System and method for automated detection of clinical deterioration events - Patents.com

Abstract

The present disclosure relates to a clinical support system and related method for automatic real-time detection of clinical deterioration events in patients. Existing clinical support systems typically rely on simple threshold-based alarm generation. This often results in too many alarms, including false alarms, and alarm fatigue or ignored alarms for medical personnel. The disclosed system and method provide an improved alternative to existing clinical support systems since it incorporates a clinically validated computer-implemented subroutine that provides higher predictive value to medical personnel. The subroutine utilizes clinically evaluated thresholds and durations to reduce false alarms while maintaining the most relevant alarms, i.e., alarms that require clinical action. The present disclosure further relates to a computer program configured to provide automatic real-time detection of clinical deterioration events in patients by performing the disclosed method.

Description

[Technical field]

[0001] The present disclosure relates to a computer-implemented method configured for automatic real-time detection of clinical deterioration events in a patient. The present disclosure further relates to a system for performing the disclosed method. [Background technology]

[0002] Practices rely on manual recording of 12-hourly physiological patient data and manual calculation of an Early Warning Score (EWS) or similar "risk score" to detect patients in need. Selected physiological patient data (blood pressure, respiratory rate, heart rate / pulse rate, and temperature) are traditionally described as vital signs. However, the term vital signs is based on biomarkers for bodily conditions that have historically been measurable (starting with pulse rate and temperature centuries ago). Peripheral monitoring of blood pressure and oxygen saturation has become reliable and commercially available, so this is also commonly included as a vital sign. Some hospitals also include parameters such as mental status, pain, urine output, blood glucose, and end-tidal CO2 as vital signs and in the EWS.

[0003] The problem with the EWS approach and similar “track-and-trigger” systems is that some serious events occur between routine measurements undetected. Although EWS has a mandatory requirement that registrations be more frequent depending on the severity of the most recent measurements, EWS has not been found to have any effect on complications or survival, despite intensive resource allocation. For example, manual routine EWS measurements have been shown to detect only 5% of serious cases of severe desaturation when compared with continuous 24 / 7 automated measurements. In layman’s terms, the main causes of low and slow detection rates, despite large resource allocations, are the manual and infrequent collection of patient data, which is insufficient to detect clinical deterioration before the event. Manual clinical interpretation of individual parameter thresholds is not sensitive or sophisticated enough to capture events in a timely manner. Trends and combinations of vital signs physiological parameters are too complex to be assessed manually and are susceptible to the experience or attention of the individual assessor (e.g., nurse, physician). It is now possible to measure a variety of other biomarkers of physical condition, including advanced analysis of heart rate and rhythm and peripheral perfusion to further describe circulatory function. It is therefore highly appropriate that new biomarkers of physical condition are added to the traditional list of vital signs introduced above. In the following sections, the term "vital signs" refers to this open-ended list of biomarkers for assessing the physical condition of a patient.

[0004] In general, clinical decision support systems (CDSS) are focused on using knowledge management in such a way that they achieve clinical advice for patient care based on multiple claims of patient data. A problem encountered within many clinical support systems is that alarm generation is based on simple threshold alerts, leading to exhaustion of healthcare personnel (nurses, doctors, etc.) by a high volume of alarms, many of which are false alarms. This phenomenon, also known as alarm fatigue, is a well-known and highly recognized challenge associated with patient monitoring, where alarms are typically either muted, thresholds are ignored, or simply go unnoticed.

[0005] Thus, what is needed is a system and method that provides an alternative to simple threshold monitoring for predicting adverse patient events. In particular, what is needed is a clinical decision support system where alarm generation is based on intelligent algorithms such that true alarms are maximized and false alarms are reduced, while still providing clinically actionable vital signs alarms. Summary of the Invention

[0006] Currently, postoperative patient monitoring relies on a simple model of intermittent bedside monitoring in hospitals and early warning scores (EWS). Systems are needed to improve patient monitoring by facilitating continuous and predictive monitoring.

[0007] The present disclosure addresses the above-mentioned problems by providing a system and method for automatic and continuous detection of clinical deterioration events in a patient. An advantage of the disclosed system and method is that it provides continuous (ideally 24 / 7) real-time monitoring of a patient and an alarm is generated if the patient has a deterioration event. The method includes executing one or more computer-implemented clinically validated automatic deterioration event subroutines, each subroutine configured to determine a specific clinical deterioration event in the patient.

[0008] In particular, the present disclosure relates to a computer-implemented method configured for automated real-time detection of clinical deterioration events in a patient, the method comprising: receiving continuously a plurality of different vital sign data from a plurality of sensors worn by the patient, the vital sign data being selected from the group of electrocardiogram (ECG), photoplethysmogram (PPG), heart rate (HR), respiration rate (RR), blood pressure (e.g., systolic blood pressure, SBP), cardiac rhythm, ischemic electrocardiogram response, peripheral body temperature, peripheral skin conductance, 3D body position and acceleration, pulse rate, peripheral perfusion index, peripheral oxygen saturation (SpO2), and subcutaneous glucose concentration, and optionally / alternatively, vital sign data such as blood pressure can be estimated from other measured vital sign data, e.g., heart rate (HR), respiration rate (RR), blood oxygen saturation (SpO2), and pulse rate (PR), as disclosed herein, if these vital sign data are easier to measure than blood pressure; analyzing the vital signs data to identify artifacts; discarding one or more data samples associated with the identified artifact in the vital signs data to continuously acquire valid patient vital signs parameters; executing one or more computer implemented clinically significant deterioration event subroutines, each subroutine configured to receive one or more of the significant vital sign parameters and determine a specific clinical deterioration event in the patient, the clinical deterioration event being determined by: Bradypnea / apnea based on available heart rate and respiratory rate parameters; Tachypnea based on valid respiratory rate; Hypoventilation based on effective respiratory rate and percutaneous arterial oxygen saturation; Desaturation based on valid percutaneous arterial oxygen saturation; Sinus tachycardia based on effective heart rate, Bradycardia based on effective heart rate, Hypotension based on valid or estimated systolic blood pressure; Circulatory collapse based on valid heart rate and valid or estimated systolic blood pressure; Cardiac arrest based on valid cardiac rhythm and pulse rate; Hypertension based on valid or estimated systolic blood pressure; Atrial fibrillation based on a valid cardiac rhythm; Premature ventricular contractions based on a valid cardiac rhythm, and Ventricular tachycardia / ventricular fibrillation based on a valid cardiac rhythm; performing a process selected from the group consisting of providing an alarm if at least one of the deterioration events is detected by one of the clinically enabled automatic deterioration event subroutines; This includes the steps:

[0009] Optionally and / or alternatively, one or more computer-implemented clinically valid deterioration events subroutines can be executed based on predicted vital signs, for example as shown in Example 5, where the vital sign parameters heart rate and respiration rate are predicted based on modeling of the valid vital sign parameters HR and RR, for example machine learning, in particular a multivariate autoregressive (MAR) model. Any vital sign parameter can be predicted based on this approach, i.e., predicted for at least 5, 10, 15, 30, 45 or even 60 minutes. Each subroutine based on predicted vital signs can then be configured to receive one or more of the predicted vital sign parameters and determine a specific predicted clinical deterioration event in the patient, where the clinical deterioration event can be selected from the group of clinical deterioration events listed above. In this way, an alarm can be generated based on the predicted data, i.e., before a deterioration event actually occurs and / or predicting whether a deterioration event is likely to occur in the near future.

[0010] Preferably, all vital signs are received continuously. Data may be transmitted at different sampling frequencies and may be transmitted in blocks (sampling). Also, as described in Example 3 herein, blood pressure (both systolic and diastolic) may be estimated with other measured vital signs, such as HR, RR, SpO2, and PR, so blood pressure may be measured cuffless or non-invasively and may be based on valid vital signs data, and therefore may be validly substituted for a valid systolic blood pressure measurement used in the subroutine as described herein.

[0011] Each clinically useful deterioration event subroutine is associated with one or more criteria that determine whether an alarm needs to be generated. The criteria include thresholds and duration(s), which are clinically determined and evaluated by a physician, thus reducing the amount of alarms generated and generating more critical alarms. Thus, the disclosed system and method provides much more predictive value than existing systems, as it provides alarms of events that require clinical action from medical personnel while at the same time significantly reducing the amount of false alarms. This is accomplished by designing multiple deterioration event subroutines, also referred to as predictive computer algorithms.

[0012] The present disclosure further relates to a system for automated detection of clinical deterioration events in a patient, the system comprising: one or more sensors configured to automatically monitor a patient's vital signs data (e.g., heart rate, respiration rate, heart rate variability, body temperature, oxygen saturation, and blood pressure), the one or more sensors further configured to wirelessly transmit the vital signs data to a server and / or gateway; a first server for receiving and storing vital signs data, the first server having a computer program thereon, the computer program being configured to execute a method of the present disclosure to provide automated detection of clinical deterioration events in a patient; and one or more gateways configured to provide wireless communication links between the sensor(s) and the first server.

[0013] The system described herein is configured to provide automated detection of clinical deterioration events in a patient by performing the methods of the present disclosure. The system of the present disclosure and its functionality are illustrated in Figures 1-3 and are further described in the detailed description of the present invention.

[0014] The present disclosure relates to a computer program having instructions which, when executed by a computing device or system, cause the computing device or system to perform the methods disclosed herein to provide automated real-time detection of clinical deterioration events in a patient.

[0015] The disclosed system and method thus provides continuous 24 / 7 monitoring of a patient, where clinical deterioration events in the patient are automatically detected and reported through intelligent alarm generation. Specifically, this is accomplished by executing a number of deterioration event subroutines that receive input from one or more sensors associated with the patient, and which are configured to provide an alarm in the event of a clinical deterioration event in the patient.

[0016] Thus, the systems and methods of the present disclosure provide a significant improvement over existing clinical support systems, which typically rely heavily on simple thresholds for generating alarms.

[0017] The present disclosure further relates to a system for identifying unauthorized access of an online service account, comprising a non-transitive computer-readable storage device for storing instructions, which when executed by a processor, perform a method for identifying unauthorized access of an online service account according to the described method. The system may comprise a mobile device having a processor and a memory and adapted to perform the method, but can also be a stationary system, or a system operating in a central location, and / or a remote system, for example with cloud computing. The present invention further relates to a computer program having instructions, which when executed by a computing device or system, cause the computing device or system to identify unauthorized access of an online service account according to the described method. A computer program in this context shall be interpreted broadly and shall include, for example, a program running on a PC, or software designed to run on a smartphone, tablet computer, or other mobile device. Computer programs and mobile applications include free software and software requiring purchase, and also software distributed via distribution software platforms such as the Apple App Store, Google Play, and Windows Phone Store. [Brief description of the drawings]

[0018] [Figure 1] 1 illustrates an embodiment of a system for automatic detection of clinical deterioration events in a patient according to the present disclosure. [Diagram 2] FIG. 1 shows a block diagram illustrating the overall functionality of the systems and methods as disclosed herein. [Diagram 3] 1 illustrates another representation of an embodiment of a system according to the present disclosure. [Figure 4] 13 illustrates an embodiment of an ECG pre-processing subroutine in accordance with the present disclosure. [Diagram 5]13 illustrates an embodiment of an SpO2 pre-processing subroutine configured to assess the quality of SpO2 values ​​from a patient-worn pulse oximeter. [Figure 6] 13 shows a block diagram of a slow breathing subroutine, according to one embodiment. [Figure 7] 13 shows a block diagram of a tachypnea subroutine, according to one embodiment. [Figure 8] 13 shows a block diagram of a hypoventilation subroutine, according to one embodiment. [Figure 9] 13 shows a block diagram of a desaturation subroutine according to one embodiment. [Figure 10] 13 shows a block diagram of a desaturation subroutine according to one embodiment. [Figure 11] 13 shows a block diagram of a desaturation subroutine according to one embodiment. [Figure 12] 13 shows a block diagram of a desaturation subroutine according to one embodiment. [Figure 13] 13 shows a block diagram of a sinus tachycardia subroutine, according to one embodiment. [Figure 14] 13 shows a block diagram of a sinus tachycardia subroutine, according to one embodiment. [Figure 15] 13 shows a block diagram of a bradycardia subroutine, according to one embodiment. [Figure 16] 13 shows a block diagram of a bradycardia subroutine, according to one embodiment. [Figure 17] 13 shows a block diagram of a low blood pressure subroutine, according to one embodiment. [Figure 18] 13 shows a block diagram of a hypertension subroutine, according to one embodiment. [Figure 19] 13 shows a block diagram of a low blood pressure subroutine, according to one embodiment. [Figure 20] 13 shows a block diagram of a hypertension subroutine, according to one embodiment. [Figure 21] 13 shows a block diagram of a circulatory collapse subroutine, according to one embodiment. [Figure 22] 13 shows a block diagram of an atrial fibrillation subroutine, according to one embodiment. [Diagram 23] FIG. 1 shows a block diagram illustrating the overall functionality of the disclosed systems and methods. [Figure 24] A block diagram illustrating how electrocardiogram (ECG) and photoplethysmogram (PPG) data are handled is shown. [Diagram 25] FIG. 1 shows a diagram of the deep generative model (DGM) used in Example 1. [Figure 26] 1 shows a diagram of (a) an inference model and (b) a generative model of the proposed network used in Example 1. [Figure 28] 1 shows selected sample input segments and corresponding reconstructions used in Example 1. [Figure 29] 1 shows the distribution of samples for the test set in the latent space used in Example 1. [Diagram 30] 1 depicts the steps of the algorithm used in Example 2. [Figure 31A] 13 shows the relationship between the prediction window and the overlap window used in Example 2. [Figure 31B] 1 shows an extraction of a control sample used in Example 2. [Diagram 32] 1 shows the patient enrollment process applied in Example 4. [Diagram 33] Visualization of the overnight extraction process. The patient has an SAE on day 2 (red line), so the night before that (in green) is selected. The other nights (in grey) are discarded. [Diagram 34] Visualization of the overnight extraction process. All nights are included (in blue) to represent patients without SAEs. [Diagram 35] 1 shows a time series of patient vital signs used to fit the MAR model used in Example 5. [Diagram 36] 13 shows a probabilistic graphical model of the implemented pooled MAR model used in Example 5. [Figure 37]Figure 1 shows the setup used to evaluate the model in Example 5 on a new patient. At each step, a prediction (right box) is made based on the data available in the model window (left box). The window is then advanced by 10 minutes and the process is repeated. [Figure 38] FIG. 13 shows a visualization of the response of a hierarchical AR model fitted to HR and RR data, as used in Example 5. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0019] The present disclosure relates to a computer-implemented method configured for the automated real-time detection of clinical deterioration events in a patient.

[0020] Vital Signs Data In a preferred embodiment, the first step of the method is to receive a plurality of different vital sign data from a plurality of sensors worn by the patient. The vital sign data may be selected from the group of electrocardiogram (ECG), photoplethysmogram (PPG), heart rate (HR), respiratory rate (RR), blood pressure (e.g., systolic blood pressure, SBP), cardiac rhythm, ischemic electrocardiogram response, peripheral temperature, peripheral skin conductance, 3D body position and acceleration, pulse rate (PR), peripheral perfusion index, peripheral oxygen saturation (SpO2) (e.g., derived from PPG), and subcutaneous glucose concentration. The vital sign data may be received continuously or at a predefined time interval, such as every minute. The vital signs may have a fixed sampling frequency, some of which may have block-wise sampling. As an example, the ECG may be delivered every first 10 seconds per minute, i.e., the next package of 10-second samples is sent 50 seconds later, and so on.

[0021] Artifact removal In a preferred embodiment, the next step of the method is to analyze the vital signs data to identify artifacts in the data. Artifacts and noise are preferably dealt with on a vital signs basis whenever necessary. Artifacts should be understood as erroneous data that display non-physical values ​​arising from external factors that affect the (sensor's) measurement such that the measurement is prevented or changes from its true value. One example is a movement-related artifact, e.g., if a patient suddenly exercises, this may affect some of the measured vital signs data. Another example is that if one of the sensors is misplaced, it may not be possible to measure the intended vital signs. As an example, if a pulse oximeter measures a negative value or SpO2 above 100%, such data points are considered artifacts and are therefore removed. Thus, the next step of the method is preferably to discard one or more data samples (i.e., data point sets) associated with the identified artifacts in the vital signs data in order to obtain valid patient vital signs parameters. Artifacts and noise are estimated by a global approach that looks for anomalous deviations as a function of time, amplitude and frequency content.

[0022] ECG Preprocessing In a preferred embodiment, the method further comprises executing an ECG pre-processing subroutine configured to assess the quality of ECG data from an ECG sensor worn by the patient. Some vital sign data, such as RR interval (RRI), PP interval (PPI), HR, cardiac rhythm and RR, are estimated / calculated based on R-peak detection in ECG data, i.e., valid cardiac rate and valid cardiac rhythm are commonly based on ECG data. RR interval (RRI) and PP interval (PPI) respectively represent beat-to-beat intervals extracted from ECG and PPG signals. R-peak is understood to have its common meaning, i.e., the maximum amplitude of the R-wave in the QRS complex of an electrocardiogram. The system associated with the method of the present disclosure is preferably configured to automatically provide such vital sign data, i.e., to automatically perform R-peak detection in the ECG data. However, when the ECG data is noisy or distorted, the detection of the R-peak(s) may be flawed, leading to erroneous estimation of RRI, HR and RR. Therefore, the system is preferably further configured to stream portions of the ECG data. The purpose of the ECG Pre-Processing subroutine is to evaluate whether the streaming portion of the ECG data (also referred to as ECG samples) is of acceptable quality. The output of the ECG Pre-Processing subroutine are several parameters (goodForHR, goodForAF, goodForRR, goodForMorph) that can take on a value of either 1 or 0. A value of 1 indicates that the ECG sample is good enough to be used to derive vital signs data (HR, RR, and / or RRI) and can be used as an input to the clinically valid deterioration events subroutine. The vital signs data is then referred to as valid vital signs data. Conversely, a value of 0 indicates that the associated ECG sample and / or derived vital signs data should be discarded and need not be applied to the deterioration events subroutine.The parameters goodForHR and goodForRR mean that the associated ECG sample is of sufficiently good quality (for a value of 1) to be used to estimate the patient's heart rate (HR), heart rhythm, and respiration rate (RR), respectively, based on the ECG sample. Similarly, the parameter goodForAF means that the associated ECG sample is of sufficiently good quality to be used in the atrial fibrillation (AF) subroutine. The parameter goodForMorph indicates that the ECG sample is sufficiently good (with respect to noise, e.g., quantified by signal-to-noise ratio) to calculate other values ​​from the ECG morphology. In one embodiment, the ECG pre-processing subroutine: receiving ECG values ​​from an ECG sensor, the ECG values ​​including an ECG timestamp, an ECG sample, a heartbeat timestamp, an RR interval, and / or a QRS amplitude; Discarding ECG values ​​that are outside of a predefined threshold, such as ECG<0 or ECG>4000 (numbers representing 12-bit resolution, i.e., realistic cardiac values), and / or discarding ECG values ​​that are overlapped; performing linear interpolation of missing ECG values; Using the QRS timestamp to determine the QRS index in the ECG (i.e., the number of predominant cardiac peak samples within a given time window, such as a 10 second window); comparing the heart rate and / or the QRS amplitude and / or the ratio between a subselection of the QRS amplitude and the heart rate to one or more predefined thresholds and discarding one or more of the ECG values ​​if the thresholds are exceeded; performing band pass filtering of the ECG values; correcting the R-peak(s) from the ECG values ​​and / or normalizing each heart beat from the ECG values ​​based on heuristic rules; calculating an ecgTemplate (i.e., an average ECG cardiac cycle calculated from the first cardiac cycle in the cardiac signal) or a previously received ecgTemplate; Calculating the correlation coefficient (cc); comparing the average cc and / or number of cc to one or more predefined thresholds and discarding one or more of the ECG values ​​if they exceed these thresholds; calculating the deviation between the corrected R peak(s) and the uncorrected R peak(s) and storing this deviation as HBdev value(s); comparing the HBdev value(s) to one or more predefined thresholds and discarding ECG values ​​associated with the HBdev value(s) that exceed the one or more predefined thresholds; This includes the steps:

[0023] SpO2 Pretreatment In a preferred embodiment, the method further includes executing an SpO2 pre-processing subroutine configured to evaluate the quality of SpO2 data from a pulse oximeter worn by the patient to obtain valid SpO2 data. A system associated with the method of the present disclosure is preferably configured to receive SpO2 data from the pulse oximeter at a given sampling frequency, e.g., 1 Hz. A given time segment (e.g., one minute long) of data including several SpO2 values ​​may be represented by an average SpO2 sample (i.e., a value of 1 representing the oxygen level for a one minute interval), and may be transferred, e.g., every minute, to one or more servers storing computer programs for executing the method of the present disclosure. In one embodiment, the SpO2 pre-processing subroutine: receiving SpO2 data (e.g., SpO2 data samples) from a pulse oximeter worn by the patient; removing SpO2 data or SpO2 data values ​​in the SpO2 data sample that are above one or more predefined thresholds, such as SpO2<0 and / or SpO2>100; Removing duplicate SpO2 data; Removing SpO2 data having a difference of more than 4 per second; Extrapolating missing SpO2 data; Optionally, calculating an average SpO2 value from the SpO2 data; This includes the steps:

[0024] Deteriorating Events Subroutine In a preferred embodiment, the next step of the method is to execute one or more computer-implemented clinically valid deterioration event subroutines. Preferably, each subroutine is configured to receive one or more of the valid vital sign parameters and determine a specific clinical deterioration event in the patient. The clinical deterioration event may be selected from the group of bradypnea / apnea, tachypnea, hypoventilation, desaturation, sinus tachycardia, bradycardia, hypotension, circulatory collapse, hypertension, atrial fibrillation, ventricular extrasystole, ventricular tachycardia / fibrillation (VT / VF), cardiac arrest, cardiac ischemia, hypoperfusion index, and acute stress. Each deterioration event subroutine is described in further detail below.

[0025] Slow breathing subroutine The method of the present disclosure may include a bradypnea subroutine configured to determine bradypnea / apnea. According to one embodiment, the bradypnea subroutine comprises: receiving valid HR and RR values; comparing the valid HR and RR values ​​to one or more predetermined bradypnea thresholds; providing an alarm when the HR and RR values ​​exceed a predetermined bradypnea threshold for a predetermined duration; This includes the steps:

[0026] The one or more slow breathing thresholds may be selected from the group of HR>10 bpm, HR>15 bpm, HR>20 bpm, HR>25 bpm, RR≦3 bpm, RR≦5 bpm, RR≦10 bpm, RR≦15 bpm, and / or combinations thereof. In a preferred embodiment, the slow breathing subroutine includes slow breathing thresholds HR>20 and RR≦5. The predetermined duration may be selected from the group of ≧1 minute, ≧2 minutes, ≧3 minutes, ≧5 minutes, or ≧10 minutes. In a preferred embodiment, the slow breathing subroutine provides an alarm if HR>20 and RR≦5 for more than 1 minute.

[0027] Tachypnea Subroutine The methods of the present disclosure may include a tachypnea subroutine configured to determine tachypnea. According to one embodiment, the tachypnea subroutine comprises: Receiving valid RR value(s); comparing the effective RR value(s) with a predetermined tachypnea threshold; Providing an alarm if the RR value(s) exceeds a tachypnea threshold for a predetermined duration.

[0028] The predetermined tachypnea threshold may be selected from the group of RR≧20 bpm, RR≧24 bpm, RR≧28 bpm, and / or combinations thereof. In a preferred embodiment, the tachypnea subroutine includes a tachypnea threshold: RR≧24 bpm. The predetermined duration may be selected from the group of ≧1 minute, ≧2 minutes, ≧3 minutes, ≧5 minutes, ≧10 minutes. A duration of ≧5 minutes is preferred. In a preferred embodiment, the tachypnea subroutine provides an alarm if RR≧24 bpm for more than 5 minutes.

[0029] Hypoventilation Subroutine The method of the present disclosure may include a hypoventilation subroutine configured to determine hypoventilation. According to one embodiment, the hypoventilation subroutine comprises: receiving valid RR and SpO2 value(s); comparing each of the available RR and SpO2 values ​​to one or more hypoventilation thresholds; Providing an alarm if the RR and SpO2 values ​​exceed the hypoventilation threshold(s) for a predetermined duration; This includes the steps:

[0030] The hypoventilation threshold may be selected from the group of RR<15 bpm, RR<13 bpm, RR<11 bpm, RR<9 bpm, SpO2<92%, SpO2<90%, SpO2<88%, SpO2<86%, and / or combinations thereof. In a preferred embodiment, the hypoventilation threshold includes RR<11 bpm and SpO2<88%. The predetermined duration may be selected from the group of ≧1 minute, ≧2 minutes, ≧3 minutes, ≧5 minutes, or ≧10 minutes. A duration of ≧5 minutes is preferred. In a preferred embodiment, the hypoventilation subroutine provides an alarm if RR<11 bpm and SpO2<88% for more than 5 minutes.

[0031] Desaturation Subroutine The method of the present disclosure may include a desaturation subroutine configured to determine the desaturation. According to one embodiment, the desaturation subroutine comprises: receiving a valid SpO2 value; comparing the valid SpO2 value to one or more predetermined SpO2 thresholds; providing an alarm if the SpO2 value exceeds the SpO2 threshold(s) for a predetermined duration t; This includes the steps:

[0032] The predetermined SpO2 thresholds may include SpO2<92%, SpO2<88%, SpO2<85%, SpO2<80%, and / or any combination thereof. The predetermined duration may be selected from the group of: ≧1 minute, ≧5 minutes, ≧10 minutes, ≧30 minutes, or ≧60 minutes. In a preferred embodiment, the desaturation subroutine: SpO2 < 92% for t ≥ 60 min, or SpO2 < 88% for t ≥ 10 min, or SpO2 < 85% for t ≥ 5 min, or SpO2 < 80% for t ≥ 1 min, Provide an alarm if

[0033] sinus tachycardia The methods of the present disclosure may include a sinus tachycardia subroutine configured to determine sinus tachycardia, the subroutine comprising: receiving valid HR value(s); comparing the valid HR value to one or more predetermined sinus tachycardia thresholds; providing an alarm if the HR value exceeds a sinus tachycardia threshold(s) for a predetermined duration t; This includes the steps:

[0034] The one or more predetermined sinus tachycardia thresholds may be selected from the group of HR > 100 bpm, HR > 111 bpm, HR > 120 bpm, or HR > 130 bpm. The predetermined duration may be selected from the group of ≥ 5 minutes, ≥ 10 minutes, ≥ 30 minutes, ≥ 60 minutes, or ≥ 80 minutes. In a preferred embodiment, the sinus tachycardia subroutine provides an alarm if HR > 130 bpm for t > 30 minutes, or if HR > 111 bpm for t > 60 minutes.

[0035] Bradycardia Subroutine The methods of the present disclosure may include a bradycardia subroutine configured to determine bradycardia, the subroutine comprising: receiving valid HR value(s); comparing the effective HR value to one or more predetermined bradycardia thresholds and / or ranges; Providing an alarm if the HR value exceeds the bradycardia threshold(s) and / or range for a predefined duration t; This includes the steps:

[0036] The bradycardia threshold / range may be selected from the group of HR<40 bpm, HR<30 bpm, HR<25 bpm, 25 bpm≦HR≦45 bpm, or 30 bpm≦HR≦40 bpm. The predetermined duration may be selected from the group of ≧1 min, ≧2 min, ≧3 min, ≧5 min, or ≧10 min. In a preferred embodiment, the bradycardia subroutine provides an alarm if HR<30 bpm for t≧1 min, or if 30 bpm≦HR≦40 bpm for t≧5 min. Preferably, an alarm is provided only if the parameter goodForHR is equal to 1.

[0037] Hypotension Subroutine The disclosed method may include a hypotension subroutine configured to determine hypotension, the subroutine comprising: receiving valid SBP value(s); comparing the valid SBP value to one or more predetermined hypotensive thresholds; providing an alarm if the SBP value exceeds the hypotensive threshold(s) during one or more consecutive measurements and / or for a predetermined duration t; This includes the steps:

[0038] The hypotension threshold may be selected from the group of SBP<91 mmHg, SBP<80 mmHg, SBP<70 mmHg, SBP<60 mmHg, and / or combinations thereof. In a preferred embodiment, the hypotension subroutine provides an alarm if SBP<91 mmHg or if SBP<70 mmHg between two consecutive measurements.

[0039] circulatory collapse The methods of the present disclosure may include a circulatory collapse subroutine configured to determine circulatory collapse, the subroutine comprising: receiving valid SBP and HR value(s); comparing the valid SBP and HR values ​​to one or more predetermined SBP and HR thresholds; providing an alarm when at least one of the SBP threshold(s) and the HR threshold is exceeded for a predetermined duration t; This includes the steps:

[0040] The predetermined SBP threshold(s) may be selected from the group of SBP<110mmHg, SBP<100mmHg, and / or SBP<90mmHg. The predetermined HR threshold may be selected from the group of HR>110bpm, HR>120bpm, HR>130bpm, HR<60bpm, HR<50bpm, HR<40bpm, and / or combinations thereof. The predetermined duration may be selected from the group of ≧1 minute, ≧5 minutes, ≧10 minutes, ≧30 minutes, or ≧60 minutes. In a preferred embodiment, the circulatory collapse subroutine: SBP < 100 mmHg and HR > 110 bpm for t ≥ 30 min, or SBP < 100 mmHg and HR > 130 bpm for t ≥ 5 min, or SBP < 100 mmHg and HR < 50 bpm during t ≥ 30 min; Provide an alarm if

[0041] Cardiac Arrest Subroutine The method of the present disclosure may include a cardiac asystole subroutine configured to determine cardiac asystole, the subroutine comprising: receiving ECG data samples; detecting a QRS complex in the ECG data sample; providing an alarm if there is no detected QRS complex for a predetermined duration t1 and / or if there is no detected pulse for a predetermined duration t2; This includes the steps:

[0042] The predetermined durations t1 and t2 may be greater than 10 seconds, or greater than 15 seconds, or greater than 20 seconds, or greater than 25 seconds, or greater than 30 seconds.

[0043] Hypertension Subroutine The disclosed method may include a hypertension subroutine configured to determine hypertension, the subroutine comprising: receiving valid SBP value(s); comparing the valid SBP value to one or more predetermined hypertension thresholds; providing an alarm if the SBP value exceeds a hypertension threshold(s) during one or more consecutive measurements and / or for a predetermined duration t; This includes the steps:

[0044] The hypertension threshold(s) may be selected from the group of SBP≧180 mmHg, SBP≧190 mmHg, SBP≧200 mmHg, SBP≧210 mmHg, SBP≧220 mmHg, and / or combinations thereof. The predetermined duration may be selected from the group of ≧1 min, ≧5 min, ≧10 min, ≧30 min, or ≧60 min. In a preferred embodiment, the hypertension subroutine provides an alarm if SBP≧180 mmHg for t≧60 min or if SBP≧220 mmHg for at least one measurement.

[0045] Atrial Fibrillation Subroutine The methods of the present disclosure may include an atrial fibrillation subroutine configured to determine atrial fibrillation, the subroutine comprising: receiving valid RRI values ​​and / or RRI samples; comparing the effective RRI value with one or more predefined RRI threshold(s); Removing RRI values ​​that exceed a predefined RRI threshold(s); Optionally, comparing the amount of RRI values ​​(e.g. a length of an array storing the RRI values) with a second RRI threshold; Calculating the normalized difference of RRI (NDR) and storing these as NDR values; Optionally, removing NDR values ​​that fall outside a predefined percentile; providing the NDR value to a supervised learning model, e.g., a support vector machine (SVM) model, trained and configured to determine atrial fibrillation (AF) based on the NDR value; Providing an alarm in case of atrial fibrillation; This includes the steps:

[0046] The predefined RRI threshold(s) may be selected from the group of RRI<300, RRI<200, RRI<150, RRI>2500, RRI>3000, or RRI>3500. The second RRI threshold may be that the size of the RRI array storing the RRI values ​​is greater than 15, or greater than 20, or greater than 25, or greater than 30. The atrial fibrillation subroutine preferably includes the step of calculating a normalized difference of the valid RRI values ​​and storing the calculated normalized difference value in the set of stored NDR values. The atrial fibrillation subroutine may further include the step of removing NDR values ​​that fall outside a predefined percentile of values. The predefined percentiles may include the 10th percentile and the 90th percentile, whereby NDR values ​​that are less than the 10th percentile and / or greater than the 90th percentile are removed from the set of stored NDR values. The atrial fibrillation subroutine may further include providing the set of stored NDR values ​​to a support vector machine (SVM) model configured to determine the presence of atrial fibrillation. In one embodiment, the SVM model was trained separately for binary classification (of atrial fibrillation) using a radial basis function (RBF) kernel. The misclassification cost was set to be proportional to the number of training samples per class. The feature NDR was extracted from multiple RRI samples, each sample including RRI values ​​from a one-minute time span. These RRI samples were fed to the SVM model for AF detection.

[0047] The method of the present disclosure may include yet another atrial fibrillation subroutine configured to determine atrial fibrillation, the subroutine comprising: receiving ECG data; Optionally, executing an ECG pre-processing subroutine to obtain valid ECG data based on the received ECG data, preferably ECG pre-processing as disclosed herein; providing the ECG data (validated) to a semi-supervised learning model, e.g., a deep generative model based on a neural network, trained and configured to determine atrial fibrillation (AF) based on the ECG data (validated) (as exemplified in Example 1); Providing an alarm in case of atrial fibrillation; This includes the steps:

[0048] The semi-supervised learning model may be trained on less than 50% labeled data, preferably less than 40% labeled data, more preferably less than 30% labeled data, even more preferably less than 20% labeled data, and most preferably less than 10% labeled data.

[0049] VT / VF subroutine A reference subroutine for ventricular fibrillation (VF) detection can be found in Ibtehaz et al., "VFPred: A fusion of signal processing and machine learning techniques in detecting ventricular fibrillation from ECG signals", Biomedical Signal Processing and Control 49 (2019), pp. 349-359. The VFPred algorithm can detect VF, including classes VF and non-VF, and can be extended by classes VT / VF and non-VT / VF, thereby detecting both VF and ventricular tachycardia (VT).

[0050] Serious Adverse Events Traditional bedside monitoring systems have proven difficult for long-term monitoring of post-operative patients, as the majority of post-operative patients are ambulatory. With the disclosed approach using wearable sensors and advanced data analytics, these patients would greatly benefit from continuous and predictive monitoring.

[0051] A Serious Adverse Event (SAE) is any untoward medical occurrence or effect at any dose, including any of the following undesired or unintended effects: Death (regardless of cause) Life-threatening leading to hospitalization or prolongation of an existing hospitalization; The subject becomes permanently or significantly impaired or incapacitated associated with congenital malformations or birth defects, Qualifies as an "other" important medically significant event or condition, e.g., the event may endanger the subject or require intervention (e.g., intensive care in an emergency room or at home) to prevent one of the outcomes listed above.

[0052] Examples of SAEs are pneumonia, wound infection, anastomotic leak, pneumothorax, hemorrhage, heart attack, pulmonary embolism, delirium, syncope, stroke, transient ischemic attack, respiratory failure, atelectasis, pneumothorax, pleural effusion, pulmonary embolism, heart failure, deep vein thrombosis, nonfatal cardiac arrest, troponin elevation, myocardial infarction, atrial fibrillation, atrial flutter, ventricular tachycardia, other supraventricular tachyarrhythmias, second degree atrioventricular block, third degree atrioventricular block, urinary tract infection, sepsis, septic shock, surgical site infection, major bleeding, drains, acute renal failure, hypoglycemia, diabetic ketoacidosis, intestinal obstruction, fracture, opioid toxicity, reoperation, and death.

[0053] SAEs such as atrial fibrillation, atrial flutter, ventricular tachycardia, other supraventricular tachyarrhythmias, second degree atrioventricular block, and third degree atrioventricular block are examples of worsening events, which can both be referred to as clinical worsening events and thereby detected according to the approach of the present disclosure, and can be referred to as SAEs because they also fall within the definition of an SAE above.

[0054] Example 2 discloses a machine learning based detection of serious adverse events (SAEs), where a support vector machine model is trained with valid data. The features input to the model are extracted from the time series of four vital signs parameters HR, RR, SpO2 and sysBP, from which clinical deterioration events are extracted as trends in the data time series. However, the model may equally well be trained based on features selected from one or more of the specific clinical deterioration events disclosed herein. That is, once the model is trained as described in Example 2, the input to predict SAEs is vital signs data as disclosed herein, and the detection of one or more clinical deterioration events.

[0055] The application of the approach also disclosed in Example 2 applies to clinical deterioration events detected according to the approach of the present disclosure, i.e. clinical deterioration events and / or SAEs can potentially be predicted and preferably prevented by using the detection of clinical deterioration events as disclosed herein, i.e. by the application of machine learning and continuous vital signs monitoring of the (post-operative) patient.

[0056] Nighttime Monitoring As disclosed in Example 4, overnight monitoring of a patient can improve prediction of SAE, in particular patients with increased heart rate and respiratory rate as well as slight drops in oxygen saturation during sleep at night (e.g., from midnight to 6AM) are at increased risk of developing SAE during the next day compared to their normal vital signs parameters. This can be improved by combining monitoring with a sleep stage detector, e.g., based on EEG measurements, so that only sleep vital signs data is used for overnight analysis, knowing when the patient is asleep. Observation of abnormalities in the patient's nighttime period may trigger an alarm or pre-alarm, so that the patient is investigated by adjusting one or more of the subroutine thresholds to generate an alarm closer to the next day and / or earlier.

[0057] Alarm occurs The disclosed method is configured to provide an alarm if at least one deterioration event is detected by one of the deterioration event subroutines described above. Each subroutine receives one or more valid vital sign parameters (such as RR, HR, SpO2, and SBP) and provides an alarm if the monitored parameter(s) exceed one or more predefined thresholds for a predefined duration, as described in further detail with respect to each subroutine. Preferred values ​​of the different thresholds and durations for alarm generation associated with the different subroutines are summarized in the table below. [Table 1]

[0058] system The present disclosure further relates to a system for automated detection of clinical deterioration events in a patient, the system comprising: one or more sensors configured to automatically monitor a patient's vital signs data (e.g., heart rate, respiration rate, heart rate variability, body temperature, oxygen saturation, and blood pressure), the one or more sensors further configured to wirelessly transmit the vital signs data to a server and / or gateway; a first server for receiving and storing vital signs data, the first server having a computer program thereon, the computer program being configured to execute a method of the present disclosure to provide automated detection of clinical deterioration events in a patient; one or more gateways configured to provide wireless communication links between the sensor(s) and the first server; Includes.

[0059] Sensors The sensors worn by the patient are preferably selected from the group of electrocardiogram (ECG) sensors, pulse oximeters, oscillometric blood pressure monitors, peripheral skin conductance sensors, 3D accelerometers, peripheral thermometers, and continuous glucose monitors. The sensors are preferably wireless wearable sensors configured for wireless communication with one or more gateways or servers. Data from the sensors may be streamed at a predefined streaming interval to conserve battery consumption and data storage. The streaming interval may vary from sensor to sensor. As an example, the streaming interval for an ECG sensor may be every 2 minutes, every minute, or every 30 seconds. Additionally, different time interval data may be selected and streamed from each sensor. For example, ECG data may be collected continuously, but only 10 seconds of ECG data may be selected and streamed every minute. Preferably, the patient's heart rate and temperature are received continuously. Preferably, respiration rate is received as a 10 second average and may be streamed continuously or at a predefined interval, such as every 10 seconds. Peripheral oxygen saturation and perfusion index are preferably measured (and streamed) every second, and blood pressure is preferably measured (and streamed) every 15 or 30 minutes.

[0060] Patient Gateway A patient gateway is to be understood herein as an electronic device configured to communicate with one or more sensors and / or a server. An example of a patient gateway is a tablet computer. The gateway is preferably placed near the patient, e.g., at the patient's bedside, so that wireless signals from the sensors can reach the gateway. The system preferably includes a patient gateway for each patient. Wireless communication between the sensors and the patient gateway(s) can be any suitable wireless standard, such as Bluetooth, Bluetooth Low Energy (BLE), Ultra Wideband (UWB), Wi-Fi, IEEE 802.11ah (Wi-Fi HaLow), GSM, 4G, 5G, or other similar techniques.

[0061] server A server is to be understood as a computer or computer program that provides services (e.g., computation) to other programs or devices. The server of the system of the present disclosure is preferably a cloud server, i.e. located remotely from the rest of the system and accessible via the Internet. The subroutines of the present disclosure preferably form part of a computer program stored on one or more servers, such as a cloud server. In a preferred embodiment, the computer program including the one or more subroutines is stored on a first server. The first server is preferably configured to communicate with a patient gateway. The communication is preferably encrypted and may be wired or wireless. The wireless communication between the patient gateway and the first server may be of any suitable wireless standard mentioned with respect to the sensor and the patient gateway(s).

[0062] The system may further include a second server, preferably configured to provide an alarm (e.g., in the form of a push notification) to a remote device (such as a computer, a smartphone, or a tablet computer) when the system detects a clinical deterioration event or a medical complication.

[0063] Detailed Description of the Drawings 1 shows an embodiment of a system according to the present disclosure. In this embodiment, the system includes one or more wireless sensors, a patient gateway, a first server configured to receive data from the patient gateway and execute the methods and subroutines of the present disclosure herein, and a second server for providing an alarm (e.g., a push notification) when a clinical deterioration event in a patient is identified. The alarm is preferably transmitted to an external device of the system, such as a smartphone.

[0064] FIG. 2 shows a block diagram illustrating the overall functionality of the system and method as disclosed herein. A patient's multiple vital signs (e.g., heart rate, respiration rate, blood pressure, RR interval, oxygen saturation, etc.) are received from one or more sensors. The system is illustrated herein with three sensors: a Lifetouch Blue device (a combination ECG sensor and accelerometer), a Nonin WristOx sensor (a wireless pulse oximeter), and a blood pressure cuff. Some of the vital signs parameters are provided as direct inputs to one or more subroutines configured for event detection. Other parameters are provided as inputs to an ECG pre-processing subroutine configured to assess the quality of the ECG data and / or to an SpO2 pre-processing subroutine configured to assess the quality of the SpO2 value received from the pulse oximeter. The ECG pre-processing subroutine assigns a value, either 1 or 0, indicating the quality of the parameter concerned (e.g., AF, HR, RR, or Morph). A value of 1 indicates that the parameter is good enough to be used in the worsening event subroutine.

[0065] FIG. 3 shows another representation of an embodiment of a system according to the present disclosure. In this embodiment, the system includes one or more wireless sensors, a patient gateway, and a first server (referred to herein as a Lifeguard server) configured to receive data from the patient gateway and execute the methods and subroutines of the present disclosure herein. In this embodiment, the sensors are configured to wirelessly communicate with the patient gateway using Bluetooth or Bluetooth low energy. Further, the patient gateway is configured to wirelessly communicate with the first server using WiFi and / or GSM. The first server may be configured to provide a website that can be accessed by one or more external devices, such as a computer, tablet, smartphone, or similar device.

[0066] FIG. 4 illustrates an embodiment of an ECG Pre-Processing subroutine according to the present disclosure. In general, several vital sign parameters are derived from ECG data received by an ECG sensor. The purpose of the ECG Pre-Processing subroutine is to evaluate the quality of the ECG data so that the derived values ​​(e.g., RRI, HR, and RR) can be validated and used as inputs in the Deterioration Event subroutine. The flow diagram details how the ECG Pre-Processing subroutine determines whether the derived values ​​are good enough to be used as inputs to the Deterioration Event subroutine. This is done by performing a series of calculations on the ECG data and comparing the ECG data and / or the calculated values ​​from the ECG data to one or more thresholds, such that the parameters goodForHR, goodForAF, goodForRR, and goodForMorph are assigned values ​​of 0 or 1.

[0067] FIG. 5 illustrates an embodiment of the SpO2 pre-processing subroutine configured to evaluate the quality of the SpO2 values ​​from a patient-worn pulse oximeter. The purpose of the SpO2 pre-processing subroutine is to remove SpO2 values ​​that are considered to be non-physical. Examples of non-physical SpO2 values ​​are SpO2<0% and SpO2>100%. Such values ​​are preferably removed from the SpO2 data. Another example is that the difference in SpO2 values ​​per second should not exceed 4 percentage points. The SpO2 pre-processing subroutine is configured to remove such SpO2 values ​​before calculating the average SpO2 value.

[0068] 6 shows a block diagram of the slow breathing subroutine, according to one embodiment, which is configured to provide an alarm if HR>20 bpm and RR≦5 bpm for more than 1 minute.

[0069] 7 shows a block diagram of the tachypnea subroutine, according to one embodiment, in which the tachypnea subroutine is configured to provide an alarm if RR≧24 for more than 5 minutes.

[0070] 8 shows a block diagram of the hypoventilation subroutine, according to one embodiment, in which the hypoventilation subroutine is configured to provide an alarm if RR<11 bpm and SpO2<88% for more than 5 minutes.

[0071] 9 shows a block diagram of the desaturation subroutine, according to one embodiment, in which the desaturation subroutine is configured to provide an alarm if SpO2<80% for more than 1 minute.

[0072] 10 shows a block diagram of the desaturation subroutine, according to one embodiment, in which the desaturation subroutine is configured to provide an alarm if SpO2<85% for more than 5 minutes.

[0073] 11 shows a block diagram of the desaturation subroutine, according to one embodiment, in which the desaturation subroutine is configured to provide an alarm if SpO2<88% for more than 10 minutes.

[0074] 12 shows a block diagram of the desaturation subroutine, according to one embodiment, in which the desaturation subroutine is configured to provide an alarm if SpO2<92% for more than 60 minutes.

[0075] 13 shows a block diagram of a sinus tachycardia subroutine, according to one embodiment, in which the sinus tachycardia subroutine is configured to provide an alarm if HR > 111 for more than 60 minutes.

[0076] 14 shows a block diagram of a sinus tachycardia subroutine, according to one embodiment, in which the sinus tachycardia subroutine is configured to provide an alarm if HR>130 for more than 30 minutes.

[0077] 15 shows a block diagram of the bradycardia subroutine, according to one embodiment, in which the bradycardia subroutine is configured to provide an alarm if HR<30 for more than 1 minute.

[0078] 16 shows a block diagram of the bradycardia subroutine, according to one embodiment, in which the bradycardia subroutine is configured to provide an alarm if 30 bpm≦HR≦40 bpm for more than 5 minutes.

[0079] 17 shows a block diagram of the hypotension subroutine, according to one embodiment, in which the hypotension subroutine is configured to provide an alarm if SBP<70 mmHg.

[0080] 18 shows a block diagram of the hypertension subroutine, according to one embodiment, in which the hypertension subroutine is configured to provide an alarm if SBP > 220mmHg.

[0081] 19 shows a block diagram of the hypotension subroutine, according to one embodiment, in which the hypotension subroutine is configured to provide an alarm if SBP<91 mmHg between two consecutive measurements.

[0082] 20 shows a block diagram of the hypertension subroutine, according to one embodiment, in which the hypertension subroutine is configured to provide an alarm if SBP > 180mmHg between two consecutive measurements.

[0083] 21 shows a block diagram of the circulatory collapse subroutine, according to one embodiment, which is configured to provide an alarm if SBP<100 mmHg and HR>110 bpm for t≧30 minutes, or if SBP<100 mmHg and HR>130 bpm for t≧5 minutes, or if SBP<100 mmHg and HR<50 bpm for t≧30 minutes.

[0084] 22 shows a block diagram of an atrial fibrillation subroutine, according to one embodiment, in which the atrial fibrillation subroutine is configured to provide an alarm in the event of irregular RR intervals.

[0085] FIG. 23 shows a block diagram illustrating the overall functionality of the disclosed system and method. The block diagram shows what kind of received data (heart rate data, respiratory rate, SpO2, blood pressure) is used as input to different subroutines. The deterioration event subroutine is configured to determine whether an alarm needs to be given based on the received input. Information regarding the alarm (e.g., yes / no) can be stored in a database and then broadcast to a server or one or more external electronic devices.

[0086] FIG. 24 shows a block diagram illustrating a method for handling electrocardiogram (ECG) and photoplethysmogram (PPG) data. Preferably, ECG and PPG data are not streamed continuously as this would consume too much battery power on the wireless sensor. Rather, ECG and PPG data may be streamed in bundles of data, and the data stream may be started and stopped, for example, by a timer, during execution of the method of the present disclosure. The streamed ECG / PPG data may be stored in a cache (i.e., memory), which may be cleared from time to time, such as when new heartbeat data is received.

[0087] Example 1 - Detection of atrial fibrillation from an ECG Atrial fibrillation (AF) is the most common cardiac arrhythmia and is associated with a six-fold higher risk of stroke and a two-fold higher risk of death. According to the National Health Service (NHS), AF is the most common cardiac rhythm abnormality affecting over a million people in the UK alone. Atrial fibrillation is classified as a tachyarrhythmia, where the electrical impulse is not initiated at the sinus node, but instead initiates with fibrillation waves in the atria. Atrial fibrillation may also be characterized as an irregular rhythm with loss of P waves in the ECG signal. Preliminary studies have shown that atrial fibrillation is common in post-operative cancer patients. In the disclosed approach, the ECG is available from continuous bedside monitoring, providing the possibility of autonomous analysis of the ECG, thereby providing the possibility of detecting atrial fibrillation, as shown in this example.

[0088] Typically, deep neural networks are trained in a fully supervised manner, requiring a large amount of labeled data. There is a huge amount of medical data, but only a small amount of it is labeled. This can be used for semi-supervised learning, where an unsupervised model is co-trained on a large amount of unlabeled data with a supervised model that is trained on a smaller amount of labeled data. The neural network used in this example is therefore trained in a semi-supervised manner, where both labeled and unlabeled data are used. This allows the neural network to learn features from a larger dataset, where the segments are not necessarily labeled. The model is built as a convolutional neural network, utilizing the ResNet architecture.

[0089] The input to the model is a 10 second segment from a single lead ECG. The classification model used after model training is complete includes an encoder (see FIG. 25) and a classifier. The output of the classification model is the probability that the ECG signal from the input segment is indicative of an atrial fibrillation rhythm. The latent space and the decoder (see FIG. 25) are only used to train the unsupervised portion of the model.

[0090] The data used in this project were obtained from the publicly available MIT-BIH Atrial Fibrillation Database (AFDB). The AFDB contains 25 recordings from different subjects of 10 time duration each (2 contain only the location of the QRS complex, not the waveform). The remaining 23 recordings contain ECG signals obtained from two leads. Each signal was digitized using a sampling frequency of 250 Hz and 12-bit resolution within a ±10 mV range. Unaudited annotation of the QRS complexes is available along with manual annotation into the following subcategories: atrial fibrillation, atrial flutter, AV junction rhythm, and sinus rhythm (SR).

[0091] Each ECG recording was divided into non-overlapping segments of 10 seconds to avoid parts of the same segment being present in both the labeled and unlabeled datasets. Labels were assigned based on the annotation file available with the data and divided into AF and non-AF. In conditions where multiple labels were present in the same segment, the label present in the majority segment was used for the entire segment. For both training and test sets, the data were stratified by downsampling of the majority class. The dataset was split into a training set containing 90% of the segments and a test set containing the remaining 10%. To remove DC offsets and any baseline drift before normalization, a high-pass filter was used with a cutoff frequency of 0:5 Hz and a filter order of 5. All segments were downsampled to 100 Hz.

[0092] A variational autoencoder is an unsupervised generative model consisting of two neural networks, an inference model, an encoder, and a generative model, a decoder. The encoder maps input samples to lower dimensional latent variables that the decoder maps to reconstructions of the input samples. A variational autoencoder is built on probability theory and Bayes' rule. In a variational autoencoder, the inference model is defined as q-(zjx) and the generative model is defined as p(xjz). By including a label variable y in the model, a semi-supervised generative probabilistic model can be achieved. In this model, the inference model Q is defined as q-(zjx;y)q-(yjx), where each term is defined as:

number

[0093] A Gaussian distribution q(z|x;y) is used to estimate the final layer of the model with mean μ φ , and log variance

number

number

[0094] The objective of optimizing the parameters θ and φ is to maximize the log-likelihood log p(x). This is achieved by using Jensen's inequality to obtain an evidence lower bound function, which can be optimized. In the unlabeled case, the lower bound is given as

number

number

[0095] In the lower bound, the contributions of z and y in the unlabeled case and z in the labeled case are marginalized. In the unlabeled case, y is treated as a latent variable and sampled by summing over the two classes, and the integral of z is approximated by sampling from a Gaussian distribution in the latent space. For labeled data, optimization of the label y is done using binary cross-entropy.

[0096] In addition to the lower bounds defined in equations (5) and (6), an extra loss was introduced: the standard deviations of the input signal and the reconstruction were subtracted and the absolute value of the difference was taken. This was introduced to help the decoder to perform better reconstruction. For the classifier, we used the binary cross-entropy loss.

[0097] To further aid in training the DGM, we introduced two warmups, defined as delays and linear gradients to a maximum: one for KL divergence, with a delay of 25 epochs and a max weight of 0.1 for 100 epochs, and one for classification loss, with a delay of 0 and a max weight of 0.5 for 40 epochs. These were introduced to not overly inhibit the generative part of the network initially, before pushing towards classification and standard normal distribution of z.

[0098] Figure 25 shows a diagram of the deep generative model used in Example 1. 1D Conv: one-dimensional convolution layer. FC layer: fully connected layer.

[0099] Figure 26 shows a diagram of (a) the inferential model and (b) the generative model of the proposed network used in Example 1. The grey nodes indicate known data, and the partially colored nodes labeled y highlight the semi-supervised aspects of the model.

[0100] The deep generative model (DGM) can be divided into three parts: the encoder, the classifier, and the decoder. The encoder was constructed by a residual network (ResNet) architecture consisting of four blocks, each containing three convolutional layers and one residual connection. ResNet has shown superiority in other image classification tasks when compared with classical convolutional networks. To increase the receptive field of the network, dilations of 2, 4, and 8 were applied to the three layers in each block, respectively. Max pooling was performed at the end of each block using a kernel size of 3 and a stride of 3 to reduce the signal size per block by a factor of three. The kernel size and stride were 3 and 1, respectively, for all convolutional layers, and the number of output channels was fixed per block at 32, 32, 64, and 64, respectively, for the four blocks. Two fully connected layers were applied at the ends of the blocks with sizes of 1,000 and 500. The decoder and classifier were constructed as simple fully connected neural networks (CNNs). The decoder consisted of an input layer, four hidden layers with 4,096 nodes each, and an output layer. The classifier consisted of three layers with 500, 200, and 200 nodes respectively, and a binary softmax function as output. All layers except the output layer used rectified linear units as activation functions, with batch normalization and dropout (p=0:3). The mA diagram of the model is shown in Figure 2.

[0101] To demonstrate the feasibility of using a semi-supervised approach, the proposed DGM was tested against a conventional combined neural network (CNN) identical to the encoder+classifier of the DGM. The setups were constructed with different percentages of unlabeled and labeled data, where the labeled data was used to train both the DGM and the supervised part of the CNN, and the unlabeled part of the data was used to train only the unsupervised part of the DGM. In this way, a "titration curve" style setup was obtained, mimicking cases where different amounts of labeled data may be the acquired data. The models were trained using 1%, 5%, 10% and 50% of the data in the setups as labeled and the rest as unlabeled. For each setup, it was ensured that the data in the training and test sets were the same for both the DGM and the CNN. Furthermore, the random seeds were fixed to remain as similar as possible between runs. After balancing the classes, a total of 111,894 segments were available in the training set. The test set consisted of 12,434 segments, which were also balanced. Each training phase of the DGM consisted of 50 epochs, with the labeled data rotated to correspond to the amount of unlabeled data. The smaller amount of data per epoch when training CNNs allowed them to train for more epochs and to convergence instead, as fewer weight updates would be required if they were only allowed to train for 50 epochs.

[0102] result The results of training DGM and CNN using different amounts of labeled data are shown in the table below. [Table 2]

[0103] The best results are obtained by DGM in the semi-supervised approach using 50% of the labeled data. The input segments of selected samples and the corresponding reconstructions are shown in Figure 28, and the distribution of the samples of the test set in the latent space is shown in Figure 29. The results show the maximum performance of the model on a stratified test set of 98.8% with a sensitivity of 98.9% and a specificity of 98.8%. This was obtained using 50% of the training data as labeled data and 50% as unlabeled data. The results in the above table show that the proposed semi-supervised approach achieves higher performance in all test cases, with the most notable difference in the cases with a lower amount of labeled data. Comparing the results with different amounts of labeled data also shows that the semi-supervised approach with 5% labeled data is better than the best performance obtained by the fully supervised approach with 50% labeled data. Even in the case of 1% labeled data, DGM achieved a 94.0% accuracy rate.

[0104] Example 2 - Prediction of Serious Adverse Events (SAEs) from Vital Signs In general, monitoring of postoperative patients is important to prevent serious adverse events (SAEs) that increase morbidity and mortality, but current monitoring of postoperative patients relies on intermittent bedside monitoring. The approach of the present disclosure facilitates continuous and predictive monitoring, thus improving patient management. This example demonstrates machine learning-based prediction of SAEs in postoperative patients based on vital signs acquired by wearable sensors, and shows that SAEs can be predicted with a high AUROC that reached 93% by monitoring only four common vital signs. As in this example, by using descriptive statistics extracted from trends as features and machine learning techniques based on SVM to reduce the complexity of the algorithm, it consumes less battery power, which is very important for wearable systems.

[0105] This example shows classification of "SAE" vs. "No SAE" within 2 hours (prediction window) based on the last 10 hours of recording (observation window). First, trends in the vital signs time series were extracted using moving averages to remove noise. Descriptive statistics were then calculated from the trends of each modality and concatenated into feature vectors. Finally, machine learning based on support vector machines was used to predict SAE.

[0106] During the study, vital signs of heart rate, respiratory rate, and blood oxygen saturation were continuously acquired by wearable devices, and blood pressure was measured intermittently in 453 postoperative patients. Data acquisition was managed by the Isansys patient status engine.

[0107] The study was conducted at Rigshospitalet and Bispebjerg Hospital, Copenhagen, Denmark, from February 2018 to August 2020. Four hundred fifty-three postoperative patients (278 men and 175 women) participated in the study. The mean age was 71 years (range: 60-93 years) and the mean monitoring time was 79 hours (range: 0.73-168.8 hours). Patients in the study had a wide range of clinical SAEs, ranging from neurological, respiratory, circulatory, infectious and other complications. Information on SAEs was registered by physicians.

[0108] Vital signs HR, RR and SpO2 were acquired continuously by wearable sensors and BP was measured intermittently. Acquisition of vital signs was managed by an Isansys patient status engine (PSE) (Isansys Lifecare Ltd). An Isansys Lifetouch was attached to the patient's chest to acquire a single-lead ECG with a sampling frequency of 1000 Hz, from which HR in beats per minute and RR in breaths per minute were derived. A pulse oximeter (Nonin Model 3150 WristOx2) was attached to a finger for acquisition of a photoplethysmogram (PPG) with a sampling frequency of 75 Hz, from which SpO2 was derived as a percentage. Data from the wearable sensors and derived values ​​were first transmitted via Bluetooth to a gateway in the PSE placed near the patient's bed and then to the hospital server, which stored the data in the patient database via WIFI every minute. Systolic blood pressure (sysBP) in mmHg was measured intermittently using a Meditech BlueBP-05. These sysBP measurements were entered into the gateway by the medical staff and then automatically transmitted to the patient database. HR, RR, SpO2 and sysBP were synchronized by their timestamps.

[0109] Prediction of serious adverse events The prediction of SAE can be considered as a classification problem that aims to classify between "SAE" and "No SAE" over a period of time (prediction window), such as hours, based on the last recording (observation window). As shown in Figure 31A, 2 hours were selected for the prediction window and 10 hours for the observation window. In this study, samples of SAE caused by neurological, respiratory, circulatory, infectious and other complications were extracted from the patient database. These extracted SAE samples were considered as the "SAE class". Control class samples were extracted from patients who did not have SAE during the monitoring period. A classifier for prediction of SAE was trained from these two classes. The prediction of SAE was based on features extracted from the trends of the four time series HR, RR, SpO2 and sysBP and the classification made using a support vector machine (SVM) model. Figure 30 shows the steps of the applied algorithm.

[0110] 1) Extraction of SAE and control classes: SAE classes were identified based on SAE timestamps. To account for class imbalance, SAE classes were oversampled. SAE class samples were extracted as 8-h time series of vital signs overlapping from 2 h before the SAE timestamp to 12 h before the SAE timestamp. Four samples were extracted for each SAE, as shown in Figure 31A. Control class samples were extracted from patients who did not have an SAE during vital signs monitoring in the hospital, and the monitoring duration was at least 8 h. Figure 31B shows the extraction of control samples. Samples were extracted during the entire monitoring period to cover all possible patient conditions.

[0111] 2) Feature Extraction: Selection of discriminative features is usually important for predicting SAE. One or more clinical deterioration events often precede SAE and can be extracted from vital signs as demonstrated in the approach of the present disclosure. In this example, trends of time series of HR, RR, SpO2 and sysBP were extracted using moving average with a sliding window of 60 minutes. In this example, trends were considered to represent deterioration. Four descriptive statistics (maximum, minimum, mean, and standard deviation) were then calculated as features from the trends of each modality. Features from each modality were concatenated into one feature vector. The size of each feature vector was 16.

[0112] 3) SVM Classification: The SVM model used in this example is a supervised machine learning algorithm for solving classification and regression problems. It has shown good generalization properties in many applications. The basic idea is to construct an optimal hyperplane for linearly separable patterns. The optimal hyperplane is the one that has the maximum margin between the two classes. For non-linearly separable patterns, one solution is to transform the original data into a higher or indefinite dimensional space, and then find a separating hyperplane in the transformed space by using a kernel function. Given a training set (x i ;y i ), i=1,...,N, and x i ∈R n and y i ={±1}, x i is a data point, and y i is that point x i The output of the classifier is defined as y(x i )=sign[w T φ(x i )+b] The function is x i into a higher dimensional space, where w is the weight vector and b is the bias of the hyperplane. Standard SVM requires the solution of the following optimization problem:

number

number

number

[0113] result The performance of the classifier with three-fold cross-validation is summarized in Table I. The accuracy rate, sensitivity, specificity, PPV, NPV, and area under the receiver operating characteristic curve (AUROC) are relatively close among the three tests. On average, the classifier achieved 89% accuracy rate, 80% sensitivity, 93% specificity, 82% PPV, 92% NPV, and 93% AUROC. Furthermore, Figure 29 shows the receiver operating characteristic curves (ROC) of the three tests. On average, the AUROC of 93% indicates the good discriminatory power of the classifier. These findings are promising and demonstrate the feasibility of predicting SAE from vital signs obtained with wearable devices and intermittent measurements.

[0114] Example 3 - Non-invasive blood pressure estimation based on vital signs This example demonstrates a new non-invasive method of estimating a patient's blood pressure (BP) without the need for a regular cuff. It is based on measured vital signs and the application of artificial intelligence, in particular a trained machine learning model, also disclosed herein. The new BP estimation does not require the usual strict synchronization between wearable devices. Since vital signs can be acquired continuously and in real time, the BP estimation of the present disclosure can also be provided in real time, for example by the application of a trained machine learning model.

[0115] Blood pressure (BP) is a hemodynamic variable important in the assessment and diagnosis of conditions such as stroke and cardiovascular disease. BP can vary dramatically from beat to beat and minute to minute. It is important to continuously monitor BP in postoperative patients. Currently BP is often monitored continuously with an invasive arterial catheter, for example in critically ill patients in the ICU. This method has the risk of infection and requires clinical manipulation. Outside the ICU, BP is measured by cuff-based devices, but only intermittently. Inflation / deflation often causes discomfort / pain to the patient and prevents the patient from resting. Therefore, cuffless BP estimation is preferred. Many cuffless BP estimations are based on features that require synchronization between the electrocardiogram (ECG) and the photoplethysmogram (PPG). Synchronization between ECG and PPG often causes problems, since ECG and PPG are recorded from two different devices. In this study, we propose a novel method to estimate BP based on the vital signs heart rate (HR), respiratory rate (RR), blood oxygen saturation (SpO2) and pulse rate (PR), which are calculated independently and are not affected by synchronization.

[0116] method A total of 498 postoperative patients participated in the study. After major abdominal cancer surgery, patients were readmitted to a general ward where their vital signs were monitored for up to 4 days with the approach described herein. Serious adverse events due to a wide range of complications were collected for up to 30 days. Two wearable devices were attached to the patients to continuously acquire vital signs. One was an Isansys Lifetouch on the chest to acquire a single-lead ECG, from which HR and RR were derived. The other was a pulse oximeter on a finger to acquire a PPG, from which SpO2 and PR were derived. Systolic blood pressure (SBP) and diastolic blood pressure (DBP) were measured intermittently by a Meditech BlueBP-05. Wireless acquisition and transmission of vital signs was managed by an Isansys patient status engine (PSE) (Isansys Lifecare Ltd). HR, RR, SpO2, PR, SBP and DBP were synchronized through their timestamps.

[0117] A random forest with 200 trees was applied to estimate DBP and SBP, although other models could be used. First, 3-h time series of HR, RR, SpO2 and PR before BP measurements were extracted, from which descriptive statistics such as mean, standard deviation and range were calculated as features. Regression models were then trained on the data from the first day for each patient. The trained models were tested by the data from the next day. To evaluate the estimation performance, the mean absolute error (MAE) and standard deviation (STD) of the error were used.

[0118] result The estimated performance is shown in Table I. According to the Association for the Advancement of Medical Instrumentation (AAMI) standards, for both DBP and SDP, the MAE should be less than or equal to 5 mmHg, and the STD should be less than or equal to 8 mmHg. In this example, the STD for DBP met the standard and the MAE was close to the standard. The STD for SBP was closer to the standard and the MAE was higher. [Table 4]

[0119] Example 4 - Prediction of SAE from nighttime vital signs The period immediately after surgery is critical for patients because they are prone to infections and other types of complications, i.e. serious adverse events (SAEs). Impending complications may be detectable during the night before, as they may alter circadian rhythms. This example provides a predictive model that is able to classify nocturnal vital signs according to whether they precede a serious adverse event or come from patients who have not had any complications, based on data from 450 postoperative patients. The predictive model is compared to a random classifier to demonstrate applicability.

[0120] The circadian clock, an autonomous molecular mechanism, is found in all mammalian cells and regulates bodily functions such as hormone secretion, immune responses and sleep / wake phases. These normal changes in cardiovascular function can be accompanied by adverse events, for example, as the incidence of myocardial infarction or sudden cardiac death has been found to be elevated in the early morning compared to the nighttime. Protection may not be optimal, as antiarrhythmic mechanisms such as increasing heart rate variability can be constrained by disease. As demonstrated herein, serious adverse events can be preceded by changes in nocturnal vital signs, arising from deactivation of the patient's parasympathetic nervous system. The nighttime presents an opportunity to observe physiological baselines and make comparisons between patients who develop and do not develop complications, as heart rate and respiration rate reliably reach their nadir during sleep. This can lead to increases in heart rate, respiration rate and blood pressure.

[0121] Vital Signs Monitoring Vital signs were acquired according to the procedures described herein. In particular, systolic blood pressure (SBP) and diastolic blood pressure (DBP) values ​​in mmHg were automatically recorded using a Meditech BlueBP-05. Heart rate (HR) and respiratory rate (RR) were acquired once per minute from an Isansys Lifetouch single-lead ECG using a sampling rate of 1000 Hz at the patient's chest. Pulse rate measurements were taken at the patient's arm. Photoplethysmograms (PPG) were measured at a rate of 75 Hz using a pulse oximeter (Nonin model 3150 WristOx2). SpO2 values ​​were determined therefrom.

[0122] Modeling The problem was modeled as a binary classification task. Nights, defined as the range from midnight to 6AM, were extracted from the continuous measurements and labeled as either 1 if they preceded an SAE or 0 if they did not. Figure 33 shows the extraction process based on four idealized days of heart rate measurements. As shown, heart rate typically decreases during the night and increases during the day. SAEs were always assigned to the day they were recorded, and nights in which an SAE occurred at night were discarded. Nights were also rejected if data collection began or ended at night. At least 60 valid data points for heart rate values ​​were required per night to ensure that a minimum amount of usable data was present. This modality was selected because it was expected to have some predictive validity after consultation by a clinician. The process is shown in Figure 32.

[0123] Feature Extraction The procedure described in the previous section provides one 360 ​​min × 6 modality vector for each night. For each modality with at least one recorded data point, missing values ​​are filled by first rounding up and next rounding down. To smooth the data and correct for measurement errors, a moving average is calculated using the nearest 10 values. Nine features are calculated for each night: mean, median, standard deviation (STD), maximum, minimum, kurtosis, skewness, 10th and 90th percentiles. The mean, median and standard deviation can be useful to detect abnormal vital sign values, such as elevated heart rate or unstable respiratory rate. Maximum and minimum values ​​measured during the night can indicate unusual events, such as hypoxemic episodes in the case of SpO2. Kurtosis measures the weight of the distribution in the tails relative to the center, while skewness evaluates its asymmetry. The 10th and 90th percentiles provide information about the distribution. At each time step, the static variables age, sex, height, weight, whether the patient smokes, number of packs smoked, and amount of alcohol consumed per week are added. This results in a total of (9 x 6) + 7 = 61 features. Missing features, which arise when not a single value is recorded for each modality during the night, are filled by mean value imputation.

[0124] classification After feature extraction, the data is split into training and test sets using 5-fold cross-validation. Each split is used once for testing, while the remaining 4 splits constitute the training set. This procedure is performed 10 times and the average value is calculated from the evaluation metrics. To correct for imbalances in the dataset, a synthetic minority oversampling technique (SMOTE) is applied to the training but not to the test set. The implementation used in this example is provided by the imbalanced-learn library, making both classes of equal size. XGBoost was chosen as the classification algorithm because it has been shown to deliver state-of-the-art performance while running faster than most other solutions. In short, the algorithm works by gradient tree boosting. A dataset with n examples and m features.

number

number

number

number

number

number

[0125] This prediction approach can be compared to two implementations of a random classifier, such as those provided by scikit-learn's Dummy classifier class. The first, a uniform version, represents a simple coin flip, selecting classes with equal probability. The second, a stratified version, selects classes with the same probability as they are presented to the classifier in the labeled training output set, and therefore selects the majority class if it is selected more frequently.

[0126] result To evaluate the effectiveness of this approach, standard performance measures were calculated: accuracy indicates what proportion of the test data was correctly predicted, while recall, precision and F1 score provide information about the type of error. Additionally, the ROC-AUC value is calculated by integrating over the area under the curve.

[0127] After filtering the data as described in the previous section, 184 nights preceding SAE and 475 nights of patients without SAE remained. In the table below, the nightly averages for both classes are shown with the standard deviation of the averages in brackets. On average, on the night before an SAE, patients have higher heart and respiratory rates as well as slightly lower oxygen saturation compared to patients without complications. However, due to the large standard deviation, it is also clear that a simple threshold-based classifier is not sufficient to solve this task, which reaffirms the machine learning-based approach of the present disclosure. [Table 5]

[0128] Comparing the average percentage of missing data in the table below, we can see that the SAE group had more missing values ​​than the non-SAE group, which may be because critically ill patients more frequently removed the measurement devices, especially the finger oxygen saturation sensors. [Table 6]

[0129] Performance of complication prediction models The following table presents the performance of the classifier compared to two random baseline models. The model in this example achieved an F1 score of 0.49, precision of 0.58, accuracy of 0.75, and ROC-AUC score of 0.65, all better than the baseline. However, this classifier performs poorly in the recall metric due to the wide standard deviation as presented earlier. [Table 7]

[0130] summary This example has some limitations. All data used were from a single cohort for both training and validation. Models trained on data from various institutions will generalize better and be more useful in clinical situations. In addition, the fact that SAEs used as outcome measures in this study have different causes and severity and may lead to changes in vital signs in any way may explain the heterogeneous results. There was also a notable difference between the night before the SAE and the night without SAE in terms of missing data. The night before the severe event had a higher proportion of missing data for all modalities. Another problem is that we assumed that the patient was asleep based on the time of day. If the patient is awake during the night, the patient's vital signs may change, making prediction more difficult. Combining the algorithm in this example with a sleep stage detector, for example based on EEG measurements, may significantly improve its predictive ability. Despite these limitations, this example further demonstrates that vital sign monitoring is an important tool in predicting SAEs and that overnight monitoring can further improve prediction of SAEs, in some cases even hours before an SAE occurs, making the disclosed approach an important step forward in understanding disease progression during sleep.

[0131] Patient Data Data for this project were acquired from February 2018 to August 2020 at Rigshospitalet and Bispebjerg Hospital, Copenhagen, Denmark. 450 patients (275 men, 175 women) had a mean age of 71 years (range 60-93 years). On average, 80 hours of data (range 12-169 hours) were recorded. Serious adverse events (SAEs) are as defined according to the guidelines as any medical occurrence that results in death, hospitalization of the subject, persistent or significant disability or incapacity of the subject, is associated with congenital malformation or birth defect, or qualifies as "other important medically significant event or condition." These events were recorded by the attending clinician and entered into a database. Written informed consent was obtained from all patients participating in the study.

[0132] Example 5 - Continuous vital signs prediction As disclosed herein, continuous monitoring of vital signs has improved the basis of data analysis relative to the standard of care. This embodiment relates to predicting vital signs, i.e., not only providing an alarm if deterioration occurs, but actually predicting whether it is likely to occur in the near future.

[0133] This embodiment uses a multivariate autoregressive (MAR) model to generate predictive projections of vital sign parameters based on past measurements. Predicting vital signs can be useful in identifying deviations from normal physiology that may occur in the near future.

[0134] Multivariate Autoregressive (MAR) Model Set of variables y=y1,...,y N Considering that, each element y t =[y t1 ,...,y tm ] is the response at time t, N is the signal length, and m is the number of modalities in the signal. The response at time t, as defined by the MAR model, is given by:

number

[0135] Due to the nature of vital signs signals, the physiological expectation of the temporal evolution of the signal is that its value will return to some patient-specific baseline value due to homeostasis. It can be advantageous to construct a model that includes a "pull" towards the baseline value. This can be achieved by creating a MAR model centered on the intercept parameter. A typical implementation is to use the mean μ of the signal as a y The model is centered on the response y t is instead obtained from

number

[0136] data In this example, data from an observational study of 500 postoperative cancer patients monitored for up to 4 days after major abdominal surgery were used. Data were acquired from February 2018 to August 2020 at Rigshospitalet and Bispebjerg Hospital, Copenhagen, Denmark. Patients were monitored by a single-lead ECG patch (Lifetouch Blue), a wrist-worn pulse oximeter (Nonin WristOx2), and a cuff-based blood pressure monitor (TM-2441). From these sensors, the following modalities were available: heart rate (HR) (1 / 60Hz), respiratory rate (RR) (1 / 60Hz), peripheral oxygen saturation (SpO2) (1 / 60Hz), and systolic and diastolic blood pressure (measured every 30 minutes). All data were transmitted to a central server by the Isansys Patient Status Engine. A subset of measurements was selected from the cohort to perform parameter inference in the model and evaluate the prediction accuracy. Only measurements of HR and RR values ​​were used. The extracted data was ensured to have no missing values ​​of HR or RR during the period. To fit the model, the subset consisted of 150 minutes of simultaneous HR and RR measurements from 8 different randomly selected patients. This gave a total of 20 hours of data for inference. The time series used is shown in Figure 35, which shows the time series vital signs of the patients used to fit the MAR model. To test the prediction accuracy of the model, 400 minutes from 5 different patients were used. It was ensured that the patients in the subset used for evaluation were different from the subset used for inference.

[0137] Fitting the model The MAR model was constructed as a pooled model. A pooled model defines a model in which the same parameters are fitted across several different data sources, in this case the vital sign signals of different patients. This results in a single set of model parameters that are used for all future patients. For the MAR model, this means that the parameters α, β and Σ are kept equal for all patients P. A probabilistic graphical model of the implemented pooled MAR model is shown in the graph in Figure 36.

[0138] For this model, the prior distributions of the parameters were kept non-informative and given by normal distributions. If the intercept α was used as a global baseline, its value was chosen to reflect a common baseline value for heart rate and respiratory rate. For heart rate, the mean was set to 70, and for respiratory rate, the mean was set to 12. All parameters in β were set to have a prior distribution following a standard normal distribution. The model is summarized below.

number

[0139] Inferring model parameters The aim of fitting the model is to establish model parameters θ to fit the target distribution p(θ|y). This is done either by exact or approximate inference depending on the dimensionality of the problem at hand. Due to the computational complexity of exact inference, the current problem is challenging for an exact approach. Figure 37 shows the setup used to evaluate the model on a new patient. At each step, a prediction (right box) is performed based on the data available in the model window (left box). The window is then advanced by 10 minutes and the process is repeated.

[0140] Instead, we use approximate inference in the form of Markov Chain Monte Carlo (MCMC) sampling. MCMC sampling is a common method based on repeatedly drawing samples of θ from an approximate distribution and updating these to successively improve the approximation to the target distribution. The idea is that, similar to Bayesian simulation, a collection of simulated draws from p(θ|y) summarizes the posterior density. MCMC sampling is therefore useful for sampling from Bayesian posterior distributions, where it is difficult to infer θ exactly from p(θ|y). Due to the random initialization of the sampling algorithm, samples have a transition period from the initialization to the posterior distribution. To account for this, we define a warm-up period and reject samples from this. To ensure that the sampling is stabilized on the posterior distribution, we sampled from multiple independent chains, and thus convergence is measured using a diagnostic measure R that compares the within-chain and between-chain variances. b The idea is that the individual chains are not mixed and therefore do not approach the target distribution, but the variance of all the mixed chains should be larger than the variance of the individual chains. As the individual chains converge, Rb-→1, and Vehtari et al.

[0141] Evaluation of prediction accuracy ​To evaluate the model's prediction accuracy, the model was applied to data from five unobserved patients. A window matching the lag parameter K=20 was provided to the model to generate a 15-minute forecast. The forecasted segments were compared to the true values ​​within the window. In this case, the root mean square error (RMSE) was used to quantify the accuracy of the expected values ​​within the forecast window relative to the original signal. The window was advanced by 10 minutes and the process was repeated for the entire time series. The setup from the first step to the next step is shown in Figure 3.

[0142] result The results of the evaluation of the prediction accuracy are shown in the table below. The model parameters α, β, and Σ are R [Table 8]

[0143] Summary Number ​Predicting future deviations in vital signs such as heart rate and respiratory rate is difficult as the nature of the signal can imply rapid changes that are not known in advance. A sudden activation of the patient will cause a change in the patient's vital signs, but it is not possible to predict before the activation occurs. The difficulty of capturing this can be seen in Figure 38, which shows a visualization of the response of a hierarchical AR model fitted to HR and RR data. In the legend "·" is the original signal time series and "-" is the expected value of the AR model. (Grey area, left side): Prediction interval of the AR model (95%). (Green area, right side): Prediction interval of the prediction from the AR model (95%), i.e., rapid changes within the prediction window are not captured by the prediction.

[0144] The model proposed in this example shows promising results when applied to different patients. The range of subsets used for the evaluation shows that the model provides good predictions for both low and high values ​​of HR and RR. However, variability occurring across multiple days and under different circumstances is hardly evaluated, and most likely there are rare events that are not represented in the evaluation. In this example, the model was implemented in a pooled configuration, which has advantages in clinical settings. Since the pooled model relies on a single set of parameters that cover all patients, there is no need to perform inference of parameters per patient, which is computationally resource-demanding when done in an iterative Bayesian approach. This can also be a disadvantage of the pooled model compared to other configurations, such as separate or hierarchical models, where patient-specific variability can be incorporated into the model. It can be an advantage when it is difficult to fit the model to the diversity in the data with different patients. However, since there is no clear patient-specific deviation, the use of these models must be restrained from increasing computational requirements.

[0145] Another aspect of the natural representation of vital signs is heteroscedasticity, which is not included in this example but is assumed to be present. The current model assumes that the data is homoscedastic within each modality, i.e., the data has the same variance across patients and time locations. It is evident from the plots in FIG. 38 that the first, second and fourth plots from the top have very little variance in the data, while the third and fifth have a large variance, indicating that in practice the homoscedasticity assumption of the model does not hold. Two solutions to achieve heteroscedasticity could be 1) to model the variance in an autoregressive manner similar to the current modeling of the mean, or 2) to model the variance in a hierarchical manner, where the parameters α, β and Σ depend on the variance in the data.

[0146] The construction of a model to create a forecast leads to the question of how to use the forecast. Due to the nature of the signal, which involves rapid changes, the notion that it is possible to predict far into the future is unrealistic. Instead, it may be advantageous to use the forecast as a baseline prediction and evaluate deviations from this based on the true values ​​in the actual setup. Quantifying the rapid changes from the forecasted values ​​may be a way to use the model to detect deviations in a real-time setting.

[0147] This example shows that it is possible to predict / project time series of vital signs HR and RR based on previous measurements, for example by using a pooled MAR model. On average, fairly low RMSEs of 11.4 bpm for HR and 3.3 bpm for RR were achieved, although the deviation in prediction accuracy within the prediction window between patients was large (see, for example, FIG. 38). Thus, although the work is based on a small subset of patient data, this example demonstrates promising results for predicting vital signs in a clinical setting.

Claims

1. A computer-implemented method for automatically and real-time detecting clinical deterioration events in a patient, comprising: Continuously receiving a plurality of different vital sign data from a plurality of sensors worn by the patient, wherein the vital sign data includes electrocardiogram (ECG), photoplethysmogram (PPG), heart rate (HR), respiratory rate (RR), blood pressure, and peripheral oxygen saturation (SpO ２ ), and the receiving analyzing the vital sign data to identify artifacts; discarding one or more data samples associated with the identified artifacts in the vital sign data to continuously obtain valid patient vital sign parameters; executing clinically valid deterioration event subroutines implemented on a plurality of computers, each subroutine receiving one or more of the valid vital sign parameters and configured to determine specific clinical deterioration events in the patient, wherein the deterioration event subroutines include: bradypnea / apnea based on valid heart rate and respiratory rate parameters; tachypnea based on valid respiratory rate; hypoventilation based on valid respiratory rate and transcutaneous arterial oxygen saturation; desaturation based on valid transcutaneous arterial oxygen saturation; sinus tachycardia based on valid heart rate; and bradycardia based on valid heart rate; hypotension based on valid systolic blood pressure or estimated systolic blood pressure; hypertension based on valid systolic blood pressure or estimated systolic blood pressure; executing, including; providing an alarm when at least one of the deterioration events is detected by one of the clinically valid automatic deterioration event subroutines; including, a method.

2. The deterioration event subroutines include: circulatory collapse based on valid heart rate and systolic blood pressure; cardiac arrest based on valid cardiac rhythm and pulse rate; atrial fibrillation based on ECG and / or valid cardiac rhythm; ventricular premature contractions based on valid cardiac rhythm; ventricular tachycardia / ventricular fibrillation based on valid ECG and / or valid cardiac rhythm; myocardial ischemia based on valid ischemic ECG response; low perfusion index based on valid peripheral perfusion index; and acute stress based on valid peripheral skin conductance and peripheral body temperature of the extremities; selected from the group of, the method according to claim 1.

3. The sensor is selected from the group consisting of an electrocardiogram (ECG) sensor, a pulse oximeter, an oscillometric blood pressure monitor, a peripheral skin conductance sensor, a 3D accelerometer, a peripheral thermometer, and a continuous glucose monitor, the method according to claim 1 or 2.

4. The sensor is a wireless sensor, the method according to claim 1 or 2.

5. The method according to claim 4, wherein the data from the sensor is streamed every about 1 minute or every about 30 seconds.

6. The method according to claim 3, further comprising an ECG preprocessing subroutine configured to evaluate the quality of the ECG data from the ECG sensor worn by the patient to obtain valid vital sign parameters.

7. The ECG preprocessing subroutine receiving an ECG value from the ECG sensor, the ECG value including an ECG timestamp, an ECG sample, a heartbeat timestamp, an R-R interval, and / or a QRS amplitude, said receiving discarding ECG values that are outside a predetermined threshold and / or discarding duplicate ECG values performing linear interpolation of ECG missing values determining a QRS index in the ECG (i.e., the time related to the position of the R peak in the QRS complex) using a heartbeat timestamp (hbTS) comparing the heart rate and / or the QRS amplitude and / or the ratio between a sub-selection of the QRS amplitude and the heart rate with one or more predetermined thresholds, and discarding one or more of the ECG values if the threshold is exceeded performing bandpass filtering of the ECG values correcting the R peak(s) in the QRS complex from the ECG values and / or normalizing each heartbeat from the ECG values calculating an average ECG cardiac cycle or a previously received average ECG cardiac cycle calculating a correlation coefficient (cc) comparing the average of the cc and / or the number of the cc with one or more predetermined thresholds, and discarding one or more of the ECG values if the threshold is exceeded calculating the deviation between the corrected R peak(s) and the uncorrected R peak(s), storing the deviation as an HBdev value(s), and comparing the HBdev value(s) with one or more predetermined thresholds, and discarding the ECG values associated with the HBdev value(s) that exceed the one or more predetermined thresholds The method according to claim 6, comprising.

8. The method is an effective SpO ２ To obtain a value, the SpO from a pulse oximeter worn by the patient ２ A SpO configured to evaluate the quality of the value ２ Further includes a SpO2 preprocessing subroutine, and the SpO2 preprocessing subroutine is receiving an SpO₂ value from the pulse oximeter removing SpO₂ values that exceed one or more predetermined thresholds removing duplicate SpO₂ values removing SpO₂ values having a difference of more than 4 per second extrapolating the missing SpO₂ value, and optionally, calculating an average SpO₂ value from the SpO₂ values, comprising the method according to claim 7.

9. The slow breathing subroutine receives valid HR and RR values, compares the valid HR and RR values with one or more predetermined slow breathing thresholds, and provides an alarm when the HR and RR values exceed the predetermined slow breathing threshold for a predetermined duration, comprising the method according to claim 1.

10. The rapid breathing subroutine receives valid RR value(s), compares the valid RR value(s) with a predetermined rapid breathing threshold, and provides an alarm when the RR value(s) exceed the rapid breathing threshold for a predetermined duration, comprising the method according to claim 1.

11. The hypoventilation subroutine Receive valid RR and SpO ２ values (multiple possible), said effective RR and SpO ２ comparing each of said values with one or more hypoventilation thresholds, and said RR and SpO ２ providing an alarm when the values exceed the hypoventilation threshold(s) for a predetermined duration comprising the method according to claim 1.

12. The saturation decrease subroutine Receive a valid SpO ２ value, the effective SpO ２ value is compared with one or more predetermined SpO ２ threshold values, and The SpO ２ value provides an alarm when it exceeds the SpO ２ threshold value(s) for a predetermined duration t. comprising the method according to claim 1.

13. The atrial tachycardia subroutine receives valid HR value(s), compares the valid HR value with one or more predetermined atrial tachycardia thresholds, and provides an alarm when the HR value exceeds the atrial tachycardia threshold(s) for a predetermined duration t, comprising the method according to claim 1.

14. The bradycardia subroutine receives valid HR value(s), compares the valid HR value with one or more predetermined bradycardia thresholds and / or ranges, and provides an alarm when the HR value exceeds the bradycardia threshold(s) and / or range for a predetermined duration t, comprising the method according to claim 1.

15. The hypotension subroutine receives valid SBP value(s), compares the valid SBP value with one or more predetermined hypotension thresholds, and provides an alarm when the SBP value exceeds the hypotension threshold(s) during one or more consecutive measurements and / or for a predetermined duration t, comprising the method according to claim 1.

16. The circulatory collapse subroutine receives valid SBP and HR value(s), compares the valid SBP and HR values with one or more predetermined SBP thresholds and HR thresholds, and Providing an alarm when at least one of the SBP threshold(s) and the HR threshold is exceeded for a predetermined duration t. The method according to claim 1, comprising.

17. The cardiac arrest subroutine is Receiving ECG data samples, Detecting QRS complexes in the ECG data samples, and Providing an alarm when there are no detected QRS complexes for a predetermined duration t1 and / or when there are no detected peripheral pulses for a predetermined duration t2. The method according to claim 2, comprising the steps of.

18. The hypertension subroutine is Receiving valid SBP value(s), Comparing the valid SBP value with one or more predetermined hypertension thresholds, and Providing an alarm when the SBP value exceeds the hypertension threshold(s) during one or more consecutive measurements and / or for a predetermined duration t. The method according to claim 1, comprising.

19. The atrial fibrillation subroutine is Receiving valid RRI values, Comparing the valid RRI values with one or more predetermined first atrial fibrillation criteria, wherein the criteria specify that the valid RRI falls below or above one or more predetermined thresholds, said comparing, Optionally, storing valid values that fall within the first atrial fibrillation criteria in an array, Optionally, comparing the amount of the values stored in the array with a predetermined second atrial fibrillation criterion regarding the size of the array, and Providing an alarm when the first atrial fibrillation criterion and / or the second atrial fibrillation criterion is satisfied. The method according to claim 2, comprising.

20. Detecting and / or predicting one or more serious adverse events based on one or more of the determined clinical deterioration events, as claimed in claim 1 or 2.

21. An automatic real-time detection system for clinical deterioration events in a patient, when executed by a processor, Continuously receiving a plurality of different vital sign data from a plurality of sensors worn by the patient, wherein the vital sign data includes electrocardiogram (ECG), photoplethysmogram (PPG), heart rate (HR), respiratory rate (RR), blood pressure (e.g., systolic blood pressure, SBP), and peripheral oxygen saturation (SpO2), said receiving. Analyzing the vital sign data to identify artifacts, Discarding one or more data samples related to identified artifacts in the vital sign data to continuously obtain valid patient vital sign parameters, Executing clinically valid deterioration event subroutines implemented on multiple computers, each subroutine receiving one or more of the valid vital sign parameters and configured to determine specific clinical deterioration events in the patient, the deterioration event subroutines being, Bradypnea / apnea based on valid heart rate and respiratory rate parameters, Tachypnea based on valid respiratory rate, Hypoventilation based on valid respiratory rate and transcutaneous arterial oxygen saturation, Saturation decrease based on valid transcutaneous arterial oxygen saturation, Sinus tachycardia based on valid heart rate, and Bradycardia based on valid heart rate, Hypotension based on valid systolic blood pressure or estimated systolic blood pressure, Hypertension based on valid systolic blood pressure or estimated systolic blood pressure, including executing, Providing an alarm if at least one of the deterioration events is detected by one of the clinically valid automated deterioration event subroutines, A system including a non-transitory computer-readable storage medium for storing instructions for executing a method including.

22. An automated detection system for clinical deterioration events in a patient, One or more sensors configured to automatically monitor the vital sign data of the patient, the one or more sensors including an electrocardiogram (ECG) sensor, a photoplethysmogram (PPG) sensor, a heart rate (HR) sensor, a respiratory rate (RR) sensor, a blood pressure sensor, and a peripheral oxygen saturation (SpO₂) sensor, and further configured to wirelessly transmit the vital sign data to a server and / or gateway, the one or more sensors, A first server for receiving and storing the vital sign data, the first server being, Continuously receiving multiple different vital sign data from multiple sensors worn by the patient, the vital sign data including electrocardiogram (ECG), photoplethysmogram (PPG), heart rate (HR), respiratory rate (RR), blood pressure, and peripheral oxygen saturation (SpO₂), Analyzing the vital sign data to identify artifacts, To continuously obtain valid patient vital sign parameters, discard one or more data samples related to identified artifacts in the vital sign data, Execute clinically valid deterioration event subroutines implemented on multiple computers, each subroutine configured to receive one or more of the valid vital sign parameters to determine specific clinical deterioration events in the patient, the deterioration event subroutines being, Bradypnea / apnea based on valid heart rate and respiratory rate parameters, Tachypnea based on valid respiratory rate, Hypoventilation based on valid respiratory rate and transcutaneous arterial oxygen saturation, Saturation decrease based on valid transcutaneous arterial oxygen saturation, Sinus tachycardia based on valid heart rate, and Bradycardia based on valid heart rate, Hypotension based on valid systolic blood pressure or estimated systolic blood pressure, and Hypertension based on valid systolic blood pressure or estimated systolic blood pressure, Determine, Provide an alarm if at least one of the deterioration events is detected by one of the clinically valid automated deterioration event subroutines, the detection being based on one or more valid vital sign parameters exceeding one or more predetermined thresholds for a predetermined duration, Configured as, a first server, and Including, a system.

Images