Estimation of Cardiogenic Artifacts in Ventilator Airway Pressure and Flow for Automatic Detection and Resolution of Patient-Ventilator Asynchrony

Abstract

A method for identifying the occurrence of patient-ventilator asynchrony (PVA) includes using a cardiogenic index generated from multiple respiratory attribute signals measured by the ventilator. The method can be implemented in a ventilator by a controller including a machine learning model trained to use the cardiogenic index and other features extracted from the ventilator waveform to identify the occurrence of PVA. The ventilator with the controller can alert a caregiver in real time that PVA has occurred so that ventilator settings can be adjusted.

Description

[Technical Field]

[0001] This patent application claims priority under 35 U.S.C. § 119 to U.S. Provisional Patent Application No. 63 / 074,550, filed September 4, 2020, the contents of which are incorporated herein by reference. [Background technology]

[0002] The disclosed concepts relate to methods for detecting patient-ventilator asynchrony (PVA), and more particularly, to methods for detecting non-airway factors indicative of PVA in order to identify and resolve PVA. Summary of the Invention [Problem to be solved by the invention]

[0003] Each year, approximately 800,000 patients receive ventilatory support in intensive care units (ICUs) across the United States. Mechanical ventilation is typically initiated when a patient is unable to maintain adequate ventilation or oxygenation, and therefore gas exchange, on their own. Ventilators are designed to provide critical, life-saving support to patients, but learning how to operate and adjust ventilator settings to accommodate complex patient conditions is a challenging task for even the most experienced intensivists and respiratory therapists. A key factor in identifying optimal ventilator settings is avoiding patient-ventilator asynchrony (PVA). PVA occurs when the ventilator and patient operate out of phase or in opposition to each other. PVA is estimated to be present in as many as 88% of mechanically ventilated patients. The occurrence of PVA can impact the outcomes and quality of life of mechanically ventilated patients. Some of the potential risks associated with PVA reported in the scientific literature are, among others, a three-fold increase in duration of ventilation, an eight-fold increase in tracheotomy rate, and a two-fold increase in ICU length of stay.

[0004] Given the significant increase in risk outcomes that PVA can cause, rapid detection of PVA is important. However, the current primary method for detecting PVA involves caregivers manually analyzing ventilator waveforms, which is not suitable for real-time detection of PVA. Therefore, there is room for improvement in methods and systems for detecting PVA. [Means for solving the problem]

[0005] It is therefore an object of the present invention, in one embodiment, to provide a method for identifying the occurrence of patient-ventilator asynchrony (PVA) by measuring a plurality of respiratory attribute signals of a patient receiving respiratory assistance from a ventilator, generating a cardiogenic index from the plurality of respiratory attribute signals, and analyzing the cardiogenic index to determine whether PVA has occurred.

[0006] In another embodiment, a ventilator includes a controller, an inflow path configured to supply air from the ventilator to a patient's airway, an outflow path configured to receive exhaled air from the patient's airway, and a plurality of sensors configured to measure data related to a respiratory attribute of the patient, wherein the controller is configured to receive data measured by the plurality of sensors, generate a cardiogenic index based on the data measured by the plurality of sensors, and analyze the cardiogenic index to determine whether PVA has occurred.

[0007] These and other objects, features, and characteristics of the present disclosure, as well as the method of operation and function of the associated elements of construction and combination of parts and economies of manufacture, will become more apparent from a consideration of the following description and appended claims, taken in conjunction with the accompanying drawings, all of which form part of this specification. Like reference numerals indicate corresponding parts in the various drawings. It is to be expressly understood, however, that the drawings are for the purpose of illustration and description only and are not intended as a definition of the limits of the invention. [Brief explanation of the drawings]

[0008] [Figure 1] Figure 1 shows a graph of waveforms to indicate the occurrence of a malfunctioning PVA. [Figure 2] FIG. 2 shows a graph of waveforms to illustrate the occurrence of a mis-triggered PVA. [Figure 3] FIG. 3 shows a graph including both the ventilator airway pressure waveform and a photoplethysmogram (PPG) signal to illustrate the relationship between the airway pressure waveform and the PPG signal. [Figure 4] FIG. 4 is a flowchart including steps of a method for calculating a cardiogenic index from a ventilator waveform, according to an exemplary embodiment of the disclosed concept. [Figure 5A] FIG. 5A shows graphs of waveforms corresponding to the steps of the method shown in the flowchart of FIG. [Figure 5B] FIG. 5B shows graphs of waveforms corresponding to the steps of the method shown in the flowchart of FIG. [Figure 5C] FIG. 5C shows graphs of waveforms corresponding to the steps of the method shown in the flowchart of FIG. [Figure 5D] FIG. 5D shows graphs of waveforms corresponding to the steps of the method shown in the flowchart of FIG. [Figure 6] FIG. 6 is a flowchart including steps of a method for calculating a cardiogenic index from a ventilator PPG waveform, according to an exemplary embodiment of the disclosed concepts. [Figure 7A] FIG. 7A shows graphs of waveforms corresponding to the steps of the method shown in the flowchart of FIG. [Figure 7B] FIG. 7B shows graphs of waveforms corresponding to the steps of the method shown in the flowchart of FIG. [Figure 7C] FIG. 7C shows graphs of waveforms corresponding to the steps of the method shown in the flowchart of FIG. [Figure 8]FIG. 8 is a flowchart of a process for dynamically assessing and adjusting ventilator settings while the ventilator is in use using any of the techniques shown in the flowcharts of FIGS. 4 and 6. [Figure 9] FIG. 9 is a schematic diagram of a ventilator including a machine learning model for detecting PVA, in accordance with an exemplary embodiment of the disclosed concepts. [Figure 10] FIG. 10 is a flowchart of a method for training a machine learning model to detect PVA in ventilator waveforms, according to an exemplary embodiment of the disclosed concepts. DETAILED DESCRIPTION OF THE INVENTION

[0009] In the specification, unless the context clearly dictates otherwise, the absence of a plural statement includes the plural.

[0010] In the specification, when two or more parts or components are said to be "coupled," it means that those parts are joined or operate together, either directly or indirectly, i.e., through one or more intermediate parts or components, so long as they are in communication.

[0011] In the specification, "number" means an integer of one or more (i.e., plural).

[0012] As used herein, a "controller" refers to any of a number of programmable analog and / or digital devices (including associated memory portions and / or storage units) capable of storing, retrieving, executing, and processing data (e.g., software routines and / or information used by such routines), including, but not limited to, a field programmable gate array (FPGA), a complex programmable logic device (CPLD), a programmable system on a chip (PSOC), an application specific integrated circuit (ASIC), a microprocessor, a microcontroller, a programmable logic controller, or any other suitable processing device or equipment. The memory units may be any one or more of various types of internal and / or external storage media providing storage registers, i.e., non-transitory machine-readable media, for storage of data and program code, such as, for example, a computer's internal storage area, including, but not limited to, RAM, ROM, EPROM, EEPROM, and FLASH, which may be volatile or non-volatile memory.

[0013] As used herein, "machine learning model" means a software system that develops and builds mathematical models based on sample data, known as "training data," to make predictions or decisions without being explicitly programmed to do so, including, but not limited to, a software system that is trained to recognize patterns from a set of training data and then develops algorithms to recognize patterns from the training data set in other data sets.

[0014] As used herein, "cardiogenic artifact" refers to distortion of airway pressure and flow waveforms due to mechanical activity from the heart and its pulsatile blood flow.

[0015] As used herein, "cardiogenic index" refers to a measure, eg, a signal, that indicates the amount of cardiogenic artifact present in both the ventilator pressure and flow waveforms over a period of time and over the collection of samples.

[0016] Directional expressions in the specification, such as, but not limited to, top, bottom, left, right, above, below, front, back, and derivatives thereof, relate to the orientation of the elements as illustrated and do not limit the claims, unless expressly stated otherwise.

[0017] The disclosed concepts, as described in more detail herein in connection with various specific exemplary embodiments, provide methods and systems for the automatic detection and resolution of PVA in a patient receiving mechanical ventilation. Depending on the needs of a particular patient, the ventilator is configured in either an assist mode or a mandatory mode. In assist mode, the ventilator is configured to deliver breaths to the patient's airway upon detecting a perceived inspiratory effort from the patient. One method of detecting inspiratory effort is to establish a baseline esophageal pressure for the patient and monitor the patient's ongoing esophageal pressure so that any increase from the baseline esophageal pressure is classified as an inspiratory effort. In mandatory mode, the ventilator is configured to deliver breaths at predetermined regular intervals.

[0018] The disclosed concepts address the detection and resolution of two specific types of PVA that can occur when a ventilator is set to assist mode: auto-triggering and ineffective triggering. Auto-triggering occurs when the ventilator delivers a breath regardless of the patient's inspiratory effort. Ineffective triggering occurs when the ventilator does not deliver a breath despite the patient's inspiratory effort.

[0019] FIG. 1 is a graph showing a patient's airway pressure waveform 1 along with a simultaneous patient's esophageal pressure waveform 2 to demonstrate the occurrence of a malfunctioning PVA. The plateau region 3 in the esophageal pressure waveform 2 indicates the patient's inspiratory effort. The airway pressure waveform 1 shows the time-varying patient airway pressure resulting from the ventilator delivering a breath to the patient. When the ventilator is operating as intended, the increasing slope 4 in the ventilator waveform 1 should coincide with the end of the plateau region 3, indicating that the ventilator is delivering a breath to the patient as a result of detecting the end of the patient's inspiratory effort. Highlighted around the 10-second mark on the x-axis is a malfunctioning PVA peak 6 in the ventilator waveform 1. Peak 6 is classified as a malfunctioning PVA because the ventilator delivered a breath to the patient while the patient was still in the middle of an inspiratory effort, rather than at the end of that effort. Plateau region 3' has more obvious pressure fluctuations than other plateau regions 3, which is ostensibly caused by a malfunctioning PVA, indicating that the malfunctioning PVA may have been caused by an overly sensitive ventilator setting.

[0020] FIG. 2 is a graph showing a patient's airway pressure waveform 11 along with a concurrent patient's esophageal pressure waveform 12 to demonstrate the occurrence of mis-triggered PVA. The dome-shaped region 13 of the esophageal pressure waveform 12 has a demonstrable increase from the patient's baseline esophageal pressure 14, indicating the patient's inspiratory effort. The airway pressure waveform 11 shows the time-varying patient airway pressure resulting from the ventilator delivering a breath to the patient. When the ventilator is operating as intended, the onset of the initial increasing slope of the sharkfin waveform 15 in the ventilator waveform 11 coincides with the onset of the initial increasing slope of the dome-shaped region 13, indicating that the ventilator is delivering a breath to the patient as a result of detecting the onset of the patient's inspiratory effort. Highlighting the x-axis around the 13-second mark indicates the presence of a period of mis-triggered PVA 17 in the airway pressure waveform 11, coinciding with the dome-shaped region 16. Section 17 is classified as a mis-triggered PVA because, even though dome-shaped region 16 indicates an increase from the patient's baseline esophageal pressure 14 around the 13-second mark to indicate the onset of an inspiratory effort, there is no significant increase in the airway pressure waveform 17 during section 17, indicating that the ventilator did not deliver the recommended breath to the patient. Dome-shaped region 16 has a significantly smaller amplitude than the other dome-shaped regions 13, ostensibly due to a mis-triggered PVA, indicating that the mis-triggered PVA may have been caused by overly sensitive ventilator settings.

[0021] Anecdotal evidence suggests that (1) cardiogenic oscillations present in the ventilator airway pressure waveform often trigger ventilator breath delivery, and that these oscillations account for most malfunctions and asynchronies, and (2) the presence of a significant cardiogenic component in the ventilator airway pressure waveform data often results in false mis-trigger detection or makes mis-trigger identification extremely difficult. Figure 3 is a graph showing a ventilator waveform 21 along with a simultaneous photoplethysmogram (PPG) signal 22 generated by a pulse oximeter to illustrate the relationship between the cardiogenic signal (in the form of a PPG signal) and the ventilator airway pressure signal. The PPG signal is not available at the ventilator, hence why the PPG signal 22 is provided by the pulse oximeter. The ventilator waveform 21 shows the time-varying patient airway pressure resulting from the ventilator delivering a breath to the patient. Displaying the PPG signal 22 simultaneously with the ventilator waveform 21 highlights the significant decrease in the ventilator waveform 21 synchronized with the peak in the PPG signal 22, suggesting an association between cardiogenic artifact and changes in ventilator-based airway pressure.

[0022] Engineers typically consider cardiogenic artifacts to be noise and attempt to remove them before any processing of the airway pressure waveform. This disclosure assumes that because their presence results in a significant number of asynchronous events, cardiogenic artifacts should be treated as signals of interest rather than noise. In addition, removing cardiogenic artifacts from ventilator waveforms is challenging because (1) there is significant overlap between ventilator and cardiogenic signals of interest, and (2) the body conditions associated with ventilator signals and cardiogenic interactions are highly non-stationary, so any strategy for detecting the overlap between ventilator and cardiogenic signals of interest requires frequent updates. Therefore, it is an objective of this disclosure to provide an algorithm for calculating the cardiogenic components present in ventilator waveforms as separate components of these waveforms and quantifying the separate components with a cardiogenic index so that PVA can be detected and resolved using the cardiogenic index. Computing cardiogenic artifacts as separate components of the ventilator waveform represents an improvement over prior art systems and methods for identifying PVA.

[0023] 4 is a flow chart of a method 100 for calculating a cardiogenic index from cardiogenic artifacts present in ventilator airway pressure and flow waveforms. Waveforms corresponding to the steps of method 100 are shown in FIGS. 5A-5D. At 101, the data points comprising the ventilator pressure waveform are calculated using the following equation (1):

number

number

[0024] At 102, the data points that make up the ventilator flow waveform are calculated using the following equation (2):

number

number

[0025] The normalized pressure waveform 121 and the normalized flow waveform 122 are plotted on the x-coordinate x i The data point is represented by the x-coordinate x on the normalized flow waveform 122. i102 represents the patient's airway data for the same time period, from the same sample number (or same point in time relative to the original ventilator waveform) as the data points in 101, and vice versa. It will be understood that the data points making up the ventilator flow waveform will be normalized to 101 rather than 102, and that the data points making up the ventilator pressure waveform will be normalized to 102 rather than 101, without departing from the disclosed concepts.

[0026] In 103, the standardized waveform z Py(t) and z Q(t) are summed to produce the normalized cumulative waveform z(t). Py(t) and z Q(t) Similarly, the x-axis of the normalized cumulative waveform is in sample number, and the y-axis is unitless. Py(t) and z Q(t) The sum of z(t) emphasizes the physiological signals synchronized between the pressure and flow waveforms, assuming that the cardiogenic signals are synchronized between the pressure and flow waveforms, but other noise sources are not. Thus, the cumulative waveform z(t) is the sum of the pressure waveform z(t) and the physiological signals synchronized between the pressure and flow waveforms. Py(t) or flow waveform z Q(t) This should result in a stronger cardiogenic component than either of these alone.

[0027] At 104, the power spectrum P(f) of the cumulative waveform z(t) is calculated by performing a fast Fourier transform of z(t), with frequency measured in Hertz. Referring to FIG. 5C, an exemplary power spectrum 123 of a normalized cumulative waveform is shown, generated by summing a normalized pressure waveform 121 and a normalized flow waveform 122. Empirical results from retrospective clinical datasets in which ventilator pressure and flow settings were fine-tuned based on concurrent PPG signals suggest that the power of cardiogenic signals is the most prevalent component of the cumulative ventilator pressure-flow waveform in the 1.2 Hz to 20 Hz frequency range. The 1.2 Hz to 20 Hz frequency region is indicated by line 124 in FIG. 5C. Datasets collected from a larger number of patients and / or using different models of ventilators show cardiogenic signals dominating in frequency regions other than the 1.2 Hz to 20 Hz region. If the cardiogenic signal is actually found to dominate in a frequency range other than the 1.2 Hz to 20 Hz range for a given set of conditions, then the lower and upper frequency boundaries of the frequency range in which the cardiogenic signal actually dominates should be replaced with frequency boundaries of 1.2 Hz and 20 Hz, respectively, in steps 105 and 106 (as described in further detail herein), and it is understood that doing so is within the scope of the disclosed concept.

[0028] In 105, the normalized power distribution value of P(f) in the frequency range from 1.2 Hz to 20 Hz is calculated by the following equation (3):

number

number

[0029] At 106, a cardiogenic index in the form of a standardized cardiogenic signal is reconstructed by performing an inverse fast Fourier transform only on the normalized distribution values ​​of P(f) in the 1.2 Hz to 20 Hz region calculated at 105. Referring to FIG. 5D, a graph of an exemplary standardized cardiogenic signal 125 reconstructed from the normalized distribution values ​​of the power spectrum 123 in the 1.2 Hz to 20 Hz region is shown, plotted on the same scale x-axis as the standardized pressure waveform 121 and flow waveform 122. It will be appreciated that normalizing the distribution values ​​of P(f) in the 1.2 Hz to 20 Hz region at 105 facilitates the use of the standardized cardiogenic signal as a cardiogenic index, since the values ​​of the standardized cardiogenic signal are always kept between absolute values ​​of 0 and 1. The standardized cardiogenic signal 125 is represented by an x-coordinate x on the standardized cardiogenic signal waveform 125. iThe data point is the x-coordinate x on the normalized pressure waveform 121. i and the x-coordinate x on the standardized flow waveform 122 i 12 shows the extracted cardiogenic components of the combined normalized pressure waveform 121 and flow waveform 122, such that the extracted cardiogenic components correspond to the same time points or same sample numbers as the data points in FIG. 12 and vice versa.

[0030] To the extent that a ventilator capable of generating a PPG signal is or becomes available, the power of the cardiogenic signal can be estimated directly from the PPG signal provided by the ventilator (i.e., without performing steps 101 through 105 of method 100 to extract cardiogenic signal data from the ventilator's airway pressure and flow waveforms) without departing from the scope of the disclosed concepts. Referring to FIG. 6, in a further non-limiting exemplary embodiment of the disclosed concepts, if the ventilator is capable of generating a PPG signal, the cardiogenic index can be calculated using method 150 instead of method 100. FIG. 6 is a flowchart of method 150, and FIGS. 7A through 7C show graphs of waveforms associated with the steps of method 150.

[0031] At 151, individual heart beats in the PPG signal are identified using zero-crossing, thresholding, or any other suitable technique for detecting individual data in a continuous signal. Such techniques often involve the use of a signal mask, e.g., a zero-crossing mask. For brevity, step 151 is discussed below with respect to the application of a mask; however, it is understood that the use of the term "mask" is not intended to be limiting and does not preclude the use of techniques for detecting individual heart beats in a PPG signal that do not involve a signal mask. FIG. 7A is a graph of a theoretical ventilator-delivered PPG signal 171 and a zero-crossing PPG signal 172 resulting from filtering the PPG signal 171 with a zero-crossing mask. Each peak in the zero-crossing PPG signal 172 identifies the start of an individual heart beat in the PPG signal 171.

[0032] After individual heartbeats are identified in the PPG signal by applying a signal mask (or using other suitable techniques) at 152, the mask-filtered PPG signal is cross-correlated with other non-PPG ventilator waveforms, such as, but not limited to, airway pressure and flow waveforms. Because the expiratory phase is less corrupted by ventilator structures than the inhalation phase, it is preferable to cross-correlate the mask-filtered PPG signal specifically with the expiratory phase of the non-PPG ventilator waveforms. Figures 7B and 7C show graphs of cross-correlated PPG signals 182, 183 resulting from cross-correlating a zero-crossing PPG signal 172 with the expiratory phase of a standardized airway pressure waveform 173 and flow waveform 174, lagging the zero-crossing PPG signal 172 by a time shift 181. The time shift 181 can be understood by observing the first peaks 192, 193 of the cross-correlated PPG signals 182, 183 relative to the first peak 191 of the zero-crossing PPG signal 172. It is understood that the first peaks 192, 193 are offset from the first peak 191 by the time shift 181, and that all subsequent nth peaks of the cross-correlated PPG signals 182, 183 are likewise offset by the time shift 181 from the corresponding nth peak of the zero-crossing PPG signal 172. Like the peaks of the zero-crossing PPG signal 172, the peaks (e.g., first peaks 192, 193) of the cross-correlated PPG signals 182, 183 indicate the start of individual heartbeats.

[0033] At 153, a cardiogenic index is generated by determining an association between the cardiogenic data in the PPG signal and at least one of other non-PPG ventilator waveforms (i.e., airway pressure or flow waveforms). In a first non-limiting example, breath delivery and non-breath delivery regions of the non-PPG waveform are identified, and a signal-to-noise ratio (SNR) generated by comparing values ​​from the breath delivery region with values ​​from nearby non-breath delivery regions can be used as the cardiogenic index. First, breath delivery intervals in the non-PPG waveform corresponding to time intervals during which the ventilator is delivering breaths to the patient are identified. The heartbeat immediately preceding each breath delivery interval is considered to be the heartbeat of interest, and the interval in the non-PPG waveform corresponding to the heartbeat of interest is considered to be the interval of interest. For example, because the onset of each heartbeat in a cross-correlated PPG signal, such as cross-correlated PPG signals 182 and 183, is indicated by a peak spanning a theoretically infinitesimal, infinitesimal, time interval, it is understood that when referring to an interval of interest in a non-PPG signal, a determination must be made as to how long the interval of interest spans. A 50-millisecond interval is suggested as being of sufficient duration to yield enough data to generate a meaningful cardiogenic index, but it is understood that longer or shorter time intervals may be used without departing from the scope of the disclosed concepts. Referring to FIG. 7B and using airway pressure waveform 173 as a non-limiting example of a non-PPG waveform used to calculate SNR, peak 195 on cross-correlated PPG signal 182 corresponds to the breath delivery interval of airway pressure waveform 173 (which is determined exogenously), and peak 196 on cross-correlated PPG signal 182 indicates the onset of the heartbeat of interest relative to peak 195. Incidentally, the breath delivered during the interval associated with peak 195 is PVA. A 50 millisecond (or other selected time period) interval 197 of the airway pressure waveform 173 beginning where the airway pressure waveform 173 coincides with peak 196 is the interval of interest.

[0034] The SNR is determined by comparing (1) the extent of distortion of the non-PPG signal values ​​in an interval of interest with (2) the values ​​of the non-PPG signal in a region not coinciding with a heartbeat and closely adjacent to the interval of interest. Referring again to Figure 7B, region 198 of airway pressure waveform 173 is a region adjacent to the interval of interest (interval 197) that is not coinciding with a heartbeat. An exemplary SNR of airway pressure waveform 173 is determined by comparing the extent of distortion of the airway pressure values ​​in interval 197 (the interval of interest) with the airway pressure values ​​in region 198 (the region adjacent to the interval of interest that is not coinciding with a heartbeat).

[0035] The extent of distortion of the non-PPG signal during the interval of interest is significantly greater immediately prior to the onset of a PVA than prior to the onset of a correctly delivered ventilator-delivered breath. Thus, the SNR calculated for a PVA-related region of the non-PPG signal is readily distinguishable from the SNR calculated for a correctly delivered ventilator-delivered breath-related region of the non-PPG signal. This distinguishability makes the SNR suitable for use as a cardiogenic index for detecting PVA.

[0036] In another non-limiting example of how the cardiogenic index is calculated in 153, another type of SNR different from the type described in the previous non-limiting example can be determined and used as the cardiogenic index. First, the dominant frequency spectrum of the PPG signal and its harmonics are determined, for example, by spectral analysis. Then, for a selected non-PPG ventilator waveform (e.g., airway pressure waveform or flow waveform), the SNR can be calculated for different regions of the non-PPG waveform by obtaining the ratio of the dominant PPG frequency component to the dominant non-PPG frequency component present in each region of the waveform. The ratio of the dominant PPG frequency component to other frequency components is significantly greater before the occurrence of PVA than before the occurrence of a correctly delivered ventilator-delivered breath. The SNR calculated for a PVA-related region of the non-PPG waveform is easily distinguishable from the SNR calculated for a non-PPG waveform region associated with a correctly delivered ventilator-delivered breath, making the SNR suitable for use as a cardiogenic index for detecting PVA.

[0037] Method 100 will be referenced throughout this disclosure as a means for calculating a cardiogenic index, for example, with respect to the description of methods 200 and 300, and it will be understood that method 150 may be used in place of method 100 for any ventilator capable of providing a PPG without departing from the scope of the disclosed concepts. Additionally, the PPG analysis described with respect to method 150 may be performed using any other physiological signal (as long as the physiological signal is available on the ventilator), including, for example, without limitation, heart rate information, such as an electrocardiogram. Furthermore, throughout this disclosure, breathing attributes are referred to, particularly with respect to performing steps of method 100, and it will be understood that if method 150 is performed instead of method 100, the breathing attributes should be read to include PPG data (or data from any other physiological signal including heart rate information).

[0038] The utility of the cardiogenic index lies in how it is used in the ongoing process of dynamically assessing and adjusting ventilator settings to maximize the effectiveness of a patient's respiratory therapy. FIG. 8 is a flowchart of the above-described process 200 of dynamically assessing and adjusting ventilator settings. Process 200, in an exemplary embodiment, is performed by a controller included in a ventilator that delivers respiratory therapy to a patient. At 201, initial ventilator settings are selected by a caregiver. At 202, patient respiratory attributes are measured by the ventilator. Examples of respiratory attributes measured at 202 include, but are not limited to, airway pressure and airway flow. At 203, a cardiogenic index is calculated based on the respiratory attributes measured at 202 using method 100 (or, if PPG data is available at the ventilator, using method 150). In an exemplary embodiment, the cardiogenic index is calculated continuously while the ventilator is in use by a controller included in the ventilator, calculated on a rolling 15-second breath-by-breath basis, such that the cardiogenic index at any given moment reflects the patient's respiratory characteristic data from the previous 15 seconds. It is understood that the cardiogenic index may be calculated on a rolling basis for time periods less than or greater than 15 seconds without departing from the disclosed concepts. As described in more detail herein with respect to FIG. 10 and shown at 204 and 205, in an exemplary embodiment, a machine learning model included in the controller analyzes multiple respiratory features generated from the respiratory attributes measured in 202, including the cardiogenic index, to determine whether a mistrigger or malfunction has occurred. If the machine learning model determines at 204 and 205 that PVA has occurred, the controller issues a notification of asynchrony detection at 206. For example, without limitation, the controller may sound a tone, send a written message on the ventilator screen, or a combination thereof, to alert the caregiver that a PVA has occurred. It is understood that if a PVA is not detected, process 200 returns to 202.

[0039] After notification of asynchrony detection is issued at 206, processing continues to 207 where the controller either issues pre-programmed recommended adjustments to the ventilator settings or the caregiver can make their own decisions regarding adjustments that need to be made to the ventilator settings. After the adjustments are recommended by the controller or determined by the caregiver at 207, processing continues to 208 where the ventilator settings are adjusted according to the adjustments determined at 207. Processing then returns to 202. It is understood that method 100 can be performed continuously throughout the period of time that the ventilator is in use.

[0040] Method 100 represents an improvement over existing methods for identifying PVA, which do not use cardiogenic artifact to identify the occurrence of PVA. Using cardiogenic artifact to classify ventilator-delivered breaths, or the lack thereof, as PVA allows for near-real-time notification of ventilator-delivered PVA occurrences, reducing or eliminating the need for respiratory therapists, intensivists, or other caregivers to manually determine the occurrence of PVA on a breath-by-breath basis. Thus, ventilator settings can be adjusted more timely to prevent PVA recurrence and minimize patient discomfort and other undesirable consequences that often result from prolonged undetected PVA. Additionally, use of method 100 facilitates more meaningful monitoring of long-term trends (up to a 76-hour timescale) resulting from adjustments made to a patient's respiratory therapy, allowing caregivers to better understand how adjustments to ventilator settings affect trends in asynchrony.

[0041] In a non-limiting exemplary embodiment, a ventilator controller executes a machine learning model to perform process 200, particularly steps 204 through 205, where the machine learning model is trained to use cardiogenic indices calculated from ventilator waveform data using method 100. Training a machine learning model to use cardiogenic artifacts present in ventilator waveforms to detect PVA represents another improvement over prior art systems and methods for identifying PVA.

[0042] FIG. 9 is a simplified schematic diagram of an exemplary ventilator 30 according to one non-limiting exemplary embodiment of the disclosed concepts. The ventilator 30 includes a controller 31 that accepts input from and sends output to a user interface 32. In the exemplary embodiment, as shown in FIG. 9, the machine learning model 33 is software embedded in the controller 31. The user interface 32 includes a screen that allows a caregiver to instruct the controller 31 to select settings for the ventilator 30 via an input mechanism included in the user interface 32, and also displays output from the controller 31, such as a waveform of the patient's respiratory activity. Airflow generated by the ventilator 30 and delivered to the patient's airway 41 exits the ventilator 30 via an inflow path, and air exhaled through the patient's airway 41 is returned to the ventilator via an outflow path. The ventilator 30 can include multiple inflow sensors 34 or outflow sensors 35 configured to measure airway pressure, flow, and / or other respiratory attributes. It is understood that the inflow sensor 34 and outflow sensor 35 are in direct or indirect communication with the controller 31 to enable the controller 31 to monitor and analyze the data measured by the sensors 34, 35.

[0043] FIG. 10 is a flowchart of a method 300 used to train a machine learning model, such as model 33 shown in FIG. 9, to detect mistriggered or malfunctioning PVA using a cardiogenic index, according to an exemplary embodiment of the disclosed concepts. At 301, a training data set is collected. In one non-limiting exemplary embodiment, the training data set is collected from ventilator waveform data of several patients as well as from a laboratory ventilator simulating a patient's respiratory effort. In another non-limiting exemplary embodiment, the training data set is collected exclusively from patient ventilator waveform data. It is understood that the more patients whose data are used to train the machine learning model, the more accurate the machine learning model will be. While it may be preferable to use data from at least 20 patients, there is no required minimum number of patients from which data should be collected. It is also understood that data for different respiratory volumes may be collected from each patient without departing from the scope of the disclosed concepts. For example, without limitation, respiratory data for 100 breaths from one patient may be used, while respiratory data for 45 breaths from another patient may be used.

[0044] At 302, breath segmentation is performed on the patient ventilator waveforms of the training dataset, where the patient waveforms are manually analyzed and labeled to indicate the start of inspiration and the start of expiration across each waveform, as well as labeling any observed PVAs (this step can be omitted if ventilator flags identifying when a breath is delivered are available). Breath labeling is preferably performed by an expert. Non-limiting examples of experts include intensivists, respiratory therapists, and anesthesiologists. A perceived PVA is preferably labeled as a PVA only if two or more experts agree that a PVA has occurred. It is understood that breath segmentation does not need to be performed on the patient's breathing effort simulated in the laboratory, as the simulating program can provide the breath segmentation information. At 303, the training set patient ventilator waveforms are checked for artifacts and quality by manually removing mandatory breaths used to estimate lung function or to purge ventilator sensors (as this can cause malfunctions and mis-triggered PVA in the context of assist mode ventilation), as well as any other signals that are attributable to ventilator sensors rather than patient breathing.

[0045] At 304, feature extraction is performed on the training data set, where several features used to classify asynchronous events are calculated for each breath by collecting all data from the breath and mapping them to a small-valued vector. A cardiogenic index is one such feature extracted at 304 using method 100 or method 150. Other features that may be extracted at 304 include, for example, but are not limited to, the ratio between inspiration duration and expiration duration (used to classify malfunctioning PVA), the estimated patient inspiratory effort at the beginning of the inspiration phase (also used to classify malfunctioning PVA), and the gradient of mean airway pressure of the estimated patient forced effort during the expiration phase (used to classify mis-triggered PVA). While some of the features extracted at 304 are specific to detecting either malfunctions or mis-triggered events, the cardiogenic index is extracted as a feature used to classify both malfunctions and mis-triggered events. In laboratory studies, using the cardiogenic index in addition to other features used to classify malfunctions and mis-trigger events increased the effectiveness of the machine learning model compared to studies in which the cardiogenic index was not used as a classification feature.

[0046] At 305, a classifier is trained, where the machine learning model is provided with a training data set that can analyze the segmented breaths and labeled PVAs, along with the extracted features, to detect patterns indicative of both malfunctioning and mis-triggered PVAs, with the goal of learning to identify PVAs in raw ventilator waveform data. As a first step, cross-validation may be performed, where the data in the training set is divided into multiple groups. Data from some of these groups is used to train the machine learning model on pattern recognition, while raw ventilator data from other groups is used to test the machine learning model's pattern recognition capabilities. This process is repeated multiple times, alternating between training and testing groups. In the initial stages of cross-validation, it may be preferable to use only patient data to train the machine learning model, and then, in later stages, augment the training data with simulated data from a laboratory ventilator (if simulated data is used). At 306, validation is performed, where test data not used in the training set is provided to the machine learning model to analyze the occurrence of PVAs, and the performance of the machine learning model is manually evaluated.

[0047] In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word "comprises" or "including" does not exclude the presence of elements or steps other than those listed in a claim. In a device claim enumerating several means, several of these means may be embodied by one and the same item of hardware. The absence of a plurality of elements does not exclude the presence of a plurality of elements. In any device claim enumerating several means, several of these means may be embodied by one and the same item of hardware. The mere fact that several elements are recited in mutually different dependent claims does not indicate that these elements cannot be used in combination.

Claims

1. 1. A method for identifying an occurrence of patient-ventilator asynchrony (PVA), the method comprising: measuring, with one or more sensors, a plurality of respiratory attribute signals in the time domain, including airway pressure and / or flow waveforms, of a patient receiving respiratory assistance from a ventilator; generating, by a controller, a cardiogenic index from the plurality of respiratory attribute signals; and analyzing, by the controller, at least the cardiogenic index and determining, based on the analysis, whether PVA has occurred. and The step of generating the cardiogenic index comprises: normalizing the plurality of respiratory attribute signals to generate a plurality of normalized respiratory signals; summing the plurality of standardized respiratory signals such that a data value at a given x-axis value from each of the standardized respiratory signals is added to corresponding data values ​​at the same given x-axis value from all other standardized respiratory signals to generate a cumulative respiratory signal; generating a power spectrum by converting the cumulative respiratory signal from a time domain to a frequency domain signal; normalizing the power distribution values ​​of the power spectrum within a predetermined cardiogenic frequency range; and generating a cardiogenic index in the form of a standardized cardiogenic signal by converting the normalized power distribution values ​​from the predetermined cardiogenic frequency range into a time domain signal; and The determining step includes: comparing the cardiogenic index to a plurality of predetermined conditions indicative of the occurrence of PVA; and making a determination that the PVA has occurred if the cardiogenic index satisfies the plurality of predetermined conditions indicative of the occurrence of the PVA. A method comprising:

2. 2. The method of claim 1, wherein the lower limit of the cardiogenic frequency range is 1.2 Hz and the upper limit of the cardiogenic frequency range is 20 Hz.

3. The step of normalizing the power distribution values ​​of the power spectrum includes: obtaining the sum of all of the power distribution values ​​contained in a first Nyquist zone; and Dividing each power distribution value that falls within the cardiogenic frequency range by the sum of all of the power distribution values ​​that fall within the first Nyquist zone.

2. The method of claim 1 , further comprising:

4. 4. The method of claim 3, wherein an upper limit of the first Nyquist zone is an upper limit of the cardiogenic frequency range.

5. The method of claim 1 , wherein the plurality of respiratory attribute signals includes a photoplethysmogram (PPG) signal.

6. The method of claim 1 , wherein the PVA is one of a malfunctioning PVA or a mis-triggered PVA.

7. training a machine learning model to make the determination that PVA has occurred if the cardiogenic index satisfies the plurality of predetermined conditions indicative of the occurrence of PVA. The method of claim 1 further comprising:

8. A controller having a machine learning model. an inlet pathway configured to supply air from the ventilator into the patient's airway; an outflow pathway configured to receive exhaled air from the patient's airway; and a plurality of sensors configured to measure a plurality of respiratory attribute signals in the time domain, including airway pressure waveforms and flow waveforms of a patient receiving respiratory assistance from the ventilator; In the ventilator, the controller is configured to receive respiratory attribute signals measured by the plurality of sensors; the controller is configured to generate a cardiogenic index based on the respiratory attribute signals measured by the plurality of sensors, and analyze at least the cardiogenic index to determine whether PVA has occurred; generating the cardiogenic index normalizing the plurality of respiratory attribute signals to generate a plurality of normalized respiratory signals; summing the plurality of standardized respiratory signals such that a data value at a given x-axis value from each of the standardized respiratory signals is added to corresponding data values ​​at the same given x-axis value from all other standardized respiratory signals to generate a cumulative respiratory signal; generating a power spectrum by converting the cumulative respiratory signal from a time domain to a frequency domain signal; normalizing the power distribution values ​​of the power spectrum within a predetermined cardiogenic frequency range; and generating a cardiogenic index in the form of a standardized cardiogenic signal by converting the normalized power distribution values ​​from the predetermined cardiogenic frequency range into a time domain signal; Having that, the machine learning model is trained to recognize conditions indicative of the occurrence of PVA; The machine learning model is trained to determine whether PVA has occurred using at least the cardiogenic index. Respirator.

9. 9. The ventilator of claim 8, wherein the controller is configured to issue a notification of asynchrony detection if the machine learning model determines that PVA has occurred.

10. 9. The ventilator of claim 8, wherein the machine learning model is trained to distinguish between malfunctioning and mis-triggering PVAs.

11. 10. The ventilator of claim 8, wherein the machine learning model is trained to distinguish between a malfunctioning PVA and a mis-triggered PVA.

Images