算资源的空间冗余允许安全性和非安全性关键应用程序共存，从而提供适当的分区机制，能够用于实现高可靠性，可安全认证和重认证的多核计算处理系统。
根据本公开实施例，该调度器例如可以由中央处理器CPU实现，该调度器仅用于任务调度而不参与任何运算，所有运算任务由运算器执行，可以降低系统功耗。
根据本公开实施例，除调度器311外，固定部分310还可以包括运行时环境312、操作系统313、输入输出管理314等功能模块，以保障处理组件实现必要的功能。例如，运行时环境RTE(Runtime Environment)是为了聚合不同的设备，构建了可安装的客户端驱动程序加载程序，并充当用户程序和实际实现之间的代理，这样，可以顺利调用不同供应商的OpenCL实现，而不会发生任何冲突。该些模块可以通过软件的方式实现，可以通过可编程硬件的方式来实现，也可以通过软硬结合的方式实现，本公开对此不做限定。
根据本公开实施例，该运算器可以为支持PCIe的加速器。运算器可以包括以下一种或多种：通用图形处理组件(GPGPU，General-purpose computing on graphics processing units)、现场可编程逻辑门阵列(FPGA，Field Programmable Gate Array)、基于现场可编程逻辑门阵列的集成芯片。
其中，FPGA是作为特殊应用电路领域中的一种半定制电路而出现的，既解决了全定制电路的不足，又克服了原有可编程逻辑器件门电路数有限的 缺点。其一个显著的特点是低功耗。GPGPU由于现代图形处理组件强大的并行处理能力和可编程流水线，令流处理组件可以处理非图形数据。特别在面对单指令流多数据流(SIMD)，且数据处理的运算量远大于数据调度和传输的需要时，GPGPU在性能上大大超越了传统的中央处理器。
根据本公开实施例，在上述FPGA的基础上，通过仿真技术验证件FPGA制成芯片的可行性，从而将FPGA制成体积更小的基于现场可编程逻辑门阵列的集成芯片，可以进一步降低处理器的尺寸、重量和功耗。
根据本公开实施例，运算器中的FPGA是基于已训练的神经网络模型定制的。定制FPGA的过程包括：构造基于神经网络的健康状态识别模型，使用心电数据作为训练数据训练并压缩所述健康状态识别模型，基于压缩后的健康状态模型定制可编程逻辑门阵列。其中，压缩的步骤可以包括剪枝或参数二值化，剪枝减少了神经元之间的关联，参数二值化使神经元方便地实现为硬件电路，二者极大地降低了神经网络的计算量，也降低了被部署的FPGA的复杂度，能够提高计算效率并降低功耗。
根据本公开实施例，所述二值神经网络的权值为-1或1，由此，神经元之间的权值乘法计算可以简化为位运算，通过补码即可实现乘以-1的运算，极大地提高了模型计算的效率。
根据本公开实施例，所述二值神经网络模型可以通过以下方法训练得到：
构造实数型神经网络模型，该实数型神经网络模型的隐藏层的权值和激活值为实数型数据；
使用训练数据训练并压缩该实数型神经网络模型，得到二值神经网络模型。
根据本公开实施例，可以首先在具有较高运算能力的平台上训练一个实数型神经网络，该实数型神经网络例如可以包括一个输入层，四个隐藏层和一个输出层，通过压缩该神经网络得到小型的二值神经网络模型，再基于压缩后的二值神经网络模型设计FPGA。
根据本公开实施例，所述使用训练数据训练并压缩所述实数型神经网络模型，得到二值神经网络模型，包括：
获取训练数据；
交替执行训练操作和压缩操作，直至满足收敛条件，得到二值神经网络 模型，其中，在所述训练操作中，通过所述训练数据更新模型中的权值；在所述压缩操作中，通过剪枝和/或二值化压缩模型的大小。
根据本公开实施例，获取训练数据包括：
获取不同健康状态的人的心电数据；
对所述心电数据进行预处理得到训练样本，所述预处理包括噪声处理、基线漂移处理、心拍分割、特征预提取和数据重采样中的至少一种；
对所述训练样本标注得到样本标签，所述样本标签表示所述心电数据对应的不同健康状态，将所述训练样本和样本标签确定为训练数据。
根据本公开实施例，可以先对多模态生理信号(包括心脏异常患者和正常人的多模态生理信号)进行去噪处理，如可以通过带阻滤波器去除信号中的工频干扰噪声、通过低通滤波器消除肌电干扰噪声、通过IIR零相移数字滤波器纠正基线漂移等，本发明实施例对去除噪声的方式不加以限制。
根据本公开实施例，对于其中的心电信号，在去噪处理后，可以对其分割心拍得到每个心电信号的多个心拍信号，对心拍信号进行数据重采样获得的采样数据作为训练样本，心拍信号所属的心电信号属于心电异常患者还是正常人作为训练样本的样本标签，训练样本和样本标签即构成训练数据。
根据本公开实施例，在获取训练数据之后，可以通过所述训练数据训练所述实数型神经网络模型，以更新实数型神经网络模型的实数型权值，然后对所述实数型神经网络模型进行剪枝，并将所述实数型权值二值化，得到二值神经网络模型，并通过所述训练数据训练所述二值神经网络模型直至收敛。以上通过一次压缩得到二值神经网络模型并训练至收敛，或者，也可以分多个轮次进行压缩，每次压缩可以对部分节点进行剪枝和/或二值化，压缩后继续进行训练，然后进入下一轮压缩，直至模型大小和预测准确度都满足预定的收敛条件。
根据本公开实施例，由于该FPGA是基于特定神经网络定制的，其相对于通用处理组件而言具有更小的体积、更低的功耗和更高的效率。根据本公开实施例的处理组件的综合指标可以达到如下水平：
表1
Figure PCTCN2020137983-appb-000001
根据本公开实施例，处理组件的电源接口例如可以采用两个TPS54386双3A非同步转换器，以确保稳定且充足的双电源。
根据本公开实施例，所述处理组件还可以包括供电管理单元，被配置为基于当前任务负载调整被供电的运算器的数量。例如，在当前任务负载不小于第一阈值时，可以启用全部的运算器；在当前任务负载小于第一阈值时，可以根据当前任务负载的情况确定启用预算器的数量。可以根据当前任务负载与第一阈值的比例确定启用运算器的数量，也可以划定区间，根据预定的对应关系确定启用运算器的数量。通过供电管理单元，本公开实施例的处理器可以有效地降低其功耗。
根据本公开实施例所描述的供电管理单元、预处理单元或特征预提取单元可以通过软件的方式实现，也可以通过可编程硬件的方式来实现。电源管理单元、预处理单元或特征预提取单元例如可以设置在调度器中，这些单元或模块的名称并不构成对该单元或模块本身的限定。
根据本公开实施例，该输出组件可以至少具有第一输出模式和第二输出模式，所述处理组件被配置为：
在所述健康状态为第一类状态的情况下，控制所述输出装置按照第一输出模式输出预警信息；
在所述健康状态为第二类状态的情况下，控制所述输出装置按照第二输出模式输出预警信息。
图4示出根据本公开实施例的主控设备400的示意图。
如图4所示，该主控设备400例如可以包括屏幕410、指示灯421、422、423、按键431、432等。当然，以上输入/输出单元的种类和数量只是示例性的，可以设置更多或更少的输入/输出单元，或者设置不同种类的输入/输出单元，也可以省略以上输入/输出单元中的一种或多种。
根据本公开实施例，输出装置可以具有不同的输出模式，用于提示不同的健康状态类型。例如，第一输出模式可以是指示灯421或屏幕410点亮，还可以通过屏幕410显示提示信息，第二输出模式可以指示灯422或屏幕410点亮，并且发出声音提示(设有声音播放装置的情况下)。在本公开另一些实施例中，第一输出模式可以是指示灯423或屏幕410闪烁，第二输出模式可以是指示灯423或屏幕410常亮。在本公开又一些实施例中，第一输出模式可以是本地提示，第二输出模式可以是将预警信息发送到预定设备，以提醒预定联系人。当然，在其他的实施例中，无论是否健康状态的类型如何，都可以将预警信息发送到预定设备，以提醒预定联系人。
根据本公开实施例，所述输出装置包括三色指示灯，例如，指示灯421、422、423可以分别为绿色指示灯、黄色指示灯和红色指示灯，或者，在同一个指示灯内可以集成三种不同颜色的灯珠。所述处理组件被配置为：
在所述健康状态为正常状态的情况下，控制绿色指示灯点亮；
在所述健康状态为异常状态的情况下，控制黄色指示灯点亮；
在所述健康状态为紧急状态的情况下，控制红色指示灯点亮。
一些相关算法处理心电信号得到患者的疾病名称，例如心律失常、心房颤动、心力衰竭、心源性猝死等。如果将这些内容直接用于健康监测设备的提示，缺乏相关知识的患者不能很好地采取应对措施。根据本公开实施例提供的技术方案，使用者无需掌握医学知识，即可了解监测的结果，从而采取合适的应对措施。例如，发现红色指示灯点亮的情况可立即呼叫救护车，发现黄色指示灯点亮的情况可自行去医院就诊，对于绿色指示灯电亮的情况则无需采取措施。
根据本公开实施例提供的技术方案，该健康监测设备能够及时对使用者的健康状态进行预警；通过小型、可以随身携带的处理组件对多模态生理信号进行本地分析可以很好的解决网络不稳定的问题。使用者通过佩戴可穿戴 设备检测多模态生理信号，信号通过本机植入的算法在本机实时预警，可以帮助居民、患者、健康及卫生管理人员实现突发性心血管疾病预警、解决大量健康管理问题。移动终端实现智能医疗能够全方位实时检测患者生理信号，捕捉到常规短时检测记录不到的阵发性心律异常，同时还能够避免收到医生技术水平、熟练程度、疲劳程度等影响导致的诊断失误。
以上描述仅为本公开的较佳实施例以及对所运用技术原理的说明。本领域技术人员应当理解，本公开中所涉及的发明范围，并不限于上述技术特征的特定组合而成的技术方案，同时也应涵盖在不脱离所述发明构思的情况下，由上述技术特征或其等同特征进行任意组合而形成的其它技术方案。例如上述特征与本公开中公开的(但不限于)具有类似功能的技术特征进行互相替换而形成的技术方案。

Claims (10)

  1. 一种健康监测设备，包括：
    输出组件；
    可穿戴组件，由导电材料和绝缘材料混合编织而成，所述导电材料在所述可穿戴组件上形成多个电极以及在所述多个电极之间形成多模态生理信号采集电路，所述多模态生理信号至少包括心电信号和生物阻抗信号；
    处理组件，包括调度器和多个运算器，其中，所述调度器被配置为接收所述多模态生理信号，调度所述多个运算器以确定使用者的健康状态信息，并基于所述健康状态信息控制所述输出组件输出预警信息，所述运算器包括可编程逻辑门阵列，所述运算器内固化地设置有二值神经网络模型的任务运算单元，所述二值神经网络模型的隐藏层的权值和激活值为二值数据，用于基于多模态生理信号确定健康状态信息；以及
    电池，用于向所述可穿戴组件、处理组件以及输出组件供电。
  2. 根据权利要求1所述的设备，其特征在于，所述调度器包括中央处理器，所述运算器还包括通用图形处理器和基于可编程逻辑门阵列定制的集成芯片，所述调度器和运算器部署于印制电路板。
  3. 根据权利要求1所述的设备，其特征在于，所述处理组件还包括：
    预处理单元，被配置为对所述多模态生理信号进行预处理，所述预处理包括噪声处理、基线漂移处理、心拍分割和数据重采样中的至少一种。
  4. 根据权利要求1所述的设备，其特征在于，还包括输入组件，所述处理组件还被配置为基于用户输入的身高体重信息，确定用户的体表面积。
  5. 根据权利要求4所述的设备，其特征在于，所述多模态生理信号采集电路还用于采集心率、射血前期、心室射血时间、平均动脉压、左心收缩时间、心收缩力指数、中心静脉压、肺动脉阻塞压中的部分或全部。
  6. 根据权利要求5所述的方法，其特征在于，处理组件还包括特征预提取单元，被配置为执行以下一种或多种：
    基于所述生物阻抗信号确定每搏输出量和/或每搏指数；
    基于所述射血前期和心室射血时间确定射血分数；
    基于所述每搏输出量和射血分数确定左室舒张末期容积；
    基于所述每搏输出量和心率确定心输出量；
    基于所述心输出量和体表面积确定心脏指数；
    基于所述心脏指数、平均动脉压和肺动脉阻塞压确定左心做功指数；
    基于所述心输出量、平均动脉压和中心静脉压确定外周血管阻力；
    基于所述心脏指数、平均动脉压和中心静脉压确定外周血管阻力指数。
  7. 根据权利要求6所述的设备，其特征在于，所述二值神经网络模型的输入包括心电信号、生物阻抗信号、左室排血功能指标、心机收缩功能指标、前负荷指标以及后负荷指标，其中：
    所述左室排血功能指标包括每搏输出量、每搏指数、心输出量、心脏指数中的至少一种；
    所述心肌收缩指标包括心收缩力指数、左心做功指数、左心收缩时间、射血分数中的至少一种；
    所述前负荷指标包括中心静脉压和/或左室舒张末期容积；
    所述后负荷指标包括外周血管阻力和/或外周血管阻力指数。
  8. 根据权利要求1所述的设备，其特征在于，所述二值神经网络模型通过以下方法训练得到：
    构造实数型神经网络模型，所述实数型神经网络模型的隐藏层的权值和激活值为实数型数据；
    交替执行训练操作和压缩操作，直至满足收敛条件，得到二值神经网络模型，其中，在所述训练操作中，通过所述训练数据更新模型中的权值；在所述压缩操作中，通过剪枝和/或二值化压缩模型的大小。
  9. 根据权利要求1～8任一项所述的设备，其中，所述输出组件具有第一输出模式和第二输出模式，所述处理组件被配置为：
    在所述健康状态为第一类状态的情况下，控制所述输出装置按照第一输 出模式输出预警信息；
    在所述健康状态为第二类状态的情况下，控制所述输出装置按照第二输出模式输出预警信息。
  10. 根据权利要求9所述的设备，其中，所述输出装置包括三色指示灯，所述处理组件被配置为：
    在所述健康状态为正常状态的情况下，控制绿色指示灯点亮；
    在所述健康状态为异常状态的情况下，控制黄色指示灯点亮；
    在所述健康状态为紧急状态的情况下，控制红色指示灯点亮。
PCT/CN2020/137983 2020-10-26 2020-12-21 健康监测设备 WO2022088442A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011158916.1 2020-10-26
CN202011158916.1A CN112263262B (zh) 2020-10-26 2020-10-26 健康监测设备

Publications (1)

Publication Number Publication Date
WO2022088442A1 true WO2022088442A1 (zh) 2022-05-05

Family

ID=74341197

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/137983 WO2022088442A1 (zh) 2020-10-26 2020-12-21 健康监测设备

Country Status (2)

Country Link
CN (1) CN112263262B (zh)
WO (1) WO2022088442A1 (zh)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113288103B (zh) * 2021-04-01 2024-06-18 安徽通灵仿生科技有限公司 一种无创心排量监测系统与方法
CN113567737B (zh) * 2021-06-28 2023-01-13 东北大学 电压数据的采样方法和肺通气监测装置
WO2023115558A1 (en) * 2021-12-24 2023-06-29 Mindamp Limited A system and a method of health monitoring
CN117524489B (zh) * 2024-01-08 2024-03-08 吉林大学 患者的健康监测管理系统及方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH11128187A (ja) * 1997-10-24 1999-05-18 Gunze Ltd 心電図用電極、心電図測定用着衣及び心電図測定システム
JP5305396B2 (ja) * 2009-07-09 2013-10-02 国立大学法人大阪大学 多電極付き織物
CN205359464U (zh) * 2016-01-29 2016-07-06 江苏思维森网络技术有限公司 婴幼儿心电及呼吸信号检测的纤维传感器
CN109846464A (zh) * 2018-12-29 2019-06-07 曹乃钊 血液流动力学参数的无创测量系统和数据处理系统
CN110200620A (zh) * 2019-06-06 2019-09-06 中山大学孙逸仙纪念医院 多模态心脏数据检测方法、装置、存储介质及终端设备
CN111743532A (zh) * 2020-06-20 2020-10-09 成都吱吖科技有限公司 一种独立穿戴式心电数据监测系统及其检测方法

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101357062B (zh) * 2008-08-22 2010-11-03 华南理工大学 基于容积脉搏波检测的生命体征参数测评装置
US10089577B2 (en) * 2016-08-05 2018-10-02 Xilinx, Inc. Binary neural networks on progammable integrated circuits
CN110123306A (zh) * 2019-04-30 2019-08-16 中山大学孙逸仙纪念医院 一种基于穿戴式设备的心肺数据采集检测系统
CN110974172A (zh) * 2019-12-13 2020-04-10 北京理工大学 一种实时生理参数的测量系统

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH11128187A (ja) * 1997-10-24 1999-05-18 Gunze Ltd 心電図用電極、心電図測定用着衣及び心電図測定システム
JP5305396B2 (ja) * 2009-07-09 2013-10-02 国立大学法人大阪大学 多電極付き織物
CN205359464U (zh) * 2016-01-29 2016-07-06 江苏思维森网络技术有限公司 婴幼儿心电及呼吸信号检测的纤维传感器
CN109846464A (zh) * 2018-12-29 2019-06-07 曹乃钊 血液流动力学参数的无创测量系统和数据处理系统
CN110200620A (zh) * 2019-06-06 2019-09-06 中山大学孙逸仙纪念医院 多模态心脏数据检测方法、装置、存储介质及终端设备
CN111743532A (zh) * 2020-06-20 2020-10-09 成都吱吖科技有限公司 一种独立穿戴式心电数据监测系统及其检测方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
NAFIUL RASHID ET AL.: "Energy-efficient Real-time Myocardial Infarction Detection on Wearable Devices", 2020 42ND ANNUAL INTERNATIONAL CONFERENCE OF THE IEEE ENGINEERING IN MEDICINE & BIOLOGY SOCIETY (EMBC, 24 July 2020 (2020-07-24), pages 4648 - 4651, XP033815293, DOI: 10.1109/EMBC44109.2020.9175232 *

Also Published As

Publication number Publication date
CN112263262A (zh) 2021-01-26
CN112263262B (zh) 2022-05-06

Similar Documents

Publication Publication Date Title
WO2022088442A1 (zh) 健康监测设备
CN105726011B (zh) 一种健康监控服装
WO2022088443A1 (zh) 边缘计算节点的制造方法和边缘计算节点
US20130281815A1 (en) Wearable remote electrophysiological monitoring system
CN108320814A (zh) 一种基于穿戴式智慧衣的老人家庭远程监护系统
US20210338190A1 (en) Multi-sensor device for monitoring health
CN205458634U (zh) 一种具有健康监护及反馈治疗功能的手机
CN106618532A (zh) 一种用于心电与血压及脉搏特征参数采集的穿戴装置
Sánchez-Tato et al. Health@ Home: A telecare system for patients with chronic heart failure
CN105748061A (zh) 一种无线柔性心电采集装置
CN205493820U (zh) 模块化智能内衣
Laudato et al. ATTICUS: Ambient-intelligent Tele-monitoring and Telemetry for Incepting and Catering over hUman Sustainability
CN112037916A (zh) 共享多功能防猝死生理信息检测系统及其方法
WO2022088441A1 (zh) 健康监测设备
KR20150138736A (ko) 심음, 또는 폐음을 이용한 사용자의 건강 상태 정보 제공 장치 및 방법
WO2022228191A1 (zh) 腕式可穿戴设备及腕式可穿戴设备的控制方法
CN111973164A (zh) 智能可穿戴监测设备、系统及血压估算模型的训练方法
CN207886204U (zh) 一种心电监测装置
Luprano et al. HeartCycle: advanced sensors for telehealth applications
CN109431477A (zh) 医用心肺监护系统和医用心肺监护方法
Chowdhury et al. Future techniques and perspectives on implanted and wearable heart failure detection devices
CN110025301A (zh) 智能心血管监测穿戴装置
Liu et al. Design of real-time ECG monitoring wearable equipment based on the Holter system
CN220833040U (zh) 智能心电胸带
CN218458103U (zh) 一种基于电阻抗成像的智能监护可穿戴系统

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20959599

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20959599

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 12.09.2023)