Ai computed blood pressure parameters from non-invasive signals

Abstract

Cerebral autoregulation is a physiological process that ensures stable and adequate blood flow to the brain, despite fluctuations in systemic blood pressure. Non-invasive measurement of cerebral blood flow and related parameters is complex and challenging, often requiring an invasive arterial line (A-line). Using concurrent PPG and ECG signals as inputs to one or more machine learning models, methods and systems non-invasively determine blood pressure (BP) output, including waveform morphology, MAP trending, and predicted BP in mmHg, and real-time quality metrics associated therewith. Based on the BP output, along with other inputs, non-invasive monitoring of cerebral autoregulation is provided.

Description

Attorney Docket No. A0012962W001Al COMPUTED BLOOD PRESSURE PARAMETERS FROM NON-INVASIVE SIGNALSCROSS-REFERENCE TO RELATED APPLICATION

[0001] This application claims the benefit of U.S. Provisional Patent Application Serial No. 63 / 703,080, filed October 3, 2024, the entire content of which is incorporated herein by reference.FIELD

[0002] The present technology is generally related to non-invasive methods for determining patient parameters, such as blood pressure; in particular, the present technology is related to utilizing artificial intelligence (Al) for computing blood pressure and related parameters from PPG and ECG signals.BACKGROUND

[0003] Arterial blood pressure (ABP or BP) is the pressure exerted by circulating blood against the walls of the arterial blood vessels. BP is usually expressed in terms of the systolic blood pressure (SBP), which is the maximum pressure during one heartbeat, over diastolic blood pressure (DBP), which is the minimum pressure between two heartbeats. The unit of measurement for BP is millimeters of mercury (mmHg) above atmospheric pressure. Other common metrics for BP include mean arterial pressure (MAP), which represents the average BP for an individual during a single cardiac cycle, and MAP trending, which represents an individual’s MAP over time. The continuous monitoring of BP can be an important indicator of other physiological conditions, such as heart rate variability (HRV), risk of cardiovascular events, sleep apnea, and cerebral autoregulation. However, non-invasive methods for the accurate determination of continuous BP are challenging, often relying on an invasive arterial line (A-line).SUMMARY

[0004] The present technology relates to methods and systems for determining BP waveform morphology, MAP trending, predicted BP in mmHg, and real-time quality metrics associated with these computations, using concurrent PPG and ECG signals as inputs to one or more machine learning models.

[0005] In one aspect, a system is provided. The system includes at least one processor and memory storing instructions that, when executed by the at least one processor, cause the systemAttorney Docket No. A0012962W001 to perform a set of operations. The set of operations include receiving an electrocardiogram (ECG) signal and a photoplethysmography (PPG) signal, encoding the ECG signal using a first encoder to generate a first embedding, and encoding the PPG signal using a second encoder to generate a second embedding, where the first encoder and the second encoder operate in parallel . The set of operations further includes inputting the first embedding to a first task model configured to generate a first prediction of a physiological parameter and inputting the second embedding to a second task model configured to generate a second prediction of the physiological parameter. Additionally, the set of operations includes decoding the first embedding and the second embedding to generate a blood pressure output and, based on a variance between the first prediction of the physiological parameter and the second prediction of the physiological parameter, determining a quality metric for the blood pressure output.

[0006] In additional aspects, the set of operations includes issuing an alert regarding the blood pressure output based on the variance being greater than a variance threshold. Further, the set of operations includes concatenating the first embedding and the second embedding and decoding the concatenated first embedding and second embedding. In aspects, the blood pressure output is one of a series of blood pressure values, a blood pressure waveform, or a trend of a mean arterial pressure (MAP). In further aspects, an indication of patient cerebral autoregulation is based at least in part on the blood pressure output. Additionally, the physiological parameter is one of a heart rate, a respiratory rate, or a heart rate variability (HRV). In yet additional aspects, the ECG signal is concurrent with the PPG signal. Further, the blood pressure output is continuous. In additional aspects, the first encoder, the second encoder, and the decoder are included in a machine learning (ML) model, where in some aspects the ML model includes one or more of a transformer system, a convolutional neural network (CNN), or a generative model. The set of operations further includes determining a difference between a measured value of the physiological parameter and at least one of the first prediction of the physiological parameter and the second prediction of the physiological parameter. Based on the difference, the set of operations includes determining a status of a sensor that recorded the measured value of the physiological parameter, and / or based on the difference being greater than an error threshold, issuing an alert regarding the status of the sensor.

[0007] In another aspect, a system is provided. The system includes at least one processor and memory storing instructions that, when executed by the at least one processor, cause the system to perform a set of operations. The system further includes a first encoder configuredAttorney Docket No. A0012962W001 to encode an electrocardiogram (ECG) signal and a second encoder configured to encode a photoplethysmography (PPG) signal, where the first encoder and the second encoder operate in parallel. Additionally, the system includes a first task model configured to process the encoded ECG signal to generate a first prediction of a physiological parameter and a second task model configured to process the encoded PPG signal to generate a second prediction of the physiological parameter, where there is a variance between the first prediction and the second prediction. The system further includes a decoder configured to decode the encoded ECG signal and the encoded PPG signal to generate a blood pressure output, where a quality of the blood pressure output is based on the variance.

[0008] In further aspects, the system includes a sensor configured to measure a value of the physiological parameter, where there is a difference between the measured value of the physiological parameter and at least one of the first prediction of the physiological parameter or the second prediction of the physiological parameter. In aspects, a status of the sensor is based on the difference and, when the difference is greater than an error threshold, an alert is issued regarding the status of the sensor.

[0009] In yet another aspect, a method of generating a blood pressure output is provided. The method includes encoding an electrocardiogram (ECG) signal using a first encoder to generate a first embedding and, in parallel, encoding a photoplethysmography (PPG) signal using a second encoder to generate a second embedding. The method further includes inputting the first embedding to a first task model configured to generate a first prediction of a physiological parameter and inputting the second embedding to a second task model configured to generate a second prediction of the physiological parameter. Additionally, the method includes decoding the first embedding and the second embedding to generate a blood pressure output and, based on a variance between the first prediction of the physiological parameter and the second prediction of the physiological parameter, determining a quality metric for the blood pressure output.

[0010] In additional aspects, the method includes, based on the variance being greater than a variance threshold, issuing an alert regarding the blood pressure output. Additionally, the blood pressure output is one of a series of blood pressure values, a blood pressure waveform, or a trend of a mean arterial pressure (MAP). Further, an indication of patient cerebral autoregulation is based at least in part on the blood pressure output.Attorney Docket No. A0012962W001

[0011] The details of one or more aspects of the disclosure are set forth in the accompanying drawings and the description below. It should be understood that various aspects disclosed herein may be combined in different combinations than the combinations specifically presented in the description and accompanying drawings. Other features, objects, and advantages of the techniques described in this disclosure will be apparent from the description and drawings, and from the claims.BRIEF DESCRIPTION OF DRAWINGS

[0012] Non-limiting and non-exhaustive examples are described with reference to the following Figures.

[0013] FIG. 1 illustrates an overview of an example system for non-invasive patient monitoring, according to aspects described herein.

[0014] FIGS. 2A-2C illustrate an overview of example machine learning (ML) architectures for non-invasive patient monitoring, according to aspects described herein.

[0015] FIG. 3 illustrates an overview of an example ML architecture for non-invasive patient monitoring, according to aspects described herein.

[0016] FIGS. 4A-4J illustrate examples of input and output of an example ML model for non-invasive patient monitoring, according to aspects described herein.

[0017] FIGS. 5A-5B illustrate an overview of an example method for non-invasive patient monitoring, according to aspects described herein.DETAILED DESCRIPTION

[0018] Continuous monitoring of BP can be an important indicator of other physiological conditions, such as heart rate variability (HRV), risk of cardiovascular events, sleep apnea, and cerebral autoregulation. Cerebral autoregulation is a physiological process that ensures stable and adequate blood flow to the brain, despite fluctuations in systemic blood pressure. Non- invasive measurement of the adequacy of cerebral blood flow and related parameters is complex and challenging, often requiring an invasive arterial line (A-line). To address these issues, concurrent PPG and ECG signals are used as inputs to one or more machine learning models, which generate BP output including BP waveform morphology, MAP trending, predicted BP in mmHg, and real-time quality metrics associated therewith. In turn, MAP trending may be used as input for non-invasively monitoring cerebral autoregulation.Attorney Docket No. A0012962W001

[0019] An electrocardiogram (ECG) is a non-invasive bioelectric technology for measuring the electrical activity of the heart using small electrodes placed outside the body. Signals from the electrodes are recorded by an electrocardiograph and may be printed or displayed as a waveform. In examples, an ECG signal corresponds to electrical data recorded by the electrocardiograph over time, which may be represented as a waveform in millivolts per second (mV / s). A photoplethysmogram (PPG) is a non-invasive optical technique used to measure blood volume changes in microvascular tissue. PPG is often measured using a pulse oximeter that illuminates the skin with low-intensity infrared (IR) light. Since light is better absorbed by blood than surrounding tissues, changes in blood flow during each heartbeat can be detected as changes in light intensity. In aspects, a PPG signal can be represented as a voltage signal that is proportional to blood volume. The PPG signal can be displayed as a waveform, with a pulsatile component (referred to as the AC component corresponding to changes in light absorption due to synchronous fluctuations in blood volume with each heartbeat) superimposed onto a baseline component (referred to as the DC component corresponding to a baseline level of light absorption influenced, for example, by the subject’s total blood volume, tissue characteristics, thermoregulation, and sympathetic nervous system). In aspects, BP influences both PPG and ECG signals.

[0020] As noted above, methods and systems described herein may be implemented by one or more machine learning (ML) models. ML models are based on various algorithms, such as decision trees, neural networks (NN), convolutional neural networks (CNN), transformers, support vector machines (SVM), and the like, that can be trained to recognize patterns or make predictions based on input data. Training a ML model involves utilizing a dataset (e.g., training data) to enable the ML model to learn relationships between input variables (e.g., input features) and output variables (e.g., labels, if available). Lor example, training data used to prepare a ML model can include labeled data (e.g., supervised learning), unlabeled data (e.g., unsupervised learning), or a combination of labeled and unlabeled data (e.g., semi-supervised learning). The performance of a trained ML model can be evaluated using a validation dataset and, in some cases, further training can be conducted. Once trained and validated, a ML model is able to make predictions and / or identify patterns in new data (e.g., data not included in the training dataset and / or the validation dataset).

[0021] In examples, a generative model (also generally referred to herein as a type of machine learning (ML) model) may be used according to aspects described herein and may generate any of a variety of output types (and may thus be a multimodal generative model, inAttorney Docket No. A0012962W001 some examples). For example, the generative model may include a generative transformer model and / or a large language model (LLM), a generative image model, a generative audio model, or the like. Example generative models include, but are not limited to, Bidirectional Encoder Representations from Transformers (BERT), Megatron-Turing Natural Language Generation model (MT-NLG), Generative Pre-trained Transformer 3 (GPT-3), Generative Pretrained Transformer 4 (GPT-4), BigScience BLOOM (Large Open-science Open-access Multilingual Language Model), DALL-E, DALL-E 2, Stable Diffusion, or Jukebox.

[0022] FIG. 1 illustrates an overview of an example system 100 for non -invasive patient monitoring, according to aspects described herein. As illustrated, system 100 includes a subject 102, e.g., a human individual or an animal, who is being monitored by one or more monitoring devices 104A-104B. In communication with monitoring devices 104A-104B via network 110, system 100 further includes computing device 120, model server 122, and database 112.

[0023] Monitoring devices 104A-104B may be portions of the same monitoring device having multiple capabilities or different monitoring devices each having different capabilities. For example, monitoring device 104A may be an electrocardiograph, which non-invasively records electrical activity of the heart of subject 102 over a period of time. Monitoring device 104A may be a stationary device, a portable device, or a wearable device, for example, and may comprise or be communicatively coupled to one or more computing devices (e.g., computing device 120 and / or model server 122) via network 110, which may comprise a local area network (LAN), a wide area network (WAN) (e.g., the Internet), a wireless network, a cellular network (e.g., radio access network (RAN)), or any combination thereof, among other examples. In examples, monitoring device 104A utilizes small electrodes (e.g., placed on the arms, legs, and chest of subject 102) to detect electrical signals produced by the heart during each cardiac cycle. The electrical signals may be transmitted (e.g., via lead wires or wirelessly) to monitoring device 104A, which may process the electrical signals to output ECG signal 106 in millivolts per second (mV / s).

[0024] In further examples, monitoring device 104B may be a pulse oximeter, which operates by illuminating the skin with low-intensity infrared (IR) light to detect changes in light intensity corresponding to changes in blood flow during each heartbeat over a period of time. Monitoring device 104B may be a stationary device, a portable device, or a wearable device, for example, and may comprise or be communicatively coupled to one or more computing devices (e.g., computing device 120 and / or model server 122) via network 110. MonitoringAttorney Docket No. A0012962W001 device 104B may process the changes in light intensity over a period of time and output PPG signal 108, which may be represented as a voltage signal proportional to blood volume. The PPG signal 108 can be displayed as a waveform, with a pulsatile component (referred to as the AC component corresponding to changes in light absorption due to synchronous fluctuations in blood volume with each heartbeat) superimposed onto a baseline component (referred to as the DC component corresponding to a baseline level of light absorption influenced, for example, by the subject’s total blood volume, tissue characteristics, thermoregulation, and sympathetic nervous system)..

[0025] In aspects, the ECG signal 106 (e.g., ECG waveform) may be recorded concurrently with the PPG signal 108 (e.g., PPG waveform). In this case, the ECG signal 106 may be recorded over a period of time (or window) and the PPG signal 108 may be recorded over a concurrent period of time (or concurrent window). In further aspects, the period of time may contain several heartbeats so that ML model 146, described further below, can derive a relationship between BP, ECG signal 106, and PPG signal 108 over several cardiac cycles. In still further aspects, ECG signal 106 and PPG signal 108 may be recorded continuously or periodically. In this case, a plurality of ECG signals 106 may be recorded by monitoring device 104A over a plurality of time periods, and a plurality of PPG signals 108 may be recorded by monitoring device 104B over a plurality of concurrent time periods, where each time period and concurrent time period spans a plurality of cardiac cycles (e.g., 5 seconds, 10 seconds, 15 seconds, 20 seconds, or any other suitable time period). In some aspects, the plurality of time periods and the plurality of concurrent time periods may be sequential, overlapping, or intermittent, for example. The ECG signal 106 and the PPG signal 108 may be stored as signal data 114 on database 112 and / or sent to ML manager 156 of computing device 120, for example.

[0026] Database 112 may implement any suitable mechanism for facilitating storage, retrieval, modification, and / or deletion of various types of data that are measured, recorded, computed, etc., according to aspects described herein. As illustrated, database 112 may store signal data 114, prediction data 116, and blood pressure output 118. In aspects, signal data 114 may include ECG signal data, PPG signal data, and / or encoded forms of ECG signals and PPG signals. In future aspects, prediction data 116 may include first predictions of a physiological parameter and second predictions of a physiological parameter, where the physiological parameter is derivable from ECG and PPG signals (e.g., heart rate, respiratory rate, etc.). Moreover, database 112 may also store other forms of data (not shown), including but notAttorney Docket No. A0012962W001 limited to training data, validation data, related data (e.g., oxygen saturation (SpCh), bicarbonate saturation, partial pressure of carbon dioxide (PaCCh), and the like), context (e.g., input features for prompt generation), curated databases (e.g., patient specific data, population specific data, physiological parameter specific data, etc.), downstream data (e.g., cerebral autoregulation data derived from BP output 118), or any combination thereof.

[0027] Computing device 120 includes various computing components (e.g., system memory 124, processor(s) 130, interface(s) 132, output device(s) 134) and applications or programing modules (e.g., ML manager 156). Computing device 120 may be a server computing device, a personal computing device, a tablet computing device, a mobile computing device (e.g., cellular telephone), a wearable computing device (e.g., smart watch), an Intemet-of-Things (loT) device (e.g., appliance, camera), or any combination thereof. Computing device 120 may include at least one processor 130 and a system memory 124. Depending on the configuration and type of computing device 120, system memory 124 may include an operating system 126 and storage 128, such as volatile storage (e.g., random access memory, RAM), non-volatile storage (e.g., read-only memory, ROM), flash memory, or any combination of thereof. Interface(s) 132 associated with computing device 120 may allow communication via network 110 with other computing devices, e.g., monitoring devices 104A- 104B, database 112, and / or model server 122. Example interface(s) 132 include, but are not limited to, radio frequency (RF) transmitters, receivers, and / or transceivers; universal serial bus (USB); parallel and / or serial ports; high-definition multimedia interface (HDMI), Ethernet; Wi-Fi; Bluetooth; and any combination thereof. Output device(s) 134, such as printer(s), display(s), speaker(s), etc., enable the computing device 120 to convey data, computations, predictions, metrics, notifications, alerts, or any combination thereof, to clinicians, developers, or other users (not shown).

[0028] For example, signal receiver 136 of ML model 156 may receive ECG signal 106 and PPG signal 108 from monitoring devices 104A-104B and / or database 112. In some cases, signal receiver 136 may normalize, filter, clean, augment, or otherwise process the ECG signal 106 and / or the PPG signal 108 before sending ECG signal 106 and PPG signal 108 to model server 122 for encoding. In other cases, model server 122 may process the ECG signal 106 and / or the PPG signal 108 before encoding.

[0029] As described above, computing device 120 may execute or host various applications and / or program modules. For example, computing device 120 may host ML manager 156,Attorney Docket No. A0012962W001 which may orchestrate the execution of various ML models to compute a BP output from ECG signal 106 and PPG signal 108 inputs, for example. It should be appreciated that the disclosed aspects may be implemented according to any of a variety of paradigms. In examples, any stage of input processing may be performed client side (e.g., by ML manager 156), server side (e.g., by ML server 122), or any combination thereof. Lor instance, model manager 156 may perform a first ML evaluation associated with signal input to provide a model output (e.g., a heart rate prediction or a respiratory rate prediction), while a second ML evaluation (e.g., computation of BP output) may be performed by ML server 122. In aspects, ML manager 156 may communicate with model server 122 via one or more application programing interfaces (APIs) and / or network 110, among other examples, and may send and receive various inputs, outputs, context, prompts, or any combination thereof, with model server 122.

[0030] Model server 122 may be hosted on one or more server computing devices (not shown) and may be configured with processors and memory for executing any number of different ML models. Lor example, model server 122 may execute neural networks (NNs), convolutional neural networks (CNNs), transducer models, single-task models, multimodal generative models (or foundation models), language models, speech models, video models, audio models, and any combination thereof. As used herein, a foundation model is a model that is pre-trained on broad data that can be adapted to a wide range of tasks (e.g., models capable of processing various different tasks or modalities). In examples, a foundation model may have been trained using training data having a plurality of content types. Thus, for input associated with given content of a content type, an ML model may generate model output having any of a variety of associated content types. It will be appreciated that model server 122 may execute any of the various types of ML models, including but not limited to those described above. In aspects, model server 122 executes at least ML model 146, which may be a CNN or a transformer model, for example, and task models 152A-152B, which may be models that have been finetuned (e.g., for a specific context, specific datatype, and / or a specific output), among other examples.

[0031] As described above, model server 122 may be in communication with ML manager 156 of computing device 120. Lor example, signal receiver 156 of ML manager 156 may communicate ECG signal 106 and PPG signal 108 to ML model 146 of model server 122. As illustrated ML model 146 includes first encoder 148A, second encoder 148B, and decoder 150; however, ML model 146 may include additional modules or components. In aspects, as further described with reference to PIG. 2, first encoder 148A and second encoder 148B are parallelAttorney Docket No. A0012962W001 encoder stacks, with the first encoder 148A processing (e.g., encoding) ECG signal 106 in parallel with the second encoder 148B processing (e.g., encoding) PPG signal 108. The first encoder 148A and second encoder 148B may be independent; or, as further described with reference to FIG. 2B, first encoder 148A and second encoder 148B may be linked to share information at different encoding layers. Additionally or alternatively, as further described with reference to FIG. 2C, outputs from intermediate encoding layers of first encoder 148A and / or second encoder 148B may be fed into decoder 150.

[0032] Upon encoding ECG signal 106 by the first encoder 148A, and PPG signal 108 by the second encoder 148B, the encoded ECG signal is concatenated to the encoded PPG signal and fed as input to decoder 150. Decoder 150 then reconstructs the encoded ECG signal concatenated to the encoded PPG signal to predict a BP output (e.g., BP output 118). As described above, the BP output 118 may include, but is not limited to, an indication of BP waveform morphology, an indication of mean arterial pressure (MAP) trending, and / or predicted values of BP in mmHg. For example, ML model 146 may receive as input a short concurrent window of ECG and PPG signals (e.g., ECG signal 106 and PPG signal 108) and may output a corresponding window of BP output 118, e.g., as a sequence of values using a Sequence-to-Sequence (Seq2Seq) model. In further examples, ML model 146 may receive as input a set of ECG signals 106 collected over a plurality of windows and a set of PPG signals 108 collected over a plurality of concurrent windows and may output a set of corresponding windows of BP output 118. In aspects, the BP output 118 may be displayed and / or printed (e.g., by BP output manager 142 via output devices 134). In some examples, the sequence of values associated with BP output 118 may be displayed and / or printed; in other examples, the last value (e.g., most recent value) of the sequence of values associated with BP output 118 may be displayed and / or printed. Additionally, BP output manager 142 may store BP output 118 on database 112, transmit BP output 118 to cerebral autoregulation monitor 144, and / or transmit BP output 118 (e.g., via interfaces 132) to one or more other applications and / or other computing devices (e.g., a mobile computing device, a medical device, a server, etc.) (not shown).

[0033] In aspects, cerebral autoregulation monitor 144 may be associated with ML manager 156, or may be an independent application, program module, algorithm or ML model. Cerebral autoregulation monitor 144 may receive BP output 118 (e.g., from BP output manager 142) and may process BP output 118, along with other inputs (e.g., cerebral blood flow, cerebral oxygen saturation, etc.), to determine an indication of cerebral autoregulation associated withAttorney Docket No. A0012962W001 subject 102. As described above, cerebral autoregulation is a physiological process that ensures stable and adequate blood flow to the brain, despite fluctuations in systemic blood pressure. Non-invasive measurement of cerebral blood flow and related parameters is complex and challenging, often requiring an invasive arterial line (A-line). By substantially continuously recording and feeding ECG signals 106 and PPG signals 108 into ML model 146, BP output 118 may be substantially continuously computed. For example, BP output 118 may include, but is not limited to, a substantially continuous indication of BP waveform morphology, a substantially continuous trending of mean arterial pressure (MAP), and / or substantially continuous predictions (e.g., sequential values) of BP in mmHg. Based on the substantially continuous BP output 118, and additional inputs such as oxygen saturation, cerebral autoregulation may be non-invasively monitored in near real-time.

[0034] In some examples, BP output 118 may be evaluated for quality (e.g., by quality metric determiner 140) prior to displaying, printing, or transmitting BP output 118. For example, a loss function (e.g., mean squared error (MSE), Kullback-Leibler divergence, cosine proximity, or the like) may be used to compare an actual blood pressure of the subject 102 measured during the same time period represented by the BP output 118. If the BP output 118 diverges from the actual blood pressure by more than a divergence threshold, it may be determined that a first quality metric of the BP output 118 is less than a quality threshold. If the first quality metric is less than the quality threshold, a first alert may be issued that warns a clinician of potential unreliability of the BP output 118 and / or recommends alternative means for obtaining BP for the subject 102. Additionally, the first alert may recommend evaluating monitoring devices 104A-104B for proper functioning, for example. In other examples, if the first quality metric is less than the quality threshold, the BP output 118 may be dropped (e.g., deleted) and / or it may be determined that additional training of ML model 146 may be warranted.

[0035] Model server 122 also includes one or more task models 152A-152B. In some examples, task models 152A-152B may each be specialized for performing a specific task, such as generating a specific output from a specific type of input. For instance, first task model 152Amay receive the encoded ECG signal from first encoder 148Aof ML model 146. Decoder 154A associated with first task model 152Amay process (e.g., decode) the encoded ECG signal to output a first prediction of a physiological parameter. In aspects, the physiological parameter may be any parameter that can be derived from an ECG signal, including but not limited to heart rate, respiratory rate, heart rate variability (HRV), for example. In other examples, firstAttorney Docket No. A0012962W001 task model 152A may reconstruct the ECG signal 106 from the encoded ECG signal. Potentially more relevant to the training phase, the encoded ECG signal may also be utilized to predict an anatomical site of an arterial line (A-line), such as radial, ulnar, etc.

[0036] In parallel (or concurrently), second task model 152B may receive the encoded PPG signal from second encoder 148B of ML model 146. That is, the encoded ECG signal and the encoded PPG signal should represent data recorded over a concurrent time period. Decoder 154B associated with second task model 152B may process (e.g., decode) the encoded PPG signal to output a second prediction of the physiological parameter. In aspects, the second prediction may be for the same physiological parameter over the same time period as the first prediction. That is, in examples, if the first prediction is a first heart rate over a time period, the second prediction is a second heart rate over the same time period; if the first prediction is a first respiratory rate over a time period, the second prediction is a second respiratory rate over the same time period; and if the first prediction is a first HRV over a time period, the second prediction is a second HRV over the same time period. In other examples, second task model 152B may reconstruct the PPG signal 108 from the encoded PPG signal and / or may utilize the encoded PPG signal to predict an anatomical site of an arterial line (A-line), such as radial, ulnar, etc.

[0037] In some examples, decoder 154A may generate a set of first predictions for the physiological parameter based on a plurality of ECG signals 106 recorded over a plurality of time periods, and decoder 154B may generate a set of second predictions for the physiological parameter based on a plurality of PPG signals 108 recorded over a plurality of concurrent time periods. The first prediction (or set of first predictions) and the second prediction (or set of second predictions) may be stored on database 112 (e.g., as prediction data 116) and / or sent to quality metric determiner 140 of ML Manager 156.

[0038] In aspects, quality metric determiner 140 may receive the first prediction for the physiological parameter (e.g., a first heart rate) and the second prediction for the physiological parameter (e.g., a second heart rate) from model server 122 or database 112. Quality metric determiner 140 may compare or otherwise determine a variance between the first prediction and the second prediction. If the variance is greater than a variance threshold, it may be determined that a second quality metric of the BP output 118 is less than the quality threshold. If the second quality metric is less than the quality threshold, a second alert may be issued that warns a clinician that the BP output 118 may be unreliable and / or recommends alternativeAttorney Docket No. A0012962W001 means for obtaining BP for the subject 102. Additionally, the second alert may recommend evaluating monitoring devices 104A-104B for proper functioning, for example. In aspects, the second alert may be the same as or different from the first alert. In other examples, if the second quality metric is less than the quality threshold, the BP output 118 may be dropped (e.g., deleted) and / or it may be determined that additional training of ML model 146 may be warranted. In some examples, the loss function for evaluating the BP output 118, as described above, may be augmented with a term for predicting a physiological parameter from the encoded ECG signal and / or the encoded PPG signal to evaluate the quality of the BP output 118.

[0039] In further aspects, quality metric determiner 140 may receive the first prediction for the physiological parameter, the second prediction for the physiological parameter, and a measured value of the physiological parameter. Quality metric determiner 140 may compare or otherwise determine a difference (or divergence) between the measured value of the physiological parameter and at least one of the first prediction of the physiological parameter and the second prediction of the physiological parameter. Based on the difference, quality metric determiner 140 may determine a status of a sensor that recorded the measured value of the physiological parameter. For example, if the difference is greater than an error threshold, a third alert may be issued that recommends evaluation of the sensor.

[0040] As should be appreciated, FIG. 1 is provided for purposes of illustration and the described structures and features are non-limiting, i.e., structures and / or functionalities may be added or removed from the described system without departing from the disclosure herein.

[0041] FIG. 2A illustrates an overview of an example ML architecture 200A for non- invasive patient monitoring, according to aspects described herein.

[0042] As illustrated, ML architecture 200A includes a first encoder 206A in parallel with a second encoder 206B, the output of which is fed to decoder 214. As should be appreciated, many ML models incorporate encoders and decoders. For example, one or more encoders (e.g., first encoder 206A and second encoder 206B) compress input data into a latent space representation in which similar data points are closer together in space. A decoder (e.g., decoder 214) then reconstructs the input from the latent space representation. Encoders are useful for removing noise from images or signals, identifying unusual patterns in data, and / or reducing data features while preserving essential information. In examples, Sequence-to- Sequence (Seq2Seq) ML models utilize encoders to process an input sequence into a contextAttorney Docket No. A0012962W001 vector and decoders to generate an output sequence based on the context vector. As illustrated, first encoder 206A processes ECG signal input 202 into a first context vector (e.g., encoded ECG 210) and, in parallel, second encoder 206B processes PPG signal input 204 into a second context vector (e.g., encoded PPG 212). As further illustrated, encoded ECG 210 is concatenated to encoded PPG 212 and fed into decoder 214, which generates an output sequence (e.g., BP output 218) by decoding the concatenated first and second context vectors.

[0043] More advanced neural networks include convoluted neural networks (CNNs) and transformers. CNNs process input data through multiple layers, including convolutional layers that apply filters to detect features, pooling layers that reduce dimensionality of the features, and fully connected layers that make final predictions as output. As illustrated, first encoder 206A includes multiple encoding layers, including encoding layer 208A-1, encoding layer 208B-1, to encoding layer 208N-1, and second encoder 206B includes multiple encoding layers, including encoding layer 208A-2, encoding layer 208B-2, to encoding layer 208N-2. Similarly, decoder 214 has multiple decoding layers, including decoding layer 216A, decoding layer 216B, to decoding layer 216N. In some aspects, each encoding layer may be associated with a corresponding decoding layer (as illustrated by FIG. 2C); in other aspects, more complex linkages between encoding layers and decoding layers may be configured (not shown). As such be appreciated, the architecture of the encoders 206A-206B and decoder 214 may be designed in any suitable configuration to optimize the BP output 218.

[0044] Transformers use self-attention techniques to weigh the importance of various portions of the input data. In aspects, self-attention involves modifying vector representations to capture context in the input data. In some examples, an encoder can extract and map characteristics of the input data to high-dimensional vectors called embeddings. Depending on the type of data, embeddings can represent image characteristics (e.g., texture, edges, comers, color, etc.), audio characteristics (e.g., pitch, tone, rhythm, etc.), waveform characteristics (e.g., time domain features, frequency domain features, time -frequency domain features, etc.), semantic meaning (e.g., based on word sequence or position, word frequency, context, grammatical information, etc.), relational information (e.g., node hierarchy, social interactions, etc.), and the like. In some examples, the first context vector (e.g., encoded ECG 210) may be a first embedding and the second context vector (e.g., encoded PPT 212) may be a second embedding.Attorney Docket No. A0012962W001

[0045] FIG. 2B illustrates an overview of an example ML architecture 200B for non- invasive patient monitoring, according to aspects described herein.

[0046] Similar to FIG. 2A, ML architecture 200B includes first encoder 206A in parallel with second encoder 206B, the output of which is fed to decoder 214. As with FIG. 2A, first encoder 206A processes ECG signal input 202 into encoded ECG 210 and, in parallel, second encoder 206B processes PPG signal input 204 into encoded PPG 212. As further illustrated, encoded ECG 210 is concatenated to encoded PPG 212 and fed into decoder 214, which generates an output sequence (e.g., BP output 218) from the concatenated first and second context vectors.

[0047] As with ML architecture 200A, ML architecture 200B includes first encoder 206A with multiple encoding layers (e.g., encoding layer 208A-1, encoding layer 208B-1, to encoding layer 208N-1) and second encoder 206B with multiple encoding layers (e.g., encoding layer 208A-2, encoding layer 208B-2, to encoding layer 208N-2). However, in the case of ML architecture 200B, the encoding layers of first encoder 206A share information (e.g., interim output, context, features, etc.) with the encoding layers of second encoder 206B. In this way, the encoders of ML architecture 200B may be better able to recognize patterns, filter out noise, synchronize data and / or processing, and the like. As with ML architecture 200A, decoder 214 includes decoding layer 216A, decoding layer 216B, to decoding layer 216N.

[0048] FIG. 2C illustrates an overview of an example ML architecture 200C for non- invasive patient monitoring, according to aspects described herein.

[0049] Similar to FIGS. 2A-2B, ML architecture 200C includes first encoder 206A in parallel with second encoder 206B, the output of which is fed to decoder 214. As with FIGS. 2A-2B, first encoder 206A processes ECG signal input 202 into encoded ECG 210 and, in parallel, second encoder 206B processes PPG signal input 204 into encoded PPG 212. As further illustrated, encoded ECG 210 is concatenated to encoded PPG 212 and fed into decoder 214, which generates an output sequence (e.g., BP output 218) by decoding the concatenated first and second context vectors.

[0050] As with ML architectures 200A-200B, ML architecture 200C includes first encoder 206A with multiple encoding layers (e.g., encoding layer 208A-1, encoding layer 208B-1, to encoding layer 208N-1) and second encoder 206B with multiple encoding layers (e.g., encoding layer 208A-2, encoding layer 208B-2, to encoding layer 208N-2). However, in theAttorney Docket No. A0012962W001 case of ML architecture 200C, individual encoding layers of first encoder 206A and second encoder 206B share information (e.g., interim output) with decoder 214. In some aspects, encoding layers 208A-1 to 208N-1 and encoding layers 208A-2 to 208N-2 may share information with corresponding decoder layers 216A-216N (shown); in other aspects, encoding layers 208A-1 to 208N-1 and encoding layers 208A-2 to 208N-2 may share information with decoder 214 in an interleaved fashion (not shown). In still other aspects, encoding layers 208A-1 to 208N-1 and encoding layers 208A-2 to 208N-2 may share interim output that is fed as context (e.g., via prompts associated with encoded ECG 210 and / or encoded PPG 212) to decoder 214 (not shown). As should be appreciated, the ML architecture 200C may be designed in any suitable configuration to optimize the ability of decoder 214 to decode the embeddings (encoded ECG 210 and encoded PPG 212) and generate BP output 218. As with ML architectures 200A-200B, decoder 214 includes decoding layer 216A, decoding layer 216B, to decoding layer 216N.

[0051] As should be appreciated, FIGS. 2A-2C are provided for purposes of illustration and the described structures and features are non-limiting, i.e., structures and / or functionalities may be added or removed from the described architecture without departing from the disclosure herein. Moreover, aspects of FIGS. 2A-2C may be combined, for example, individual encoding layers of encoders 206A-206B may communicate with decoder 214 and between encoders 206A-206B.

[0052] FIG. 3 illustrates an overview of an example ML architecture 300 for non-invasive patient monitoring, according to aspects described herein.

[0053] Similar to FIGS. 2A-2C, ML architecture 300 includes first encoder 206A in parallel with second encoder 206B, the output of which is fed to decoder 214. The first encoder 206A has multiple encoding layers (e.g., encoding layer 208A-1, encoding layer 208B-1, to encoding layer 208N-1) and the second encoder 206B has multiple encoding layers (e.g., encoding layer 208A-2, encoding layer 208B-2, to encoding layer 208N-2). As with FIGS. 2A- 2B, first encoder 206A processes ECG signal input 202 into encoded ECG 210 and, in parallel, second encoder 206B processes PPG signal input 204 into encoded PPG 212.

[0054] However, ML architecture 300 illustrates one of the advantages of the parallel first and second encoders 206A-B. Since each encoder has a discreet output (e.g., encoded ECG 210 and encoded PPG 212, respectively), auxiliary task models may be included in the ML architecture 300 to process the discreet output of each encoder and generate independentlyAttorney Docket No. A0012962W001 derived and physiologically useful information. For example, encoded ECG 210 may be fed into a first auxiliary task model (e.g., first task model 320A) and encoded PPG 212 may be fed into a second auxiliary task model (e.g., second task model 320B).

[0055] As described above, task models may be specialized for performing a specific task, such as generating a specific output from a specific type of input. For instance, first task model 320A may receive encoded ECG 210 from first encoder 206A. Decoder 322A associated with first task model 320Amay process (e.g., decode) encoded ECG 210 to output a first prediction 324A of a physiological parameter. In aspects, the physiological parameter may be any parameter that can be derived from an ECG signal, including but not limited to heart rate, respiratory rate or HRV, for example. Concurrently, second task model 320B may receive encoded PPG 212 from second encoder 206B. That is, the encoded ECG 210 and the encoded PPG 212 should represent data recorded over a concurrent time period. Decoder 322B associated with second task model 320B may process (e.g., decode) encoded PPG 212 to output a second prediction 324B of the physiological parameter. In aspects, the second prediction 324B may be for the same physiological parameter over the same time period as the first prediction 324A. That is, in examples, if the first prediction 324A is a first heart rate over a time period, the second prediction 324B is a second heart rate over the same time period; if the first prediction 324A is a first respiratory rate over a time period, the second prediction 324B is a second respiratory rate over the same time period; and if the first prediction 324A is a first HRV over a time period, the second prediction 324B is a second HRV over the same time period, and so on.

[0056] As with FIGS. 2A-2C, encoded ECG 210 is concatenated to encoded PPG 212 and fed into decoder 214, which generates an output sequence (e.g., BP output 218) by decoding the concatenated first and second context vectors. Decoder 214 includes decoding layer 216A, decoding layer 216B, to decoding layer 216N.

[0057] In further aspects, a quality metric determiner 326 (similar to quality metric determiner 140) may receive the first prediction 324A for the physiological parameter (e.g., a first heart rate) from decoder 322A and the second prediction 324B for the physiological parameter (e.g., a second heart rate) from decoder 322B. Quality metric determiner 326 may compare or otherwise determine a variance between the first prediction 324A and the second prediction 324B. If the variance is greater than a quality threshold, it may be determined that a quality metric of the BP output 218 is low. If the quality metric of BP output 218 is low, variousAttorney Docket No. A0012962W001 alerts may be issued, the BP output may be dropped, and / or the ML architecture 300 may be further trained or finetuned, as described above. If the variance is not greater than the quality threshold, BP output 218 may be displayed, printed, and / or transmitted to cerebral autoregulation monitor 328. Based on BP output 218, together with related parameters 330 (e.g., cerebral blood flow, cerebral oxygen saturation, etc.), cerebral autoregulation monitor 328 may determine an indication of cerebral autoregulation for a patient (e.g., subject 102).

[0058] As should be appreciated, FIG. 3 is provided for purposes of illustration and the described structures and features are non-limiting, i.e., structures and / or functionalities may be added or removed from the described architecture without departing from the disclosure herein.

[0059] FIGS. 4A-4J illustrate examples of input and output of an example ML model for non-invasive patient monitoring, according to aspects described herein.

[0060] For example, in the top graphic 414a of FIG. 4A, first ECG input 410a is illustrated. In aspects, first ECG input 410a was processed, e.g., one or more ECG signals for a first subject (e.g., subject 102 of FIG. 1) were recorded and normalized, and represented as first ECG waveform 402a in arbitrary units (AU) over a first time period of about 15 seconds (s). As illustrated, first ECG input 410a represented a plurality of cardiac cycles for the first subject over the first time period.

[0061] In the middle graphic 416a of FIG. 4A, first PPG input 412a is illustrated. In aspects, first PPG input 412a was processed (e.g., one or more PPG signals for the first subject were recorded and normalized) and represented as first PPG waveform 404a in arbitrary units (AU) over a first concurrent time period of about 15 seconds (s). As illustrated, first PPG input 412a over the first concurrent time period represented the same plurality of cardiac cycles for the first subject as first ECG input 410a over the first time period.

[0062] In the bottom graphic 418a, first true (e.g., actual) arterial blood pressure (ABP) 406a for the first subject was measured in millimeters mercury (mmHg) over a first corresponding time period (e.g., 15s) and represented by a solid line. Additionally, first predicted ABP 408a for the first subject was output by one or more ML models in mmHg over the same first corresponding time period and represented by a dotted line. In aspects, the first time period, the first concurrent time period, and the first corresponding time period represented the same first window of time. As further illustrated by FIG. 4A, values for first true ABP 406a and first predicted ABP 408a are substantially the same or similar for the first subject over the first corresponding time period. Moreover, a first waveform morphology ofAttorney Docket No. A0012962W001 the first true ABP 406a and a second waveform morphology of the first predicted ABP 408a are substantially the same or similar over the first corresponding time period. That is, as illustrated, the one or more ML models generated a substantially accurate first predicted ABP 408a for the first subject, as well as a substantially accurate waveform morphology of first predicted ABP 408a for the first subject.

[0063] In aspects, the first ECG input 410a of FIG. 4A and the first PPG input 412a of FIG.4A were processed according to the methods and systems described herein. For example, ECG signals were recorded forthe first subject over the first time period (e.g., 15 s) and normalized. First ECG input 410a was fed to a first encoder (e.g., first encoder 148 A of FIG. 1 or first encoder 206A of FIGS. 2A-3), which output a plurality of first embeddings. PPG signals were recorded for the first subject over a first concurrent time period (e.g., 15 s) and normalized. First PPG input 412a was fed to a second encoder (e.g., second encoder 148B of FIG. 1 or second encoder 206B of FIGS. 2A-3), which output a plurality of second embeddings. The plurality of first embeddings were concatenated to the plurality of second embeddings and fed to a decoder (e.g., decoder 150 of FIG. 1 or decoder 214 ofFIGS. 2A-3). The decoder processed the concatenated first embeddings and second embeddings to generate a first predicted ABP 408a in mmHg over a first corresponding time period (e.g., 15 s). During the same first corresponding time period (15 s), a first true ABP 406a of the first subject was measured by any suitable means (e.g., using a sphygmomanometer).

[0064] Similar to FIG. 4A, the top graphic 414b of FIG. 4B illustrates second ECG input 410b represented as second ECG waveform 402b in arbitrary units (AU) over a second time period of about 15 seconds (s). As illustrated, second ECG input 410b represented a plurality of cardiac cycles forthe second subject over the second time period. The middle graphic 416b of FIG. 4B illustrates second PPG input 412b represented as second PPG waveform 404b in arbitrary units (AU) over a second concurrent time period of about 15 seconds (s). As illustrated, second PPG input 412b over the second concurrent time period represented the same plurality of cardiac cycles for the second subject as second ECG input 410b over the second time period.

[0065] In the bottom graphic 418b, second true ABP 406b for the second subject was measured in millimeters mercury (mmHg) over a second corresponding time period (e.g., 15s) and represented by a solid line. Additionally, second predicted ABP 408b forthe second subject was output in mmHg by one or more ML models over the same second corresponding timeAttorney Docket No. A0012962W001 period and represented by a dotted line. As with FIG. 4A, FIG. 4B illustrates that values for second true ABP 406b and second predicted ABP 408b are substantially the same or similar for the second subject over the second corresponding time period. Moreover, a first morphology of the second true ABP 406b and a second morphology of the second predicted ABP 408b for the second subject are substantially the same or similar over the second corresponding time period. That is, as illustrated, the one or more ML models generated a substantially accurate second predicted ABP 408b for the second subject, as well as a substantially accurate waveform morphology of second predicted ABP 408b for the second subject.

[0066] In aspects, the second ECG input 410b of FIG. 4B and the second PPG input 412b of FIG. 4B were recorded for the second subject and processed according to the methods and systems described herein. Further, second ECG input 410b and second PPG input 412b were fed into one or more ML models and second predicted ABP 408b was output from the one or more ML models according to the methods and systems described herein.

[0067] Similar to FIGS. 4A-4B, the top graphic 414c of FIG. 4C illustrates third ECG input410c represented as third ECG waveform 402c in arbitrary units (AU) over a third time period of about 15 seconds (s). As illustrated, third ECG input 410c represented a plurality of cardiac cycles for the third subject over the third time period. The middle graphic 416c of FIG. 4C illustrates third PPG input 412c represented as third PPG waveform 404c in arbitrary units (AU) over a third concurrent time period of about 15 seconds (s). As illustrated, third PPG input 412c over the third concurrent time period represented the same plurality of cardiac cycles for the third subject as third ECG input 410c over the third time period.

[0068] In the bottom graphic 418c, third true ABP 406c for the third subject was measured in millimeters mercury (mmHg) over a third corresponding time period (e.g., 15s) and represented by a solid line. Additionally, third predicted ABP 408c for the third subject was output in mmHg by one or more ML models over the same third corresponding time period and represented by a dotted line. As with FIGS. 4A-4B, FIG. 4C illustrates that values for third true ABP 406c and third predicted ABP 408c are substantially the same or similar for the third subject over the third corresponding time period. Moreover, a first morphology of the third true ABP 406c and a second morphology of the third predicted ABP 408c for the third subject are substantially the same or similar over the third corresponding time period. That is, as illustrated, the one or more ML models generated a substantially accurate third predicted ABP 408c forAttorney Docket No. A0012962W001 the third subject, as well as a substantially accurate waveform morphology of third predicted ABP 408c for the third subject.

[0069] In aspects, the third ECG input 410c and the third PPG input 412c of FIG. 4C were recorded for the third subject and processed according to the methods and systems described herein. Further, third ECG input 410c and third PPG input 412c were fed into one or more ML models and third predicted ABP 408c was output from the one or more ML models according to the methods and systems described herein.

[0070] Similar to FIGS. 4A-4C, the top graphic 414d of FIG. 4D illustrates fourth ECG input 410d represented as fourth ECG waveform 402d in arbitrary units (AU) over a fourth time period of about 15 seconds (s). As illustrated, fourth ECG input 410d represented a plurality of cardiac cycles for the fourth subject over the fourth time period. The middle graphic 416d of FIG. 4D illustrates fourth PPG input 412d represented as fourth PPG waveform 404d in arbitrary units (AU) over a fourth concurrent time period of about 15 seconds (s). As illustrated, fourth PPG input 412d over the fourth concurrent time period represented the same plurality of cardiac cycles for the fourth subject as third ECG input 410d over the fourth time period.

[0071] In the bottom graphic 418d, fourth true ABP 406d for the fourth subject was measured in mmHg over a fourth corresponding time period (e.g., 15s) and represented by a solid line. Additionally, fourth predicted ABP 408d for the fourth subject was output in mmHg by one or more ML models over the same fourth corresponding time period and represented by a dotted line. As with FIGS. 4A-4C, FIG. 4D illustrates that values for fourth true ABP 406d and fourth predicted ABP 408d are substantially the same or similar for the fourth subject over the fourth corresponding time period. Moreover, a first morphology of the fourth true ABP 406d and a second morphology of the fourth predicted ABP 408d for the fourth subject are substantially the same or similar over the fourth corresponding time period. That is, as illustrated, the one or more ML models generated a substantially accurate fourth predicted ABP 408d for the fourth subject, as well as a substantially accurate waveform morphology of fourth predicted ABP 408d for the fourth subject.

[0072] In aspects, the fourth ECG input 410d and the fourth PPG input 412d of FIG. 4D were recorded for the fourth subject and processed according to the methods and systems described herein. Further, fourth ECG input 410d and fourth PPG input 412d were fed into oneAttorney Docket No. A0012962W001 or more ML models and fourth predicted ABP 408d was output from the one or more ML models according to the methods and systems described herein.

[0073] Similar to FIGS. 4A-4D, the top graphic 414e of FIG. 4E illustrates fifth ECG input410e represented as fifth ECG waveform 402e in arbitrary units (AU) over a fifth time period of about 15 seconds (s). As illustrated, fifth ECG input 41 Oe represented a plurality of cardiac cycles for the fifth subject over the fifth time period. The middle graphic 416e of FIG. 4E illustrates fifth PPG input 412e represented as fifth PPG waveform 404e in arbitrary units (AU) over a fifth concurrent time period of about 15s. As illustrated, fifth PPG input 412e over the fifth concurrent time period represented the same plurality of cardiac cycles for the fifth subject as fifth ECG input 41 Oe over the fifth time period.

[0074] In the bottom graphic 418e, fifth true ABP 406e for the fifth subject was measured in mmHg over a fifth corresponding time period (e.g., 15s) and represented by a solid line. Additionally, fifth predicted ABP 408e for the fifth subject was output in mmHg by one or more ML models over the same fifth corresponding time period and represented by a dotted line. As with FIGS. 4A-4D, FIG. 4E illustrates that values for fifth true ABP 406e and fifth predicted ABP 408e are substantially the same or similar for the fifth subject over the fifth corresponding time period. Moreover, a first morphology of the fifth true ABP 406e and a second morphology of the fifth predicted ABP 408e for the fifth subject are substantially the same or similar over the fifth corresponding time period. That is, as illustrated, the one or more ML models generated a substantially accurate fifth predicted ABP 408e for the fifth subject, as well as a substantially accurate waveform morphology of fifth predicted ABP 408e for the fifth subject.

[0075] In aspects, the fifth ECG input 410e and the fifth PPG input 412e of FIG. 4E were recorded for the fifth subject and processed according to the methods and systems described herein. Further, fifth ECG input 410e and fifth PPG input 412e were fed into one or more ML models and fifth predicted ABP 408e was output from the one or more ML models according to the methods and systems described herein.

[0076] Similar to FIGS. 4A-4E, the top graphic 414f of FIG. 4F illustrates sixth ECG input410f represented as sixth ECG waveform 402f in arbitrary units (AU) over a sixth time period of about 15 seconds (s). As illustrated, sixth ECG input 41 Of represented a plurality of cardiac cycles for the sixth subject over the sixth time period. The middle graphic 416f of FIG. 4F illustrates sixth PPG input 412f represented as sixth PPG waveform 404f in arbitrary unitsAttorney Docket No. A0012962W001(AU) over a sixth concurrent time period of about 15s. As illustrated, sixth PPG input 412f over the sixth concurrent time period represented the same plurality of cardiac cycles for the sixth subject as sixth ECG input 41 Of over the sixth time period.

[0077] In the bottom graphic 418f, sixth true ABP 406f for the sixth subject was measured in mmHg over a sixth corresponding time period (e.g., 15s) and represented by a solid line. Additionally, sixth predicted ABP 408f for the sixth subject was output in mmHg by one or more ML models over the same sixth corresponding time period and represented by a dotted line. As with FIGS. 4A-4E, FIG. 4F illustrates that values for sixth true ABP 406f and sixth predicted ABP 408f are substantially the same or similar for the sixth subject over the sixth corresponding time period. Moreover, a first morphology of the sixth true ABP 406f and a second morphology of the sixth predicted ABP 408f for the sixth subject are substantially the same or similar over the sixth corresponding time period. That is, as illustrated, the one or more ML models generated a substantially accurate sixth predicted ABP 408f for the sixth subject, as well as a substantially accurate waveform morphology of sixth predicted ABP 408f for the sixth subject.

[0078] In aspects, the sixth ECG input 41 Of and the sixth PPG input 412f of FIG. 4F were recorded for the sixth subject and processed according to the methods and systems described herein. Further, sixth ECG input 41 Of and sixth PPG input 412f were fed into one or more ML models and sixth predicted ABP 408f was output from the one or more ML models according to the methods and systems described herein.

[0079] Similar to FIGS. 4A-4F, the top graphic 414g of FIG. 4G illustrates seventh ECG input 410g represented as seventh ECG waveform 402g in arbitrary units (AU) over a seventh time period of about 15 seconds (s). As illustrated, seventh ECG input 410g represented a plurality of cardiac cycles for the seventh subject over the seventh time period. The middle graphic 416g of FIG. 4G illustrates seventh PPG input 412g represented as seventh PPG waveform 404g in arbitrary units (AU) over a seventh concurrent time period of about 15s. As illustrated, seventh PPG input 412g over the seventh concurrent time period represented the same plurality of cardiac cycles for the seventh subject as seventh ECG input 410g over the seventh time period.

[0080] In the bottom graphic 418g, seventh true ABP 406g for the seventh subject was measured in mmHg over a seventh corresponding time period (e.g., 15s) and represented by a solid line. Additionally, seventh predicted ABP 408g for the seventh subject was output inAttorney Docket No. A0012962W001 mmHg by one or more ML models over the same seventh corresponding time period and represented by a dotted line. As with FIGS. 4A-4F, FIG. 4G illustrates that values for seventh true ABP 406g and seventh predicted ABP 408g are substantially the same or similar for the seventh subject over the seventh corresponding time period. Moreover, a first morphology of the seventh true ABP 406g and a second morphology of the seventh predicted ABP 408g for the seventh subject are substantially the same or similar over the seventh corresponding time period. That is, as illustrated, the one or more ML models generated a substantially accurate seventh predicted ABP 408g for the seventh subject, as well as a substantially accurate waveform morphology of seventh predicted ABP 408g for the seventh subject.

[0081] In aspects, the seventh ECG input 410g and the seventh PPG input 412g of FIG. 4G were recorded for the seventh subject and processed according to the methods and systems described herein. Further, seventh ECG input 410g and seventh PPG input 412g were fed into one or more ML models and seventh predicted ABP 408g was output from the one or more ML models according to the methods and systems described herein.

[0082] Similar to FIGS. 4A-4G, the top graphic 414h of FIG. 4H illustrates eighth ECG input 41 Oh represented as eighth ECG waveform 402h in arbitrary units (AU) over an eighth time period of about 15 seconds (s). As illustrated, eighth ECG input 41 Oh represented a plurality of cardiac cycles forthe eighth subject overthe eighth time period. The middle graphic 416h of FIG. 4H illustrates eighth PPG input 412h represented as eighth PPG waveform 404h in arbitrary units (AU) over an eighth concurrent time period of about 15 s . As illustrated, eighth PPG input 412h over the eighth concurrent time period represented the same plurality of cardiac cycles forthe eighth subject as eighth ECG input 41 Oh over the eighth time period.

[0083] In the bottom graphic 418h, eighth true ABP 406h for the eighth subject was measured in mmHg over an eighth corresponding time period (e.g., 15s) and represented by a solid line. Additionally, eighth predicted ABP 408h forthe eighth subject was output in mmHg by one or more ML models over the same eighth corresponding time period and represented by a dotted line. As with FIGS. 4A-4G, FIG. 4H illustrates that values for eighth true ABP 406h and eighth predicted ABP 408h are substantially the same or similar forthe eighth subject over the eighth corresponding time period. Moreover, a first morphology of the eighth true ABP 406h and a second morphology of the eighth predicted ABP 408h for the eighth subject are substantially the same or similar over the eighth corresponding time period. That is, as illustrated, the one or more ML models generated a substantially accurate eighth predicted ABPAttorney Docket No. A0012962W001408h for the eighth subject, as well as a substantially accurate waveform morphology of eighth predicted ABP 408h for the eighth subject.

[0084] In aspects, the eighth ECG input 41 Oh and the eighth PPG input 412h of FIG. 4H were recorded for the eighth subject and processed according to the methods and systems described herein. Further, eighth ECG input 410h and eighth PPG input 412h were fed into one or more ML models and eighth predicted ABP 408h was output from the one or more ML models according to the methods and systems described herein.

[0085] Similar to FIGS. 4A-4H, the top graphic 414i of FIG. 41 illustrates ninth ECG input 4 lOi represented as ninth ECG waveform 402i in arbitrary units (AU) over a ninth time period of about 15 seconds (s). As illustrated, ninth ECG input 41 Oi represented a plurality of cardiac cycles for the ninth subject over the ninth time period. The middle graphic 416i of FIG. 41 illustrates ninth PPG input 412i represented as ninth PPG waveform 404i in arbitrary units (AU) over a ninth concurrent time period of about 15s. As illustrated, ninth PPG input 412i over the ninth concurrent time period represented the same plurality of cardiac cycles for the ninth subject as ninth ECG input 41 Oi over the ninth time period.

[0086] In the bottom graphic 418i, ninth true ABP 406i for the ninth subject was measured in mmHg over a ninth corresponding time period (e.g., 15s) and represented by a solid line. Additionally, ninth predicted ABP 408i for the ninth subject was output in mmHg by one or more ML models over the same ninth corresponding time period and represented by a dotted line. As with FIGS. 4A-4H, FIG. 41 illustrates that values for ninth true ABP 406i and ninth predicted ABP 408i are substantially the same or similar for the ninth subject over the ninth corresponding time period. Moreover, a first morphology of the ninth true ABP 406i and a second morphology of the ninth predicted ABP 408i for the ninth subject are substantially the same or similar over the ninth corresponding time period. That is, as illustrated, the one or more ML models generated a substantially accurate ninth predicted ABP 408i for the ninth subject, as well as a substantially accurate waveform morphology of ninth predicted ABP 408i for the ninth subject.

[0087] In aspects, the ninth ECG input 4101 and the ninth PPG input 4121 of FIG. 41 were recorded for the ninth subject and processed according to the methods and systems described herein. Further, ninth ECG input 41 Oi and ninth PPG input 412i were fed into one or more ML models and ninth predicted ABP 408i was output from the one or more ML models according to the methods and systems described herein.Attorney Docket No. A0012962W001

[0088] FIGS. 5A-5B illustrate an overview of an example method 500 for non-invasive patient monitoring, according to aspects described herein.

[0089] Method 500A of FIG. 5 A begins at operation 502, where an ECG signal and a PPG signal are received. In examples, the ECG signal (e.g., ECG signal 106 of FIG. 1) may be received from a first monitoring device such as an electrocardiograph (e.g., monitoring device 104A of FIG. 1), which uses small electrodes to record electrical activity over cardiac cycles in millivolts per second (mV / s). In further examples, the PPG signal (e.g., PPG signal 108 of FIG. 1) may be received from a second monitoring device such as a pulse oximeter (e.g., monitoring device 104B), which records relative changes in light absorption by the blood vessels over cardiac cycles to determine a voltage signal proportional to blood volume. The ECG signal and the PPG signal may be measured over a short concurrent period (or window) of time, e.g., 5 seconds (s), 10s, 15s, 30s, 60s, etc. In aspects, the period of time may contain several heartbeats so that a ML model, described further below, can derive a relationship between BP, ECG, and PPG over several cardiac cycles.

[0090] At operation 504, the ECG signal may be encoded. For example, the ECG signal (e.g., ECG signal 106 of FIG. 1) may be encoded by a first encoder (e.g., first encoder 148A of FIG. 1 or first encoder 206A of FIGS. 2A-3). The first encoder may process the ECG signal through multiple layers (e.g., encoding layers 208A-1 to 208N-1 of FIGS. 2A-3), including convolutional layers that apply filters to detect features and / or pooling layers that reduce dimensionality of the features, for example. As described above, the first encoder may process the ECG signal into a first context vector and, in some examples, the first encoder may process the ECG signal into a first high-dimensional vector, e.g., a first embedding. In aspects, the first embedding may represent the waveform characteristics of the ECG signal, such as time domain features, frequency domain features, time-frequency domain features, etc., as the first highdimensional vector. In other examples, the first encoder may be a transformer and may use self-attention techniques to weigh the importance of various portions of the ECG signal, for example, by modifying the vector representation (e.g., first embedding) of the ECG signal to include context.

[0091] In parallel with operation 504, at operation 506, the PPG signal (e.g., PPG signal 108 of FIG. 1) may be encoded. For example, the PPG signal may be encoded by a second encoder (e.g., second encoder 148B of FIG. 1 or second encoder 206B of FIGS. 2A-3). The second encoder may process the PPG signal through multiple layers (e.g., encoding layersAttorney Docket No. A0012962W001208A-2 to 208N-2 of FIGS. 2A-3), including convolutional layers that apply filters to detect features and / or pooling layers that reduce dimensionality of the features, for example. As described above, the second encoder may process the PPG signal into a second context vector and, in some examples, into a second high-dimensional vector, e.g., a second embedding. Similar to the first embedding, the second embedding may represent the waveform characteristics of the PPG signal, such as time domain features, frequency domain features, time-frequency domain features, etc., as the second high-dimensional vector. In other examples, the second encoder may be a transformer and may use self-attention techniques to weigh the importance of various portions of the PPG signal, for example, by modifying the vector representation (e.g., second embedding) of the PPG signal to include context.

[0092] At operation 508, the first embedding may be input to a first task model (e.g., first task model 152A of FIG. 1 or first task model 320A of FIG. 3). As described above, task models may be specialized for performing a specific task, such as generating a specific output from a specific type of input. For instance, the first task model may receive the first embedding and a decoder (e.g., e.g., decoder 154A of FIG. 1 or decoder 322A of FIG. 3) may process the first embedding to output a first prediction of a physiological parameter (e.g., first prediction 324A of FIG. 3). In some aspects, the physiological parameter may be any parameter that can be derived from an ECG signal, including but not limited to heart rate, respiratory rate or HRV, for example. In other aspects, the first prediction may be a reconstruction of the ECG signal from the first embedding.

[0093] At operation 510, the second embedding may be input to a second task model (e.g., second task model 152B of FIG. 1 or second task model 320B of FIG. 3). In aspects, the first embedding and the second embedding represent data recorded over a concurrent time period. The second task model may receive the second embedding and a decoder (e.g., e.g., decoder 154B of FIG. 1 or decoder 322B of FIG. 3) may process the second embedding to output a second prediction of the physiological parameter (e.g., second prediction 324B of FIG. 3). In aspects, the second prediction may be for the same physiological parameter over the same time period as the first prediction.

[0094] At operation 512, the first embedding is concatenated (or combined) with the second embedding. In this way, the characteristics or features of the ECG signal and the PPG signal are combined.Attorney Docket No. A0012962W001

[0095] Method 500B of FIG. 5B continues at operation 514, where the concatenated first embedding and second embedding is decoded. For example, a decoder (e.g., decoder 150 of FIG. 1 or decoder 214 of FIGS. 2A-3) generates an output sequence (e.g., BP output) by decoding the concatenated first embedding and second embedding. As described above, BP output may include, but is not limited to, an indication of BP waveform morphology, an indication of mean arterial pressure (MAP) trending, and / or predicted values of BP in mmHg.

[0096] At operation 516, at least one quality metric is determined for the BP output. For example, to determine a first quality metric of the BP output, a loss function (e.g., mean squared error (MSE), Kullback-Leibler divergence, cosine proximity, or the like) may be used to compare an actual blood pressure of a patient measured during the same time period represented by the BP output to determine a divergence. In further examples, to determine a second quality metric of the BP output, the first prediction of the physiological parameter may be compared to the second prediction of the physiological parameter to determine a variance.

[0097] At operation 520, it may be determined if the first quality metric and / or the second quality metric is less than a quality threshold. If the BP output diverges from the actual blood pressure by more than a divergence threshold, it may be determined that the first quality metric of the BP output is less than the quality threshold. If the first prediction varies from the second prediction by more than a variance threshold, it may be determined that the second quality metric of the BP output is less than the quality threshold. If the first quality metric and the second quality metric are not less than the quality threshold, the method may progress to operation 522. If the first quality metric and / or the second quality metric is less than the quality threshold, the method may progress to operation 526.

[0098] At operation 522, the BP output may be displayed on a display device (e.g., of output device(s) 134). In some examples, the sequence of values associated with the BP output may be displayed; in other examples, the last value (e.g., most recent value) of the sequence of values associated with BP output may be displayed.

[0099] At operation 524, based on the BP output, along with other inputs (e.g., cerebral blood flow from a transcranial doppler device, cerebral oxygen saturation from a regional oximeter, etc.), an indication of cerebral autoregulation may be determined. As described above, cerebral autoregulation is a physiological process that ensures stable and adequate blood flow to the brain, despite fluctuations in systemic blood pressure. Non-invasive measurement of cerebral blood flow and related parameters is complex and challenging, often requiring anAttorney Docket No. A0012962W001 invasive arterial line (A-line). By substantially continuously recording and feeding ECG signals and PPG signals into an ML model, BP output may be substantially continuously computed. For example, BP output may include, but is not limited to, a substantially continuous indication of BP waveform morphology, a substantially continuous trending of mean arterial pressure (MAP), and / or substantially continuous predictions (e.g., sequential values) of BP in mmHg. Based on the substantially continuous BP output, and additional inputs such as oxygen saturation, cerebral autoregulation may be non-invasively monitored in near real-time.

[0100] At operation 526, if the first quality metric and / or the second quality metric is less than the quality threshold, an alert may be issued that warns a clinician of potential unreliability of the BP output and / or recommends alternative means for obtaining BP for the patient. Additionally or alternatively, the alert may recommend evaluating monitoring devices (e.g., an electrocardiograph and / or a pulse oximeter) for proper functioning. In other examples, if the first quality metric and / or the second quality metric is less than the quality threshold, the BP output may be dropped (e.g., deleted) and / or it may be determined that additional training of the ML model may be warranted.

[0101] As should be appreciated, operations 502-526 are described for purposes of illustrating the present methods and systems and are not intended to limit the disclosure to a particular set or sequence of steps. For example, operations may be performed in a different order and more or fewer operations may be performed without departing from the scope of the present disclosure.

[0102] It should be understood that various aspects disclosed herein may be combined in different combinations than the combinations specifically presented in the description and accompanying drawings. It should also be understood that, depending on the example, certain acts or events of any of the processes or methods described herein may be performed in a different sequence, may be added, merged, or left out altogether (e.g., all described acts or events may not be necessary to carry out the techniques). In addition, while certain aspects of this disclosure are described as being performed by a single module or unit for purposes of clarity, it should be understood that the techniques of this disclosure may be performed by a combination of units or modules associated with, for example, a medical device.

[0103] In one or more examples, the described techniques may be implemented using 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 mediumAttorney Docket No. A0012962W001 and executed by a hardware -based processing unit. Computer-readable media may include non- transitory computer-readable media, which corresponds to a tangible medium such as data storage media (e.g., RAM, ROM, EEPROM, flash memory, or any other medium that can be used to store desired program code in the form of instructions or data structures and that can be accessed by a computer).

[0104] Instructions may be executed by one or more processors, such as one or more digital signal processors (DSPs), general purpose microprocessors, application specific integrated circuits (ASICs), field programmable logic arrays (FPGAs), or other equivalent integrated or discrete logic circuitry. Accordingly, the term “processor” as used herein may refer to any of the foregoing structure or any other physical structure suitable for implementation of the described techniques. Also, the techniques could be fully implemented in one or more circuits or logic elements.

[0105] The following examples are illustrative of the techniques described herein.

[0106] Example 1. A system comprising: at least one processor; and memory storing instructions that, when executed by the at least one processor, cause the system to perform a set of operations, the set of operations comprising: receiving an electrocardiogram (ECG) signal and a photoplethysmography (PPG) signal; encoding the ECG signal using a first encoder to generate a first embedding and encoding the PPG signal using a second encoder to generate a second embedding, wherein the first encoder and the second encoder operate in parallel; inputting the first embedding to a first task model configured to generate a first prediction of a physiological parameter; inputting the second embedding to a second task model configured to generate a second prediction of the physiological parameter; decoding the first embedding and the second embedding to generate a blood pressure output; and based on a variance between the first prediction of the physiological parameter and the second prediction of the physiological parameter, determining a quality metric for the blood pressure output.

[0107] Example 2. The system of Example 1, the set of operations further comprising: based on the variance being greater than a variance threshold, issuing an alert regarding the blood pressure output.

[0108] Example 3. The system of Example 1, the set of operations further comprising: concatenating the first embedding and the second embedding; and decoding the concatenated first embedding and second embedding.Attorney Docket No. A0012962W001

[0109] Example 4. The system of Example 1, wherein the blood pressure output is one of a series of blood pressure values, a blood pressure waveform, or a trend of a mean arterial pressure (MAP).

[0110] Example 5. The system of Example 1, wherein an indication of patient cerebral autoregulation is based at least in part on the blood pressure output.

[0111] Example 6. The system of Example 1, wherein the physiological parameter is one of a heart rate, a respiratory rate, or a heart rate variability (HRV).

[0112] Example 7. The system of Example 1, wherein the ECG signal is concurrent with the PPG signal.

[0113] Example 8. The system of Example 1, wherein the blood pressure output is continuous.

[0114] Example 9. The system of Example 1, wherein the first encoder, the second encoder, and the decoder are included in a machine learning (ML) model.

[0115] Example 10. The system of Example 9, wherein the ML model includes one or more of a transformer system, a convolutional neural network (CNN), or a generative model.

[0116] Example 11. The system of Example 1, the set of operations further comprising: determining a difference between a measured value of the physiological parameter and at least one of the first prediction of the physiological parameter and the second prediction of the physiological parameter; and based on the difference, determining a status of a sensor that recorded the measured value of the physiological parameter.

[0117] Example 12. The system of Example 11, the set of operations further comprising: based on the difference being greater than an error threshold, issuing an alert regarding the status of the sensor.

[0118] Example 13. A system comprising: at least one processor; and memory storing instructions that, when executed by the at least one processor, cause the system to perform a set of operations, the system further comprising: a first encoder configured to encode an electrocardiogram (ECG) signal; a second encoder configured to encode a photoplethysmography (PPG) signal, wherein the first encoder and the second encoder operate in parallel; a first task model configured to process the encoded ECG signal to generate a first prediction of a physiological parameter; a second task model configured to process the encoded PPG signal to generate a second prediction of the physiological parameter, wherein there is a variance between the first prediction and the second prediction; and a decoder configured to decode the encoded ECG signal and the encoded PPG signal toAttorney Docket No. A0012962W001 generate a blood pressure output, wherein a quality of the blood pressure output is based on the variance.

[0119] Example 14. The system of Example 13, further comprising: a sensor configured to measure a value of the physiological parameter, wherein there is a difference between the measured value of the physiological parameter and at least one of the first prediction of the physiological parameter or the second prediction of the physiological parameter.

[0120] Example 15. The system of Example 14, wherein a status of the sensor is based on the difference.

[0121] Example 16. The system of Example 15, wherein when the difference is greater than an error threshold, an alert is issued regarding the status of the sensor.

[0122] Example 17. A method of generating a blood pressure output, comprising encoding an electrocardiogram (ECG) signal using a first encoder to generate a first embedding and, in parallel, encoding a photoplethysmography (PPG) signal using a second encoder to generate a second embedding; inputting the first embedding to a first task model configured to generate a first prediction of a physiological parameter; inputting the second embedding to a second task model configured to generate a second prediction of the physiological parameter; decoding the first embedding and the second embedding to generate a blood pressure output; and based on a variance between the first prediction of the physiological parameter and the second prediction of the physiological parameter, determining a quality metric for the blood pressure output.

[0123] Example 18. The method of Example 17, further comprising: based on the variance being greater than a variance threshold, issuing an alert regarding the blood pressure output.

[0124] Example 19. The method of Example 17, wherein the blood pressure output is one of a series of blood pressure values, a blood pressure waveform, or a trend of a mean arterial pressure (MAP).

[0125] Example 20. The method of Example 17, wherein an indication of patient cerebral autoregulation is based at least in part on the blood pressure output.

Claims

Attorney Docket No. A0012962W001WHAT IS CLAIMED IS:

1. A system comprising: at least one processor (130); and memory (124) storing instructions that, when executed by the at least one processor, cause the system to perform a set of operations, the set of operations comprising: receiving (502) an electrocardiogram (ECG) signal (202) and a photoplethysmography (PPG) signal (204); encoding (504, 506) the ECG signal using a first encoder (206A) to generate a first embedding (210) and encoding the PPG signal using a second encoder (206B) to generate a second embedding (212), wherein the first encoder and the second encoder operate in parallel; inputting (508) the first embedding to a first task model (320A) configured to generate a first prediction (324A) of a physiological parameter; inputting (510) the second embedding to a second task model (320B) configured to generate a second prediction (324B) of the physiological parameter; decoding (514) the first embedding and the second embedding to generate a blood pressure output (218); and based on a variance (516) between the first prediction of the physiological parameter and the second prediction of the physiological parameter, determining (518) a quality metric for the blood pressure output.

2. The system of claim 1, the set of operations further comprising: based on the variance being greater than a variance threshold, issuing an alert regarding the blood pressure output.

3. The system of any of claims 1 to 2, the set of operations further comprising: concatenating the first embedding and the second embedding; and decoding the concatenated first embedding and second embedding.

4. The system of any of claims 1 to 3, wherein the blood pressure output is one of a series of blood pressure values, a blood pressure waveform, or a trend of a mean arterial pressure (MAP).Attorney Docket No. A0012962W0015. The system of any of claims 1 to 4, wherein an indication of patient cerebral autoregulation is based at least in part on the blood pressure output.

6. The system of any of claims 1 to 5, wherein the physiological parameter is one of a heart rate, a respiratory rate, or a heart rate variability (HRV).

7. The system of any of claims 1 to 6, wherein the ECG signal is concurrent with the PPG signal.

8. The system of any of claims 1 to 7, wherein the blood pressure output is continuous.

9. The system of any of claims 1 to 8, wherein the first encoder, the second encoder, and the decoder are included in a machine learning (ML) model.

10. The system of claim 9, wherein the ML model includes one or more of a transformer system, a convolutional neural network (CNN), or a generative model.

11. The system of any of claims 1 to 10, the set of operations further comprising: determining a difference between a measured value of the physiological parameter and at least one of the first prediction of the physiological parameter and the second prediction of the physiological parameter; and based on the difference, determining a status of a sensor that recorded the measured value of the physiological parameter.

12. The system of claim 11, the set of operations further comprising: based on the difference being greater than an error threshold, issuing an alert regarding the status of the sensor.Attorney Docket No. A0012962W00113. A method of generating a blood pressure output, comprising encoding (504, 506) an electrocardiogram (ECG) signal (202) using a first encoder (206A) to generate a first embedding (210) and, in parallel, encoding a photoplethysmography (PPG) signal (204) using a second encoder (206B) to generate a second embedding (212); inputting (508) the first embedding to a first task model (320A) configured to generate a first prediction (324A) of a physiological parameter; inputting (510) the second embedding to a second task model (320B) configured to generate a second prediction (324B) of the physiological parameter; decoding (514) the first embedding and the second embedding to generate a blood pressure output (218); and based on a variance (516) between the first prediction of the physiological parameter and the second prediction of the physiological parameter, determining (518) a quality metric for the blood pressure output.

14. The method of claim 13, further comprising: based on the variance being greater than a variance threshold, issuing an alert regarding the blood pressure output.

15. The method of any of claims 13 to 14, wherein an indication of patient cerebral autoregulation is based at least in part on the blood pressure output.

Images