Waveform-based hemodynamic instability alert

Abstract

The controller 150 for waveform-based hemodynamic instability warning includes a memory 151 that stores instructions and a processor 152 that executes the instructions. The controller 150 implements a process including receiving an electrocardiogram waveform via a first interface 153 that interfaces with at least one electrocardiogram monitor 155 monitoring a patient and identifying heartbeats from the electrocardiogram waveform. The process further includes separating the heartbeats into a plurality of first time windows, extracting features of the heartbeats in each of the first time windows as first extracted features for each first time window, and generating generated features spanning a second time window that includes the plurality of first time windows based on the first extracted features. The process further includes applying trained artificial intelligence to the generated features, predicting hemodynamic instability in the patient based on applying the trained artificial intelligence to the generated features, and outputting an alert warning of hemodynamic instability based on the predicted hemodynamic instability.

Description

[Technical Field]

[0001] FIELD OF THE INVENTION

[0001] The present invention relates to waveform-based hemodynamic instability warnings. [Background technology]

[0002]

[0002] Hemodynamic instability (HI) is a condition that can be defined as a deficiency in blood flow to organs and tissues, resulting in an inability to meet metabolic demands. Hemodynamic instability reflects problems in the circulatory system and can lead to cellular dysfunction and death. Critically ill patients with reduced cardiac function are at increased risk of circulatory failure and hemodynamic instability. In critical care settings, such as intensive care units (ICUs), interventions for patients with hemodynamic instability often include fluid resuscitation to increase preload, administration of vasopressors to increase peripheral resistance (afterload) to maintain systemic blood pressure, and / or administration of inotropes to increase cardiac contractility.

[0003] Previous studies on predicting hemodynamic instability have typically relied on nurse-recorded vital signs (e.g., temperature, pulse rate, respiratory rate, and / or blood pressure) and clinical measurements used to classify time segments during an ICU stay as stable or unstable. In one study to predict hemodynamic instability, nurse-recorded vital signs and clinical measurements were applied to a model, resulting in an area under the receiver operating characteristic (AUROC) of 0.82. In another study to predict hemodynamic instability, removing missing data from consideration resulted in an AUROC of 0.87. In another study to predict bleeding in a surgical ICU, bedside monitor vital signs were added, resulting in an AUROC of 0.92. The bedside monitor vital signs are considered more frequent than nurse-recorded vital signs. In yet another study to predict hemodynamic instability, ECG waveforms were used, albeit only in simulated settings for healthy volunteers, resulting in an AUROC of 0.86 to 0.88.

[0004]

[0004] Continuous arterial blood pressure (ABP) waveforms convey relevant hemodynamic information. Arterial blood pressure has been used in ICU settings to monitor cardiovascular dysfunction and responses to various interventions. The two main components of an individual ABP pulse cycle are systole and diastole, separated by a dicrotic notch. Aortic valve closure is graphically reflected as a dicrotic notch in the ABP pulse cycle. Useful features, such as cardiac output (CO), as a measure of cardiac function, can be estimated from the ABP waveform. CO is a function of heart rate, preload, contractility, and afterload. Other measures, such as the maximum first derivative of the ABP pulse cycle, serve as indirect measures of cardiac contractility. ABP waveforms have been used to predict hypotensive events in ICU settings. For example, in one study to predict hypotensive events over the next 15 minutes, features extracted from the ABP waveform resulted in an AUROC ranging from 0.91 to 0.97. Summary of the Invention [Problem to be solved by the invention]

[0005] It is an object of the present invention to provide a waveform-based warning of hemodynamic instability. [Means for solving the problem]

[0006] According to one aspect of the present disclosure, an apparatus includes a first interface, a memory, and a processor. The first interface interfaces with at least one electrocardiogram monitor monitoring a patient. The memory stores instructions. The processor executes the instructions. When executed by the processor, the instructions cause the apparatus to identify a plurality of heartbeats from an electrocardiogram waveform received from the at least one electrocardiogram monitor via the first interface; separate the plurality of heartbeats into a plurality of first time windows; and extract features of the heartbeats in each of the first time windows as first extracted features for each first time window. The instructions also cause the apparatus to generate generated features across a second time window that includes a plurality of first time windows based on the first extracted features; apply trained artificial intelligence to the generated features; predict hemodynamic instability for the patient based on applying the trained artificial intelligence to the generated features; and output an alert warning of hemodynamic instability based on predicting the hemodynamic instability.

[0007] According to another aspect of the present disclosure, a method includes receiving a plurality of electrocardiogram waveforms via a first interface that interfaces with at least one electrocardiogram monitor monitoring a patient; identifying a plurality of heart beats from the electrocardiogram waveforms; separating the plurality of heart beats into a plurality of first time windows; and extracting features of the heart beats in each of the first time windows as first extracted features for each first time window. The method also includes generating generated features over a second time window that includes a plurality of the first time windows based on the first extracted features; applying trained artificial intelligence to the generated features; predicting hemodynamic instability for the patient based on applying the trained artificial intelligence to the generated features; and outputting an alert warning of the hemodynamic instability based on the predicted hemodynamic instability.

[0008] According to yet another aspect of the present disclosure, a tangible, non-transitory computer-readable storage medium stores a computer program. When executed by a processor, the computer program causes a system including the tangible, non-transitory computer-readable storage medium to: identify a plurality of heartbeats from an electrocardiogram waveform received from at least one electrocardiogram monitor via a first interface; separate the plurality of heartbeats into a plurality of first time windows; and extract features of the heartbeats in each of the first time windows as first extracted features for each first time window. The computer program also causes the system to: generate generated features over a second time window comprising a plurality of first time windows based on the first extracted features; apply trained artificial intelligence to the generated features; predict hemodynamic instability for a patient based on applying the trained artificial intelligence to the generated features; and output an alert warning of hemodynamic instability based on the prediction of hemodynamic instability.

[0009]

[0008] The illustrative embodiments are best understood from the following detailed description when read in conjunction with the accompanying drawings. It is emphasized that the various features are not necessarily drawn to scale. In fact, dimensions may be arbitrarily increased or decreased for clarity of illustration. Where applicable and practical, like reference numerals refer to like elements. [Brief explanation of the drawings]

[0010] [Figure 1] FIG. 1 illustrates a system for waveform-based hemodynamic instability warning, according to a representative embodiment. [Figure 2] FIG. 2 shows a processing pipeline and extracted features for waveform-based hemodynamic instability warning, according to a representative embodiment. [Figure 3]

[0011] FIG. 3 illustrates a method for waveform-based hemodynamic instability warning, according to a representative embodiment. [Figure 4]

[0012] FIG. 4 illustrates a computer system for waveform-based hemodynamic instability warning, according to a representative embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0011]

[0013] In the following detailed description, for purposes of explanation and not limitation, exemplary embodiments disclosing specific details are presented to provide a thorough understanding of embodiments in accordance with the present teachings. Descriptions of known systems, devices, materials, methods of operation, and methods of manufacture may be omitted so as not to obscure the description of the exemplary embodiments. Nevertheless, systems, devices, materials, and methods within the understanding of those skilled in the art are within the scope of the present teachings and can be used in accordance with the exemplary embodiments. It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting. Defined terms are to be given the technical and scientific meaning of the defined terms as commonly understood and accepted in the art of the present teachings.

[0012]

[0014] In this specification, terms such as "first," "second," and "third" may be used to describe various components or parts, but it should be understood that these components or parts should not be limited by these terms. These terms are used only to distinguish one component or part from another. Thus, a first component or part described below could be called a second component or part without departing from the teachings of the present invention.

[0013]

[0015] The terms used herein are for the purpose of describing particular embodiments only and are not intended to be limiting. As used in this specification and the appended claims, the singular forms "a," "an," "the," and / or "comprising" and / or similar terms, when used herein, specify the presence of stated features, elements, and / or components, but do not exclude the presence or addition of one or more other features, elements, components, and / or groups thereof. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items.

[0014]

[0016] Unless otherwise specified, when a component or part is said to be "connected," "coupled," or "adjacent" to another component or part, it will be understood that the component or part may be directly connected or coupled to the other component or part, or there may be intervening components or components. That is, these and similar terms encompass situations where one or more intermediate components or components may be used to connect the two components or parts. However, when a component or part is said to be "directly connected" to another component or part, this only includes situations where the two components or parts are connected to each other without any intermediate or intervening components or components.

[0015]

[0017] As such, the present disclosure is intended to derive one or more of the advantages specifically described below through one or more of its various aspects, embodiments, and / or specific features or subelements. For purposes of explanation and not limitation, exemplary embodiments disclosing specific details are described to provide a thorough understanding of embodiments in accordance with the present teachings. However, other embodiments consistent with the present disclosure that depart from the specific details disclosed herein remain within the scope of the appended claims. Furthermore, descriptions of well-known devices and methods may be omitted so as not to obscure the description of the exemplary embodiments. Such methods and devices are within the scope of the present disclosure.

[0016]

[0018] As described herein, ECG and / or ABP waveforms can be used to predict hemodynamic instability without or in combination with nurse-recorded vital signs and clinical measurements. According to embodiments herein, if a baseline model for hemodynamic instability prediction is established using only clinical measurements and nurse-recorded vital signs, a 5% improvement in AUROC can be obtained along with a 9% improvement in area under the precision-recall (PR) curve (AUPRC) when features extracted from ECG and ABP waveforms are added. When features extracted from ECG and ABP waveforms are used alone without any clinical measurements or nurse-recorded vital signs, a 4% improvement in AUROC and AUPRC can still be observed compared to the baseline results.

[0017]

[0019] FIG. 1 illustrates a system 100 for waveform-based hemodynamic instability warning, according to a representative embodiment.

[0018]

[0020] 1 is a system for waveform-based hemodynamic instability warning and includes components that may be located together or distributed. System 100 includes one or more ECG monitors 101, one or more ABP monitors 102, a workstation 140, a monitor (display) 155, and an artificial intelligence controller 180. An artificial intelligence training system 190 provides trained artificial intelligence to artificial intelligence controller 180.

[0019]

[0021] ECG monitor 101 is an electrocardiogram monitor. ECG monitors monitor the electrical activity of the heart by measuring the electrical current passing through the heart. Each ECG monitor can record the strength and timing of the signal on a graph and may include one or more electrodes attached to the patient's body via a patch and wires that transmit the ECG tracing signal to a receiver on workstation 140. The signal recorded by ECG monitor 101 may include P waves, QRS complexes, and T waves.

[0020]

[0022] The ABP monitor 102 is an arterial blood pressure monitor. The ABP monitor 102 monitors the patient's arterial blood pressure via a pressure transducer. The pressure transducer can be an external pressure sensor connected to the arterial blood vessel. The external pressure sensor is typically connected to the arterial blood vessel at a radial position via a fluid-filled catheter that enters the arterial blood vessel. Alternatively, the pressure transducer can be an internal pressure sensor attached to the tip of the catheter that enters the arterial blood vessel. The ABP monitor 102 can record continuous ABP signal waveforms and measure arterial blood pressure measurements. Arterial blood pressure measurements that can be measured using the ABP monitor 102 include systolic blood pressure (SPB), diastolic blood pressure (SBP), and mean arterial pressure (MAP). The ABP monitor 102 connects to a receiver on the workstation 140.

[0021]

[0023] The workstation 140 includes a controller 150, a first interface 153, a second interface 154, and a touch panel 156. The controller 150 includes a memory 151 that stores instructions and a processor 152 that executes the instructions. The controller 150 controls and implements some or all aspects of the methods attributed to the workstation 140, as described herein. The first interface 153 interfaces the ECG monitor 101 to the workstation 140. The second interface 154 interfaces the ABP monitor 102 to the workstation 140. That is, the second interface 154 interfaces one or more arterial blood pressure monitors. The first interface 153 and the second interface 154 may be ports, adapters, or other types of wired or wireless interfaces used to send and receive data. The touch panel 156 may include a touch interface that accepts input via touch, such as by direct touch, or via a mouse, keyboard, or other hand-controlled input mechanism. The touch panel 156 may also include a visual display for displaying touch input so that the user can confirm the input. Workstation 140 may also include one or more other input interfaces. The other input interfaces (not shown) of workstation 140 may include ports, disk drives, wireless antennas, or other types of receiver circuitry. The other input interfaces may also connect other user interfaces to workstation 140, such as a mouse, keyboard, microphone, video camera, touchscreen display, or other components or parts.

[0022]

[0024] Display (monitor) 155 is an electronic monitor that displays images and data visualizations. Monitor 155 can be a computer monitor, a display on a mobile device, a television, an electronic whiteboard, or other screen configured to display electronic images. Monitor 155 can also include one or more input interfaces, such as those previously described, that allow other components or parts to be connected to workstation 140, and a touchscreen that allows direct input via touch.

[0023]

[0025] Artificial intelligence controller 180 includes memory 181 for storing instructions and processor 182 for executing the instructions. Artificial intelligence controller 180 can dynamically apply trained artificial intelligence to data received from controller 150 based on ECG signals from ECG monitor 101 and ABP signals from ABP monitor 102. In some embodiments, artificial intelligence controller 180 is implemented as a component of workstation 140. In other embodiments, controller 150 and artificial intelligence controller 180 are implemented with the same, rather than different, memory and processor.

[0024]

[0026] The artificial intelligence training system 190 includes a memory 191 for storing instructions and a processor 192 for executing the instructions. The artificial intelligence training system 190 trains the artificial intelligence based on datasets from multiple different instantiations of patient monitoring to learn optimal correlations between hemodynamic instability input data and resulting outcomes. The datasets may be provided from databases such as the Medical Information Mart for Intensive Care (MIMIC) III database described elsewhere herein. The trained artificial intelligence provided by the artificial intelligence training system 190 can then be used to predict hemodynamic instability in the manner described herein and ultimately provide warnings and treatments to prevent, treat, or address the predicted hemodynamic instability. The training may be based on analysis of thousands of patients, including hundreds of hemodynamically unstable patients. The training may include binary classification, and may result in a 5% or greater improvement in AUROC when features of both ECG and ABP waveforms are considered along with laboratory results and nurse-recorded vital signs. Such training can result in an improvement of over 4% in AUROC, even when only features from both the ECG and ABP waveforms are considered.

[0025]

[0027] FIG. 2 shows a processing pipeline and extracted features for waveform-based hemodynamic instability warning, according to a representative embodiment.

[0026]

[0028] In FIG. 2 , the ABP processing pipeline in the upper left is end-to-end and includes three steps: (1) beat segmentation, (2) ABP feature extraction, and (3) signal anomaly detection. This end-to-end ABP processing pipeline is used for feature extraction from ABP waveforms. The ECG pipeline in the upper left is also end-to-end and includes three steps: (1) beat segmentation and labeling, (2) signal quality index, and (3) ECG feature extraction. Beat segmentation in the ECG pipeline may include QRS complex detection and then beat classification. Signal quality index in the ECG pipeline may include signal quality assessment. The ECG waveform in the ECG pipeline may be from the MIMIC III database and may be recorded at 125 Hz or 250 Hz. The ECG waveform may be upsampled from 125 Hz to 250 Hz to improve the temporal resolution and detectability of the QRS complex. The end-to-end ECG and ABP processing pipeline of FIG. 2 is a targeted implementation that uses waveform signals commonly available in an ICU environment without requiring any additional specialized equipment.

[0027]

[0029] In Figure 2, panel (A) on the right shows nurse-recorded vital signs, and panel (B) on the right shows vital signs on a bedside monitor. Data from the ECG pipeline is shown on the right in (C) ECG waveform HR (5 min), which is the 5-minute average heart rate (HR) extracted from the ECG waveform. Data from the ABP pipeline is shown on the right in (D) ABP waveform SBP(*), DBP(*), and MAP(*) (5 min), which are the 5-minute average systolic blood pressure (SBP), 5-minute average diastolic blood pressure (DBP), and 5-minute averaged mean arterial pressure (MAP), respectively, extracted from the ABP waveform. In Figure 2, features (A), (B), (C), and (D) are for a sample patient. Features shown in (C) for the ECG waveform are extracted from heartbeats divided into the first 5-minute time window. The features shown in (D) were extracted from an arterial blood pressure wave that was divided into pulsatile periods and then separated into a first time window of 5 minutes.

[0028]

[0030] The six-hour window shown in Figure 2 is a second time window superimposed on (A), (B), (C), and (D). The six-hour window ends at least one hour before the intervention time point indicated by the vertical line superimposed on (A), (B), (C), and (D). The six-hour window includes 72 of the first time windows, each 5 minutes long for the ECG waveform (C) and the ABP waveform (D). Features derived from the six-hour windows (A), (B), (C), and (D) are used in the hemodynamic instability prediction model shown in the lower left to classify stable and unstable patients. The derived features are derived from the time series data of (A), (B), (C), and (D), and therefore include features derived from the nurse-recorded vital signs (A), features derived from the bedside monitor vitals (B), features derived from the segmented ECG waveform (C), and features derived from the ABP waveform (D).

[0029]

[0031] Vital signs shown on the nurse-recorded vital signs shown in Figure 2 (A) and the bedside monitor vital signs shown in Figure 2 (B) include heart rate (HR), systolic blood pressure (SBP), and diastolic blood pressure (DBP). Clinical measurements include numerous features extracted from laboratory test results and ventilator measurements, including arterial blood base excess, aspartate aminotransferase (AST), band count, basophil count, calcium, CO2, creatinine, eosinophil count (EOS), fraction of inspired oxygen (FIO2), glucose, hematocrit, hemoglobin, ionized calcium, lactate, magnesium, mean airway pressure, partial pressure of carbon dioxide (PaCO2), peak inspiratory pressure, potassium, partial thromboplastin time (PTT), oxygen saturation (SAO2), sodium, bilirubin, white blood cell count, blood urea nitrogen (BUN), and central venous pressure. Clinical measurements may also include body temperature and the patient's age at admission. Clinical measurements may also include mean arterial blood pressure (MAP), calculated as (SPB+2×DBP) / 3, and shock index (SI), calculated as HR / SBP.

[0030]

[0032] FIG. 3 illustrates a method for waveform-based hemodynamic instability warning, according to a representative embodiment.

[0031]

[0033] Figure 3 illustrates a waveform-based hemodynamic instability warning process, according to a representative embodiment. The method of Figure 3 may be performed by a single device, such as controller 150 or workstation 140, by a single system, such as system 100, by or on behalf of a single entity, or by a distributed device or system, by or on behalf of multiple entities. In some embodiments, the method of Figure 3 may be performed by workstation 140 of Figure 1, if integrated with or including artificial intelligence controller 180, or by a combination of workstation 140 and artificial intelligence controller 180 of Figure 1, if workstation 140 is separate from artificial intelligence controller 180.

[0032]

[0034] At S305, the method of Figure 3 includes receiving an electrocardiogram waveform, which may be received by workstation 140 from ECG monitor 101 via first interface 153 in Figure 1.

[0033]

[0035] At S310, the method of Figure 3 includes receiving an arterial blood pressure waveform, which may be received by workstation 140 from ABP monitor 102 via second interface 154 in Figure 1.

[0034]

[0036] At S315, the method of Figure 3 includes identifying heart beats from the electrocardiogram waveform, which heart beats may be identified from the electrocardiogram waveform received at S305 and from the analysis performed by controller 150 of Figure 1.

[0035]

[0037] At S325, the method of Figure 3 includes separating the heartbeats into first time windows. The heartbeats may be separated into the first time windows by controller 150 of Figure 1. For example, the first time windows may each include a 5-minute time interval, and 72 of the first time windows may be included within each 6-hour second time window, as described herein.

[0036]

[0038] At S330, the method of Fig. 3 includes dividing the arterial blood pressure waveform into pulsatile periods. For example, the controller 150 of Fig. 1 may divide the arterial blood pressure waveform into pulsatile periods. The continuous arterial blood pressure waveform may be divided into individual pulsatile periods using a pulse detection algorithm that detects the onset of an arterial blood pressure pulse.

[0037]

[0039] Furthermore, although not shown, the pulse period may be divided into first time intervals. For example, the pulse period of the cardiac pulse (heart beat) and the arterial blood pressure waveform may be divided into the same first time window.

[0038]

[0040] At S335, the method of FIG. 3 includes excluding a heartbeat from further processing if it is noisy. The excluding at S335 may be performed by the controller 150 of FIG. 1 and may include one or more heartbeats. Accurate extraction of ECG features depends on the quality of the ECG signal, and many factors, including patient movement, poor electrode contact, misplaced electrodes, and electrical interference, can affect the quality of the ECG signal. The ECG quality assessment measure may be based on a combination of existing criteria and may take into account various types of noise present in the data. Eight types of noise that can be considered and filtered by a noise filter include low-frequency noise, high-frequency noise, flat-line ratio, peak-to-peak amplitude, spike index with slope, amplitude saturation, power line noise, and outlier ratio of detected beat-to-beat (QRS-to-QRS or RR) intervals. For each type of noise, an empirically determined threshold can be adjusted using ECG data from different groups of patients. An overall signal quality index (SQI) can be calculated using eight subindices. The signal quality can be evaluated on a second-by-second basis: for a 1-second ECG interval, the overall SQI receives a value of 1 (acceptable) only if all sub-indicators are below the adjusted thresholds.

[0039]

[0041] In S340, the method of FIG. 3 includes excluding an ABP pulse cycle from further processing if the pulse cycle is abnormal. That is, the method of FIG. 3 may include excluding at least one individual pulse cycle identified as abnormal from application of the trained artificial intelligence based on identifying at least one individual pulse phase as abnormal. The exclusion in S340 may be performed by controller 150 of FIG. 1 and may include one or more pulse cycles. A signal abnormality index (SAI) may be calculated for each arterial blood pressure pulse cycle, and abnormal arterial blood pressure pulse cycles that are deemed unusable may be discarded. An ABP pulse cycle may be discarded if any one of the following conditions is met: (2) MAP <30mmHg or >200mmHg; (3) (pulse pressure) PP <20mmHg; (4) sum of negative slopes of ABP cycles <-40mmHg / 100ms; (5) difference between SBP and DBP between adjacent ABP cycles >20mmHg; (6) difference in ABP cycle length in seconds >2 / 3 of a second; (7) inconsistent sequence of SBP, dicrotic notch, and DBP; (8) SBP <DBP;である。

[0040]

[0042] At S345, the method of FIG. 3 includes extracting features of heart beats within the first time window. The ECG waveform can be leveraged by extracting heart rate with higher temporal resolution and deriving heart rate variability (HRV), which serves as an indirect measure of autonomic regulation and vasomotor tone. Examples of features extracted from heart beats in the first time window can include averages calculated from individual heart beats, such as average heart rate. Further, examples of features extracted from heart beats within the first time window can include features across all beats within the first time window, such as heart rate variability. A commercial-grade (FDA-approved) ECG arrhythmia detection algorithm can be employed to perform ECG beat detection and classification for each ECG lead in each patient. For each detected beat, a beat location can be assigned to the QRS peak, and a beat type can be assigned to one of the following predefined labels based on the characteristics of each heart beat: These are normal beats (N), ventricular ectopic beats (V), supraventricular premature beats (S), paced beats (P), suspicious / unclassified beats (Q), and learning beats (L).

[0041]

[0043] The following ECG features can be extracted / computed for each ECG lead using beat locations and labels from only intervals with good signal quality: If ECG features are available from multiple leads, features can be selected from a single lead in the following order, for example, II, V, MCL, aVF, III, aVR, aVL, V1, V3, V5, and V2. ECG leads can be ranked in order based on availability and clinical relevance.

[0042]

[0044] Heart rate (HR) is the number of heart beats per minute. The instantaneous heart rate can be calculated using the R-R interval between two adjacent beats labeled N, V, S, P, Q, or L. An average of the instantaneous HR values ​​can be obtained over each first time window of 5 minutes. A time series of HR values ​​can be added from (A) nurse-recorded vital signs, (B) bedside monitor vital signs measured every minute, and (C) ECG waveforms averaged over 5-minute windows.

[0043]

[0045] Another extracted ECG feature can be heart rate asymmetry (HRA), which quantifies rapid accelerations and decelerations of heart rate over a period of time and succinctly expresses the imbalance as a single index.

[0044]

[0046] An additional extracted ECG feature can be heart rate variability (HRV). HRV is the variability of the time intervals between adjacent normal heart beats. HRV measures over a 5-minute time window are calculated using only beats labeled "N." HRV measures are grouped into time, frequency, and nonlinear domains, as described below. Time-domain HRV measures include the standard deviation of NN intervals (SDNN), the root mean square of successive differences between normal heart beats (RMSSD), the standard deviation of the differences between successive NN intervals (SDSD), and the percentage of adjacent NN intervals that differ from each other by more than 20 milliseconds (pNN20) and 50 milliseconds (pNN50), respectively.

[0045]

[0047] Additional extracted ECG features may be frequency-domain HRV measures, which utilize a fast Fourier transform (FFT) on the NN interval time series to extract the contributions of different frequency components. These frequency components may be organized into four non-overlapping bands: very low frequency (0.0001 Hz to 0.003 Hz), infrasonic frequency (0.004 to 0.04 Hz), low frequency (0.05 to 0.15 Hz), and high frequency (0.16 to 0.4 Hz). These frequency-domain measures may require calculation for adjacent NN intervals without any gaps. Therefore, these measures may be calculated only for the longest contiguous segment of NN intervals within a 5-minute time window. To obtain a regularly and frequently sampled time series for this segment, the NN interval may be resampled using cubic interpolation at a sampling frequency of 4 Hz. The mean of the NN time series may be subtracted to remove the effects of baseline offset. A Welch periodogram may be programmed for this time series using a Hanning window. NN time series with length ≦255 samples can be discarded because the Hanning window was set to a minimum length of 256 samples. The frequency domain measures include power in four frequency bands, total power, normalized relative power in low-frequency (LF) and high-frequency (HF) bands, and the LF / HF ratio. The correlation between HRV and heart rate, calculated over a second time window of 6 hours, can also be added to the features.

[0046]

[0048] Several nonlinear HRV measures can also be included. These measures can be calculated for the Poincaré plot of the longest segment of adjacent NN intervals, similar to the frequency-domain analysis described above. An ellipse can be fitted to the points on the Poincaré plot, and the width (SD1) and length (SD2) of the ellipse can be calculated, along with the SD1 / SD2 ratio.

[0047]

[0049] The number of premature ventricular contractions can be tracked by tracking the number of premature ventricular contractions (PVCs) and supraventricular premature systolic contractions within each 5-minute window. The frequency of PVCs is a predictor of heart failure and death due to reduced left ventricular ejection fraction.

[0048]

[0050] At S350, the method of FIG. 3 includes extracting arterial blood pressure features in a first time window. The ABP waveform can be exploited by extracting blood pressure measurements with relatively high temporal resolution as well as morphological features of the ABP. The morphological features can be used to estimate CO, cardiac contractility, and other indices of hemodynamics.

[0049]

[0051] For each ABP cycle, SBP can be calculated as the local maximum in the ABP cycle, and DBP can be calculated as the first point of zero slope (gradient) when crossing the ABP cycle from right to left. DBP can be calculated in this way without using a minimum. MAP and pulse pressure (PP) can be calculated from SBP and DBP. Two approaches can be used to identify the location of the dicrotic notch: the first is as 0.3 × T (T is the length of the ABP cycle in seconds), and the second is as the first zero slope crossing following SBP. Both approaches locate the dicrotic notch where there is no clear division between systole and diastole. Therefore, the dicrotic notch is located as the first zero slope crossing following t, where t is selected as round(0.3 × T). This combined approach reduces errors when locating the dicrotic notch assessed by visual assessment. Additionally, if no SBP, DBP, or dicrotic notch is found, the ABP pulsatile period can be excluded from the calculation in S340.

[0050]

[0052] Morphological features derived include time to SBP, time to DBP, and time to dicrotic notch (all from beat onset), areas under the systolic, diastolic, and total ABP cycle, and ABP cycle length. The maximum slope (first derivative of the ABP cycle) (dP / dt)max is calculated as an indirect measure of cardiac contractility used in titrating inotropic drugs. These features include:

number

[0051]

[0053] At S357, the method of FIG. 3 includes extracting features over a second time window. The extracted ECG and ABP waveform features are a time series such as those shown in FIG. 2C and 2D. The second time window may include a plurality of the first time windows in time sequence as shown in FIG. 2, for example, 72 first time windows totaling 6 hours. The features extracted over the second time window may be extracted by calculating time series features such as the mean, standard deviation, and slope of the extracted features over the first time windows included in the second time window.

[0052]

[0054] Derived features (DFs), which can be extracted over a second time window, can be extracted from the time series in the first time window to capture relevant information regarding hemodynamic instability. As previously shown in Figure 2, S357 summarizes only the available time series over a six-hour window, rather than the data from the one-hour window before intervention. This provides physicians with sufficient time to intervene. To capture the distribution characteristics of the time series, the mean, standard deviation, median, minimum, and maximum values ​​can be calculated for each time series within the six-hour window. The slope and intercept of each time series are calculated to capture trend information and baseline shifts, respectively. Trend features can be calculated for the mean-subtracted time series to enable meaningful comparisons between patients. Finally, gradients and approximate entropies can be calculated to capture the regularity and variability of the time series. The pattern length for similarity search is set to a default of 2 for both approximate entropy and sample entropy.

[0053]

[0055] In one embodiment, multiple second time windows can be used and the multiple second time windows can overlap. For example, if the second time window is 6 hours, a different second time window can start every 30 minutes, such that the features generated in S357 can be generated every 30 minutes, for example.

[0054]

[0056] 3 includes applying trained artificial intelligence to the extracted features. The trained artificial intelligence may be trained by artificial intelligence training system 190 and provided to artificial intelligence controller 180. The trained artificial intelligence may be applied by artificial intelligence controller 180 or, if artificial intelligence controller 180 is provided by controller 150, by workstation 140.

[0055]

[0057] At S370, the method of FIG. 3 includes predicting hemodynamic instability. For example, hemodynamic instability may be predicted one hour or more in advance. The prediction may include a confidence level, such as 30%, 60%, or 80%. The prediction may also specify a particular time or time period during which hemodynamic instability is predicted to begin or become detectable.

[0056]

[0058] At S380, the method of FIG. 3 includes determining whether the prediction of hemodynamic instability exceeds a threshold. If the prediction of hemodynamic instability does not exceed the threshold (S380=NO), the process of FIG. 3 returns to S305 and S310. For example, the threshold may be 50% or 75%. In some embodiments, multiple thresholds may be used. For example, a first threshold of 25% may be considered a warning threshold at which clinical staff should be prepared to take action in response to the prediction, and a second threshold of 50% may be considered an action threshold at which clinical staff should take action in response to the prediction. The warning may include a prior prediction of when the patient will enter a stage of hemodynamic instability based on an estimated likelihood of passing a predetermined threshold.

[0057]

[0059] If the prediction of hemodynamic instability exceeds a threshold (S380=Yes), the method of Figure 3 includes outputting an alarm warning of hemodynamic instability. Additional actions to be taken when the prediction exceeds a threshold may include initiating therapy or preparation for therapy.

[0058]

[0060] Additionally, the method of FIG. 3 can use trained artificial intelligence. The training of the trained artificial intelligence can be based on a large patient population with relevant data, such as patients whose information is stored in the MIMICIII database. The MIMICIII database contains both electronic health records (EHRs) and waveform records for a matched subset of patients. The EHRs include entries for various lab test results, medications, fluid intake, and nurse-recorded vital signs. The waveform records include multiparameter physiological signals, including ECG and ABP. The MIMICIII database also includes vital signs generated by bedside monitors, such as heart rate and systolic blood pressure (SBP), available per minute.

[0059]

[0061] The MIMIC III database contains EHR data from 46,520 unique ICU patients. The MIMIC III database also includes matched physiological waveforms and bedside monitor vital signs for a subset of 10,282 patients. The process of training the AI ​​may include first identifying hemodynamically unstable patients who received strong or weak interventions related to hemodynamic instability. Strong interventions may include the use of vasopressors such as dobutamine, dopamine, epinephrine, norepinephrine, phenylephrine, vasopressin, and isuprel. Weak interventions may include the use of medications such as fluid therapy, packed red blood cells, and lidocaine. In the MIMIC III database, 15,713 patients were identified with one or more strong interventions, and 9,949 patients were identified with one or more weak interventions. Patients with only weak interventions may be excluded from further processing. The remaining 20,858 patients who did not receive any interventions may be identified as stable patients. Filtering can be performed for stable and unstable patients with waveforms, resulting in 4,460 stable patients with waveforms and 3,037 unstable patients with waveforms. Selection can also be limited to unstable patients who had either laboratory, nurse-recorded vital signs, bedside monitor vital signs, ECG, or ABP in the time window from admission to the very first significant intervention. The selection of the second time window can be 6 hours up to 1 hour before the intervention in some embodiments described herein. For stable patients, the 6-hour window can be randomly preselected. Stable and unstable patients with similar levels of care (e.g., at least one nurse-recorded HR and SBP is available within the 6-hour window) can be used as a training baseline, resulting in 880 unstable patients and 2,501 stable patients in a group experiment.

[0060]

[0062] The problem of predicting hemodynamically stable from unstable patients can be approached as a binary classification problem. Gradient-boosted trees (XGBoost) models can be used as boosting models using an ensemble of decision trees. At each iteration, a decision tree can be added to the ensemble of existing trees to correct for errors from the previous iteration, with the goal of minimizing a global loss function. The gradient-boosted trees configuration available in Python can be used. Hyperparameter tuning can be used for a regularization lambda ranging from 10-4 to 1, a learning rate ranging from 10-4 to 0.3, and a maximum depth for each decision tree ranging from 1 to 5. All other settings can be set to default. Stratified five-fold nested cross-validation can be performed for all patients. Hyperparameter tuning can be performed on four training folds, with testing performed on a held-out fifth fold. The average AUROC and AUPRC curves across the five test folds can be reported. The standard errors across the five folds can also be reported.

[0061]

[0063] Results from two experiments are reported. In Experiment I, the performance of adding features at higher temporal resolution is compared to the baseline. The baseline model may utilize only the last entries of clinical measurements and nurse-recorded vital signs (at the end of the second time window). Using the last entries in the 6-hour window after forward input of both clinical measurements and nurse-recorded vital signs, we can make a best effort to minimize the number of missing features per patient. This experiment includes data from all stable and stable patients. This experiment simulates a real-world ICU environment where the hemodynamic instability prediction model leverages what is available in the ICU to improve predictive performance.

[0062]

[0064] The baseline model (using clinical measurements and vital signs) may have an AUROC of 0.89 and an AUPRC of 0.79. The availability of these features is primarily dependent on the caregiver and indirectly represents clinician concerns or other trends. To the baseline model, features derived from nurse-recorded vital signs (NurseDF) and features derived from bedside monitor vital signs (MonitorDF) can be added. Adding the vital sign derived features to the baseline model improves the AUROC by 2%, while there is a more visible change in the AUPRC. A P-value from performing a paired t-test of the AUROC (from five folds) against the baseline model may be reported. The performance improvement from including nurse-recorded vital signs shows greater variance compared to the bedside monitor vital signs. Next, features at higher temporal resolution are added to the baseline model. Features derived from the ABP waveform, ECG waveform are added, and features from both the ABP and ECG waveforms are combined. These high-temporal resolution features can include basic vital signs, along with other ABP- and ECG-specific features. When including both ECG and ABP features, a 4% improvement in AUROC and a 7% improvement in AUPRC can be obtained compared to the baseline model.

[0063]

[0065] In Experiment I, the relative contribution of different feature sets is unknown because their availability across patients is uneven. For example, only one-third of patients have ABP waveforms, but almost all patients have ECG waveforms. To compare the relative performance of different feature sets, Experiment II is performed, which includes only patients who had data from all five feature sets within the time window, and each feature set is evaluated separately. This resulted in only 243 unstable patients and 230 stable patients. All other details in Experiment II are the same as in Experiment I.

[0064]

[0066] Using the last entry of clinical measurements and nurse-recorded vital signs yields an AUROC of 0.82, a 7% decrease compared to Experiment I, which used a much larger number of patients. Using nurse-recorded vital signs and monitored vital signs does not significantly improve the AUROC. Finally, evaluating waveform features yields an AUROC of 0.84 for ABP and 0.85 for ECG + ABP. Using ECG features alone reduces performance, but it is not statistically significant (P = 0.36). A slight improvement in performance can be obtained when using ABP alone versus in combination with ECG. This suggests that the prediction model can exploit information available in ABP but not ECG (e.g., cardiac output or contractility). Furthermore, none of the performances differ significantly from the baseline model. The AUPRC follows a very similar trend, but is superior to the AUROC due to the presence of many more positives (243 unstable patients) compared to negative examples. The results of this experiment show that even in the absence of clinical measurements and nurse-recorded vital signs, objective measures such as vital signs and waveform features from a bedside monitor convey relevant information about hemodynamic instability.

[0065]

[0067] As a final analysis, the clinical relevance of these features is examined to determine the degree to which these waveform features represent the underlying systems (i.e., cardiac function, autonomic tone, and vasomotor tone). Feature importance is examined by ranking them using the XGBoost model from Experiment II across all five CV folds. To examine feature importance, a small cohort is selected because all patients have all five feature groups and the proportion of stable versus unstable patients is roughly equal (230 vs. 243). However, within each feature group, there may be an imbalance in feature availability that affects importance. To correct for this, the importance of each feature is subtracted from the importance of its missing pattern. A data matrix is ​​first created by setting features to 1 if present and 0 otherwise, which simply encodes missing patterns. This binary data matrix is ​​then used in the predictive model to calculate feature importance for missing patterns. Following this, a corrected feature importance score is calculated, ranging from "1.0" (if the feature is highly ranked and its missing pattern has little or no impact), to "0.0" (if the importance of the feature is completely overwhelmed by the importance of its missing pattern), to "-1.0" (if the feature is not important, but its missing pattern conveys information about stable versus unstable patients).

[0066]

[0068] Clinical measurements / vital sign features include SBP, MAP, SI, hematocrit, hemoglobin, PaCO2, FiO2, base excess, and mean airway pressure, many of which are clinically relevant in hemodynamic monitoring. When using only the ABP waveform, systole, diastole, shock index, CO2 (Zander), dicrotic notch amplitude, ABP pulsatile cycle length, and approximate entropy of the minimum of MAP, as well as the area under the systole, can also be used. When using only the ECG waveform, median LF, ULF power; minimum LF and VLF power; approximate entropy (capturing the regularity of the heart rate signal), and Porta's index can also be used. The LF component is generally accepted as an indirect measure of both sympathetic and vagal nervous system activity, and heart rate has been shown to be influenced by both components. There are significant differences in frequency-domain HRV measures, particularly VLF, LF, and HF, between stable and unstable patients during hemodialysis-induced hypotension.

[0067]

[0069] FIG. 4 illustrates a computer system for waveform-based hemodynamic instability warning, according to a representative embodiment.

[0068]

[0070] FIG. 4 illustrates a computer system on which a method for waveform-based hemodynamic instability warning is implemented, according to another representative embodiment.

[0069]

[0071] The computer system 400 of Figure 4 illustrates a complete set of components for a communications or computing device. However, the "controller" described herein may be implemented with less than the set of components of Figure 4, such as, for example, a combination of memory and a processor. The computer system 400 may include some or all of the elements of one or more components of the systems for waveform-based hemodynamic instability warning described herein. However, any such device need not necessarily include one or more of the elements described for the computer system 400, and may include other elements not described.

[0070]

[0072] 4, computer system 400 includes a set of software instructions that can be executed to cause computer system 400 to perform some or all of the functions of any of the computer-based functions disclosed herein. Computer system 400 can operate as a stand-alone device or can be connected to other computer systems or peripheral devices, for example, using network 401. In an embodiment, computer system 400 performs logical operations based on digital signals received via an analog-to-digital converter.

[0071]

[0073] In a networked configuration, computer system 400 operates in the capacity of a server, as a client user computer in a server-client user network environment, or as a peer computer system in a peer-to-peer (or distributed) network environment. Computer system 400 may be implemented as or incorporated into a variety of devices, such as workstation 140, artificial intelligence controller 180 and / or artificial intelligence training system 190 of FIG. 1, a fixed computer, a mobile computer, a personal computer (PC), a laptop computer, a tablet computer, or other machine capable of executing a series of software instructions (sequential or otherwise) that specify actions to be taken by the machine. Computer system 400 may be incorporated as or incorporated within a device that forms an integrated system including additional devices. In one embodiment, computer system 400 may be implemented using electronic devices that provide voice, video, or data communications. Furthermore, although computer system 400 is shown as a single entity, the term "system" is intended to include any collection of systems or subsystems that individually or cooperatively execute a group or groups of software instructions to perform one or more computer functions.

[0072]

[0074] As shown in FIG. 4, computer system 400 includes processor 410. Processor 410 may be considered a representative example of processor 152 of controller 150 of FIG. 1, processor 182 of artificial intelligence controller 180 of FIG. 1, and / or processor 192 of artificial intelligence training system 190 of FIG. 1. Processor 410 executes instructions for implementing some or all aspects of the methods and processes described herein. Processor 410 is tangible and non-transitory. As used herein, the term "non-transitory" should be interpreted as a characteristic of a state that will last for a period of time, rather than a permanent characteristic of a state. The term "non-transitory" specifically negates transitory characteristics, such as characteristics of a carrier wave or signal, or other forms that exist only temporarily at any place and at any time. Processor 410 is an article of manufacture and / or a machine part. Processor 410 is configured to execute software instructions to perform the functions described in various embodiments herein. Processor 410 may be a general-purpose processor or part of an application-specific integrated circuit (ASIC). The processor 410 may be a microprocessor, microcomputer, processor chip, controller, microcontroller, digital signal processor (DSP), state machine, or programmable logic device. The processor 410 may also be logic circuitry, including a programmable gate array (PGA), such as a field programmable gate array (FPGA), or other types of circuitry including discrete gate and / or transistor logic. The processor 410 may be a central processing unit (CPU), a graphics processing unit (GPU), or both. Furthermore, any processor described herein may include multiple processors, parallel processors, or both. Multiple processors may be included in or combined with a single device or multiple devices.

[0073]

[0075] As used herein, the term "processor" encompasses an electronic component capable of executing a program or machine-executable instructions. References to a computing device having a "processor" should be interpreted to include two or more processors or processing cores, such as in a multi-core processor. A processor may also refer to a collection of processors within a single computer system or distributed across multiple computer systems. The term computing device should also be interpreted to include a collection or network of computing devices, each containing one or more processors. A program comprises software instructions that are executed by one or more processors, which may be within the same computing device or distributed across multiple computing devices.

[0074]

[0076] Computer system 400 further includes main memory 420 and static memory 430, with the memories in computer system 400 communicating with each other and with processor 410 via bus 408. Either or both of main memory 420 and static memory 430 may be considered representative examples of memory 151 of controller 150 of FIG. 1, memory 181 of artificial intelligence controller 180 of FIG. 1, and / or memory 191 of artificial intelligence training system 190 of FIG. 1. Main memory 420 and static memory 430 may store instructions used to implement some or all aspects of the methods and processes described herein. Memory, as described herein, is a tangible, non-transitory, computer-readable storage medium for storing data and executable software instructions and is non-transitory while the software instructions are stored therein. As used herein, the term "non-transitory" should be interpreted as a characteristic of a state that will last for a period of time, rather than as a permanent characteristic of a state. The term "non-transitory" specifically denies transitory characteristics, such as characteristics of a carrier wave or signal, or other forms of existence that are only temporary at any time or place. Main memory 420 and static memory 430 are articles of manufacture and / or machine components. Main memory 420 and static memory 430 are computer-readable media from which data and executable software instructions can be read by a computer (e.g., processor 410). Each of main memory 420 and static memory 430 may be implemented as one or more of random access memory (RAM), read-only memory (ROM), flash memory, electrically programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, a hard disk, a removable disk, tape, a compact disk-read-only memory (CD-ROM), a digital versatile disk (DVD), a floppy disk, a Blu-ray disk, or some other form of storage medium known in the art. These memories may be volatile or non-volatile, secured and / or encrypted, unsecured and / or unencrypted.

[0075]

[0077] A memory is an example of a computer-readable storage medium. Computer memory is any memory that can be directly accessed by a processor. Examples of computer memory include, but are not limited to, RAM memory, registers, and register files. References to "computer memory" or "memory" should be interpreted as possibly including multiple memories. A memory may be multiple memories within the same computer system, for example. A memory may also be multiple memories distributed among multiple computer systems or computing devices.

[0076]

[0078] As shown, computer system 400 further includes a video display device, such as a liquid crystal display (LCD), organic light emitting diode (OLED), flat panel display, solid state display, or cathode ray tube (CRT). Additionally, computer system 400 includes input device 460, such as a keyboard / virtual keyboard, a touch-sensitive input screen, or voice input with voice recognition, and cursor control device 470, such as a mouse or touch-sensitive input screen or pad. Computer system 400 also optionally includes a disk drive unit 480, a signal generating device 490, such as a speaker or remote control, and / or a network interface device 440.

[0077]

[0079] 4, the disk drive unit 480 includes a computer-readable medium 482 having one or more sets of software instructions 484 (software) stored thereon. The set of software instructions 484 are read from the computer-readable medium 482 and executed by the processor 410. Furthermore, the software instructions 484, when executed by the processor 410, perform one or more steps of the methods and processes described herein. In one embodiment, the software instructions 484 reside, in whole or in part, within the main memory 420, the static memory 430, and / or the processor 410 during execution by the computer system 400. Furthermore, the computer-readable medium 482 may include or receive and execute the software instructions 484 in response to propagated signals, causing devices connected to the network 401 to communicate voice, video, or data over the network 401. The software instructions 484 may be transmitted or received over the network 401 via the network interface device 440.

[0078]

[0080] In one embodiment, dedicated hardware implementations, such as application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), programmable logic arrays, and other hardware components, are constructed to perform one or more of the methods described herein. One or more embodiments described herein may perform functions using two or more specific interconnected hardware modules or devices, along with associated control and data signals that can communicate between and through those modules. Thus, this disclosure encompasses software, firmware, and hardware implementations. Nothing in this application should be interpreted as being implemented or capable of being implemented solely in software rather than in hardware, such as a tangible, non-transitory processor and / or memory.

[0079]

[0081] According to various embodiments of the present disclosure, the methods described herein can be implemented using a hardware computer system executing a software program. Further, in exemplary, non-limiting embodiments, implementations can include distributed processing, component / object distributed processing, and parallel processing. A virtual computer system process can perform one or more of the methods or functions described herein, and a processor described herein can be used to support a virtual processing environment.

[0080]

[0082] Experiments described herein demonstrated that adding ABP and ECG waveforms to clinical measurements and nurse-recorded vital signs achieved an AUROC of 0.93, while using the ABP and ECG waveforms separately achieved an AUROC of 0.85. Thus, waveform-based hemodynamic instability alerts help ensure adequate tissue perfusion and early detection of inadequate perfusion and end-organ dysfunction. ECG and / or ABP waveforms can be used alone or together with clinical tests and nurse-recorded vital signs to predict hemodynamically unstable patients. The end-to-end waveform processing pipeline described herein, including beat segmentation, signal quality assessment, and feature extraction, can lead to useful predictions of hemodynamic instability. Predicting patients likely to be hemodynamically unstable allows for early intervention, which generally leads to better patient outcomes. Nevertheless, waveform-based hemodynamic instability alerts are not limited in their application to the specific details described herein and are instead applicable to other embodiments in which other details are feasible.

[0081]

[0083] While the waveform-based hemodynamic instability warning has been described above with reference to several exemplary embodiments, it is understood that the terms used are terms of illustration and description, rather than terms of limitation. Changes may be made within the scope of the appended claims, as presently written and as amended, without departing from the scope and spirit of the waveform-based hemodynamic instability warning aspects. While the waveform-based hemodynamic instability warning has been described with reference to particular means, materials, and embodiments, the waveform-based hemodynamic instability warning is not intended to be limited to the details disclosed. Rather, the waveform-based hemodynamic instability warning is intended to cover all functionally equivalent structures, methods, and uses, as fall within the scope of the appended claims.

[0082]

[0084] The illustrations of the embodiments described herein are intended to provide a general understanding of the configuration of various embodiments. These illustrations are not intended to be a complete description of all of the elements and features of the disclosure described herein. Many other embodiments may become apparent to those skilled in the art upon review of the disclosure. Other embodiment illustrations may be utilized and derived from the disclosure, such that structural and logical substitutions and changes may be made without departing from the scope of the disclosure. Furthermore, the illustrations are merely representational and may not be drawn to scale. Certain proportions within the illustrations may be exaggerated, while other proportions may be minimized. Therefore, the disclosure and the figures should be considered illustrative, not limiting.

[0083]

[0085] One or more embodiments of the present disclosure may be referred to herein individually and / or collectively by the term "invention" for convenience only, without any intention to intentionally limit the scope of the present application to any particular invention or inventive concept. Moreover, although specific embodiments have been illustrated and described herein, it should be understood that any subsequent configurations designed to achieve the same or similar purpose may be substituted for the specific embodiment shown. The present disclosure is intended to cover any and all subsequent adaptations or modifications of the various embodiments. Combinations of the above embodiments and other embodiments not specifically described herein will be apparent to those of skill in the art upon reviewing the description.

[0084]

[0086] This Abstract of the Disclosure is provided for purposes of compliance with 37 C.F.R. and is submitted with the understanding that it will not be used to interpret or limit the scope or meaning of the claims. Moreover, in the foregoing Detailed Description, various features may be grouped together or described in a single embodiment for the purpose of streamlining the disclosure. This disclosure should not be interpreted as reflecting an intention that the claimed embodiments require more features than are expressly recited in each claim. Rather, as the following claims reflect, inventive subject matter may be directed to less than all of the features of any of the disclosed embodiments. Accordingly, the following claims are hereby incorporated into the Detailed Description, with each claim standing on its own as defining separately claimed subject matter.

[0085]

[0087] The foregoing description of the disclosed embodiments is provided to enable those skilled in the art to practice the concepts described in the present disclosure. Accordingly, the subject matter disclosed above should be considered illustrative and not limiting, and the appended claims are intended to cover all such modifications, enhancements, and other embodiments that fall within the true spirit and scope of the present disclosure. Accordingly, to the maximum extent permitted by law, the scope of the present disclosure should be determined by the broadest permissible interpretation of the following claims and their equivalents, and should not be limited or restricted by the foregoing detailed description.

Claims

1. a first interface for interfacing with at least one electrocardiogram monitor for monitoring a patient; a memory for storing instructions; and a processor that executes said instructions wherein the instructions, when executed by the processor, cause the device to: identifying a plurality of heart beats from an electrocardiogram waveform received from the at least one electrocardiogram monitor via the first interface; Separating the plurality of heart beats into a plurality of first time windows; extracting a feature of the heartbeat in each of the first time windows as a first extracted feature for each first time window; generating time series features over a second time window including a plurality of the first time windows based on the first extracted features; applying trained artificial intelligence to the time series features; predicting the likelihood that hemodynamic instability in the patient will begin or become detectable after a predetermined time or period based on applying the trained artificial intelligence to the time series features; and outputting an alarm warning of hemodynamic instability based on the prediction of the possibility; Device.

2. a second interface for interfacing with an arterial blood pressure monitor monitoring the patient, wherein the instructions when executed by the processor further cause the device to: dividing an arterial blood pressure waveform received from the arterial blood pressure monitor via the second interface into individual pulsatile periods, wherein each of the first time windows includes a plurality of the individual pulsatile periods; and extracting features of the arterial blood pressure waveform in each of the first time windows as second extracted features for each first time window, wherein the time series features over the second time windows are further based on the second extracted features.

10. The apparatus of claim 1.

3. 2. The device of claim 1, wherein the alert includes predicting in advance when the patient will enter a phase of hemodynamic instability based on estimating when the likelihood of the patient's hemodynamic instability onset or becoming detectable is expected to exceed a predetermined threshold.

4. 2. The device of claim 1, wherein the instructions, when executed by the processor, further cause the device to sequentially identify the plurality of heart beats from the electrocardiogram waveform for a plurality of the first time windows within the second time window.

5. 10. The device of claim 1, wherein the device comprises a monitor, the device further comprising a display for displaying the alert.

6. 1. A method of operating an apparatus including a processor and a first interface for interfacing with at least one electrocardiogram monitor for monitoring a patient, the method comprising: receiving, by the processor, a plurality of electrocardiogram waveforms via the first interface; the processor identifying a plurality of heart beats from the electrocardiogram waveform; the processor separating the plurality of heart beats into a plurality of first time windows; extracting, by the processor, a feature of the heartbeat in each of the first time windows as a first extracted feature for each first time window; generating, by the processor, time series features based on the first extracted features over a second time window comprising a plurality of the first time windows; the processor applying trained artificial intelligence to the time series features; predicting, based on the processor applying the trained artificial intelligence to the time series features, the likelihood that hemodynamic instability in the patient will begin or become detectable after a predetermined time or period; and outputting an alarm warning of the hemodynamic instability based on the processor predicting the likelihood. A method of operating a device comprising:

7. the device includes a second interface for interfacing with an arterial blood pressure monitor that monitors the patient; receiving, by the processor, an arterial blood pressure waveform via the second interface; the processor dividing the arterial blood pressure waveform into individual pulse periods, the first time windows each including a plurality of the individual pulse periods; and the processor extracting features of the arterial blood pressure waveform in each of the first time windows as second extracted features for each first time window, wherein the time series features over the second time windows are further based on the second extracted features.

7. The method of claim 6, further comprising:

8. the processor labeling each of the plurality of heartbeats with one of a plurality of predetermined labels based on characteristics of each of the plurality of heartbeats; and the processor excluding at least one heartbeat of the plurality of heartbeats from application of the trained artificial intelligence based on applying a noise filter to the plurality of heartbeats.

7. The method of claim 6, further comprising:

9. the processor identifying at least one individual heartbeat cycle as abnormal; and and excluding, based on the processor identifying the at least one individual heartbeat cycle as abnormal, the at least one individual heartbeat cycle identified as abnormal from application of the trained artificial intelligence.

8. The method of claim 7, further comprising:

10. 7. The method of claim 6, wherein extracting the heart rate features in each time window comprises extracting heart rate variability as the first extracted feature for each time window.

11. A tangible, non-transitory computer-readable storage medium storing a computer program, the computer program being, when executed by a processor, capable of executing a program in a system including the tangible, non-transitory computer-readable storage medium, identifying a plurality of heart beats from an electrocardiogram waveform received via the first interface from at least one electrocardiogram monitor; Separating the plurality of heart beats into a plurality of first time windows; extracting a feature of the heartbeat in each of the first time windows as a first extracted feature for each first time window; generating time series features over a second time window including a plurality of first time windows based on the first extracted features; applying trained artificial intelligence to the time series features; predicting the likelihood that hemodynamic instability in the patient will begin or become detectable after a predetermined time or period based on applying the trained artificial intelligence to the time series features; and outputting an alarm warning of hemodynamic instability based on the prediction of the possibility; A tangible, non-transitory computer-readable storage medium.

12. The computer program, when executed by the processor, further provides the system including the tangible non-transitory computer-readable storage medium with: dividing an arterial blood pressure waveform received from an arterial blood pressure monitor via a second interface into individual pulsatile periods, wherein each of the first time windows includes a plurality of the individual pulsatile periods; and extracting features of the arterial blood pressure waveform in each of the first time windows as second extracted features for each first time window, wherein the time series features over the second time windows are further based on the second extracted features.

12. The tangible, non-transitory computer-readable storage medium of claim 11.

13. The computer program, when executed by the processor, further provides the system including the tangible non-transitory computer-readable storage medium with: labeling each of the plurality of heartbeats with one of a plurality of predetermined labels based on characteristics of each of the plurality of heartbeats; and excluding at least one heartbeat of the plurality of heartbeats from application of the trained artificial intelligence based on applying a noise filter to the plurality of heartbeats; 12. The tangible, non-transitory computer-readable storage medium of claim 11.

14. The computer program, when executed by the processor, further provides the system including the tangible non-transitory computer-readable storage medium with: Identifying at least one individual heartbeat cycle as abnormal; and and based on identifying the at least one individual heartbeat cycle as abnormal, excluding the at least one individual heartbeat cycle identified as abnormal from application of the trained artificial intelligence.

12. The tangible, non-transitory computer-readable storage medium of claim 11.

15. 12. The tangible, non-transitory computer-readable storage medium of claim 11, wherein the tangible, non-transitory computer-readable storage medium is provided as a component of a monitor that displays the warning.

Images