Methods and devices for intelligent respiratory monitoring using electrocardiogram, respiratory acoustics, and chest acceleration.

Abstract

Breathing is fundamental to life. Continuous monitoring of respiratory rate and pattern is crucial for detecting the onset of respiratory failure; however, this vital sign cannot be monitored in most hospital wards. Respiratory distress is caused by dysfunction of oxygenation (insufficient oxygen intake) or dysfunction of ventilation (insufficient carbon dioxide removal). This invention discloses a method and apparatus for monitoring respiratory rate by extracting information from chest acceleration, electrocardiogram, and breath sounds. The invention also discloses a method for quantifying the degree of respiratory deterioration by defining a smart respiratory index, which is defined by combining at least three parameters extracted from physiological measurements such as electrocardiogram, breath sounds, and chest acceleration. These methods are implemented by a small wireless patch attached to the upper part of the patient's chest and communicate with external proprietary software and a monitor via Bluetooth.

Description

Technical Field

[0001] This invention generally relates to apparatus and methods for determining the quality and force of a subject's breathing, as well as respiratory rate (RR). More specifically, this invention relates to a wireless apparatus and method for obtaining an index representing the probability of respiratory deterioration. Background Technology

[0002] Breathing is fundamental to life. The lungs are responsible for breathing, which is the process of supplying the body with oxygen and expelling carbon dioxide.

[0003] Respiratory distress is caused by impaired oxygenation (insufficient oxygen intake) or impaired ventilation (insufficient carbon dioxide removal).

[0004] These forms of dysfunction can manifest as irregular respiratory rate and / or abnormal breathing patterns. Continuous monitoring of RR and pattern is crucial for detecting the onset of respiratory failure; however, most hospital wards fail to properly monitor this vital sign. When RR is measured, it is done infrequently. Patients in general wards are typically checked only every six to eight hours, and there is often significant human error because nurses often observe the patient's breathing for 15 seconds and then multiply the result by 4 to obtain the RR per minute. The World Health Organization recommends calculating respiratory rates per minute. If the RR is above or below the acceptable range of 12-16 breaths per minute, it indicates a breathing problem.

[0005] Current devices used to measure RR are very expensive and require a cable connection to the patient, so they are only used in specialized units such as intensive care units (ICUs).

[0006] Patients with respiratory rate (RR) between 25 and 29 breaths per minute (RR) account for 21% of hospital deaths (Respiratory rate: the neglected vital sign. Michelle A Cretikos, Rinaldo Bellomo, KenHillman, Jack Chen, Simon Finfer and Arthas Flabouris. MJA 2008; 188: 657-659). The best way to remain vigilant about a patient's clinical status and reduce complications is through continuous monitoring of RR and tidal volume (TV). The coronavirus pandemic (Covid-19) has clearly demonstrated the necessity of respiratory monitoring in nursing homes and even home care settings outside of hospitals.

[0007] Joseph et al.'s U.S. patent application, "Acoustic sensor and ventilation monitoring system" (US2020 / 0054277 A1), discloses a method for monitoring respiration using an acoustic measurement device. After a broad description of respiratory physiology and pathology, the authors propose a device integrating two components. One component is attached to the patient, while the second component is attached to the first. The first component consists of a sound sensor, an accelerometer, and a transmitter. The second component consists of a rechargeable battery. Multiple sensors can be integrated into the device to measure temperature, heart rate, and oxygen saturation. The device can also connect to a smartwatch. Using information from different sensors, they propose a risk factor index to determine changes in respiratory function in mammals due to physiological changes and drug or alcohol abuse. They describe a quadratic equation to calculate their respiratory risk factor index. The main differences between this and our current patent application are, firstly, that Joseph et al. only used an accelerometer to assess body movement, while we use an accelerometer to estimate respiratory rate; and secondly, the formulas for the respiratory risk index and the smart respiratory index are significantly different. Furthermore, in the patent application by Joseph et al., respiratory rate was assessed by respiratory acoustics, while in this patent application, respiratory rate is assessed by chest acceleration measured by an accelerometer and the first derivative of the breath sound envelope.

[0008] Both Joseph's patent application and this patent application use a sensor as a microphone and an accelerometer to assess breathing sounds and movement. However, the parameters analyzed by Joseph et al. in those two works, as well as the derivation formula used to calculate the risk factor index, are completely different from the SRI of this disclosure. Joseph et al. used an accelerometer to assess the patient's body movement, while in this disclosure, the accelerometer is used to calculate respiratory rate.

[0009] Joseph et al. used breathing sounds to calculate respiratory rate and TV, while this disclosure uses envelopes to assess the variability of TV.

[0010] Finally, for Joseph et al., the combined scores for physical movement and speech were the highest, while slow breathing and lack of movement were the lowest. The US patent application "Portable device with multiple integrated sensors for vital signs scanning" (US2015 / 0313484 A1) discloses a portable device with multiple integrated sensors. The patient application of this invention differs; it does not use a thermometer or photoplethysmography (PPG), and this invention defines an index of respiratory quality.

[0011] The U.S. patent application “Mobile frontend system for comprehensive cardiac diagnosis” (US2015 / 0065814 A1) is significantly different from this patent application. The purpose of US2015 / 0065814 is to provide a comprehensive diagnosis of heart problems.

[0012] US patent application “Mesh network personal emergency appliance” (US2008 / 0001735 A1) discloses a system that includes one or more wireless nodes forming a wireless mesh network. The difference in this application is that it does not form a wireless mesh network.

[0013] The U.S. patent application “Monitoring, predicting and treating clinical episodes” (US2008 / 0275349 A1) discloses a device for sensing physiological parameters of a subject and capable of sensing a wide range of body movements, which is significantly different from this patent application, which does not disclose sensing a wide range of body movements.

[0014] The US patent application “Physiological acoustic monitoring system” (US8821415B2) discloses a method for assessing respiratory rate using acoustic signals; however, that application uses only one or two acoustic sensors (microphones), and is therefore completely different from the present invention, which also uses an electrocardiogram (ECG) and an accelerometer.

[0015] US Patent “METHODS AND SYSTEMS FOR MONITORING RESPIRATION” (US 6,918,878B2) discloses a method for determining a patient’s respiratory rate, which includes several components. Respiratory rate can be determined by measuring the S2 split of the heart. The S2 split can be identified by observing the timing of heart sounds. Other respiratory-related information, such as respiratory phasing and the occurrence of apnea, can also be identified. This type of respiratory monitor may be useful for monitoring subacute patients and outpatients. The sensor for the respiratory monitor and the electrode for the ECG monitor can be combined into a single probe.

[0016] This patent application does not include S2 splitting and is therefore different from patent application US 6,918,878.

[0017] US Patent “System and method for monitoring respiratory rate measurements” (US20180214090A1) discloses a system and method for determining multi-parameter confidence in respiratory rate measurements using multiple physiological parameter inputs. This disclosure differs from this patent application because this application combines respiratory acoustics with ECGR-R interval variations (Rpeaks). Summary of the Invention

[0018] In a preferred embodiment, patches (1, 13, 27, 42) are attached to the patient's chest, and the patches (1, 13, 27, 42) include at least a sensor for electrocardiogram (ECG), a sensor for measuring respiratory rate (RR) by an accelerometer, and a sensor for measuring RR by a microphone. The ECG (2, 14, 29) is further processed using a heart rate (HR) extraction algorithm, for example, but not limited to using Fast Fourier Transform (FFT) (6, 19, 34) to extract the HR of the ECG, and Rpeaks (7, 20, 35) are also determined by FFT (spectral analysis).

[0019] Breathing sounds (breathing acoustics) from microphones (3, 15, 30) are processed using envelope waveform breathing extraction formulas (8, 21, 36) to obtain the respiratory rate, termed RespR. Formulas (9, 22, 37) are used to calculate tidal volume variability (TVv). Chest acceleration signals (4, 16, 31) are analyzed using a Hilbert transform model to assess breathing, termed RespRacc (10, 23, 38). The relationship between the three parameters Rpeaks, RespR, and RespRacc may change as breathing worsens; therefore, cross mutual information (5, 18, 32) is calculated to generate the variable CMIbreath. The parameters extracted from the measurements are fed into a classifier, which may be, but is not limited to, an Adaptive Fuzzy Neural Network System (ANFIS) (11, 25, 40). The output of the classifier is termed the Smart Breathing Index (SRI) (12, 26, 41).

[0020] RespR is estimated using the envelope formula.

[0021] The acoustic signal recorded by the microphone is input into a spline function or other curve function to evaluate the amplitude envelope, such as... Figure 7 As shown. RespR can be calculated by calculating the peak value in the envelope curve. See also: Figure 6 Examples are shown in the text.

[0022] Estimation of tidal volume variability (TVv)

[0023] Previously, the relationship between respiratory airflow F and the energy E of respiratory (tracheal) sound could be expressed as A = kF α The power law best fits the value of the exponent, where k and α are constants, and different research teams have proposed different values ​​for this exponent. This amplitude-airflow relationship of sound has been used for respiratory monitoring, particularly for qualitative and quantitative assessment of respiratory airflow and for estimation of continuous respiratory rate.

[0024] However, we found that the estimation can be improved by adding the derivative of the breath sound envelope to the equation, and therefore the following flow equation is defined:

[0025] F = k1 max (dA envelope / dt) + k2A β

[0026] Therefore, the volume can be estimated as the integral of the airflow F over the inhalation time.

[0027]

[0028] In this paper, tidal volume variability is defined as the change over time; for example, if the tidal volume increases from 10 to 12, the tidal volume variability is 20%.

[0029] RespRacc is determined by applying the Hilbert transform to the acceleration signal.

[0030] A novel method for extracting respiration from acceleration signals using Hilbert vibrational decomposition (HVD) is proposed. The maximum energy component of the acquired acceleration signal is proportional to the respiration signal.

[0031] Determination of the Smart Respiratory Index (SRI)

[0032] In one embodiment, SRI is a linear or quadratic function of RespR, HR, HRv (heart rate variability), and TVv. The formula could be:

[0033] SRI =K1 * RespR+k2 * HR+k3 * HRv+k4 * TVv+K5 * RespR * HR * HRv *TVv,

[0034] A quadratic equation can contain more quadratic factors.

[0035] The constants k1 to k4 should be within the following range:

[0036] 0.30 <k1<0.7,

[0037] 0.2 <k2<0.4,

[0038] 0.05 <k3<0.2,

[0039] 0.01 <k4<0.1。

[0040] The parameters are combined using a classifier to define the SRI.

[0041] like Figure 1 As shown, in the second embodiment, the device uses a classifier, such as, but not limited to, an ANFIS model, to combine parameters to define the SRI definition. Parameters extracted from at least three sensors (ECG, respiratory sounds, chest acceleration) are used as input to the Adaptive Fuzzy Neural Network System (ANFIS). In the third embodiment, as... Figure 2 As shown, the device uses an ANFIS model to combine parameters to define the SRI. Parameters extracted from at least four sensors (ECG, respiratory sounds, chest acceleration, and pulse oximeter) are used as inputs to the Adaptive Fuzzy Neural Network System (ANFIS).

[0042] In the fourth embodiment, as Figure 3As shown, the device uses an ANFIS model to combine parameters to define the SRI. Parameters extracted from at least four sensors (ECG, breath sounds, chest acceleration, pulse oximeter) as well as patient demographic data (age, sex, height, weight) and clinical data (chronic obstructive pulmonary disease, asthma, sympathetic nervous system disorders, atrial fibrillation, beta-blockers, pacemaker) are used as inputs to the adaptive fuzzy neural network system (ANFIS).

[0043] ANFIS Overview

[0044] ANFIS is a hybrid of fuzzy logic systems and neural networks. It does not assume any mathematical function that controls the relationship between inputs and outputs. ANFIS employs a data-driven approach, where training data determines the system's behavior.

[0045] ANFIS's five layers have the following functions:

[0046] In layer 1, each cell stores three parameters to define a bell-shaped membership function. Each cell is connected to exactly one input cell and calculates the membership degree of the obtained input value.

[0047] In Layer 2, each rule is represented by a unit in Layer 2. Each unit is connected to units in the previous layer that are related to the premises of that rule. The input to a unit is its membership degree, which is multiplied to determine the degree to which the rule it represents is implemented.

[0048] In layer 3, for each rule, there is a unit that calculates its relative realization degree using a normalization equation. Each unit is connected to all rule units in layer 2.

[0049] In layer 4, each cell is connected to all input cells, and each cell is connected to exactly one cell in layer 3. Each cell computes the output of the rule.

[0050] In layer 5, the output unit calculates the final output by summing all the outputs of layer 4.

[0051] ANFIS applies the standard learning process of neural network theory. Backpropagation is used to learn the premise parameters, i.e., membership functions, while least squares estimation is used to determine the coefficients of the linear combination in the rule conclusions. The learning process involves two passes. In the first pass, the forward pass, the input pattern is passed through, and the optimal conclusion parameters are estimated through an iterative least mean squares process, while the premise parameters remain unchanged in the current loop of the training set. In the second pass, the backpropagation, the pattern is passed through again, and in this pass, the backpropagation is used to modify the premise parameters, while the conclusion parameters remain unchanged. This process is then iterated according to the required number of training epochs. If the premise parameters are initially chosen appropriately based on expert knowledge, usually one training epoch is sufficient, because the LMS algorithm can determine the optimal conclusion parameters in one pass, and if the premise parameters do not change significantly under gradient descent, then LMS will not produce any additional results in calculating the conclusion parameters. For example, in a system with two inputs and two rules, rule 1 is defined as:

[0052] If x is A and y is B, then f1 = p1x + q1y + r1

[0053] Where p, q, and r are all linear, they are called conclusion parameters or unique conclusions. The most common is first-order f, because higher-order Sugeno fuzzy models introduce significant complexity without offering any obvious advantages.

[0054] Number of categories

[0055] The input to the ANFIS system is fuzzified into multiple predefined categories. The number of categories should be greater than or equal to two. The number of categories can be determined using different methods. In traditional fuzzy logic, categories are defined by experts. This method can only be applied when experts clearly understand where the landmarks between two categories can be placed. ANFIS optimizes the placement of the landmarks; however, if the initial values ​​of the parameters defining the categories are close to the optimal values, gradient descent will reach its minimum more quickly. By default, the initial landmarks of ANFIS are chosen by dividing the interval from the minimum to the maximum value of all data into n equidistant intervals, where n is the number of categories. The number of categories can also be chosen using various clustering methods or Markov models, by visually determining a sufficient number of categories by plotting the data as a histogram, or by sorting using fuzzy inductive reasoning (FIR). This invention chooses the default value of ANFIS and, during the validation phase, found that more than three categories lead to instability; therefore, two or three categories are used.

[0056] Number of inputs

[0057] The number of categories and the number of inputs both increase the complexity of the model, i.e., the number of parameters. For example, a system with 4 inputs, each fuzzified into 3 categories, consists of 36 premise (non-linear) parameters and 405 conclusion (linear) parameters, calculated using the following two formulas:

[0058] Premise = Number of categories × Number of inputs × 3

[0059] Conclusion = Number of categories 输入的数量 × (Number of inputs + 1)

[0060] To obtain a meaningful solution for the parameters, the number of input-output pairs should typically be much larger than the number of parameters (at least 10 times).

[0061] Stability Standards

[0062] Unfortunately, there is currently no definition of a stability standard for neural fuzzy systems. The most useful tool for ensuring stability is experience gained through testing on specific datasets using certain neural fuzzy systems, such as ANFIS, and using extreme data obtained, for example, through simulation.

[0063] Number of training cycles

[0064] ANFIS uses the root mean square error (RMSE) to validate training results, and the RMSE validation error can be calculated after each training epoch based on a set of validation data. An epoch is defined as one update of the premise and conclusion parameters. Increasing the number of epochs generally reduces the training error. Attached Figure Description

[0065] Figure 1 The extracted parameters are fed into ANFIS, which is a hybrid of neural network and fuzzy logic system. The input includes at least the following three parameters: HR (6), HRv (7), RespR (8), TVv (9), RespRacc (10), and cross-information (CMIbreath) (5) between HR, HRv, and RespRacc. The output of the ANFIS model (11) is the Smart Breath Index (SRI) (12), which is a unitless number from 0 to 100, where a decreasing value corresponds to the degree of deterioration of the patient's respiratory function.

[0066] Figure 2The extracted parameters are fed into ANFIS, a hybrid of neural networks and fuzzy logic systems. The input includes at least the following three parameters: HR (19), HRv (20), RespR (21), TVv (22), RespRacc (23), pulse oximeter (17), and cross-information (CMIbreath) between HR, HRv, and RespRacc (18). The output of the ANFIS (25) model is the Smart Breath Index (SRI) (26), which is a unitless number from 0 to 100, where decreasing values ​​correspond to the degree of deterioration in the patient's respiratory function.

[0067] Figure 3 The extracted parameters are fed into an ANFIS, which is a hybrid of a neural network and a fuzzy logic system. The input includes at least the following three parameters: HR (34), HRv (35), RespR (36), TVv (37), RespRacc (38), pulse oxygen saturation (39), cross-interaction information between HR, HRv and RespRacc (CMIbreath) (32), and demographic data such as sex, age and body mass index (BMI). The output of the ANFIS model (40) is the Smart Breath Index (SRI) (41), which is a unitless number from 0 to 100, where a decreasing value corresponds to the degree of deterioration of the patient's respiratory function.

[0068] Figure 4 This diagram shows how the breathing patch (42) is attached to a person's chest and shows the location of the patch.

[0069] Figure 5 The patch consists of an amplifier (43) for ECG (51), an accelerometer (52), a microphone (53), a radio transmitter module (e.g., a Bluetooth Low Energy module (46)), a battery (45), and four electrodes (47-50) that are also used to attach the patch to the patient.

[0070] Figure 6 This figure illustrates the digital processing of the acquired signals, such as ECG (54), breath sounds (55), and chest movements (56). The figure also shows the parameters obtained from each signal: HR and HRv from ECG (54), RespR and TVv from microphone (55), and RespRacc from accelerometer (56).

[0071] Figure 7 This figure shows a schematic diagram of the cyclic breathing sound of inhalation and exhalation, and the amplitude envelope is also drawn.

[0072] Figure 8This table shows the relationship between clinical status and the Smart Breathing Index (SRI). The SRI is a progressive scale, where 100 corresponds to normal respiratory function, while decreasing values ​​reflect deterioration of respiratory function, and 0 is used when breathing stops.

[0073] Figure 9 This diagram illustrates one of the graphical user interfaces (GUIs) of the display, where SRI is the most important parameter, and therefore the largest. The GUI also displays RR and HR.

Claims

1. A method for determining the respiratory status of a subject, the method comprising the following steps: a. Measure an electrocardiogram (ECG); b. Measure chest movement and acceleration; c. Measuring respiratory acoustics; d. Calculate heart rate (HR) and heart rate variability (HRv) from electrocardiogram using fast Fourier transform or Choi-Williams distribution; e. Calculate tidal volume variability (TVv) based on the amplitude and first derivative of the amplitude envelope in respiratory acoustics; f. Calculate respiratory rate RespR based on chest movement and acceleration; g. Calculate the cross-information between heart rate variability (HRv) on electrocardiogram, respiratory acoustics, chest movement, and tidal volume variability (TVv) as inputs for the Smart Respiratory Index (SRI); h. Using an adaptive fuzzy neural network system or any other classifier, combine at least three parameters extracted from electrocardiogram, respiratory acoustics, chest movement, and their cross-information to form a respiratory quality level; SRI is given by the following formula: SRI = k1*RespR+k 2* HR+k3*HRv+k4*TVv+k5*RespR*HR*HRv*TVv, in, The ranges for k1 (between 0.3 and 0.7), k2 (between 0.2 and 0.4), k3 (between 0.05 and 0.2), and k4 (between 0.01 and 0.1). SRI of 99 corresponds to normal breathing, SRI between 98 and 75 corresponds to a slight deterioration in breathing, SRI between 74 and 50 corresponds to a severe deterioration, and SRI less than 49 corresponds to a life-threatening condition.

2. The method of claim 1, wherein step a is characterized by measuring via a patch consisting of two or more electrodes to determine an electrocardiogram located in the upper or lower chest of the subject.

3. The method of claim 1, wherein step b is characterized by recording chest acceleration using an accelerometer integrated in the patch, thereby calculating respiratory rate based on the acceleration.

4. The method of claim 1, wherein step c is characterized by recording respiratory acoustics, inhalation and exhalation via a microphone integrated in the patch, thereby calculating the respiratory rate.

5. The method according to claim 1, wherein step e is characterized in that the flow rate is estimated using the following formula: F=k1 max (dA 包络 / dt) + k2A β , Where F is the flow rate and A is the amplitude of the respiratory acoustics, therefore tidal volume is the integral of the flow rate over time. ， Tidal volume variability is defined as the change in tidal volume over time.

6. The method of claim 1, wherein step f is characterized by extracting respiration from acceleration by applying a Hilbert transform to the acceleration signal to calculate the respiration frequency, wherein the maximum energy component of the acquired acceleration signal is proportional to the respiration frequency.

7. The method of claim 1, wherein step g is characterized by calculating cross-information between heart rate variability extracted from electrocardiogram, features extracted from respiratory acoustics, and features extracted from acceleration.

8. The method of claim 1, wherein step h is characterized by heart rate variability extracted from electrocardiogram and tidal volume variability calculated from respiratory acoustics, and cross-mutual information used as input to an adaptive fuzzy neural network system or other classifier, and its output being an indicator of respiratory quality.

9. The method according to any one of claims 1 to 8, wherein the method is integrated in a wireless patch comprising interconnected electrocardiogram sensors and amplifiers, a microphone, an accelerometer, a battery, and a radio transmitter including a Bluetooth Low Energy module for transmitting data to an external observer device.

10. An apparatus for determining the respiratory status of a subject, the apparatus comprising: a. A sensor used to measure an electrocardiogram (ECG); b. Sensors used to measure chest movement and acceleration; c. Sensors used to measure respiratory acoustics; d. A microprocessor configured to: i. Calculate heart rate variability (HR) and heart rate variability (HRv) from electrocardiogram using fast Fourier transform or Choi-Williams distribution; ii. Calculate tidal volume variability (TVv) based on the amplitude and first derivative of the amplitude envelope in respiratory acoustics; iii. Calculate respiratory rate (RespR) based on chest movement and acceleration; iv. Calculate the cross-information between heart rate variability (HRv) on electrocardiogram, respiratory acoustics, chest movement, and tidal volume variability (TVv) as inputs for the Smart Respiratory Index (SRI); v. Using an adaptive fuzzy neural network system or any other classifier, combine at least three parameters extracted from electrocardiogram, respiratory acoustics, chest movement, and their cross-information to form an index of respiratory quality level; SRI is given by the following formula: SRI = k1*RespR+k 2* HR+k3*HRv+k4*TVv+k5*RespR*HR*HRv*TVv, in, The ranges for k1 (between 0.3 and 0.7), k2 (between 0.2 and 0.4), k3 (between 0.05 and 0.2), and k4 (between 0.01 and 0.1). SRI of 99 corresponds to normal breathing, SRI between 98 and 75 corresponds to a slight deterioration in breathing, SRI between 74 and 50 corresponds to a severe deterioration, and SRI less than 49 corresponds to a life-threatening condition.

11. The apparatus of claim 10, wherein calculating the tidal volume variability TVv further comprises estimating the flow rate using the following formula: F= k1 max (dA 包络 / dt) + k2A β , Where F is the flow rate and A is the amplitude of the respiratory acoustics, therefore tidal volume is the integral of the flow rate over time. ， Tidal volume variability is defined as the change in tidal volume over time.

12. The apparatus of claim 10, wherein calculating the respiratory frequency RespR further comprises extracting the breath from the acceleration by applying a Hilbert transform to the acceleration signal to calculate the respiratory frequency, wherein the maximum energy component of the acquired acceleration signal is proportional to the respiratory frequency.

13. The apparatus of claim 10, wherein calculating cross-mutual information further comprises calculating cross-mutual information between heart rate variability extracted from electrocardiogram, features extracted from respiratory acoustics, and features extracted from acceleration.

14. The apparatus of claim 10, wherein the at least three parameters extracted in combination further include heart rate variability extracted from an electrocardiogram and tidal volume variability calculated from respiratory acoustics, and cross-mutual information used as input to an adaptive fuzzy neural network system or other classifier, and the output of which is an indicator of respiratory quality.

15. The device according to any one of claims 10 to 14, wherein the device is integrated in a wireless patch comprising interconnected electrocardiogram sensors and amplifiers, a microphone, an accelerometer, a battery, and a radio transmitter including a Bluetooth Low Energy module for transmitting data to an external observer device.

Images