Calibration of continuous non-invasive blood pressure monitoring using artificial intelligence

Abstract

A system for continuous noninvasive blood pressure monitoring can include processing circuitry configured to determine calibration data for a continuous noninvasive blood pressure model at a calibration point, receive a PPG signal from an oximetry sensing device at a particular time after the calibration point, derive values for a set of indicators of a patient from the PPG signal, and determine a blood pressure of the patient at the particular time using the continuous noninvasive blood pressure model and based at least in part on the calibration data determined at the calibration point, the values for the set of indicators, and a time elapsed since the calibration point into the continuous noninvasive blood pressure model at the particular time.

Description

[0001] This application claims priority to U.S. Patent Application No. 16 / 723,678, filed December 20, 2019, entitled “CALIBRATION FOR CONTINUOUSNON-INVASIVE BLOOD PRESSURE MONITORING USING ARTIFICIAL INTELLIGENCE”. Background Technology

[0002] Unlike standard occluded cuff technology, continuous noninvasive blood pressure monitoring (CNIBP) systems allow for continuous tracking of a patient's blood pressure. In some instances, a pulse oximetry sensor can be placed on the patient to measure photoplethysmography (PPG) signals, and these PPG signals can be used to estimate the patient's blood pressure. CNIBP monitoring systems can be periodically recalibrated based on blood pressure measurements taken by noninvasive blood pressure monitoring systems, such as pneumatic cuff blood pressure monitors. Summary of the Invention

[0003] This disclosure describes example apparatus, systems, and techniques for performing continuous noninvasive blood pressure monitoring on a patient in a manner that can improve the accuracy of continuous noninvasive blood pressure monitoring by using PPG signals. Instead of using only PPG signals to determine a patient's blood pressure, this disclosure describes example techniques for determining calibration data during periodic recalibration of the continuous noninvasive blood pressure monitoring system and using the most recently determined calibration data and PPG signals to determine (e.g., derive) a patient's blood pressure. Furthermore, in some aspects, this disclosure describes a continuous noninvasive blood pressure model that is trained using machine learning with such determined calibration data, enabling the model to receive the most recently determined calibration data and PPG signals to more accurately determine a patient's blood pressure.

[0004] By determining calibration data during periodic recalibration of the continuous noninvasive blood pressure monitoring system, and using this calibration data to train the continuous noninvasive blood pressure model and as input to the model, the apparatus, system, and technique of this disclosure can improve the accuracy of continuous noninvasive blood pressure monitoring algorithms and enable continuous noninvasive blood monitoring devices to present more accurate information about a patient's blood pressure. The presentation of more accurate information, compared to continuous noninvasive monitoring systems that do not use such calibration data, can lead clinicians to make more informed decisions.

[0005] In some instances, a method includes determining calibration data for a continuous noninvasive blood pressure model during calibration, wherein the calibration data includes: receiving a patient's blood pressure measurement from a blood pressure sensing device at a calibration point; receiving a first photoplethysmography (PPG) signal from an oxygen saturation sensing device at the calibration point; and deriving a first value for a set of patient parameters from the first PPG signal. The method also includes receiving a second PPG signal from the oxygen saturation sensing device at a specific time after the calibration point. The method further includes deriving a second value for the set of patient parameters from the second PPG signal. The method also includes determining the patient's blood pressure at the specific time using the continuous noninvasive blood pressure model and based at least in part on the calibration data determined at the calibration point, the second value for the set of parameters, and the time elapsed since entering the continuous noninvasive blood pressure model from the calibration point at the specific time.

[0006] In some instances, a system includes a blood pressure sensing device, an oxygen saturation sensing device, and processing circuitry configured to: determine calibration data for a continuous noninvasive blood pressure model at a calibration point by at least the following methods: receiving a patient's blood pressure measurement from the blood pressure sensing device at the calibration point; receiving a first photoplethysmography (PPG) signal from the oxygen saturation sensing device at the calibration point; and deriving a first value for a set of patient parameters from the first PPG signal; receiving a second PPG signal from the oxygen saturation sensing device at a specific time after the calibration point; deriving a second value for the set of patient parameters from the second PPG signal; and determining the patient's blood pressure at the specific time using the continuous noninvasive blood pressure model and based at least in part on the calibration data determined at the calibration point, the second value for the set of parameters, and the time elapsed since entering the continuous noninvasive blood pressure model from the calibration point at the specific time.

[0007] In some instances, a non-transitory computer-readable storage medium includes instructions that, when executed, cause processing circuitry to: determine calibration data for a continuous non-invasive blood pressure model at a calibration point by at least the following manner: receiving a patient's blood pressure measurement at the calibration point; receiving a first photoplethysmography (PPG) signal at the calibration point and deriving a first value for a set of patient parameters from the first PPG signal; receiving a second PPG signal at a specific time after the calibration point; deriving a second value for the set of patient parameters from the second PPG signal; and using the continuous non-invasive blood pressure model and based at least in part on the calibration data determined at the calibration point, the second value for the set of parameters, and the time elapsed since entering the continuous non-invasive blood pressure model from the calibration point at the specific time, determining the patient's blood pressure at the specific time.

[0008] Details of one or more examples are illustrated in the accompanying drawings and the following description. Other features, objectives, and advantages will become apparent from the description and figures, as well as Examples 1-20. Attached Figure Description

[0009] Figure 1 This is a conceptual block diagram illustrating a continuous non-invasive blood pressure (CNIBP) monitoring device.

[0010] Figure 2 The illustration shows details of an example training system 200, which can perform training by... Figure 1 The training of the CNIBP model used by the CNIBP device shown.

[0011] Figure 3A and 3B The illustration shows an example of the discrepancy between blood pressure measurements and PPG signals.

[0012] Figure 4 The illustration shows an example graph of blood pressure changing over time, determined using the CNIBP monitoring algorithm, which is periodically calibrated with blood pressure measured by a non-invasive blood pressure monitoring system.

[0013] Figure 5 The diagram shows... Figure 1 and Figure 2 An example deep learning architecture for the CNIBP model.

[0014] Figure 6A and 6B The diagram illustrates what can be used for training. Figure 1 and Figure 2 Example features of the example CNIBP model.

[0015] Figure 7A and 7B An example diagram is shown, which depicts the use of... Figure 1 and Figure 2 The CNIBP model is used to determine a patient's blood pressure to improve accuracy.

[0016] Figure 8 The diagram shows... Figure 1 and Figure 2 An alternative instance of the deep learning architecture for the CNIBP model.

[0017] Figure 9 It is used as a diagram. Figure 1 and Figure 2 A flowchart of an example method for determining patient blood pressure using the CNIBP model 124. Detailed Implementation

[0018] Figure 1 This is a conceptual block diagram illustrating an example of a continuous non-invasive blood pressure monitoring device 100. The continuous non-invasive blood pressure (CNIBP) monitoring device 100 includes processing circuitry 110, memory 120, control circuitry 122, user interface 130, sensing circuitry 140 and 142, and sensing devices 150 and 152. Figure 1 In the illustrated example, the user interface 130 may include a display 132, an input device 134, and / or a speaker 136, which may be any suitable audio device configured to generate and output noise. In some instances, the CNIBP monitoring device 100 may be configured to determine and output (e.g., for display on the display 132) the continuous blood pressure of patient 101, for example, during medical procedures or for longer-term monitoring, such as in the intensive care unit (ICU) and general postoperative monitoring. Clinicians can receive information about the patient's continuous noninvasive blood pressure through the user interface 130 and adjust treatment or therapy for patient 101 based on the continuous noninvasive blood pressure information.

[0019] The processing circuitry 110 described herein, along with other processors, processing circuits, controllers, control circuits, etc., may include one or more processors. The processing circuitry 110 may include any combination of integrated circuits, discrete logic circuits, analog circuits (e.g., one or more microprocessors), digital signal processors (DSPs), application-specific integrated circuits (ASICs), or field-programmable gate arrays (FPGAs). In some instances, the processing circuitry 110 may include multiple components, such as one or more microprocessors, one or more DSPs, one or more ASICs, or one or more FPGAs, and any combination of other discrete or integrated logic circuits and / or analog circuits.

[0020] Control circuitry 122 may be operatively coupled to processing circuitry 110. Control circuitry 122 is configured to control the operation of sensing devices 150 and 152. In some instances, control circuitry 122 may be configured to provide timing control signals to coordinate the operation of sensing devices 150 and 152. For example, sensing circuitry 140 and 142 may receive one or more timing control signals from control circuitry 122, which may be used by sensing circuitry 140 and 142 to turn the respective sensing devices 150 and 152 on and off, for example, to periodically collect calibration data using sensing devices 150 and 152. In some instances, processing circuitry 110 may use timing control signals to operate synchronously with sensing circuitry 140 and 142. For example, processing circuitry 110 may synchronize the operation of analog-to-digital converters and demultiplexers with sensing circuitry 140 and 142 based on timing control signals.

[0021] The memory 120 can be configured to store, for example, monitored physiological parameter values, such as blood pressure values, oxygen saturation values, peripheral oxygen saturation values, or any combination thereof. The memory 120 can also be configured to store calibration data, which is periodically collected by the CNIBP monitoring device 100.

[0022] In some instances, memory 120 may store program instructions, such as neural network algorithms. The program instructions may include one or more program modules executable by processing circuitry 110. For example, memory 120 may store a continuous noninvasive blood pressure (CNIBP) model 124, which may be a model trained through machine learning to continuously and noninvasively determine the blood pressure of patient 101. When executed by processing circuitry 110, program instructions such as program instructions for CNIBP model 124 may enable processing circuitry 110 to provide the functionality attributed to it herein. The program instructions may be embodied in software, firmware, and / or RAMware. Memory 120 may include any one or more of volatile, non-volatile, magnetic, optical, or electrical media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), electrically erasable programmable ROM (EEPROM), flash memory, or any other digital media.

[0023] User interface 130 may include display 132, input device 134, and speaker 136. In some instances, user interface 130 may include fewer or more components. User interface 130 is configured to present information to a user (e.g., a clinician). For example, user interface 130 and / or display 132 may include a monitor, cathode ray tube display, flat panel display such as a liquid crystal (LCD) display, plasma display, light-emitting diode (LED) display, and / or any other suitable display. In some instances, user interface 130 may be a multi-parameter monitor (MPM) or other physiological signal monitor used in a clinical or other setting, a personal digital assistant, mobile phone, tablet computer, laptop computer, any other suitable computing device, or any combination thereof, with a built-in display or a separate display.

[0024] In some instances, processing circuitry 110 may be configured to present a graphical user interface to a user via a user interface 130, such as a display 132. The graphical user interface may include, via the display 132, indications of values ​​for one or more physiological parameters of the patient, such as, for example, blood pressure values, oxygen saturation values, information about autoregulation status (e.g., brain autoregulation status), pulse rate information, respiratory rate information, other patient physiological parameters, or combinations thereof. The user interface 130 may also include means for projecting audio to the user, such as a speaker 136.

[0025] In some instances, processing circuitry 110 may also receive input signals from an additional source (not shown), such as a user. For example, processing circuitry 110 may receive input signals from input device 134, such as a keyboard, mouse, touchscreen, button, switch, microphone, joystick, touchpad, or any other suitable input device or combination of input devices. The input signals may contain information about patient 101, such as physiological parameters, treatments provided to patient 101, or similar information. Additional input signals may be used by processing circuitry 110 in any determination or operation performed by processing circuitry 110.

[0026] In some instances, if processing circuitry 110 determines that the patient 101 is in an abnormal state, processing circuitry 110 may present a notification indicating the abnormal state. This notification may include visual, auditory, tactile, or somatosensory notifications (e.g., alarm signals) indicating the abnormal state. In some instances, processing circuitry 110 and user interface 130 may be part of the same device or housed within a enclosure (e.g., a computer or monitor). In other instances, processing circuitry 110 and user interface 130 may be separate devices configured to communicate via wired or wireless connections.

[0027] Sensing circuits 140 and 142 are configured to receive signals (“physiological signals”) indicating physiological parameters from corresponding sensing devices 150 and 152 and transmit the physiological signals to processing circuit 110. Sensing devices 150 and 152 may include any sensing hardware configured to sense a patient’s physiological parameters, such as, but not limited to, one or more electrodes, a photoreceiver, a blood pressure cuff, etc. The sensed physiological signals may include signals indicating physiological parameters from the patient, such as, but not limited to, blood pressure, blood oxygen saturation (e.g., pulse oxygen saturation and / or local oxygen saturation), blood volume, heart rate, and respiration. For example, sensing circuits 140 and 142 may include, but are not limited to, blood pressure sensing circuits, blood oxygen saturation sensing circuits, blood volume sensing circuits, heart rate sensing circuits, temperature sensing circuits, electrocardiogram (ECG) sensing circuits, electroencephalogram (EEG) sensing circuits, or any combination thereof.

[0028] In some instances, sensing circuits 140 and 142 and / or processing circuit 110 may include signal processing circuit 112, which is configured to perform any suitable analog conditioning on the sensed physiological signal. For example, sensing circuits 140 and 142 may transmit an unchanged (e.g., raw) signal to processing circuit 110. Processing circuit 110, such as signal processing circuit 112, may be configured to modify the raw signal into a usable signal by, for example, filtering (e.g., low-pass, high-pass, band-pass, notch, or any other suitable filtering), amplifying, performing operations on the received signal (e.g., differentiation, averaging), performing any other suitable signal conditioning (e.g., converting a current signal to a voltage signal), or any combination thereof. In some instances, the conditioned analog signal may be processed by an analog-to-digital converter of signal processing circuit 112 to convert the conditioned analog signal into a digital signal. In some instances, signal processing circuit 112 may operate in either analog or digital form to separate different components of the signal. In some instances, signal processing circuitry 112 may perform any suitable digital conditioning on the converted digital signal, such as low-pass, high-pass, band-pass, notch filtering, averaging, or perform any other suitable filtering, amplification, operation, or any combination thereof. In some instances, signal processing circuitry 112 may reduce the number of samples in the digital detector signal. In some instances, signal processing circuitry 112 may remove darkness or ambient light effects on the received signal. Additionally or alternatively, sensing circuits 140 and 142 may include signal processing circuitry 112 to modify one or more original signals and transmit one or more modified signals to processing circuitry 110.

[0029] Oxygen saturation sensing device 150 (also referred to herein as blood oxygen saturation sensing device 150) is configured to generate an oxygen saturation signal indicating blood oxygen saturation within a region of patient 101, including veins, arteries, and / or capillaries. For example, oxygen saturation sensing device 150 may include a sensor configured to noninvasively generate a plethysmography (PPG) signal. An example of such a sensor may be one or more blood oxygen saturation sensors (e.g., one or more pulse oximetry sensors) placed at one or more locations on patient 101, such as at the fingertip of patient 101, earlobe of patient 101, etc.

[0030] In some instances, the oxygen saturation sensing device 150 may be configured to be placed on the skin of patient 101 to determine local oxygen saturation in a specific tissue area, such as the frontal cortex or other brain location of patient 101. The oxygen saturation sensing device 150 may include an emitter 160 and a detector 162. The emitter 160 may include at least two light-emitting diodes (LEDs), each configured to emit light of a different wavelength, such as red or near-infrared light. As used herein, the term "light" may refer to energy generated by a radiation source and may include any wavelength within one or more of the electromagnetic radiation spectrum of ultrasound, radio, microwave, millimeter waves, infrared, visible light, ultraviolet, gamma rays, or X-rays. In some instances, light-driving circuitry (e.g., within the sensing device 150, sensing circuitry 140, control circuitry 122, and / or processing circuitry 110) may provide a light-driving signal to drive the emitter 160 and cause the emitter 160 to emit light. In some instances, the LEDs of the emitter 160 emit light in the range of approximately 600 nanometers (nm) to approximately 1000 nm. In a specific example, one LED of emitter 160 is configured to emit light at approximately 730 nm, and the other LED of emitter 160 is configured to emit light at approximately 810 nm. Other wavelengths of light may be used in other examples.

[0031] Detector 162 may include a first detection element positioned relatively "closer" (e.g., proximal) to emitter 160 and a second detection element positioned relatively "farther" (e.g., distal) from emitter 160. In some instances, the first and second detection elements may be selected to be particularly sensitive to a selected target energy spectrum of emitter 160. Multiple wavelengths of light intensity may be received at both the "closer" and "farther" locations of detector 162. For example, if two wavelengths are used, the two wavelengths can be compared at each location, and the resulting signals can be compared to achieve an oxygen saturation value that, when light passes through a region of the patient (e.g., the patient's skull), is related to other tissues (tissues other than those through which the light received at the "closer" detector passes, such as brain tissue) through which light passes. In operation, light may enter detector 162 after passing through tissues of patient 101, including skin, bone, other superficial tissues (e.g., non-brain tissues and superficial brain tissue), and / or deep tissues (e.g., deep brain tissue). Detector 162 may convert the intensity of the received light into an electrical signal. Light intensity can be directly correlated with light absorption and / or reflection in tissues. Surface data from the skin and skull can be subtracted to generate a target tissue oxygen saturation signal over time.

[0032] Oxygen saturation sensing device 150 can provide an oxygen saturation signal to processing circuitry 110 or any other suitable processing device for determining the blood pressure of patient 101. Further details of other examples of determining oxygen saturation based on optical signals can be found in commonly assigned U.S. Patent No. 9,861,317, published January 9, 2018, entitled "Methods and Systems for Determining Regional Blood Oxygen Saturation". One example of such an oxygen saturation signal could be a volumetric plethysmography (PPG) signal.

[0033] In operation, the blood pressure sensing device 152 and the oxygen saturation sensing device 150 can be placed on the same or different parts of the patient 101's body. For example, the blood pressure sensing device 152 and the oxygen saturation sensing device 150 can be physically separated from each other and can be placed separately on the patient 101. As another example, the blood pressure sensing device 152 and the oxygen saturation sensing device 150 can, in some cases, be supported by a single sensor housing. One or both of the blood pressure sensing device 152 or the oxygen saturation sensing device 150 can be further configured to measure other parameters, such as hemoglobin, respiratory rate, respiratory effort, heart rate, saturation pattern detection, response to stimuli such as the bispectral index (BIS) or electromyographic (EMG) response to electrical stimulation, etc. Although in Figure 1 The example CNIBP monitoring device 100 is shown in the image, but... Figure 1 The components shown are not intended to be limiting. Additional or alternative components and / or implementation schemes may be used in other instances.

[0034] Blood pressure sensing device 152 is configured to generate a blood pressure signal indicating the blood pressure of patient 101. For example, blood pressure sensing device 152 may include a blood pressure cuff configured for noninvasive blood pressure sensing or an arterial line configured for invasive monitoring of blood pressure in an artery of patient 101. In some instances, the blood pressure signal may include at least a portion of a waveform of the acquired blood pressure. Blood pressure sensing device 152 may be configured to generate a blood pressure signal indicating changes in patient blood pressure over time. Blood pressure sensing device 152 may provide the blood pressure signal to sensing circuitry 142, processing circuitry 110, or any other suitable processing device, which may be part of device 100 or separate from device 100, such as another device located in the same location as device 100 or remotely positioned relative to device 100.

[0035] Processing circuit 110 may be configured to receive one or more signals generated by sensing devices 150 and 152 and sensing circuits 140 and 142. Physiological signals may include signals indicating blood pressure and / or signals indicating oxygen saturation, such as PPG signals. Processing circuit 110 may be configured to determine a blood pressure value based on the blood pressure signals.

[0036] According to aspects of this disclosure, the CNIBP monitoring device 100 is configured to provide continuous noninvasive blood pressure monitoring of a patient 101. For this purpose, the CNIBP monitoring device 100 can be configured to periodically determine the blood pressure of the patient 101 by periodically receiving blood pressure readings from a blood pressure sensing device 152, for example, every minute, every five minutes, every 15 minutes, etc. For example, the CNIBP monitoring device 100 can periodically turn on or activate the blood pressure sensing device 152 so that the blood pressure sensing device 152 can measure the blood pressure of the patient 101. In another embodiment, the blood pressure sensing device 152 can continuously monitor the blood pressure of the patient 101, and the CNIBP monitoring device 100 can periodically request the blood pressure readings of the patient 101 from the blood pressure sensing device 152.

[0037] In order to provide continuous non-invasive blood pressure monitoring of patient 101 during the time interval between periodic measurements of patient 101's blood pressure using blood pressure sensing device 152, processing circuitry 110 may be configured to derive patient 101's blood pressure using CNIBP model 124 based at least in part on the PPG signal provided by oxygen saturation sensing device 150 during these time intervals.

[0038] During continuous noninvasive blood pressure monitoring of patient 101, each time CNIBP monitoring device 100 uses blood pressure sensing device 152 to determine patient 101's blood pressure, referred to herein as a calibration point, is the CNIBP monitoring device 100 calibrated using the actual blood pressure of patient 101 as measured by blood pressure sensing device 152 at each calibration point. During each calibration point, processing circuitry 110 can be configured to determine calibration data for calibrating CNIBP model 124. Calibration points can occur periodically, such as every 3 minutes, every 5 minutes, every 10 minutes, every 15 minutes, every 20 minutes, etc., and each calibration point can last for a specified time period, such as the duration of one or more cardiac cycles, a specified number of seconds (e.g., 1 second, 5 seconds, etc.), etc.

[0039] For example, at each calibration point, control circuitry 114 may be configured to send one or more timing control signals to sensing circuits 140 and 142 to turn on, activate, or otherwise receive data from corresponding sensing devices 150 and 152 to collect calibration data using sensing devices 150 and 152. Blood pressure sensing device 152 may provide a blood pressure signal, indicating the patient 101's blood pressure at the calibration point, to processing circuitry 110. Similarly, oxygen saturation sensing device 150 may provide an oxygen saturation signal in the form of a PPG signal to processing circuitry 110 at the calibration point. Processing circuitry 110 may be operable to extract features from the PPG signal, such as values ​​of a set of indicators. For example, features may include values ​​of any combination of indicators, such as PPG pulse duration, PPG relative position of the maximum upward slope of systolic blood pressure rise, PPG peak position and amplitude, PPG perfusion index, PPG baseline trend, PPG respiratory cycle information (such as respiratory rate), PPG ascending region, PPG descending region, maximum gradient of the PPG ascending slope, and one or more of the PPG baseline value. The patient's blood pressure and features extracted from the PPG signal can form calibration data collected by the processing circuit 110. The calibration point can be a specified time period, such as the time period of a single cardiac cycle, one second, or any other suitable time period. Therefore, for each feature, the feature extracted from the PPG signal can include a sequence of values ​​within the time period of the calibration point. For example, the PPG pulse duration can be a sequence of PPG pulse duration values ​​over the time period of the calibration point, the PPG ascending region can be a sequence of PPG ascending region values ​​over the time period, and so on.

[0040] Processing circuitry 110 can store calibration data in memory 120. Because the CNIBP monitoring device 100 is configured to periodically calibrate itself by determining the blood pressure of patient 101 using blood pressure sensing device 152, calibration points occur periodically, and the CNIBP monitoring device 100 can be configured to determine calibration data whenever the CNIBP monitoring device 100 performs self-calibration using the aforementioned technique. In some instances, processing circuitry 110 can be configured to overwrite the calibration data stored in memory 120 with the most recently determined calibration data, such that only the most recently determined calibration data at the most recent calibration point is stored in memory 120. In other instances, processing circuitry 110 can be configured to store calibration data determined based on multiple previous calibration points in memory 120.

[0041] As described above, between the periodic occurrences of calibration points, the CNIBP monitoring organization 100 can be configured to determine the continuous noninvasive blood pressure of patient 101 using the CNIBP model 124, without using the blood pressure sensing device 152 to determine the blood pressure of patient 101. In some instances, the CNIBP model 124 is a neural network algorithm trained by machine learning to acquire the patient's (e.g., patient 101's) PPG signal at the current time and calibration data determined at the most recent calibration point as input to determine the patient's blood pressure at the current time.

[0042] Neural network algorithms, or artificial neural networks, can include trainable or adaptive algorithms that utilize nodes with defined rules. For example, a corresponding node among multiple nodes can utilize a function such as a nonlinear function or an if-then rule to generate an output based on an input. A corresponding node among multiple nodes can be connected along edges to one or more different nodes among multiple nodes, such that the output of the corresponding node includes the input of the different nodes. The function can include parameters that can be determined or adjusted using a training set of inputs and desired outputs, such as, for example, a predetermined correlation between one or more PPG signals sensed from patient 101 or a patient group and one or more second blood pressure readings of patient 101 or a patient group measured simultaneously with the PPG signals, and learning rules, such as backpropagation learning rules. Backpropagation learning rules can utilize one or more error measurements that compare the desired output with the output generated by the neural network algorithm to train the neural network algorithm by changing parameters to minimize one or more error measurements.

[0043] The example neural network includes multiple nodes, at least some of which have node parameters. Inputs, including at least a PPG signal generated by oxygen saturation sensing device 150 or oxygen saturation sensing circuit 140 and indicating the blood oxygen saturation of patient 101, can be fed into a first node of the neural network algorithm. In some instances, the inputs may include multiple inputs, each fed to a corresponding node. The first node may include a function configured to determine the output based on the inputs and one or more adjustable node parameters. In some instances, the neural network may include a propagation function configured to determine the inputs to subsequent nodes based on the outputs and bias values ​​of previous nodes. In some instances, the learning rule may be configured to modify one or more node parameters to produce a preferred output. For example, the preferred output may be constrained by one or more thresholds and / or minimize one or more error measurements. The preferred output may include the outputs of a single node, a group of nodes, or multiple nodes.

[0044] The neural network algorithm can iteratively modify node parameters until the output includes the preferred output. In this way, processing circuitry 110 can be configured to iteratively evaluate the output of the neural network algorithm and, based on the evaluation of the output, iteratively modify at least one of the node parameters to determine the blood pressure of a patient, such as patient 101, based on the modified neural network algorithm. In some instances, compared to other techniques, the neural network algorithm enables processing circuitry 110 to more accurately determine the continuous blood pressure of patient 101 using features extracted from the PPG signal along with calibration data from the most recent calibration point and / or reduce the computation time and / or power required to determine changed blood pressure values.

[0045] According to aspects of this disclosure, processing circuitry 110 can be configured to execute CNIBP model 124 to continuously determine the blood pressure of patient 101 between calibration points. At any specific time t between calibration points, processing circuitry 110 can be configured to use CNIBP model 124 to determine the blood pressure of patient 101 at time t. At time t, processing circuitry 110 can be configured to receive the PPG signal of patient 101 from oxygen saturation sensing device 150. Processing circuitry 110 can input the PPG signal of patient 101 at time t, along with calibration data derived from the calibration point closest to time t, into CNIBP model 124, and processing circuitry 110 can be configured to execute CNIBP model 124 to determine the blood pressure of patient 101 at time t from such input.

[0046] Once the processing circuitry 110 determines the blood pressure of the patient 101, it can provide information indicating the patient 101's continuous blood pressure to an output device, such as a user interface 130. In some instances, under the control of the processing circuitry 110, the user interface 130, such as a display 132, can present a graphical user interface including information indicating the patient 101's continuous blood pressure. In some instances, the blood pressure indication of the patient 101 can include text, colors, and / or audio presented to the user. In addition to or instead of a graphical user interface, the processing circuitry 110 can be configured to generate and present information indicating the patient 101's continuous blood pressure via a speaker 136, for example, by announcing the patient 101's current blood pressure via voice.

[0047] In some instances, the CNIBP monitoring device 100, such as processing circuitry 110 or user interface 130, may include a communication interface that enables the CNIBP monitoring device 100 to exchange information with external devices. The communication interface may comprise any suitable hardware, software, or both, that allows the CNIBP monitoring device 100 to communicate with electronic circuitry, devices, networks, servers or other workstations, displays, or any combination thereof. For example, processing circuitry 110 may receive blood pressure and / or oxygen saturation values ​​from an external device via the communication interface.

[0048] The components of the CNIBP monitoring device 100, shown and described as individual components, are shown and described for illustrative purposes only. In some instances, the functionality of some components may be combined in a single component. For example, the functionality of processing circuitry 110 and control circuitry 122 may be combined in a single processor system. Furthermore, in some instances, the functionality of some components of the CNIBP monitoring device 100 shown and described herein may be divided into multiple components. For example, some or all of the functionality of control circuitry 122 may be performed in processing circuitry 110 or sensing circuitry 140 and 142. In other instances, the functionality of one or more components may be performed in a different order or may not be required.

[0049] Figure 2 The illustration shows details of an example training system 200, which can perform training by... Figure 1 The training of CNIBP model 124 is shown. Figure 2 Only one specific instance of the training system 200 is shown, and many other example devices with more, fewer, or different components can also be configured to perform operations according to the techniques disclosed herein.

[0050] Although Figure 2 In some instances, the training system 200 is shown as part of a single device, but in others, components of the training system 200 may reside within and / or be part of different devices. For example, in some instances, the training system 200 may represent a "cloud" computing system. Therefore, in these instances, Figure 2 The modules shown can span multiple computing devices. In some instances, training system 200 can represent one of multiple servers in a server cluster that constitutes a "cloud" computing system. In other instances, training system 200 can be... Figure 1 An example of the CNIBP monitoring device 100 is shown.

[0051] like Figure 2 As shown in the example, training system 200 includes one or more processors 202, one or more communication units 204, and one or more storage devices 208. Storage device 208 also includes a CNIBP model 124, a training module 212, and training data 214. Each of the one or more processors 202, one or more communication units 204, and one or more storage devices 208 can be interconnected (physically, communicatively, and / or operatively) for inter-component communication. Figure 2In this example, one or more processors 202, one or more communication units 204, and one or more storage devices 208 may be coupled through one or more communication channels 206. In some instances, the communication channel 206 may include a system bus, a network connection, an inter-process communication data structure, or any other channel used for transmitting data. The CNIBP model 124, training module 212, and training data 214 may also exchange information with each other and with other components in the training system 200.

[0052] exist Figure 2 In this example, one or more processors 202 may implement functions and / or execute instructions within the training system 200. For instance, one or more processors 202 may receive and execute instructions stored in storage device 208 that perform the functions of training module 212. These instructions executed by one or more processors 202 may cause the training system 200 to store information in storage device 208 during execution. One or more processors 202 may execute instructions of training module 212 to train CNIBP model 124 using training data 214. That is, training module 212 may be operated by one or more processors 202 to perform various actions or functions of the training system 200 described herein.

[0053] exist Figure 2 In this example, one or more communication units 204 can operatively communicate with external devices via one or more networks by transmitting and / or receiving network signals over one or more networks. For example, training system 200 can use communication unit 204 to transmit and / or receive radio signals over a radio network such as a cellular radio network. Similarly, communication unit 204 can transmit and / or receive satellite signals over a satellite network such as a Global Positioning System (GPS) network. Examples of communication unit 204 include network interface cards (e.g., Ethernet cards), optical transceivers, radio frequency transceivers, or any other type of device capable of transmitting and / or receiving information. Other examples of communication unit 204 may include near-field communication (NFC) units, Bluetooth, etc. Radio, shortwave radio, cellular data radio, wireless networks (e.g., Wi-Fi) Radio and Universal Serial Bus (USB) controller.

[0054] exist Figure 2In some instances, one or more storage devices 208 may be operable to store information for processing during operation of the training system 200. In some instances, storage device 208 may represent temporary memory, meaning that the primary purpose of storage device 208 is not long-term storage. For example, storage device 208 of training system 200 may be volatile memory configured for short-term storage of information, so that the stored contents are not retained if power is lost. Examples of volatile memory include random access memory (RAM), dynamic random access memory (DRAM), static random access memory (SRAM), and other forms of volatile memory known in the art.

[0055] In some instances, storage device 208 also refers to one or more computer-readable storage media. That is, storage device 208 can be configured to store a larger amount of information than temporary storage. For example, storage device 46 may include non-volatile memory that retains information in power-on / power-off cycles. Examples of non-volatile memory include magnetic hard disks, optical disks, floppy disks, flash memory, or electrically programmable memory (EPROM) or electrically erasable programmable memory (EEPROM). In any case, in Figure 2 The storage device 208 in the example can store program instructions and / or data associated with the CNIBP model 124, training module 212 and training data 214.

[0056] exist Figure 2 In an example, training system 200 may execute training module 212 to train CNIBP model 124 using training data 214, thereby more accurately determining a patient's blood pressure between periodic calibration points by associating one or more PPG morphological feature sequences with their respective blood pressure values. CNIBP model 124 may include a deep learning architecture, such as a recurrent neural network, convolutional neural network, etc., which includes multiple layers to progressively extract higher-level features from the inputs of CNIBP model 124.

[0057] In some instances, training module 212 trains CNIBP model 124 to use sensing circuit 140 coupled to patient 101 and oxygen saturation sensing device 150 at time t. Figure 1 The received PPG signal, calibration data determined at the nearest calibration point t, and the time elapsed from the nearest calibration point at time t are used as inputs to determine the patient 101's blood pressure at time t. The calibration data may include, for example, data obtained at the nearest calibration point from the BP sensing circuit 142 and the blood pressure sensing device 152. Figure 1 The values ​​of a set of indicators are derived from the blood pressure of patient 101 measured by the PPG signal received from the oxygen saturation sensing device 150 at the most recent calibration point.

[0058] In some instances, the training data 214 used to train the CNIBP model 124 includes data from patients 101 alone and not from other subjects. In other instances, the training data 214 may include data from a patient population, such as a calibration dataset of the patient population at the most recent calibration point, a set of times elapsed since the most recent calibration point, a set of PPG signals for the patient population, and a set of target blood pressure values ​​for the patient population. For example, each individual training data point in the training data 214 may be the patient's target blood pressure at time t, the values ​​of a set of indicators derived from the PPG signals received from the patient at time t, the values ​​of a set of indicators derived from the PPG signals received from the patient at the most recent calibration time t, the patient's blood pressure measured at the most recent calibration time t, and the correlation between time t and the time elapsed since the most recent calibration time.

[0059] In some instances, once training module 212 has trained CNIBP model 124 using training data 214, training module 212 can test CNIBP model 124 using a set of test data that CNIBP model 124 has not yet encountered to determine how closely the blood pressure determined by CNIBP model 124 based on the test data matches the expected target blood pressure of the test data. In this way, training module 212 can evaluate and further refine CNIBP model 124.

[0060] When training module 212 completes training of CNIBP model 124, CNIBP model 124 can be installed, uploaded, or otherwise transmitted to CNIBP monitoring device 100. In some instances, training module 212 can upload or otherwise transmit a copy of CNIBP model 124 to another server or cloud, and CNIBP monitoring device 100 can use CNIBP model 124 via, for example, the Internet, a VPN, or a local area network.

[0061] Figure 3A and 3B The illustration shows an example of the bias between blood pressure measurements and PPG signals. This bias can potentially be improved using the techniques disclosed herein. Figure 3A As shown, curve 300A includes a blood pressure signal 302A acquired from a patient using an invasive arterial line, while curve 300B includes a PPG signal 302B acquired from the same patient using a pulse oximeter approximately simultaneously with the blood pressure signal 302A in Figure 300A. Similarly, as shown in curve 300B, curve 300C includes a blood pressure signal 302C acquired from a patient using an invasive arterial line, while curve 300D includes a PPG signal 302D acquired from the same patient using a pulse oximeter approximately simultaneously with the blood pressure signal 302C in Figure 300C. Although Figure 3AThe PPG signal 302B is morphologically similar to the blood pressure signal 302A, but... Figure 3B The PPG signal 302D is morphologically very different from the blood pressure signal 302C.

[0062] It can be seen that accurately mapping PPG signals to blood pressure signals can be difficult. This is likely due to many confounding factors, such as one or more of the following: contact force, ambient temperature, medication, vasomotor activity, exercise, arteriosclerosis, postural changes, etc. Because of these confounding factors, continuous non-invasive blood pressure monitoring algorithms may lose their accuracy over time.

[0063] One technique for addressing the possibility that continuous noninvasive blood pressure monitoring algorithms may lose their accuracy over time involves periodically calibrating the algorithm using blood pressure measured by a noninvasive blood pressure monitoring system, such as an inflatable cuff blood pressure monitoring system.

[0064] Figure 4 Figure 400 shows an example of blood pressure variation over time determined using a continuous noninvasive blood pressure monitoring algorithm, which is periodically calibrated with blood pressure measured by a noninvasive blood pressure monitoring system. Figure 4 As shown, the CNIBP monitoring system can periodically calibrate, for example, at times t0, t1, and t2 (referred to as calibration points throughout this disclosure), using blood pressure measured by the non-invasive blood pressure monitoring system, such as blood pressure 402 determined using a continuous non-invasive blood pressure monitoring algorithm.

[0065] At time t0, the CNIBP monitoring system calibrates blood pressure 402 using blood pressure 404A measured by the CNIBP monitoring system at time t0. After calibrating blood pressure 402 at time t0, the CNIBP monitoring system uses a continuous non-invasive blood pressure monitoring algorithm to determine blood pressure 402 until time t1 is reached. At time t1, the CNIBP monitoring system calibrates blood pressure 402 using blood pressure 404B measured by the CNIBP monitoring system at time t1. After calibrating blood pressure 402 at time t1, the CNIBP monitoring system uses a continuous non-invasive blood pressure monitoring algorithm to determine blood pressure 402 until time t2 is reached. At time t2, the CNIBP monitoring system determines blood pressure 402 using blood pressure 404C measured by the CNIBP monitoring system at time t2. After calibrating blood pressure 402 at time t2, the CNIBP monitoring system uses a continuous non-invasive blood pressure monitoring algorithm to determine blood pressure 402 until the next calibration point is reached.

[0066] like Figure 4As observed, the blood pressure 402 determined using the continuous non-invasive blood pressure monitoring algorithm may drift significantly from the blood pressure measured by the non-invasive blood pressure monitoring system (e.g., in a way that could meaningfully affect the value of the blood pressure monitoring). For example, there may be a significant difference between the blood pressure 402 determined using the continuous non-invasive blood pressure monitoring algorithm at time t1 and the blood pressure 404B measured by the non-invasive blood pressure monitoring system at time t1.

[0067] One technique for compensating for this difference involves adding or subtracting the difference between the blood pressure determined using a continuous non-invasive blood pressure monitoring algorithm at the most recent calibration point and the blood pressure measured by the non-invasive blood pressure monitoring system. For example, because the blood pressure 402 determined using the continuous non-invasive blood pressure monitoring algorithm at time t1 is higher than the blood pressure 404B measured by the non-invasive blood pressure monitoring system at time t1, the instantaneous blood pressure 402 determined between times t1 and t2 can be adjusted by subtracting the difference between the blood pressure 402 determined using the continuous non-invasive blood pressure monitoring algorithm at time t1 and the blood pressure 404B measured by the non-invasive blood pressure monitoring system at time t1 from the instantaneous blood pressure 402 determined between times t1 and t2.

[0068] According to aspects of this disclosure, the CNIBP monitoring device 100 is configured to potentially determine patient blood pressure between calibration points more accurately using the CNIBP model 124. For example, at each calibration point, such as at times t0, t1, and t2, the CNIBP monitoring device 100 can determine calibration data, which may include blood pressure measured at the calibration point and values ​​of a set of indicators derived from the PPG signal received at the calibration point. To determine the patient's blood pressure over time between calibration points, such as at time t3, the CNIBP model 124 can be input with the values ​​of the set of indicators determined at time t3, the calibration data determined at time t0, and the time elapsed since the most recent calibration point at time t0 when the patient entered the CNIBP model 124 at time t3, to determine the patient's blood pressure at time t3.

[0069] Figure 5 The illustration shows an example deep learning architecture 500 for CNIBP model 124. Although Figure 5 The diagram illustrates a deep learning architecture 500, which is a Long Short-Term Memory (LSTM) deep learning architecture used to train LSTM models, but any other deep learning architecture is equally suitable for training CNIBP models 124.

[0070] like Figure 5As shown, the deep learning architecture 500 may include a sequence input layer 502, a dropout layer 504, a bidirectional long short-term memory (BiLSTM) layer 506, a dropout layer 508, a BiLSTM layer 510, a fully connected layer 512, and a regression output layer 514. The sequence input layer 502 can be connected to the dropout layer 504. The dropout layer 504 can be connected to the BiLSTM layer 506. The BiLSTM layer 506 can be connected to the dropout layer 508. The dropout layer 508 can be connected to the BiLSTM layer 510, and the BiLSTM layer 510 can be connected to the fully connected layer 512. The fully connected layer 512 can be connected to the regression output layer 514.

[0071] Sequence input layers, such as sequence input layer 502, feed sequence data into the neural network. Therefore, sequence input layer 502 receives features used to train the deep learning architecture 500, including values ​​of a set of metrics derived from the PPG signal, calibration data determined at the most recent calibration point, and the time elapsed since the most recent calibration point. The features received by sequence input layer 502 are discussed below regarding... Figure 6A and 6B Detailed description.

[0072] Dropout layers, such as dropout layer 504 and dropout layer 508, randomly set the input elements to zero with a given probability. By randomly setting the input elements to zero, dropout layers can make elements ignored during the training phase. Selectively ignoring elements during the training phase can prevent overfitting of the training data.

[0073] BiLSTM layers, such as BiLSTM layer 506 and BiLSTM layer 510, learn bidirectional long-term dependencies between time steps of time series or sequence data. These dependencies can help the network learn from the complete time series at each time step.

[0074] Fully connected layers, such as fully connected layer 512, multiply the input (e.g., from BiLSTM layer 510) by a weight matrix and then add a bias vector. Regression output layers, such as regression output layer 514, compute the half-mean squared error loss of the regression problem and output the predicted response of the trained regression network as a result of training the CNIBP model 124 with deep learning architecture 500.

[0075] To train CNIBP model 124, training system 200 ( Figure 2A set of features and associated target values ​​can be derived, and these features and associated target values ​​can be input into the CNIBP model 124 to train the CNIBP model 124 to estimate the target values ​​based on the input features. For example, to train the CNIBP model 124 to predict the blood pressure of patient 101 from the PPG signal of patient 101 (without using a blood pressure monitoring device, such as an inflatable cuff blood pressure monitoring system or an arterial line, to actually measure blood pressure), the training system 200 can extract features from the PPG signal of patient 101 or from the PPG signals of one or more other patients / subjects, and can use these features along with the associated target blood pressure to train the CNIBP model 124 to predict the blood pressure value of patient 101 from the features of the PPG signal.

[0076] For example, training system 200 can pair time-series features extracted from a patient's PPG signal (e.g., PPG signal received from a patient from a pulse oximeter) with associated target blood pressures (e.g., the patient's systolic blood pressure (SP), diastolic blood pressure (DP), mean arterial pressure (MAP), pulse pressure (PP), etc.). Training system 200 can input such feature pairs extracted from PPG signals and train CNIBP model 124 with blood pressure from a patient population as the target to associate PPG morphological feature sequences with blood pressure values.

[0077] By training the CNIBP model 124, the CNIBP model 124 can receive time-series features extracted from the PPG signal of patient 101 as input, and can determine the patient's blood pressure value based on the input features. The training system 200 can test the CNIBP model 124 using a test set of PPG feature sequences that the model has not yet encountered to determine the relevant blood pressure. The training system 200 can input the test set of PPG feature sequences into the CNIBP model 124, and can compare the blood pressure value output by the CNIBP model 124 with the expected target blood pressure value of the test set of PPG feature sequences to evaluate and further refine the CNIBP model 124.

[0078] Figure 6A and 6B The diagram illustrates what can be used for training. Figure 1 and Figure 2 Example features of CNIBP model 124. Figure 6A As shown, feature 600 can be a set of example features from the PPG signal used to train the CNIBP model 124, as described above. Similarly, feature 600 can also be an instance of input to the CNIBP model 124, which can determine the blood pressure of patient 101 based on the input (without using a blood pressure monitoring device, such as an inflatable cuff blood pressure monitoring system or an arterial line, to actually measure blood pressure).

[0079] Feature 600 can be a set of values ​​derived from the PPG signal over a specific time period. For example, feature 600 can include features 600A-600M, which can include any combination of indicators derived from the PPG signal, such as PPG pulse duration, PPG relative position of the maximum upward slope of systolic blood pressure rise, PPG peak position and amplitude, PPG perfusion index, PPG baseline trend, PPG respiratory cycle information, PPG ascending region, PPG descending region, maximum gradient of the PPG ascending slope, and PPG baseline value.

[0080] Each of features 600A-600M may include a value sequence over time for a specific indicator derived from the PPG signal. For example, feature 600A may include a value sequence for the PPG uplink region, feature 600B may include a value sequence for the PPG downlink region, feature 600C may include a value sequence for the PPG amplitude, feature 600D may include a sequence of values ​​for the maximum uplink slope of the PPG, and so on.

[0081] The value sequence over time can be a sequence of values ​​over a specified time period, such as 5 seconds, 10 seconds, 20 seconds, etc. In some instances, the specified time period can be the entire time period that has elapsed since the most recent occurrence of the calibration point, while in other instances, the specified time period can be the previous 5 seconds, 10 seconds, 20 seconds, etc. Because each of features 600A-600M includes a value sequence over a time period for a specific indicator, each of features 600A-600M can be a feature vector of the value sequence over the time period, and features 600 as a whole can be called a feature matrix composed of feature vectors (e.g., features 600A-600M).

[0082] In some instances, over a time period, the value sequence of feature 600A-600M can be calculated over the cardiac cycles of patient 101, such that each value of the value sequence of feature 600A-600M can be a value obtained in different cardiac cycles of patient 101 within that time period. In some instances, over a time period, the value sequence of feature 600A-600M can be calculated within specified time intervals, such as 1 second, 2 seconds, etc., such that the value sequence of feature 600A-600M can be values ​​obtained at a series of such time intervals within that time period.

[0083] In a further example, within this time period, the values ​​of the value sequence can be acquired at a single time point (e.g., on a single pulse), such that the value sequence of feature 600A-600M can include values ​​acquired at the time point sequence. For example, the peak value of the derivative of the pulse can be taken from the entire pulse, for example, using the pulse area, thus allowing the value sequence to be acquired over the pulse sequence, or over many pulses. The values ​​of the value sequence can also be acquired at a series of time points, for example, over multiple pulses, such as when deriving values ​​from the maximum derivative sequence obtained from the pulse sequence. Therefore, in some instances, the length of the value sequence can vary based on the number of pulses within a specified time period as a function of the patient's heart rate.

[0084] Feature 600 is derived from the PPG signal indicating the patient's blood oxygen saturation at a given time, at which point the patient's blood pressure will be determined. Therefore, during the training of the CNIBP model 124, feature 600, derived from the PPG signal received from the patient 101 at a specific time, is paired with the target blood pressure for the patient 101 at the same specific time. Similarly, feature 600, derived from the PPG signal indicating the patient 101's blood oxygen saturation at a specific time, is input into the machine-trained model to determine the patient 101's blood pressure at that specific time.

[0085] While training system 200 can train deep learning architecture 500 using feature set 600 derived from PPG signals and associated target blood pressure values, training system 200 can improve the accuracy of its generated machine-trained model by also using features derived from calibration data at the most recent calibration point to train CNIBP model 124 and by using these features derived from calibration data as input to CNIBP model 124 to determine the patient's blood pressure. Continuous noninvasive blood pressure monitoring systems, such as CNIBP monitoring device 100, can periodically determine calibration data from a patient, such as patient 101, at calibration points. Features from the calibration data may include, for example, values ​​of a set of indicators derived from the PPG signal received from the patient at the calibration point and the patient's blood pressure at the calibration point.

[0086] As discussed in this disclosure, each time the CNIBP monitoring device 100 determines calibration data from the patient 101, referred to herein as a calibration point. At each calibration point, the CNIBP monitoring device 100 receives a PPG signal from an oxygen saturation sensing device 150 that measures the oxygen saturation of the patient 101, and measures the blood pressure of the patient 101 using a blood pressure sensing device 152. The CNIBP monitoring device 100 determines values ​​for a set of indicators based on the received PPG signal, and the values ​​of the set of indicators derived from the PPG signal, along with the measured blood pressure, form the calibration data determined at the calibration point.

[0087] Because the CNIBP monitoring device 100 receives both the PPG signal from the oxygen saturation sensing device 150 and the blood pressure of the patient 101 using the blood pressure sensing device 152 at each calibration point, the characteristics of the PPG signal received from the patient 101 are effectively mapped to the patient 101's actual measured blood pressure at each calibration point. Therefore, by using the calibration data received at the most recent calibration point as additional input for training the CNIBP model 124, the CNIBP model 124 can improve its accuracy in determining the patient's blood pressure.

[0088] like Figure 6B As shown, features 602 and 606 can be features extracted from the calibration point. Specifically, feature 602 can be a set of values ​​derived from the PPG signal at the calibration point over a specific time period. For example, feature 602 can include features 602A-602M, which can include indicators derived from the PPG signal, such as PPG pulse duration, the relative position of the PPG with the maximum upward slope of the systolic blood pressure rise, the location and amplitude of the PPG peak, the PPG perfusion index, the PPG baseline trend, PPG respiratory cycle information, the PPG ascending region, the PPG descending region, the maximum gradient of the PPG ascending slope, and any combination of PPG baseline values.

[0089] Each of features 602A-602M may include a value sequence over time for a specific indicator derived from the PPG signal. The value sequence over time may be a value sequence over a specific time period, such as the cardiac cycle. Because each of features 602A-602M includes a value sequence over a time period for a specific indicator, each of features 602A-602M may be a feature vector of the value sequence over the time period, and features 602 as a whole may be referred to as a feature matrix composed of feature vectors (e.g., features 602A-602M).

[0090] Feature 602 is similar to feature 600 in that feature 600 may include a time-varying sequence of the same set of indicators as feature 600. For example, if feature 600 includes a value sequence of the PPG upward region, a value sequence of the PPG downward region, a value sequence of the PPG amplitude, and a value sequence of the PPG maximum upward slope, then feature 600 may also correspondingly include a value sequence of the PPG upward region, a value sequence of the PPG downward region, a value sequence of the PPG amplitude, and a value sequence of the PPG maximum upward slope.

[0091] Therefore, features 602 and 600 can each have the same number of feature vectors, where the feature vectors of features 602 and 600 include value sequences for the same set of indicators. For example, if feature 600 includes 5 feature vectors containing value sequences of five indicators, then feature 602 can also correspondingly include 5 feature vectors containing value sequences of the same five indicators.

[0092] Feature 606 extracted from the calibration point can be a patient's blood pressure value measured at the calibration point. Blood pressure can be any one or more of the patient's systolic (SP), diastolic (DP), mean arterial pressure (MAP), pulse pressure (PP), etc., measured using a non-invasive blood pressure monitoring system, such as an inflatable cuff blood pressure monitoring system. Feature 606 can include copies of the patient's blood pressure value for each feature of feature 600 or feature 602, such that feature 606 includes as many copies of the patient's blood pressure value as the features in feature 600 or feature 602. For example, if feature 600 includes 6 feature vectors, then feature 602 also includes 6 feature vectors, and feature 606 includes six copies of the patient's blood pressure value, one for each feature of feature 600 or feature 602. The six copies of the patient's blood pressure value can be the patient's blood pressure value measured six times repeatedly.

[0093] Feature 604 can be a value corresponding to the time elapsed since the most recent calibration point. This value can be milliseconds, seconds, minutes, etc. Because calibration points can occur periodically, such as every 3 minutes, every 5 minutes, every 15 minutes, every 30 minutes, every hour, every 2 hours, every 4 hours, every 8 hours, etc., the most recent calibration point can be the time when the processing circuit 110 uses the CNIBP model 124 to determine the patient's blood pressure without any intermediate calibration points between calibration points, and the time before the processing circuit 110 uses the CNIBP model 124 to determine the patient's blood pressure. For example, if a second calibration point is reached after a first calibration point, and if the CNIBP model 124 is used to determine the patient's blood pressure at a specific time after reaching the second calibration point but before reaching the third calibration point, then the second calibration point is the most recent calibration point at the specific time when the CNIBP model 124 uses to determine the patient's blood pressure.

[0094] Similar to feature 606, feature 604 may include copies of the time elapsed since the most recent calibration point for each feature of feature 600 or feature 602, such that feature 604 includes the same number of time elapsed since the most recent calibration point as features 600 or feature 602. For example, if feature 600 includes 6 feature vectors, then feature 602 also includes 6 feature vectors, and feature 604 includes six copies of the time elapsed since the most recent calibration point, one copy for each feature of feature 600 or feature 602. The six copies of the time elapsed since the most recent calibration point may be the time elapsed since the most recent calibration point was repeated six times.

[0095] Training system 200 can concatenate features 600, 602, 604, and 606 and use the concatenation of features 600, 602, 604, and 606 to train CINBP model 124. (See above reference.) Figure 2 As described, in some instances, the training module 212 of the training system 200 can train the CINBP model 124 using training data 214 from either patients 101 alone or from a patient population. Each piece of training data in the training data 21 may include a concatenated association of the target blood pressure with features 600, 602, 604, and 606 to train the CINBP model 124 to determine blood pressure based on features 600, 602, 604, and 606.

[0096] Similarly, the CNIBP monitoring device 100 (e.g., processing circuitry 110) can use features 600, 602, 604, and 606 as inputs to the CNIBP model 124 to determine the patient's blood pressure at time t. As described above, the CNIBP monitoring device 100 can determine feature 600 based on the PPG signal received at time t from an oxygen saturation sensing device 150 (e.g., a pulse oximeter) attached to the patient. For example, the CNIBP monitoring device 100 can determine feature 600 by deriving a set of patient parameters from the PPG signal at time t.

[0097] The CNIBP monitoring device 100 can also determine features 602 and 606 based on calibration data obtained at the nearest calibration point from time t. As described above, the nearest calibration point from time t can be a calibration point before time t and also the nearest previous calibration point from time t. The CNIBP monitoring device 100 can determine feature 602 based on the PPG signal received from the oxygen saturation sensing device 150 (e.g., pulse oximeter) attached to the patient at the nearest calibration point from time t. For example, the CNIBP monitoring device 100 can determine feature 602 by deriving a set of patient parameters from the PPG signal at the nearest calibration point.

[0098] The CNIBP monitoring device 100 may also define feature 606 as the patient's blood pressure measurement at the most recent calibration point. For example, the CNIBP monitoring device 100 may use a non-invasive blood pressure monitoring system, such as an inflatable cuff blood pressure monitoring system, to measure the patient's blood pressure. The CNIBP monitoring device 100 may also define feature 604 as the value of the time elapsed at time t since the most recent calibration point. For example, the CNIBP monitoring device 100 may obtain feature 604 by subtracting time t from the time at the most recent calibration point.

[0099] The CNIBP monitoring device 100 can use the CNIBP model 124 to determine the patient's blood pressure at time t. The CNIBP monitoring device 100 can input features 600, 602, 604, and 606 into the CNIBP model 124, and the CNIBP model 124 can output the patient's blood pressure at time t as a response.

[0100] Figure 7A and 7B Example figures 700A, 700B, 750A, and 750B are shown, depicting improved accuracy in determining a patient's blood pressure using the CNIBP model 124 disclosed herein. Figure 7A As shown, graph 700A depicts a patient's blood pressure (e.g., mean arterial pressure, systolic pressure, diastolic pressure, etc.) measured by a blood pressure sensing device (e.g., an arterial line), and the patient's blood pressure determined at least in part based on the patient's PPG signal using a CNIBP model that has not yet been trained with calibration data. Meanwhile, graph 700B depicts a patient's blood pressure (e.g., mean arterial pressure, systolic pressure, diastolic pressure, etc.) measured by a blood pressure sensing device (e.g., an arterial line), and the patient's blood pressure determined according to the techniques disclosed herein, for example, by using... Figure 5 The neural network shown in the medium-depth learning architecture 500 is based at least in part on the patient's PPG signal and the patient's blood pressure determined by the CNIBP model 124, which has not yet been trained with calibration data.

[0101] exist Figure 7A and 7B In the example, the CNIBP model 124 can be implemented using MATLAB. Furthermore, the patient's PPG signal can be acquired using a pulse oximeter, and the features of the CNIBP model 124 can be derived from the patient's 15 pulses (i.e., cardiac cycles). It can be seen that the patient's blood pressure determined by the CNIBP model 124 trained with calibration data, as shown in Figure 700B, tracks the patient's blood pressure much more closely than the patient's blood pressure determined by the CNIBP model trained without calibration data, as shown in Figure 700A.

[0102] like Figure 7BAs shown, graph 750A depicts the measurement of the difference between the blood pressure of a group of patients determined by a CNIBP model that has not been trained using calibration data and the blood pressure of patients measured by a blood pressure sensing device. Conversely, graph 750B depicts the measurement of the difference between the blood pressure of the same group of patients determined by a CNIBP model 124 trained using calibration data and the blood pressure of patients measured by a blood pressure sensing device. The measurements of the differences shown in Figures 750A and 750B can be determined, for example, by the root mean square deviation (RMSD). As can be seen from Figures 750A and 750B, the blood pressure of the group of patients determined by the CNIBP model 124 trained using calibration data is much smaller than the blood pressure of the same group of patients determined by the CNIBP model that has not been trained using calibration data.

[0103] Figure 8 The diagram shows... Figure 1 and Figure 2 This is another example of the CNIBP model 124 deep learning architecture 800. Deep learning architecture 800 differs from... Figure 5 The deep learning architecture 500 shown differs in that features 604 and 606 are separate inputs to the neural network of deep learning architecture 800, and they are subsequently combined in the network, for example, by adding layers or cascading layers. Compared to the neural network of deep learning architecture 500 in which features 600, 602, 604, and 606 are cascaded and input into the neural network of deep learning architecture 500, this allows the neural network to more easily learn the relationship between two feature sets, such as the relationship between feature 600 determined at a specific time and feature 602 determined at the most recent calibration point for determining a patient's blood pressure.

[0104] like Figure 8 As shown, the deep learning architecture 800 includes a sequence input layer 802, a sequence input layer 804, one or more Long Short-Term Memory (LSTM) layers 806, one or more LSTM layers 808, a merging layer 810, a dense layer 816, and a regression output layer 818. The sequence input layer 802 is connected to one or more LSTM layers 806. The sequence input layer 804 is connected to one or more LSTM layers 808. The one or more LSTM layers 806 and one or more LSTM layers 808 are connected to the merging layer 810. The merging layer 810 is connected to the dense layer 816. The dense layer 816 is connected to the regression output layer 818.

[0105] Sequence input layers, such as sequence input layer 802 and sequence input layer 804, input sequence data into the neural network. Therefore, sequence input layers 802 and 804 receive features used to train the deep learning architecture 800. Figure 8In the example, sequence input layer 802 receives feature 600, which contains values ​​of a set of patient indicators derived from the PPG signal received at the current time, while sequence input layer 804 receives feature 602, which contains values ​​of the same set of patient indicators derived from the PPG signal received at the most recent calibration point.

[0106] LSTM layers, such as one or more LSTM layers 806 and one or more LSTM layers 808, learn the long-term dependencies between time steps and sequence data in a time series. LSTM layers perform additive interactions, which can help improve gradient flow on long sequences during training. Figure 8 In the example, one or more LSTM layers 806 can receive input from the sequence input layer 802, and one or more LSTM layers 808 can receive input from the sequence input layer 804.

[0107] Merging layers, such as merging layer 810, can also be called append layers or concatenated layers. Merging layers add inputs from multiple neural network layers element-wise. Figure 8 In the example, the merging layer 810 can add inputs from one or more LSTM layers 806, one or more LSTM layers 808, feature 604, and feature 606. Feature 604 can be the time elapsed since the time of the most recent calibration point to the current time. Feature 606 can be the patient's blood pressure measured at the most recent calibration point.

[0108] Dense layers, such as a dense layer 816, are regular layers of neurons in a neural network. Each neuron in a neuron layer can receive input from all neurons in the previous layer. Figure 8 In the example, each neuron in the dense layer 816 can receive input from all neurons in the merged layer 810.

[0109] Regression output layers, such as regression output layer 818, compute the semi-mean squared error loss of the regression problem and output the predicted response of the trained regression network as a result of training the deep learning architecture 800. Figure 8 In the example, the regression output layer 818 receives the output of the dense layer 816 and calculates the half mean square error loss of the regression problem to output the patient's predicted blood pressure at the current time.

[0110] In some instances, instead of training the CNIBP model 124 to predict the blood pressure of patient 101 at the current time, the training system 200 can train the CNIBP model 124 to predict the blood pressure (BP) of patient 101 measured at the most recent calibration point. 校准 The difference in blood pressure (BP) between patient 101 and patient 101 at the current time. 变化The processing circuit 110 can then add the patient's blood pressure measured at the most recent calibration point to the difference between the patient's blood pressure measured at the most recent calibration point and the patient's blood pressure at the current time (BP). 校准 +BP 变 (The method) determines the patient's blood pressure at the current time.

[0111] Figure 9 This is a flowchart illustrating an example method for determining a patient's blood pressure using the CNIBP model 124. Although Figure 9 Regarding CNIBP monitoring device 100 ( Figure 1 The processing circuit 110 is described in this description, but in other instances, different processing circuits, either alone or in combination with the processing circuit 110, may perform the operation. Figure 9 Any part of the technology. Figure 9 The technique illustrated includes determining calibration data for a continuous noninvasive blood pressure model 124 at calibration points via processing circuitry 110 at least by: receiving a blood pressure signal (indicating the blood pressure of patient 101) from blood pressure sensing device 152 via processing circuitry 110 at the calibration points; receiving a first PPG signal (indicating the blood oxygen saturation of patient 101) from oxygen saturation sensing device 150 via processing circuitry 110 at the calibration points; and deriving a first value (902) of a set of indicators for patient 101 from the first PPG signal via processing circuitry. In some instances, the calibration point may encompass a specified time period, such as a single cardiac cycle.

[0112] In some instances, the continuous noninvasive blood pressure model 124 is a neural network algorithm trained on training data by machine learning, the training data including at least a calibration dataset at the most recent calibration point for patient 101 (patient-specific data) and / or patient group, a time set elapsed since the most recent calibration point, a PPG signal set for patient 101 and / or patient group, and a target blood pressure value set for patient 101 and / or patient group.

[0113] In some instances, a set of indicators for patient 101 includes one or more of the following: PPG pulse duration, PPG relative position of the maximum upward slope of systolic blood pressure rise, PPG peak position and amplitude, PPG perfusion index, PPG baseline trend, PPG respiratory cycle information, PPG ascending region, PPG descending region, maximum gradient of PPG ascending slope, or PPG baseline value.

[0114] Figure 9 The technique shown also includes receiving a second PPG signal (904) from the oxygen saturation sensing device 150 via the processing circuit 110 at a specific time after the calibration point. Figure 9The technique shown also includes deriving a second set of values ​​(906) of a set of indicators of patient 101 from the second PPG signal via processing circuitry 110. In some instances, the first values ​​of the set of indicators comprise a first plurality of value sequences over time, and the second values ​​of the set of indicators comprise a second plurality of value sequences over time.

[0115] Figure 9 The technique shown also includes determining the blood pressure (908) of the patient 101 at the specific time by using the continuous non-invasive blood pressure model 124 through the processing circuit 110 and based at least in part on the calibration data determined at the calibration point, a second value of the set of indicators, and the time elapsed since entering the continuous non-invasive blood pressure model from the calibration point at the specific time.

[0116] In some instances, the technique shown in Figure 3 also includes the processing circuit 110 periodically determining calibration data for a continuous non-invasive blood pressure model 124 at multiple calibration points by at least the following manner: periodically receiving blood pressure measurements of patient 101 from blood pressure sensing device 152 by the processing circuit 110, periodically receiving PPG signals of patient 101 from oxygen saturation sensing device 150 by the processing circuit, and periodically deriving a set of index values ​​of patient 101 from the PPG signals by the processing circuit 110, wherein the blood pressure measurements of patient 101 at the calibration points and a first set of index values ​​of patient 101 include calibration data at the most recent calibration point at a specific time among the multiple calibration points.

[0117] In some instances, Figure 9 The technology shown also includes, wherein the blood pressure measurement includes a first blood pressure measurement, the calibration point includes a first calibration point among a plurality of calibration points, the specific time includes a first specific time, and the processing circuit 110 determines calibration data of the continuous non-invasive blood pressure model 124 at a second calibration point among the plurality of calibration points by at least the following manner: at the second calibration point, the processing circuit 110 receives a second blood pressure measurement value of the patient 101 from the blood pressure sensing device 152; at the second calibration point, the processing circuit 110 receives a third PPG signal from the oxygen saturation sensing device 150; and the processing circuit 110 derives a third value of a set of indicators of the patient 101 from the third PPG signal. The second calibration point may be at a time later than the first calibration point.

[0118] In some instances, Figure 9 The technique illustrated also includes receiving a fourth PPG signal from the oxygen saturation sensing device 150 via processing circuitry 110 at a second specific time following the second calibration point, wherein the second calibration point is the most recent calibration point at the second specific time. In some instances, Figure 9 The technology shown also includes deriving a fourth value of the set of indicators of patient 101 from the fourth PPG signal via processing circuitry 110.

[0119] In some instances, Figure 9 The technique shown also includes determining the blood pressure of patient 101 at the second specific time by using the continuous non-invasive blood pressure model 124 through processing circuitry 110 and based at least in part on calibration data determined at the second calibration point, a fourth value of the set of indicators, and the time elapsed since entering the continuous non-invasive blood pressure model 124 from the second calibration point at the second specific time.

[0120] The techniques described in this disclosure, including those belonging to device 100, processing circuitry 110, control circuitry 122, sensing circuitry 140, 142, or various constituent components, can be implemented at least in part in hardware, software, firmware, or any combination thereof. For example, various aspects of the techniques can be implemented within one or more processors, including one or more microprocessors, DSPs, ASICs, FPGAs, or any other equivalent integrated or discrete logic circuitry, and any combination of such components embodied in a programmer such as a clinician or patient programmer, medical device, or other device. For example, the processing circuitry system, control circuitry system, and sensing circuitry system, as well as other processors and controllers described herein, can be implemented at least in part as or include one or more executable applications, application modules, libraries, classes, methods, objects, routines, subroutines, firmware, and / or embedded code.

[0121] In one or more instances, the functions described herein may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software, the functions may be stored as one or more instructions or code on a computer-readable medium and executed by a hardware-based processing unit. The computer-readable medium may be an article of manufacture comprising a non-transitory computer-readable storage medium encoded with instructions. Instructions embedded or encoded in an article of manufacture comprising an encoded non-transitory computer-readable storage medium may enable one or more programmable processors or other processors to implement one or more of the techniques described herein, for example, when the instructions included or encoded in the non-transitory computer-readable storage medium are executed by one or more processors. Examples of non-transitory computer-readable storage media may include RAM, ROM, programmable ROM (PROM), erasable programmable ROM (EPROM), electronically erasable programmable ROM (EEPROM), flash memory, hard disk, optical disk ROM (CD-ROM), floppy disk, magnetic tape, magnetic media, optical media, or any other computer-readable storage device or tangible computer-readable medium.

[0122] In some instances, computer-readable storage media comprises non-transitory media. The term "non-transitory" can indicate that the storage media is not implemented in a carrier wave or propagating signal. In some examples, non-transitory storage media can store data that can change over time (e.g., in RAM or cache).

[0123] The functionality described herein may be provided within dedicated hardware and / or software modules. Describing different features as modules or units is intended to highlight different functional aspects and does not necessarily imply that such modules or units must be implemented by separate hardware or software components. Rather, the functionality associated with one or more modules or units may be performed by separate hardware or software components, or integrated within shared or separate hardware or software components. Similarly, the technology may be implemented entirely within one or more circuit or logic elements.

[0124] Example 1: In a first example, a method includes determining calibration data for a continuous noninvasive blood pressure model at a calibration point, wherein determining the calibration data includes: receiving a patient's blood pressure measurement from a blood pressure sensing device at the calibration point; receiving a first photoplethysmography (PPG) signal from an oxygen saturation sensing device at the calibration point; and deriving a first value for a set of patient parameters from the first PPG signal. The method further includes receiving a second PPG signal from the oxygen saturation sensing device at a specific time after the calibration point; deriving a second value for a set of patient parameters from the second PPG signal; and determining the patient's blood pressure at the specific time using the continuous noninvasive blood pressure model and based at least in part on the calibration data determined at the calibration point, the second value for the set of parameters, and the time elapsed since entering the continuous noninvasive blood pressure model from the calibration point at the specific time.

[0125] Example 2: In some of the examples described in Example 1, the continuous non-invasive blood pressure model includes a neural network algorithm trained on training data by machine learning, the training data including at least a calibration dataset at the most recent calibration point of the patient population, a time set elapsed since the most recent calibration point, a PPG signal set of the patient population, and a target blood pressure value set of the patient population.

[0126] Example 3: In some of the examples described in Example 1 or Example 2, the continuous non-invasive blood pressure model includes a neural network algorithm trained on training data by machine learning, the training data including at least a calibration dataset at the patient's calibration point, a time set elapsed since the calibration point, the patient's PPG signal set, and the patient's target blood pressure value set.

[0127] Example 4: In some of the examples of any one of Examples 1-3, the first value of the set of indicators includes a first plurality of value sequences over time; and the second value of the set of indicators includes a second plurality of value sequences over time.

[0128] Example 5: In some of the examples described in Example 4, a patient’s set of indicators includes one or more of the following: PPG pulse duration, PPG relative position of the maximum upward slope of systolic blood pressure rise, PPG peak position and amplitude, PPG perfusion index, PPG baseline trend, PPG respiratory cycle information, PPG ascending region, PPG descending region, maximum gradient of PPG ascending slope, or PPG baseline value.

[0129] Example 6: In some examples of any one of Examples 1-5, the method further includes periodically determining calibration data of a continuous non-invasive blood pressure model at multiple calibration points by at least the following manner: periodically receiving blood pressure measurements of a patient from a blood pressure sensing device, periodically receiving PPG signals from an oxygen saturation sensing device, and periodically deriving a set of index values ​​of the patient from the PPG signals, wherein the blood pressure measurements of the patient at the calibration points and a first set of index values ​​of the patient include calibration data at the most recent calibration point at a specific time among the multiple calibration points.

[0130] Example 7: In some of the examples described in Example 6, the blood pressure measurement includes a first blood pressure measurement, the calibration point includes a first calibration point among a plurality of calibration points, and the specific time includes a first specific time. The method further includes: determining calibration data of a continuous non-invasive blood pressure model at a second calibration point among a plurality of calibration points by at least the following manner: receiving a second blood pressure measurement of the patient from a blood pressure sensing device at the second calibration point; receiving a third PPG signal from an oxygen saturation sensing device at the second calibration point; deriving a third value of a set of patient indicators from the third PPG signal; receiving a fourth PPG signal from an oxygen saturation sensing device at a second specific time after the second calibration point, wherein the second calibration point is the most recent calibration point at the second specific time; deriving a fourth value of a set of patient indicators from the fourth PPG signal; and determining the patient's blood pressure at the second specific time using the continuous non-invasive blood pressure model and at least in part based on the calibration data determined at the second calibration point, the fourth value of the set of indicators, and the time elapsed since entering the continuous non-invasive blood pressure model from the second calibration point at the second specific time.

[0131] Example 8: In some of the examples of any one of Examples 1-7, determining a patient's blood pressure at a specific time further comprises: using a continuous non-invasive blood pressure model and based at least in part on the patient's blood pressure measurement at a calibration point, a first value of the set of indicators, a second value of the set of indicators, and the time elapsed from the calibration point to the specific time, determining the predicted blood pressure change between the calibration point and the specific time; and determining the patient's blood pressure at the specific time based at least in part on the patient's blood pressure measurement at the calibration point and the predicted blood pressure change between the calibration point and the specific time.

[0132] Example 9: In some of the examples described in any one of Examples 1-8, the calibration point includes a specified time period.

[0133] Example 10: In another example, a system includes: a blood pressure sensing device; an oxygen saturation sensing device; and processing circuitry configured to: determine calibration data of a continuous noninvasive blood pressure model at a calibration point by at least the following manner: receiving a patient's blood pressure measurement from the blood pressure sensing device at the calibration point; receiving a first photoplethysmography (PPG) signal from the oxygen saturation sensing device at the calibration point; and deriving a first value for a set of patient parameters from the first PPG signal; receiving a second PPG signal from the oxygen saturation sensing device at a specific time after the calibration point; deriving a second value for the set of patient parameters from the second PPG signal; and using the continuous noninvasive blood pressure model and based at least in part on the calibration data determined at the calibration point, determining the second value for the set of parameters, and the patient's blood pressure at the specific time, representing the time elapsed since entering the continuous noninvasive blood pressure model from the calibration point.

[0134] Example 11: In some of the examples described in Example 10, the continuous non-invasive blood pressure model is a neural network algorithm trained on training data by machine learning, the training data including at least a calibration dataset at the most recent calibration point of the patient population, a time set elapsed since the most recent calibration point, a PPG signal set of the patient population, and a target blood pressure value set of the patient population.

[0135] Example 12: In some of the examples described in Example 10 or Example 11, the continuous non-invasive blood pressure model includes a neural network algorithm trained on training data by machine learning, the training data including at least a calibration dataset at the patient's calibration point, a time set elapsed since the calibration point, a set of the patient's PPG signals, and a set of the patient's target blood pressure values.

[0136] Example 13: In some of the examples of any one of Examples 10-12, the first value of the set of indicators includes a first plurality of value sequences over time; and the second value of the set of indicators includes a second plurality of value sequences over time.

[0137] Example 14: In some of the examples described in any of Examples 10-13, a set of patient parameters includes one or more of the following: PPG pulse duration, PPG relative position of the maximum upward slope of systolic blood pressure rise, PPG peak position and amplitude, PPG perfusion index, PPG baseline trend, PPG respiratory cycle information, PPG ascending region, PPG descending region, maximum gradient of the upward slope of PPG, or PPG baseline value.

[0138] Example 15: In some examples of any one of Examples 10-15, the processing circuitry is further configured to periodically determine calibration data of a continuous non-invasive blood pressure model at multiple calibration points by at least the following manner: periodically receiving blood pressure measurements of a patient from a blood pressure sensing device, periodically receiving a PPG signal of a patient from an oxygen saturation sensing device, and periodically deriving a set of index values ​​of the patient from the PPG signal; and wherein the patient's blood pressure measurements at the calibration points and a first value of the patient's set of index values ​​include calibration data at the most recent calibration point at a specific time among the multiple calibration points.

[0139] Example 16: In some of the examples described in Example 15, the blood pressure measurement includes a first blood pressure measurement, the calibration point includes a first calibration point among a plurality of calibration points, the specific time includes a first specific time, and the processing circuitry is further configured to: determine calibration data of a continuous non-invasive blood pressure model at a second calibration point among a plurality of calibration points by at least the following manner: receiving a second blood pressure measurement of the patient from a blood pressure sensing device at the second calibration point; receiving a third PPG signal from an oxygen saturation sensing device at the second calibration point and deriving a third value of a set of patient indicators from the third PPG signal; receiving a fourth PPG signal from an oxygen saturation sensing device at a second specific time after the second calibration point, wherein the second calibration point is the most recent calibration point at the second specific time; deriving a fourth value of a set of patient indicators from the fourth PPG signal; and determining the patient's blood pressure at the second specific time using the continuous non-invasive blood pressure model and at least in part based on the calibration data determined at the second calibration point, the fourth value of the set of indicators, and the time elapsed since entering the continuous non-invasive blood pressure model from the second calibration point at the second specific time.

[0140] Example 17: In some examples of any one of Examples 10-16, the processing circuitry configured to determine a patient's blood pressure at a specific time is further configured to: determine a predicted blood pressure change between a calibration point and a specific time using a continuous non-invasive blood pressure model and based at least in part on the patient's blood pressure measurement at a calibration point, a first value of the set of indicators, a second value of the set of indicators, and the time elapsed from the calibration point at the specific time; and determine the patient's blood pressure at the specific time based at least in part on the patient's blood pressure measurement at the calibration point and the predicted blood pressure change between the calibration point and the specific time.

[0141] Example 18: In some of the examples described in any of Examples 10-17, the calibration point includes a specified time period.

[0142] Example 19: In another example, a non-transitory computer-readable storage medium includes instructions that, when executed, cause processing circuitry to: determine calibration data of a continuous non-invasive blood pressure model at a calibration point by at least the following manner: receiving a patient's blood pressure measurement at the calibration point; receiving a first photoplethysmography (PPG) signal at the calibration point and deriving a first value for a set of patient parameters from the first PPG signal; receiving a second PPG signal at a specific time after the calibration point; deriving a second value for the set of patient parameters from the second PPG signal; and using the continuous non-invasive blood pressure model and based at least in part on the calibration data determined at the calibration point, the second value for the set of parameters, and the time elapsed since entering the continuous non-invasive blood pressure model from the calibration point at the specific time, determining the patient's blood pressure at the specific time.

[0143] Example 20: In some of the examples described in Example 19, the continuous non-invasive blood pressure model is a neural network algorithm trained on training data by machine learning, the training data including at least a calibration dataset at the most recent calibration point of the patient population, a time set elapsed since the most recent calibration point, a PPG signal set of the patient population, and a target blood pressure value set of the patient population.

[0144] Various examples of this disclosure have been described. Consider any combination of the described systems, operations, or functions. These and other examples are within the scope of the appended claims.

Claims

1. A system comprising: a blood pressure sensing device; an oxygen saturation sensing device; and processing circuitry configured to: determine calibration data for a continuous non-invasive blood pressure model at a calibration point at least by: receiving a blood pressure measurement of a patient from the blood pressure sensing device at the calibration point, receiving a first photoplethysmographic signal from the oxygen saturation sensing device at the calibration point, and deriving a first value of a set of indicators of a patient from the first photoplethysmographic signal; receiving a second photoplethysmographic signal from the oxygen saturation sensing device at a particular time after the calibration point; deriving a second value of the set of indicators of the patient from the second photoplethysmographic signal; and determining a blood pressure of the patient at the particular time using the continuous non-invasive blood pressure model and based at least in part on inputting the calibration data determined at the calibration point, the second value of the set of indicators, and a time elapsed since the calibration point into the continuous non-invasive blood pressure model.

2. The system of claim 1, wherein the continuous non-invasive blood pressure model comprises a neural network algorithm trained on training data by machine learning, the training data comprising at least a set of calibration data at a most recent calibration point of a population of patients, a set of times elapsed since the most recent calibration point, a set of photoplethysmographic signals of the population of patients, and a set of target blood pressure values of the population of patients.

3. The system of claim 1, wherein the continuous non-invasive blood pressure model comprises a neural network algorithm trained on training data by machine learning, the training data comprising at least a set of calibration data at a calibration point of the patient, a set of times elapsed since the calibration point, a set of photoplethysmographic signals of the patient, and a set of target blood pressure values of the patient.

4. The system of any of claims 1-3, wherein: the first value of the set of indicators comprises a first plurality of sequences of values over time; and the second value of the set of indicators comprises a second plurality of sequences of values over time.

5. The system of any of claims 1-3, wherein the set of indicators of the patient comprises one or more of: photoplethysmographic pulse duration, photoplethysmographic relative position of maximum upstroke slope of systolic pressure rise, photoplethysmographic peak position and amplitude, photoplethysmographic perfusion index, photoplethysmographic baseline trend, photoplethysmographic respiratory cycle information, photoplethysmographic upstroke region, photoplethysmographic downstroke region, maximum gradient of photoplethysmographic upstroke slope, or photoplethysmographic baseline value.

6. The system of any of claims 1 to 3, wherein the processing circuitry is further configured to: determine calibration data for the continuous non-invasive blood pressure model at a plurality of calibration points at least by: periodically receiving the blood pressure measurement of a patient from the blood pressure sensing device, periodically receiving a photoplethysmographic signal of a patient from the oxygen saturation sensing device, and periodically deriving a value of the set of indicators of a patient from the photoplethysmographic signal; and wherein the blood pressure measurement of the patient at the calibration point and the first value of the set of indicators of the patient comprise calibration data at a most recent calibration point of a particular time in the plurality of calibration points.

7. The system of claim 6, wherein the blood pressure measurement comprises a first blood pressure measurement, the calibration point comprises a first calibration point in the plurality of calibration points, the particular time comprises a first particular time, and the processing circuit is further configured to: determine calibration data for the continuous noninvasive blood pressure model at a second calibration point in the plurality of calibration points by at least: receiving a second blood pressure measurement of the patient from the blood pressure sensing device at the second calibration point, receiving a third photoplethysmographic signal from the oxygen saturation sensing device at the second calibration point, and deriving a third value of the set of indicators of the patient from the third photoplethysmographic signal; receiving a fourth photoplethysmographic signal from the oxygen saturation sensing device at a second particular time after the second calibration point, wherein the second calibration point is a most recent calibration point of the second particular time; deriving a fourth value of the set of indicators of the patient from the fourth photoplethysmographic signal; and determining a blood pressure of the patient at the second particular time using the continuous noninvasive blood pressure model and based at least in part on inputting the calibration data determined at the second calibration point, the fourth value of the set of indicators, and a time elapsed since the second calibration point into the continuous noninvasive blood pressure model at the second particular time.

8. The system of any of claims 1-3, wherein the processing circuit configured to determine a blood pressure of the patient at the particular time is further configured to: determine a predicted blood pressure change between the calibration point and the particular time using the continuous noninvasive blood pressure model and based at least in part on the blood pressure measurement of the patient at the calibration point, the first value of the set of indicators, the second value of the set of indicators, and a time elapsed since the calibration point at the particular time; and determine a blood pressure of the patient at the particular time based at least in part on the blood pressure measurement of the patient at the calibration point and the predicted blood pressure change between the calibration point and the particular time.

9. The system of any of claims 1-3, wherein the calibration point comprises a specified time period.

10. A non-transitory computer-readable storage medium comprising instructions that, when executed, cause a processing circuit to: determine calibration data for a continuous noninvasive blood pressure model at a calibration point by at least: receiving a blood pressure measurement of a patient at the calibration point, receiving a first photoplethysmographic signal at the calibration point, and deriving a first value of a set of indicators of the patient from the first photoplethysmographic signal; receiving a second photoplethysmographic signal at a particular time after the calibration point; deriving a second value of the set of indicators of the patient from the second photoplethysmographic signal; and and determining, using the continuous non-invasive blood pressure model and based at least in part on the calibration data determined at the calibration point, the second value of the set of indicators, and a time elapsed since the calibration point into the continuous non-invasive blood pressure model at the particular time, a patient blood pressure at the particular time.

11. The computer-readable storage medium of claim 10, wherein the continuous non-invasive blood pressure model is a neural network algorithm trained on training data by machine learning, the training data comprising at least a set of calibration data at a most recent calibration point of a patient population, a set of times elapsed since the most recent calibration point, a set of photoplethysmography signals of the patient population, and a set of target blood pressure values of the patient population.

Images