Medical monitoring devices and machine learning techniques for predicting physiological changes and adverse events

Abstract

A medical monitoring method may include obtaining physiological data of a patient indicating one or more sensed values of at least one physiological parameter of the patient, the one or more sensed values including a current value of the physiological parameter and / or one or more past values of the physiological parameter; generating, using a machine learning model, one or more future values of the physiological parameter, the future values of the physiological parameter being predicted values of the physiological parameter at future times, the generating being based on the physiological data and a physiological baseline of the patient; and providing visualization data to a display device configured to display a visualization of the physiological parameter, the visualization indicating at least a portion of the sensed values of the physiological parameter and at least a portion of the future values of the physiological parameter.

Description

MEDICAL MONITORING DEVICES AND MACHINE LEARNING TECHNIQUES FOR PREDICTING PHYSIOLOGICAL CHANGES AND ADVERSE EVENTSCROSS REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of U.S. Provisional Patent Application No.63 / 641,168, filed May 1, 2024, and U.S. Provisional Patent Application No. 63 / 721,249, filedNovember 15, 2024, the entire disclosure of each of which is hereby incorporated by reference herein in its entirety.BACKGROUND

[0002] During a surgical procedure, various medical devices may monitor physiological parameters of a patient and present (e.g., display) the monitored parameter values to the surgical team. For example, a patient’s blood pressure, heart rate, and / or oxygenation (e.g., oxygen saturation) can be monitored, and the current values of the monitored parameters can be displayed. In some cases, previous values of the monitored parameters during a prior time period are also displayed, often in the form of time-series graphs. Physicians routinely detect adverse events based on the displayed parameter values and initiate corrective clinical actions when such adverse events are detected.SUMMARY

[0003] In some aspects, the techniques described herein relate to a medical monitoring system including: one or more processors; and one or more computer-readable media storing instructions that, when executed by the one or more processors, cause the system to perform operations including: obtaining physiological data of a patient, the physiological data indicating one or more sensed values of at least one physiological parameter of the patient, the one or more sensed values including a current value of the physiological parameter and / or one or more past values of the physiological parameter; generating, using at least one machine learning model, one or more future values of the physiological parameter, the one or more future values of the physiological parameter being predicted values of the physiological parameter at one or more future times, the generating being based on the physiological data and a physiological baseline ofthe patient; and providing visualization data to a display device configured to display a visualization of the physiological parameter, the visualization indicating at least a portion of the one or more sensed values of the physiological parameter and at least a portion of the one or more future values of the physiological parameter.

[0004] In some aspects, the techniques described herein relate to a medical monitoring system, further including the display device.

[0005] In some aspects, the techniques described herein relate to a medical monitoring system, wherein the one or more sensed values of the physiological parameter are obtained via one or more sensors.

[0006] In some aspects, the techniques described herein relate to a medical monitoring system, further including the one or more sensors.

[0007] In some aspects, the techniques described herein relate to a medical monitoring system, wherein the one or more sensors include a blood pressure monitor, an oxygen saturation monitor, a heart rate monitor, an end tidal CO2 monitor, a temperature monitor, and / or a hydration monitor.

[0008] In some aspects, the techniques described herein relate to a medical monitoring system, wherein the visualization includes a time- varying graph of the portion of the one or more sensed values and the portion of the one or more future values of the physiological parameter.

[0009] In some aspects, the techniques described herein relate to a medical monitoring system, wherein the visualization simultaneously indicates the portion of the one or more sensed values and the portion of the one or more future values of the physiological parameter.

[0010] In some aspects, the techniques described herein relate to a medical monitoring system, wherein the machine learning model is a first machine learning model, and wherein the operations further include determining, by a second machine learning model, a probability of at least one adverse event occurring at one or more future times, wherein the determined probability is based on the physiological data and the physiological baseline of the patient.

[0011] In some aspects, the techniques described herein relate to a medical monitoring system, wherein the operations further include: determining that the determined probability exceeds a threshold probability; and alerting a user to a potential occurrence of the adverse event.

[0012] In some aspects, the techniques described herein relate to a medical monitoring system, wherein the operations further include obtaining the physiological baseline of a patient.

[0013] In some aspects, the techniques described herein relate to a medical monitoring system, wherein the physiological data are second physiological data, and wherein obtaining the physiological baseline of the patient includes: obtaining first physiological data of the patient prior to obtaining the second physiological data of the patient, wherein the first physiological data include one or more values of one or more physiological parameters of the patient; and determining the physiological baseline of the patient based on the first physiological data.

[0014] In some aspects, the techniques described herein relate to a medical monitoring system, wherein the predicted values of the physiological parameter at one or more future times include a predicted value of the physiological parameter at a time approximately five minutes subsequent to a current time.

[0015] In some aspects, the techniques described herein relate to a medical monitoring system, wherein the machine learning model includes a generative Al model.

[0016] In some aspects, the techniques described herein relate to a medical monitoring system, wherein the generative Al model includes a transformer.

[0017] In some aspects, the techniques described herein relate to a medical monitoring system, wherein the operations further include selecting the machine learning model from a plurality of machine learning models configured to generate future values of the physiological parameter, wherein the selecting the machine learning model is based on one or more characteristics of the patient.

[0018] In some aspects, a medical monitoring method includes obtaining physiological data of a patient, the physiological data indicating one or more sensed values of at least one physiological parameter of the patient, the one or more sensed values including a current value of the physiological parameter and / or one or more past values of the physiological parameter; generating, using at least one machine learning model, one or more future values of the physiological parameter, the one or more future values of the physiological parameter being predicted values of the physiological parameter at one or more future times, the generating being based on the physiological data and a physiological baseline of the patient; and providing visualization data to a display device configured to display a visualization of the physiological parameter, the visualization indicating at least a portion of the one or more sensed values of the physiological parameter and at least a portion of the one or more future values of the physiological parameter.

[0019] In some aspects, a non-transitory computer-readable storage medium stores computer-executable instructions that, when executed by one or more processors, cause the processors to perform medical monitoring operations including obtaining physiological data of a patient, the physiological data indicating one or more sensed values of at least one physiological parameter of the patient, the one or more sensed values including a current value of the physiological parameter and / or one or more past values of the physiological parameter; generating, using at least one machine learning model, one or more future values of the physiological parameter, the one or more future values of the physiological parameter being predicted values of the physiological parameter at one or more future times, the generating being based on the physiological data and a physiological baseline of the patient; and providing visualization data to a display device configured to display a visualization of the physiological parameter, the visualization indicating at least a portion of the one or more sensed values of the physiological parameter and at least a portion of the one or more future values of the physiological parameter.

[0020] An exemplary method for determining a probability of an adverse event for a patient comprises, at a computing system: receiving first physiological data of the patient, wherein the first physiological data comprises values of a plurality of physiological parameters of the patient; determining a physiological baseline of the patient based on the first physiological data; receiving second physiological data of the patient, the second physiological data comprising data obtained after the first physiological data; determining a probability of at least one adverse event occurring at one or more future times using a first trained machine learning model, wherein the determined probability is based on the second physiological data and the physiological baseline of the patient; and predicting one or more values of at least one physiological parameter at the one or more future times using a second trained machine learning model based on the second physiological data and the physiological baseline of the patient.

[0021] In some examples, the second physiological data is obtained during a surgical procedure. In some examples, the one or more future times are during the surgical procedure. In some examples, the surgical procedure is a pediatric cardiac catheterization procedure. In some examples, the first and second physiological data are obtained from a patient being monitored outside of a surgical procedure. In some examples, the physiological baseline is determined prior to at least one of insertion of a catheter into the patient, intubation of the patient, andadministration of an anesthetic to the patient. In some examples, the physiological baseline comprises a heart rate baseline, a mean blood pressure baseline, an oxygen saturation baseline, an end tidal CO2 baseline, or any combination thereof. In some examples, the second physiological data of the patient comprises one or more sets of time-series hemodynamic data.

[0022] In some examples, the one or more sets of time-series hemodynamic data comprise at least one of: a time series of heart rate measurements, a time series of mean blood pressure measurements, a time series of oxygen saturation measurements, and a time series of end tidal CO2 measurements obtained from the patient. In some examples, the one or more sets of timeseries hemodynamic data comprise a plurality of overlapping time segments of time-series hemodynamic data, wherein each of the overlapping time segments respectively include a predefined duration of hemodynamic data. In some examples, the one or more future times comprise one or more predefined times after a segment of the plurality of overlapping segments.

[0023] In some examples, the at least one adverse event comprises at least one of: a heart rate dropping below a first threshold, a blood pressure dropping below a second threshold, and an oxygen saturation dropping below a third threshold. In some examples, the first threshold comprises a predefined deviation from a heart rate baseline of the physiological baseline, the second threshold comprises a predefined deviation from a mean blood pressure baseline of the physiological baseline, and the third threshold comprises a predefined deviation from an oxygen saturation baseline of the physiological baseline. In some examples, the at least one adverse event further comprises administration of at least one resuscitative intervention comprising one or more of Epinephrine boluses, Calcium, Bicarbonate, Atropine, Ephedrine, and Phenylephrine during a surgical procedure.

[0024] In some examples, the first trained machine learning model comprises a deep neural network model. In some examples, the second trained machine learning model comprises a transformer model. In some examples, the first trained machine learning model and the second trained machine learning model both comprise a transformer model. In some examples, the first machine learning model and the second machine learning model were trained using time-series hemodynamic data and static data.

[0025] In some examples, the method comprises: determining a pre-procedure cardiac status, and wherein the probability of the occurrence of the at least one adverse event is determined based on the pre-procedure cardiac status. In some examples, the pre-procedure cardiac status isdetermined based on information associated with a surgical procedure and a health status of the patient. In some examples, the probability of the at least one adverse event occurring at one or more future times during the surgical procedure is determined by the first trained machine learning model based on the pre-procedure cardiac status, and wherein the predicted one or more values of at least one physiological parameter are predicted based on the pre-procedure cardiac status. In some examples, the at least one adverse event is predicted by the first trained machine learning model based on a physician-initiated event, wherein the physician-initiated event comprises any of an administration of an anesthetic, a balloon inflation event, and a coronary angiography event.

[0026] In some examples, the at least one physiological parameter comprises at least one of: a heart rate, an oxygen saturation, an end-tidal carbon dioxide, a mean blood pressure, a temperature, and patient levels of hydration. In some examples, the method comprises: displaying a visualization of at least one of the one or more predicted values of the at least one physiological parameter and the probability of the at least one adverse event.

[0027] An exemplary non-transitory computer readable storage medium stores instructions for predicting an occurrence of an adverse event for a patient, the instructions configured to be executed by one or more processors of a computing system to cause the system to: receive first physiological data of a patient, wherein the first physiological data comprises values of a plurality of physiological parameters of the patient; determine a physiological baseline of the patient based on the first physiological data; receive second physiological data of the patient, the second physiological data comprising data obtained after the first physiological data; determine a probability of at least one adverse event occurring at one or more future times using a first trained machine learning model, wherein the determined probability is based on the second physiological data and the physiological baseline of the patient; and predict one or more values of at least one physiological parameter at the one or more future times using a second trained machine learning model based on the second physiological data and the physiological baseline of the patient.

[0028] An exemplary system for predicting an occurrence of an adverse event for a patient, comprises one or more processors and memory storing one or more computer programs that include computer instructions, which when executed by the one or more processors, cause the system to: receive first physiological data of a patient, wherein the first physiological datacomprises a plurality of physiological parameters of the patient; determine a physiological baseline of the patient based on the first physiological data; receive second physiological data of the patient, the second physiological data comprising data obtained after the first physiological data; determine a probability of at least one adverse event occurring at one or more future times using a first trained machine learning model, wherein the determined probability is based on the second physiological data and the physiological baseline of the patient; and predict one or more values of at least one physiological parameter at the one or more future times using a second trained machine learning model based on the second physiological data and the physiological baseline of the patient.

[0029] In some examples, the system comprises: a plurality of physiological parameter sensors configured to generate the second physiological data of the patient. In some examples, the plurality of physiological parameter sensors comprises at least one of a blood pressure monitor, an oxygen saturation monitor, a heart rate monitor, an end tidal CO2 monitor, a temperature monitor, and a hydration monitor.

[0030] In some embodiments, any one or more of the characteristics of any one or more of the systems, methods, and / or computer-readable storage mediums recited above may be combined, in whole or in part, with one another and / or with any other features or characteristics described elsewhere herein.BRIEF DESCRIPTION OF THE FIGURES

[0031] A better understanding of the features and advantages of the present disclosure will be obtained by reference to the following detailed description that sets forth illustrative embodiments, in which the principles of the disclosure are utilized, and the accompanying drawings of which:

[0032] FIG. 1 illustrates an exemplary system for predicting an adverse event and / or predicting at least one future physiological parameter of a patient according to some embodiments.

[0033] FIG. 2 illustrates an exemplary process for predicting an occurrence of an adverse event occurring and / or predicting at least one future physiological parameter of a patient procedure according to some embodiments.

[0034] FIG. 3 illustrates an exemplary visualization depicting future physiological parameters of a patient during a surgical procedure according to some embodiments.

[0035] FIGS. 4A-4B illustrate a performance of an exemplary deep neural network model trained to predict adverse events according to some embodiments.

[0036] FIGS. 5A-5B illustrate a performance of an exemplary deep neural network model trained to predict adverse events limited to heart rate deceleration according to some embodiments.

[0037] FIG. 6 illustrates predictive performance of a transformer model configured to predict future physiological parameters based on synthetic training data.

[0038] FIG. 7 illustrates an exemplary timeline of events during a surgical procedure according to some embodiments.

[0039] FIG. 8 illustrates an exemplary computing system according to some embodiments.

[0040] FIG. 9 illustrates classification model performance results for different classification model configurations according to some embodiments.

[0041] FIG. 10A illustrates an AUC-ROC curve indicative of adverse event classification model performance on training data according to some embodiments.

[0042] FIG. 10B illustrates an AUC-ROC curve indicative of adverse event classification model performance on validation data according to some embodiments.

[0043] FIG. 11 A illustrates regression model performance of a baseline model, an updated model, and a tuned model each configured to predict physiological parameter values according to some embodiments.

[0044] FIG. 1 IB illustrates regression model performance of a second baseline model, a second updated model, and a second tuned model each configured to predict physiological parameter values according to some embodiments.

[0045] FIG. 12 illustrates a chart of heart rate predictions predicted by the baseline model of FIG. 11 A compared against ground truth values according to some embodiments.

[0046] FIG. 13 A illustrates heart rate predictions by the updated model of FIG. 11 A compared against ground truth values according to some embodiments.

[0047] FIG. 13B illustrates heart rate predictions by the tuned model of FIG. 11A compared against the ground truth values of FIG. 13 A according to some embodiments.

[0048] FIG. 14A illustrates heart rate predictions by the updated model of FIG. 11 A compared against ground truth values according to some embodiments.

[0049] FIG. 14B illustrates heart rate predictions by the tuned model of FIG. 11A compared against the ground truth values of FIG. 14A according to some embodiments.

[0050] FIG. 15A illustrates heart rate predictions by the updated model of FIG. 11A compared against ground truth values according to some embodiments.

[0051] FIG. 15B illustrates heart rate predictions by the tuned model of FIG. 11A compared against the ground truth values of FIG. 15A according to some embodiments.

[0052] FIG. 16A illustrates heart rate predictions by the updated model of FIG. 11 A compared against ground truth values according to some embodiments.

[0053] FIG. 16B illustrates heart rate predictions by the tuned model of FIG. 11A compared against the ground truth values of FIG. 16A according to some embodiments.

[0054] FIG. 17A illustrates model attention by the updated model of FIG. 11 A to data at different times in a time series of input data according to some embodiments.

[0055] FIG. 17B illustrates model attention by the tuned model of FIG. 11A to data at different times in a time series of input data according to some embodiments.

[0056] FIG. 18A illustrates static variable importance to the updated model of FIG. 11 A according to some embodiments.

[0057] FIG. 18B illustrates static variable importance to the tuned model of FIG. 11 A according to some embodiments.

[0058] FIG. 19A illustrates dynamic variable importance to the updated model of FIG. 11 A according to some embodiments.

[0059] FIG. 19B illustrates dynamic variable importance to the tuned model of FIG. 11 A according to some embodiments.

[0060] FIG. 20 illustrates heart rate predictions by two models compared against the ground truth values according to some embodiments.DETAILED DESCRIPTIONOverview of Some Embodiments

[0061] The rapid identification of adverse events during surgical procedures is crucial for improving patient safety and achieving positive outcomes. In the surgical environment, adverseevents, which may include physiological changes for which pharmacologic or other clinical interventions can have therapeutic benefits, can escalate quickly and lead to significant complications or even fatalities. Corrective clinical actions are generally taken after an adverse clinical event has transpired, which may be too late to prevent or mitigate harm to the patient. Early detection can allow for more effective intervention, potentially mitigating harm and improving patient outcomes.

[0062] Described herein are systems, methods, devices, and non-transitory computer readable storage media for prediction of adverse events and / or future values of physiological parameters of a patient. An exemplary system may be configured to determine a patient baseline based on physiological data (e.g., time-series physiological data). The data used to determine the baseline may be obtained prior to or during a surgical procedure, while a patient is awake or asleep in the ICU or emergency room, in-lab, and / or in a variety of other contexts. The baseline may include a baseline associated with one or more physiological parameters, such as a heart rate (e.g., the number of heart beats in a specific time period, such as the number of heart beats per minute), noninvasive mean blood pressure (NBPm), oxygen saturation, and / or end tidal carbon dioxide (CO2), status of hydration, temperature, and / or other disease specific factors. It should be understood that while reference is made to specific physiological parameters throughout, these are only meant to be exemplary and in no way limiting. Various additional or different parameters fall within the scope of this disclosure.

[0063] After determining the physiological baseline, the system may receive additional physiological data (e.g., time-series physiological data) of the patient obtained, for instance, during a surgical procedure and analyze the physiological data to predict adverse events and / or future values of physiological parameters. The system may be configured to predict intraprocedural adverse events that may correspond to above-threshold deviations from the patient physiological baseline and predict physiological parameter values at future times during a surgical procedure.

[0064] Adverse event predictions and / or physiological parameter value predictions may be made by one or more machine learning models. In some embodiments, adverse event predictions are made by a first machine learning model and physiological parameter value predictions are made by a second machine learning model that is different from the first machine learning model. One or more machine learning models used to generate adverse eventpredictions and / or physiological parameter predictions may employ a deep-learning architecture, such as a transformer architecture, that captures the time-variant and non-linear characteristics of the time-series physiological data and predicts adverse events and future physiological parameter values.

[0065] The systems and methods described herein provide a variety of technical advantages. For instance, an advantage of some embodiments of the systems described herein is that they can augment and support clinical experience, so that less experienced surgeons, anesthesiologists, and other medical professionals can avoid adverse events that a more experienced surgeon, anesthesiologist, etc. may have seen coming based on their experience, so that clinical outcomes can be improved (e.g., optimized) for each patient. Moreover, by providing predictions of physiological parameter values in addition to, or in place of, the adverse event predictions, the systems and methods described herein may enable physicians to intervene / take corrective action specifically tailored to the predicted future parameter values. For example, a given predicted future blood pressure value may be treated using one dose agent while a different predicted future blood pressure value may be treated using a different dose. Accordingly, the systems and methods described herein may augment physician knowledge and experience and enable effective, rapid, and tailored response to intra-procedural adverse events.

[0066] There is a long-felt, unsolved need in the medical device industry for medical monitoring systems that not only (1) present (e.g., display) past and current measurements of a patient’s physiological parameters, and (2) alert healthcare providers when the past and current measurements indicate that the patient is presently experiencing an adverse event (AE), but also (3) reliably predict how the patient’s physiological parameters will change in the near future (e.g., in response to events). Such events can include adverse events that have already occurred or are currently occurring, adverse events predicted to occur in the near future, and / or interventions performed by healthcare providers in response to adverse events.

[0067] Some examples of such medical monitoring systems and related techniques are described herein. In some examples, the disclosed systems constitute a technological improvement over monitoring systems that present (e.g., display) past and current measurements of a patient’s physiological parameters, and alert healthcare providers when the past and current measurements indicate that the patient is presently experiencing an adverse event (AE), because the disclosed systems can present information (e.g., visualizations) reliably indicating how thepatient’s physiological parameters will change in the near future. Based on that information, healthcare providers can infer whether an adverse event is likely to occur in the near future and can intervene in advance to prevent or mitigate the adverse event, leading to improved patient outcomes. In addition, based on the presented information, healthcare providers can infer whether such interventions are likely to be effective and, if not, healthcare providers can initiate alternative interventions sooner, leading to further improvements in patient outcomes.

[0068] Even assuming (without conceding) that existing medical monitoring systems can reliably predict some adverse events in the near future, some examples of the disclosed systems still constitute a technological improvement over adverse-event-predicting systems because the presented information (e.g., reliable predictions of future values of the patient’s physiological parameters) is much richer than an advance warning of the onset of an adverse event. For example, an advance warning of the onset of an adverse event may help healthcare providers initiate an intervention in advance, but reliable predictions of future values of the patient’s physiological parameters enable healthcare providers to infer the severity of the adverse event and the extent to which ongoing interventions are proving effective in preventing or mitigating the severity of the adverse event. Thus, relative to adverse-event-predicting systems, some examples of the disclosed systems can lead to further improvements in patient outcomes. For example, some examples of the disclosed systems can help healthcare providers calibrate and adapt their interventions in real time.

[0069] In some examples, systems and / or devices (e.g., medical monitoring systems, medical monitoring devices, medical devices, etc.) can use the techniques described herein to generate visualizations of predicted future values of a patient’s physiological parameters. For example, time varying graphs of the predicted values of physiological parameter can be generated and displayed. Such graphs may illustrate the predicted physiological parameter values during a time period extending from the current time to a specified future time (e.g., 3-10 minutes into the future, 5 minutes into the future, 10 minutes into the future, etc.). In some examples, visualizations of a patient’s predicted physiological parameter values can be displayed separately from visualizations of the patient’s past or current physiological parameter values. In some examples, visualizations of a patient’s predicted physiological parameter values can be displayed via the same user interface(s) that display visualizations of the patient’s past or currentphysiological parameter values, such that a combined visualization of past, current, and predicted future values of a physiological parameter is presented.

[0070] As noted above, the rapid and early identification of adverse events during surgical procedures is crucial for improving patient safety and achieving positive outcomes. With conventional medical monitors, it can be quite difficult for surgical teams to identify adverse events (and initiate interventions designed to mitigate the adverse events) before those adverse events occur or at their onset, because the medical monitors display measurements of past and / or current parameter values. For the same reason, it can be difficult for surgical teams to assess whether such interventions are effectively mitigating an adverse event until the adverse event has been resolved. In contrast, medical monitors that display future physiological parameter values (or combined visualizations of past, current, and predicted future values of physiological parameters) can provide a visual indication of an impending adverse event before its onset, thereby prompting the surgical team to initiate an intervention sooner than a conventional medical monitor would have prompted the team to initiate the intervention, thereby contributing to improved patient outcomes. Likewise, medical monitors that display future physiological parameter values (or combined visualizations of past, current, and predicted future values of physiological parameters) can provide a visual indication of the efficacy of an ongoing intervention while the adverse event is still being resolved, thereby prompting the surgical team to continue the intervention (if appropriate) or initiate a different intervention (if the current intervention is not resolving the adverse event). Likewise, medical monitors that predict impending adverse events and alert the surgical team to these events in advance (using visualizations or any other suitable alerting techniques) can improve patient outcomes by prompting surgical teams to initiate interventions earlier than they otherwise would.

[0071] In addition or alternatively, some embodiments may provide indications (e.g., visual indications) of an adverse event before its onset in non-surgical contexts. For example, some embodiments can predict adverse events and / or future physiological parameter values for a patient drawn from any suitable patient population (e.g., patients with or without heart disease, patients with or without anesthesia, etc.), including patients who are awake or asleep in an intensive care unit (ICU), emergency room (ER), rehabilitation facility, nursing facility, assisted living facility, and / or any other environment in which physiological parameter values are monitored.Generative Artificial Intelligence (Al) and Deep Learning (DL)

[0072] ‘Machine learning” may refer to the application of certain techniques (e.g., pattern recognition and / or statistical inference techniques) by computer systems to perform specific tasks. Machine learning techniques (automated or otherwise) may be used to build data analytics models based on sample data (e.g., “training data”) and to validate the models using validation data (e.g., “testing data”). The sample and validation data may be organized as sets of records (e.g., “observations” or “data samples”), with each record indicating values of specified data fields (e.g., “independent variables,” “inputs,” “features,” or “predictors”) and corresponding values of other data fields (e.g., “dependent variables,” “outputs,” or “targets”). Machine learning techniques may be used to train models to infer the values of the outputs based on the values of the inputs. When presented with other data (e.g., “inference data”) similar to or related to the sample data, such models may accurately infer the unknown values of the targets of the inference dataset.

[0073] The term “generative model” as used herein may generally refer to a type of machine learning model that is trained on existing data to enable the generative model to generate, based on an input or prompt, new data that shares characteristics similar to that of the training data. In some examples, a generative model may handle text. In these examples, the generative model may accept text prompts and produce text outputs. Any suitable type of Al model can be used, including predictive models, generative Al models, etc. Predictive models can analyze historical data, identify patterns in that data, and make inferences (e.g., produce predictions or forecast outcomes) based on the identified patterns. Some non-limiting examples of predictive models include neural networks (e.g., deep neural networks (DNNs), convolutional neural networks (CNNs), recurrent neural networks (RNNs), learning vector quantization (LVQ) models, etc.), regression models (e.g., linear regression models, logistic regression models, linear discriminant analysis (LDA) models, etc.), decision trees, random forests, support vector machines (SVMs), naive Bayes models, classifiers, etc.

[0074] Generative Al technology generally utilizes generative models such as Generative Adversarial Networks (GANs), transformer-based models, diffusion models (e.g., stable diffusion models), and / or Variational Autoencoders (VAEs), etc., which are based on artificial neural networks and deep learning. Deep Learning (DL) is a subset of machine learning (ML) that focuses on artificial neural networks (ANN) and their ability to learn and make decisions.Deep Learning involves the use of complex algorithms to train ANNs to recognize patterns and make predictions based on large amounts of data. A key difference between DL and traditional ML algorithms is that DL algorithms can learn multiple layers of representations, allowing them to model highly nonlinear relationships in the data. This capability makes them particularly effective for applications such as image and speech recognition, natural language processing (NPL), etc.

[0075] Most DL methods use ANN architectures, which is why DL models are often referred to as deep neural networks (DNNs). The term “deep” refers to the number of hidden layers in the neural network. For example, a traditional ANN may only contain 2-3 hidden layers, while DNNs can have as many as 150 layers (or more). DL uses these multiple layers to progressively extract higher-level features from the raw input. For example, in image processing, lower layers may identify edges, while higher layers may identify the concepts relevant to a human, such as digits or letters or faces. DL models are trained by using large sets of labeled data and ANN architectures that learn features directly from the data without the need for manual feature extraction.Hyperparameters and Hyperparameter Tuning

[0076] Hyperparameters are external configuration variables that control or guide machine learning model training. In other words, hyperparameters are parameters that control the learning process and thereby influence the ultimate structure of the model and the learned values of the model parameters. Many hyperparameters are used to guide the training of DNNs, such as the size (number of layers and number of units per layer), the learning rate (e.g., a tuning parameter in an optimization algorithm that determines the step size at each iteration while moving toward a minimum of a loss function), and initial weights of model parameters.

[0077] The process of training an ANN involves choosing hyperparameter values that control and guide the learning algorithm. The process of experimenting with different hyperparameter values to find a suitable or optimum hyperparameter set is known as hyperparameter tuning or hyperparameter optimization. Hyperparameter tuning is an important aspect of developing ML tools and Al systems, because the selected set of hyperparameters can have a significant impact on model performance and accuracy. For example, if the learning rate hyperparameter of an ANN training algorithm is too high, the model may converge too quickly with suboptimal results. On the other hand, if the learning rate is too low, training may take toolong and results may not converge. Auto ML tools may assist with or control the hyperparameter turning process.Large Language Models (LLM) and Transformer Networks

[0078] In many generative Al systems, the generative model that generates content is a large language model (LLM). A large language model (LLM) is a type of ML model that can perform a variety of natural language processing (NLP) tasks such as generating and classifying text, answering questions in a conversational manner, and translating text from one language to another. The term Targe’ refers to the number of values (parameters) the language model can change autonomously as it learns. Some LLMs have hundreds of billions of parameters. In general, LLMs are NN models that have been trained using deep learning techniques to recognize, summarize, translate, predict, and generate content using very large datasets.

[0079] Many state-of-the-art LLMs use a class of deep learning architectures called transformer neural networks (“transformer networks” or “transformers”). A transformer is a neural network that learns context and meaning by tracking relationships between data units, such as the words in a sentence. A transformer can include multiple transformer blocks, also known as layers. For example, a transformer may have self-attention layers, feed-forward layers, and normalization layers, all working together to decipher input to predict (or generate) streams of relevant output. The layers can be stacked to make deeper transformers and powerful language models.

[0080] Two key innovations that make transformers particularly adept for large language models: positional encodings and self-attention. Positional encoding embeds the order in which the input occurs within a given sequence. Rather than feeding words within a sentence sequentially into the neural network, with positional encoding, the words can be fed in non- sequentially. Self-attention assigns a weight to each part of the input data while processing it. This weight signifies the importance of that portion of the input in the context of the rest of the input. The use of the attention mechanism enables models to focus on the parts of the input that matter the most. This representation of the relative importance of different inputs to the neural network is learned over time as the model sifts and analyzes data. These two techniques in conjunction allow for analyzing the subtle ways and contexts in which distinct elements influence and relate to each other over long distances, non-sequentially. The ability to processdata non-sequentially enables the decomposition of the complex problem into multiple, smaller, simultaneous computations.

[0081] “Completion” may refer to the process of a generative model generating additional content (e.g., text) based on a provided prompt (e.g., text), e.g., providing the next word in a sentence. The additional content (e.g., text) provided by the generative model may be referred to herein as a “completion.” Completions generated by generative models may include text, audio data (e.g., speech, music, etc.), image data (e.g., images), video data (e.g., videos), time-series data, or any other suitable type of data. “Prompting” may refer to a technique in which a generative model (e.g., an LLM) is matched to a desired downstream task by formulating the task as natural language text explaining the desired behavior, such that a generative model can carry out the task by performing text completion. Often these instructions are split into a “system message” containing general task instructions providing general guidance about the desired behavior and a “prompt template” containing the portion of the prompt that contains indicator values that are substituted in each use. “Fine-tuning” may refer to the process whereby a generative model is adapted to a particular task by changing its parameters by providing prompts with desired completions.

[0082] Generative models can analyze existing content, identify patterns in the content, and combine or modify the identified patterns to generate new content. The new content can include text, images, video, music, or any other suitable type of content. Some non-limiting examples of generative models include generative adversarial networks (GANs), variational autoencoders (VAEs), autoregressive models (e.g., large language models (LLMs)), recurrent neural networks (RNNs), transformer-based models, reinforcement learning models for generative tasks, etc. Transformer-based models generally have an encoder-decoder architecture, use an attention mechanism (e.g., scaled dot-product attention, multi-head attention, masked attention, etc.) to model the relationships between different elements in a sequence of content, and perform well when processing long sequences of content. Some non-limiting examples of transformer-based models include Generalized Pre-trained Transformer 4 (GPT-4), DALL-E3, etc. Other examples of generative models with text-processing capability include Jurassic- 1, Command, and Paradigm. Generative models can benefit from hyperparameter tuning to tweak the model’s performance for desired results, as discussed above.Some Embodiments of Medical Monitoring Devices and Related Machine-Learning Techniques for Providing Advance Warning of Adverse Events

[0083] Described herein are systems, methods, devices, and non-transitory computer- readable storage media for prediction of adverse events and / or values of future physiological parameters of a patient, for instance, during a surgical procedure. An exemplary system may be configured to predict intraprocedural adverse events and / or physiological parameter values at future times during a surgical procedure using one or more machine learning models. Predictions made by the one or more machine learning models may be based on physiological data (e.g., time-series physiological data) obtained prior to and / or during the surgical procedure, as well as static procedure-specific and / or patient-specific data. In other words, the systems and methods described herein may be configured to predict adverse events alongside future physiological parameter values (e.g., hemodynamic values) utilizing both static contextual data, such as a patient’s or surgical procedure’s specific risk factors, as well as time-series physiological data acquired prior to and / or during a procedure. As used herein “static” data may refer to timeinvariant data, including patient specific features such as sex, weight, height, etc. and / or procedure specific features such as procedure type, etc. As used herein, “dynamic” data may refer to time-variant data such as heart rate, blood pressure, and so on.

[0084] Adverse events may include threshold-exceeding deviations from a patient specific physiological baseline determined by the system. For instance, adverse events may include a heart rate deceleration (as used herein, heart rate deceleration may refer to a heart rate dropping below a threshold and / or a rate of decline of a heart rate exceeding a threshold), a heart rate acceleration (as used herein, a heart rate acceleration may refer to a heart rate increasing above a threshold and / or a rate of increase in heart rate exceeding a threshold), oxygen saturation dropping below or exceeding a threshold, a rate of increase or a rate of decrease of oxygen saturation exceeding a threshold, mean blood pressure dropping below or exceeding a threshold, a rate of increase or a rate of decrease of mean blood pressure exceeding a threshold, end tidal CO2 dropping below or exceeding a threshold, and / or a rate of increase or a rate of decrease in end tidal CO2 exceeding a threshold, a hydration level dropping below a threshold value, a temperature exceeding or falling below a threshold, a rate of temperature increase or a rate of temperature decline exceeding a threshold value, and / or a variety of other disease-specific factors, any of which may be caused by the anesthesia used during the surgical procedure, thesurgical procedure itself, the insertion of a catheter, removal of the catheter, and / or other use of the catheter within the body, among several other circumstances. In some examples, an adverse event includes (or can be detected based on) a patient’s physiological parameter values (e.g., past, current, and / or predicted future physiological parameter values) satisfying one or more criteria associated with the adverse event.

[0085] The threshold values (or criteria) may be specific to different physiological parameters and / or may be percentage differences from baseline physiological parameter values. In some embodiments, an adverse event may be predicted if a patient’s heart rate is predicted to fall below a predefined heart rate that may be determined based on a heart rate baseline, an adverse event may be predicted if a patient’s oxygen saturation is predicted to fall below an oxygen saturation value that may be determined based on an oxygen saturation baseline, and so on. Adverse events may include a predicted at least 20%, at least 30%, at least 40%, and / or at least 50% change from baseline heart rate, oxygen saturation, blood pressure, and / or end tidal CO2. Adverse events may include designated severity levels (e.g., a first severity level for a predicted change of 20% to 30% from baseline, a second severity level for a predicted change of between 30% and 50% from baseline, and a third severity level for a predicted change of greater than 50% from baseline).

[0086] The physiological baseline may be determined based on one or more sets of physiological data (e.g., time-series physiological data) (e.g., heart rate data, mean blood pressure data, etc.) for a patient acquired prior to and / or during a surgical procedure. For instance, the physiological baseline may be determined based on an average heart rate, average noninvasive mean blood pressure, average oxygen saturation, and average end tidal CO2 using one or more segments of time-series physiological data including observed values of one or more of the aforementioned parameters for a patient. The baseline may be determined before a predefined event during a surgical procedure, such as before catheter insertion and / or before administration of anesthetic. By defining a patient-specific baseline, the system may more accurately determine whether monitored physiological parameters are trending toward an adverse event in light of deviations from the baseline.

[0087] While reference is made throughout to specific physiological parameters (e.g., heart rate, blood pressure, end-tidal CO2), it should be understood that many other parameters may be measured and predicted by the systems and methods described herein. For instance, the systemsand methods described herein may be configured to predict hydration, temperature, and other disease specific (or non-disease-specific) factors. It should further be understood that the systems and methods described herein may be configured to predict adverse events and / or physiological parameters for any patient population (e.g., patients with and without heart disease, patients with or without anesthesia, etc.), as well as outside of a surgical procedure altogether. For instance, the systems and methods described herein may be applied to patients who are awake or asleep in the ICU and / or ER or any other environment where physiologic parameters are measured continuously.

[0088] One or more adverse events may be predicted based on time-series physiological data obtained during a surgical procedure following determination of the patient baseline. In some embodiments, a first trained machine learning model may be configured to receive time-series physiological data including heart rate data, noninvasive mean blood pressure data, oxygen saturation data, and / or end tidal CO2 data. In some examples, data received by the first machine learning model may include an absolute value of one or more physiological parameters (e.g., heart rate, blood pressure, or any other parameter described herein) and the first machine learning model may be configured to predict adverse events based on the absolute value of the one or more physiological parameters. In some examples, data received by the first machine learning model may include a rate of change of one or more physiological parameters (e.g., heart rate, blood pressure, or any other parameter described herein), and the first machine learning model may be configured to predict adverse events based on the rate of change of one or more physiological parameters. In some examples, the data received may include a rate of acceleration or deceleration (e.g., a rate of change of a rate of change of any of the physiological parameters described herein), and the first machine learning model may be configured to predict adverse events based on the rate of change of the rate of change of the one or more physiological parameters. The first trained machine learning model may predict one or more adverse events occurring at one or more future times during the surgical procedure based on the received timeseries data. As noted above, the model may predict an adverse event when it determines, based on the physiological data of the patient received during the surgical procedure, that a patient’s physiological parameter values may drop below (or exceed) a threshold value associated with the physiological baseline of the patient. In some examples, the system may determine a probabilityof an adverse event. In some examples, the system may predict a type of adverse event. In some examples, the system may predict the occurrence of any adverse event.

[0089] One or more future physiological parameter values may be predicted based on the physiological data (e.g., time-series physiological data) received during a surgical procedure. In some embodiments, a second trained machine learning model may be configured to receive timeseries physiological data including any of the aforementioned physiological parameters received by the first machine learning model and predict one or more of the measured parameters in and / or into the future. In some examples, data received may include an absolute value of one or more physiological parameters (e.g., heart rate, blood pressure, or any other parameter described herein) and the second machine learning model may be configured to predict future physiological parameter values, rates of change of physiological parameter values, and so on as described herein, based on the absolute value of the one or more physiological parameters. In some examples, data received may include a rate of change of one or more physiological parameters (e.g., heart rate, blood pressure, or any other parameter described herein), and the second machine learning model may be configured to predict future physiological parameter values and / or rates of change of physiological parameter values based on the received rate of change of one or more physiological parameters. In some examples, the data received may include a rate of acceleration or deceleration (e.g., a rate of change of a rate of change of any of the physiological parameters described herein), and the second machine learning model may be configured to predict future physiological parameter values, rates of change of physiological parameter values, and / or acceleration or deceleration of the rates of change of physiological parameter values based on the received rate of change of the rate of change of one or more physiological parameters. In some examples, the second machine learning model is a generative model.

[0090] For instance, the second machine learning model may be configured to predict future heart rate value(s) of the patient, future mean blood pressure value(s) of the patient, future oxygen saturation value(s) of the patient, and / or future end tidal CO2 values of the patient based on the time-series data. It should be understood that the one or more machine learning models described herein may be trained to predict additional or different parameters to those described above (e.g., temperature, hydration, respiratory rate, and other disease-specific factors). The predictions generated by either or both the first and second machine learning models may bedisplayed on an interactive interface to allow users (e.g., clinicians) to monitor predicted patient physiological parameters and proactively mitigate predicted adverse events. That is, the system may be configured to generate visualizations based on the predictions, such as a visualization of monitored and predicted physiological parameter values (e.g., heart rate, mean blood pressure). The system may also generate and display visualizations and / or issue auditory, haptic, etc., alerts based on predicted adverse events.

[0091] One or more of the machine learning models may be configured to receive static data and respectively predict adverse events and / or future physiological parameter values based on both the static data and the time-series data. The static data may include a pre-procedure cardiac status determined based on information associated with a specific surgical procedure to be performed, information associated with a patient’s health status, or any combination thereof. The pre-procedure cardiac status may be a score (e.g., a numerical value from 1 to 5). The score may influence one or both of the first and second machine learning models’ predictions. For instance, a riskier procedure and corresponding score associated with such a procedure may result in an increased likelihood that the first machine learning model predicts one or more adverse events. The pre-procedure cardiac status may include and / or be obtained as described in Quinn, B. P., et al., (2022). ICU Admission Tool for Congenital Heart Catheterization (iCATCH): A Predictive Model for High Level Post-Catheterization Care and Patient Management. Pediatric critical care medicine: a journal of the Society of Critical Care Medicine and the World Federation of Pediatric Intensive and Critical Care Societies, 23(10), 822-830. https: / / doi.org / 10.1097 / PCC.0000000000003028.

[0092] To effectively capture both the time-variant and non-linear characteristics of the data, the one or more machine learning models may include a deep neural network (DNN) architecture. Deep learning models, like recurrent neural networks (RNNs), long short-term memory models (LSTMs), and transformer models, have a number of distinct advantages for time-series regression and / or classification over non-deep learning models. For example, deep learning models can model complex non-linear relationships in the data, automatically extract useful features without the need for extensive manual feature engineering, long-term dependencies, and context in the data, making them more suitable for complex time-series tasks, and can scale to handle high-dimensional data and long sequences. In contrast to conventional, non-deep learning models, the deep learning models may not assume linearity in the data, whichmay be advantageous since real-world time-series data often exhibits non-linear patterns. Additionally, unlike non-deep learning models, the deep learning models may not require the time-series data to be stationary, meaning that the statistical properties of the data can change over time, which is often the case in real-world scenarios where trends, seasonality, and other factors may cause the data to be non-stationary. Moreover, unlike conventional non-deep learning models, the deep learning models described herein may not require extensive manual feature engineering and domain expertise to identify relevant features and their interactions, and as such, can handle high-dimensional datasets and long sequences and can have the ability to capture long-term dependencies or understand the wider context in the data.

[0093] One or more of the machine learning models described herein may employ a transformer model architecture, such as a temporal fusion transformer (TFT) model, that offers unique advantages over other deep learning architectures applied to time-series physiological data. Transformer models offer a variety of technical advantages that make them powerful tools for various prediction tasks. For instance, transformer models utilize self-attention mechanisms, which allow them to weigh the importance of different points in the input data. Additionally, transformer models utilize positional encoding, meaning they process all data points simultaneously, unlike RNNs or LSTM models that process data sequentially. Transformers have shown superior performance in various tasks, from language translation to time-series prediction. See Irwan Bellow, et al., Attention Augmented Convolutional Networks, Proceedings of the IEEE / CVF International Conference on Computer Vision 3286 (2019); Eli Simhayev, Kashif Rasul & Niels Rogge, Yes, Transformers are Effective for Time Series Forecasting (+ Autoformer), HUGGING FACE (June 16, 2023), https: / / faiggi ngface. co / b i og / autoformer. They capture complex temporal dependencies without the constraints of sequentially processed models. This may allow for more flexible attention over the data, leading to more nuanced predictions. For instance, transformer models handle long sequences and high-dimensional data better relative to LSTM models and RNN models.

[0094] Additionally, transformer models enable prediction of physiological parameter values at least five minutes into the future, leveraging their superior pattern recognition and learning capabilities. Finally, transformers reduce the amount of data and time needed to train a predictive model compared to other models, while improving performance, thus improving the functioning of a computer by increasing computer processing speed, reducing processing andpower requirements, and reducing memory requirements. Accordingly, application of the transformer model to adverse event prediction and / or physiological parameter value prediction may provide for a more accurate and more interpretable system. In summary, TFTs are well suited for ingesting multivariate time series data, capturing both long-term dependencies and short-term events in the data, and their self-attention mechanism allows for dynamic focus on the critical parts of the time series. TFTs also have variable selection networks that identify and focus on the most important input features, offer higher degrees of interpretability and explainability, and can combine both static and dynamic features simultaneously. In some examples the TFT model may be used for regression and a DNN may be used for classification, but in some examples, the TFT architecture may be used for both classification and regression. It should be understood, however, that additional or different transformer model architectures to the TFT model may be implemented in the systems and methods described herein. For instance, the systems and methods described herein may utilize encoder / decoder models, auto-regressive models, a probabilistic timeseries transformer, and so on.

[0095] Additionally, while some examples provided throughout the disclosure refer to a “first” machine learning model and a “second” machine learning model, such references are meant to provide clear exemplary descriptions of systems and methods for predicting adverse events and physiological parameters (where a model may be trained for predicting adverse events and another model may be trained for predicting physiological parameter models). It should be understood that many different models may be trained for adverse event prediction (e.g., RNN models, LSTM models, transformer models, etc.), and many different models may be trained for physiological parameter prediction (e.g., RNN models, LSTM models, transformer models, etc.). In some embodiments, a single machine learning model having multiple modules may be trained for both adverse event prediction and physiological parameter prediction. For instance, a machine learning model may include multiple modules trained for different tasks (e.g., one or more modules for adverse event prediction and one or more modules trained for predicting physiological parameter values). In some examples, the one or more machine learning models and / or one or more modules of one or more machine learning models may include an ensemble model trained for different tasks.

[0096] In some examples, one or more machine learning models and / or one or more modules of one or more machine learning models may be utilized in parallel with one another. Forinstance, one or more machine learning models and / or one or more modules of one or more machine learning models may be configured to simultaneously predict adverse events and physiological parameter values. In some examples, one or more machine learning models and / or one or more modules of one or more machine learning models may be utilized in sequence. For instance, data may be input into one or more machine learning models and / or one or more modules of one or more machine learning models to predict adverse events. Once an adverse event is predicted, data may be input into one or more machine learning models and / or one or more modules of one or more machine learning models to predict physiological parameter values associated with the predicted adverse event. In some examples, the previous occurrence of an adverse event may be an input to the one or more machine learning models, and the one or more machine learning models may predict a future adverse event (or likelihood thereof) based on the previous adverse event.

[0097] The one or more machine learning models and / or one or more modules of one or more machine learning models may be trained using first training data and retrained using second training data. In some examples, one or more machine learning models and / or one or more modules of one or more machine learning models are initially trained as off-the-shelf machine learning models (e.g., transformer models) using one or more commonly utilized training data sets. In some examples, the one or more machine learning models and / or one or more modules of one or more machine learning models are then retrained / finetuned using the data described herein for adverse event prediction and / or physiological parameter prediction. In some examples, one or more machine learning models (for instance, a temporal fusion transformer model and / or DNN) may be initially trained using a subset of the parameters described herein, retrained using an expanded set of parameters, and then tuned by optimizing one or more hyperparameters (e.g., gradient clip, hidden size, hidden continuous size, attention head size, dropout range, and / or learning rate). Accordingly, the one or more machine learning models and / or one or more modules of one or more machine learning models may be trained and retrained / fine-tuned to predict adverse events and / or physiological parameter values.

[0098] In some examples, the systems and methods described herein include a continuous retrospective analysis of a preceding time-period (e.g., 10 minutes into the past, 15 minutes into the past, 20 minutes into the past, and so on) to predict the events and / or to generate physiological parameter values at a future time or time period (e.g., 5 minutes into the future, 10minutes into the future, and so on). The techniques described herein may utilize both dynamic and static features over time and may utilize a patient-specific (or population-specific) baseline value based on data specific to the patient (or population), for instance, a 10-minute window prior to catheter insertion. The patient-specific baseline may be determined by calculating the mean value of all clinical measurements obtained within a predetermined time period. The techniques described herein may employ transformer algorithms to predict adverse events and / or future physiological parameters and may generate predictions in real-time, ensuring timely and accurate results. Moreover, the techniques described herein are versatile and adaptable, capable of functioning irrespective of the population demographic or the device(s) in use (e.g., without regard to the specific sensors utilized to collect physiological or other data).

[0099] In the following description of the various embodiments, it is to be understood that the singular forms “a,” “an,” and “the” used in the following description are intended to include the plural forms as well, unless the context clearly indicates otherwise. It is also to be understood that the term “and / or” as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items. For example, “A, B, and / or C” can refer to A alone, B alone, C alone, A and B without C, A and C without B, B and C without A, or A, B, and C. It is further to be understood that the terms “includes, “including,” “comprises,” and / or “comprising,” when used herein, specify the presence of stated features, integers, steps, operations, elements, components, and / or units but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, units, and / or groups thereof.

[0100] Certain aspects of the present disclosure include process steps and instructions described herein in the form of an algorithm. It should be noted that the process steps and instructions of the present disclosure could be embodied in software, firmware, or hardware and, when embodied in software, could be downloaded to reside on and be operated from different platforms used by a variety of operating systems. Unless specifically stated otherwise as apparent from the following discussion, it is appreciated that, throughout the description, discussions utilizing terms such as “processing,” “computing,” “calculating,” “determining,” “displaying,” “generating” or the like refer to the action and processes of a computer system, or similar electronic computing device, that manipulates and transforms data represented as physical(electronic) quantities within the computer system memories or registers or other such information storage, transmission, or display devices.

[0101] The present disclosure in some embodiments also relates to a device for performing the operations herein. This device may be specially constructed for the required purposes, or it may comprise a general-purpose computer selectively activated or reconfigured by a computer program stored in the computer. Such a computer program may be stored in a non-transitory, computer readable storage medium, such as, but not limited to, any type of disk, including floppy disks, USB flash drives, external hard drives, optical disks, CD-ROMs, magnetic-optical disks, read-only memories (ROMs), random access memories (RAMs), EPROMs, EEPROMs, magnetic or optical cards, application specific integrated circuits (ASICs), or any type of media suitable for storing electronic instructions, and each connected to a computer system bus. Furthermore, the computing systems referred to in the specification may include a single processor or may be architectures employing multiple processor designs, such as for performing different functions or for increased computing capability. Suitable processors include central processing units (CPUs), graphical processing units (GPUs), field programmable gate arrays (FPGAs), and ASICs.

[0102] The methods, devices, and systems described herein are not inherently or necessarily related to any particular computer or other apparatus. Various general-purpose systems may also be used with programs in accordance with the teachings herein, or it may prove convenient to construct a more specialized apparatus to perform the required method steps. The structure for a variety of these systems will appear from the description below. In addition, the present invention is not described with reference to any particular programming language. It will be appreciated that a variety of programming languages may be used to implement the teachings of the present disclosure as described herein.

[0103] FIG. 1 illustrates an exemplary computing system 100 configured to predict adverse events and / or future physiological parameter values (e.g., during a surgical procedure) based on physiological data received from one or more sensors monitoring a patient’s physiological parameters and / or pre-procedural data (e.g., static data). In some examples, computing system 100 is a medical monitoring system, medical monitoring device, medical device, etc., or a component thereof. In some examples, computing system 100 is remote from but communicatively coupled to a medical monitoring system, medical monitoring device, medicaldevice, etc. Computing system 100 may include one or more machine learning models (e.g., machine learning model 105 and machine learning model 106) stored in memory 102 configured to be executed by one or more processors 101 of the computing system 100. At least one of the one or more machine learning models, for instance, a first machine learning model 105 may be trained to predict one or more adverse events occurring during a surgical procedure based on observed patient data obtained during the surgical procedure. Adverse events may include a physiological parameter, such as a heart rate or mean blood pressure, falling below a threshold value. Additionally, at least one of the one or more machine learning models, for instance, a second machine learning model 106, may be trained to predict future physiological parameter values (e.g., during a surgical procedure). For instance, one of the machine learning models may be trained to predict future heart rate, future mean blood pressure, etc., based on observed patient data obtained during the surgical procedure.

[0104] The predictions generated by each of the respective machine learning models may be based on observed data, which may include (without limitation) physiological parameter values obtained during a procedure. The observed data may include data associated with one or more physiological parameters obtained using one or more sensors 104 configured to monitor a patient 110 undergoing a surgical procedure. The one or more sensors 108 may include, for instance, blood pressure monitor(s), oxygen saturation monitor(s), heart rate monitor(s), carbon dioxide (CO2) monitor(s), thermometers, etc. The sensors 108 may be configured to obtain time-series data corresponding to a patient’s heart rate, non-invasive mean blood pressure, oxygen saturation, end tidal CO2, temperature, or other parameter, or any combination thereof, as described throughout. The computing system may use time-series data acquired at a first time before (e.g., during pre-op) or during the procedure to determine a patient’s physiological baseline (e.g., a baseline heart rate, baseline non-invasive mean blood pressure, baseline oxygen saturation, baseline end tidal CO2, etc.) and may use time-series data acquired at a second time during the procedure, for instance, after determination of the physiological baseline, to predict adverse events (which may be defined in terms of deviation from a baseline value) and / or to predict future physiological parameters.

[0105] The computing system 100 may include one or more input / output components 103 (e.g., user interfaces) configured to enable presentation of the predicted adverse events and / or the predicted future physiological parameter values (e.g., display of an interactive visualization ofthe predicted adverse events and / or the predicted future physiological parameter values). For instance, the input / output components 103 may be configured to generate one or more alerts when machine learning model 105 predicts an adverse event. For instance, one or more alerts may be generated if the machine learning model 105 determines that a probability of an adverse event occurring within a predefined time period in the future exceeds a threshold. As an example, if machine learning model 105 determines that a probability that the patient’s heart rate will decline more than 20% from a baseline threshold value within, for instance, five minutes following the last observed heart rate data point, then the computing system may generate an alert indicating a predicted heart rate deceleration adverse event and display the alert via input / output component 103. The alert may include one or more user selectable affordances that enable a user (e.g., physician) to obtain additional information about the predicted adverse event. For instance, upon selection of the affordance, the input / output component 103 may display an indication of a severity of the predicted adverse event. As noted, computing system 100 may also be configured to generate a visualization of predicted future physiological parameters of the patient 110 based on predictions generated by machine learning model 106. An exemplary visualization of predicted future physiological parameters of a patient is provided in FIG. 3 and discussed in additional detail below. Computing system 100 may be configured to implement an exemplary process for predicting / determining a probability of at least one adverse event occurring during a surgical procedure and / or for predicting future values of at least one physiological parameter of a patient during a surgical procedure, as described in detail below with reference to FIG. 2.

[0106] FIG. 2 illustrates an exemplary process for predicting at least one adverse event occurring for a patient and / or for predicting future values of at least one physiological parameter of the patient. The adverse events and physiological parameter values may be predicted during a surgical procedure, such as a pediatric cardiac catheterization procedure or any other surgical procedure, according to some embodiments. The adverse events and physiological parameter values may be predicted outside of a surgical procedure, for instance, for a patient in the ER, ICU, or any other environment where physiological data maybe observed (e.g., continuously observed). Process 200 is performed, for example, using one or more electronic devices implementing a software platform. In some examples, process 200 is performed using one or more electronic devices. In some embodiments, process 200 is performed using a client-serversystem, and the blocks of process 200 are divided up in any manner between the server and one or more client devices (e.g., medical monitoring devices). Thus, while portions of process 200 are described herein as being performed by particular devices, it will be appreciated that process 200 is not so limited. In process 200, some blocks are, optionally, combined, the order of some blocks is, optionally, changed, and some blocks are, optionally, omitted. In some examples, additional steps may be performed in combination with the process 200. Accordingly, the operations as illustrated (and described in greater detail below) are exemplary by nature and, as such, should not be viewed as limiting.

[0107] At block 202, an exemplary system (e.g., computing system 100 of FIG. 1 above, or physician, operator, etc.) may optionally determine a pre- procedure cardiac status (also known as a precardiac score) based on pre-procedural data (e.g., data obtained prior to or otherwise not during a surgical procedure). The pre-procedure cardiac status may be a score of one, two, three, four, or five, and may be determined based on information associated with the surgical procedure and / or a health status of the patient; although, it should be understood that other metrics are within the scope of this disclosure. For example, the pre-procedure cardiac status may be relatively higher for higher-risk surgeries than for lower-risk surgeries. The pre-procedure cardiac status may also be relatively higher for patients with one or more risk factors, such as hypertension, obesity, etc. Accordingly, the pre-procedure cardiac status may be impacted by a patient’s medical history, including EC G / EKG results, stress test results, physical examinations, or other data included in a patient’s medical record. The pre-procedure cardiac status may impact the system’s prediction of an adverse event during a surgical procedure and / or the system’s prediction of future physiological parameters such as heart rate, noninvasive mean blood pressure, etc., as described in more detail below. For instance, a relatively higher pre-procedure cardiac status may result in a higher probability of a predicted adverse event. An exemplary tool for determining pre-procedure cardiac status is provided below in Table 1 :Table 1: Pre-Procedure Cardiac Status (PCS) ToolSingle Ventricle Categories Variable Category 1 Category 2 Category 3Saturation >80% 70-80% <70%Mean PAp Normal - * Severely elevatedT, . , _ . Moderate / severeVentricular Function - — ,rdysfunctionAVVR None or mild Moderate SevereBiventricular Categories Variable Category 1 Category 2 Category 3Saturation >95% 85-95% <85%RVSp systemicAnticipated PVR Normal Moderately elevated Severely elevated„x■ ,r. Moderate / severeVentricular Function - — ,rdysfunctionSystemic AVVR None or mild Moderate Severe* Severely elevated mPAp: anticipated mPAp >20 mmHg x Anticipated PVR reference: normal being PVR <4 iWU, moderately elevated 4-8 iWU, and severely elevated >8iWULegend: PAp, pulmonary artery pressure; RVSp, right ventricular systolic pressure; PVR, pulmonary vascular resistance; AVVR, atrioventricular valve regurgitation

[0108] At block 204, the system (e.g., computing system 100) receives first physiological data of a patient. The first physiological data may be obtained using one or more sensors prior to a surgical procedure (e.g., during pre-op), in the ER or ICU (or otherwise outside of a surgical procedure), and / or during a surgical procedure prior to one or more events of the procedure, for instance, before intubation, before insertion of a catheter, before administration of anesthesia, etc. The one or more sensors may include blood pressure monitor(s), oxygen saturation monitor(s), heart rate monitor(s), carbon dioxide (CO2) monitor(s), thermometers, etc. The sensors used to obtain the first physiological data may be the same as or different from the sensors used to obtain the second physiological data. The first physiological data may include a plurality of physiological parameters of the patient, such as any of heart rate data, mean blood pressure data, oxygen saturation data, and end tidal CO2 data, temperature data, hydration data, or other disease-specific data of the patient. The first physiological data may include time-series physiological data, for instance, including a time series of heart rate measurements, a time series of mean blood pressure measurements, a time series of oxygen saturation measurements, and a time series of end tidal CO2 measurements, a time series of temperature measurements, a timeseries of hydration measurements, and / or a time series of any other physiological data obtained from the patient. The time-series hemodynamic data may include minute-to-minute data obtained using the one or more monitoring devices, although it should be understood that time-series data may be captured at any other interval (e.g., second-to-second).

[0109] The data received at block 204 may optionally be processed by the system to remove outliers and format, and segment time-series data into overlapping series (e.g., overlapping segments of a predefined duration, such as 10 minutes). For instance, for heart rate data, oxygen saturation data, and end tidal CO2 data, missing values may be imputed using the carry forward method. If a particular record is missing more than 10% of the values for any of the previously mentioned hemodynamics, the missing values may be imputed using the carry-forward method. The carry-forward method takes the most recent previous value that exists and applies it to the following time period (e.g., the following minute). The system may alternatively be configured to fill missing values using a carry-back method, imputing the missing values based on those observed at future time frames. Regarding outliers, values that fall outside the following ranges may be imputed: heart rate (0 to 300 BPM), oxygen saturation (30% to 100%), end tidal CO2 (0 mmHg to 100 mmHg), and noninvasive mean blood pressure (0 mmHg to 200 mmHg). Values outside of this range are highly unlikely to occur and may consequently be imputed.

[0110] At block 206, the system may determine a physiological baseline of the patient based on the first physiological data. The physiological baseline may include a heart rate baseline, a mean blood pressure baseline, an oxygen saturation baseline, an end tidal CO2 baseline, or any combination thereof. The physiological baseline may include a respective average of a heart rate, a mean blood pressure, an oxygen saturation, and an end tidal CO2 determined for a predefined segment of the time-series data received at block 204. For example, 10 minutes of each of heart rate data, mean blood pressure data, oxygen saturation data, and end tidal CO2 data of the patient may be received during the surgical procedure prior to one or more surgical events (e.g., a 10- minute period prior to insertion of a catheter). An average value may be determined for each of the heart rate, mean blood pressure, oxygen saturation, and end tidal CO2, and those averages may be set as the baseline for the patient during the remainder of the procedure.

[0111] As noted, in some examples the physiological baseline may be determined prior to one or more surgical events occurring during a procedure. For instance, the physiological baseline may be determined based on a segment of time-series data acquired for a period of time(e.g., a segment of time-series data of a predefined length) prior to at least one of insertion of a catheter into the patient, intubation of the patient, and / or administration of an anesthetic to the patient. The time at which the physiological baseline determination occurs may vary from patient to patient. For instance, the time at which the physiological baseline determination occurs may depend on whether a patient is classified as stable or unstable. For example, for stable patients, a physiological baseline may be determined 20 to 30 minutes post-intubation. For unstable patients, a physiological baseline may be determined, for instance, 10 minutes before administration of an anesthetic. It should be understood that using an average of a time series of measured values obtained at a predefined time prior to least one of insertion of a catheter into the patient, intubation of the patient, and administration of an anesthetic to the patient is only meant to be an exemplary method of determining a baseline. A variety of additional options for determining the physiological baseline are available. For instance, the physiological baseline may be determined for respective patient populations (e.g., grouped according to age, weight, biological sex, etc.) based on historical data or may be determined for individual patients based on measurements obtained during a pre-op to lab phase and / or at any other time. In some examples, a baseline may be obtained at any point in a time series where the data does not vary significantly (e.g., where the data varies less than 1% over a predefined interval, less than 2% over a predefined interval, less than 3% over a predefined interval, less than 4% over a predefined interval, less than 5% over a predefined interval, less than 6% over a predefined interval, less than 7% over a predefined interval, less than 8% over a predefined interval, less than 9% over a predefined interval, and / or less than 10% over a predefined interval).

[0112] At block 208, the system may receive second physiological data of the patient. The second physiological data may include static and dynamic data (e.g., time-series data), and at least a portion of the second physiological data may be received after determining the physiological baseline. The second physiological data may be obtained during a surgical procedure. The sensors used to obtain the second physiological data may be the same as or different from the sensors used to obtain the first physiological data. The sensors used to obtain the second physiological data may include blood pressure monitor(s), oxygen saturation monitor(s), heart rate monitor(s), carbon dioxide (CO2) monitor(s), thermometers, hydration monitors, etc. It should be understood that the static data may be received by the system at any time (e.g., not only after determining the physiological baseline), and the dynamic data may beobtained after obtaining the first physiological data using one or more monitoring devices, for instance, during a surgical procedure or while monitoring a patient in the ER or ICU. The static data may include the pre-procedure cardiac status determined based on pre-procedural data at block 202. The static data may include a patient’s age, a duration of the surgical procedure, ventricular function of the patient, ejection fraction of the patient, ventricular size of the patient (Z score), valvular function of the patient, regurgitation or stenosis, pulmonary vascular resistance of the patient, or any combination thereof.

[0113] The dynamic data of the second physiological data may include time-series hemodynamic data of the patient, drug administration data, and / or information associated with a physician-initiated event. Drug administration data may include minute-to-minute data on when drugs (e.g., anesthetics) are administered. In some examples, drug administration data includes dosage information for each administration. The physician-initiated event may include, for instance, a balloon inflation event, a coronary angiography event, or any other intervention that interrupts the pattern of physiologic data. The balloon inflation event or a coronary angiography event may cause a rapid increase and / or decrease in one of the physiological parameters captured in the second physiological data, which may impact the first machine learning model’s prediction of an adverse event at a future time during the procedure. Providing the model with dynamic data representative of such events may enable the model to make more accurate predictions by enabling the model to recognize whether severe fluctuations in the time-series data are likely associated with a physician-initiated event.

[0114] Like the data received at block 204, the second physiological data received at block 208 may be processed by the system to remove outliers and format, and to segment time-series data into overlapping series (e.g., overlapping segments of a predefined duration, such as 10 minutes). For instance, for heart rate data, oxygen saturation data, and end tidal CO2 data, missing values may be imputed using the carry-forward method. If a particular record is missing more than 10% of the values for any of the previously mentioned hemodynamics, the missing values may be imputed using the carry-forward method. The carry forward method takes the most recent previous value that exists and applies it to the following time period (e.g., the following minute). The system may alternatively be configured to fill missing values using a carry-back method, imputing the missing values based on those observed at future time frames (for instance, during training or retraining of the one or more machine learning models). Thesystem may also remove outliers and impute values when measured values fall outside the following ranges: heart rate (0 to 300 BPM), oxygen saturation (30% to 100%), end tidal CO2 (0 to 100), and noninvasive mean blood pressure (0 mmHg to 200 mmHg). Values outside this range are highly unlikely to occur and may consequently be imputed.

[0115] The system may further calculate values (e.g., feature values) that will be subsequently input (e.g., automatically by the computing system and / or by a user of the system) to the first and / or second machine learning models based on the second physiological data. For instance, the system may calculate deviations from the physiological baseline (e.g., absolute values of the physiological parameters at each time point in the received time-series data may be converted to deviations from the corresponding baseline value for that parameter). The system may calculate a rate of change of one or more of the physiological parameters that is subsequently input into either or both of the machine learning models. The system may calculate a rate of change of a rate of change of one or more of the physiological parameters that is subsequently input into either or both of the machine learning models. In some examples, the system may be configured to input the absolute value of the received parameters in the timeseries data to either or both of the machine learning models.

[0116] Also, like the first physiological data received at block 204, the time-series hemodynamic data may include at least one of: a time series of heart rate measurements, a time series of mean blood pressure measurements, a time series of oxygen saturation measurements, and a time series of end tidal CO2 measurements or other data described throughout obtained from the patient, for instance, during surgical procedure. The time-series hemodynamic data may be segmented into a plurality of segments of hemodynamic data having predefined durations (e.g., 1 minute, 5 minutes, 10 minutes, 20 minutes, etc.). Each set of time-series hemodynamic data may include a predefined duration of at least 1 minute, at least 2 minutes, at least 3 minutes, at least 4 minutes, at least 5 minutes, at least 6 minutes, at least 7 minutes, at least 8 minutes, at least 9 minutes, at least 10 minutes, at least 11 minutes, at least 12 minutes, at least 13 minutes, at least 14 minutes, at least 15 minutes., at least 20 minutes, at least 30 minutes, at least 40 minutes, at least 50 minutes, at least 1 hour, at least 2 hours, at least 3 hours, at least 4 hours, and / or at least 5 hours. Each set of time-series hemodynamic data may include a predefined duration of hemodynamic data of no more than 5 hours, no more than 4 hours, no more than 3 hours, no more than 2 hours, no more than 1 hour, no more than 60 minutes, no more than 50minutes, no more than 40 minutes, no more than 30 minutes, no more than 20 minutes, no more than 15 minutes, no more than 14 minutes, no more than 13 minutes, no more than 12 minutes, no more than 11 minutes, no more than 10 minutes, no more than 9 minutes, no more than 8 minutes, no more than 7 minutes, no more than 6 minutes, no more than 5 minutes, no more than 4 minutes, no more than 3 minutes, no more than 2 minutes, and / or no more than 1 minute.

[0117] The time-series hemodynamic data may be segmented into a plurality of overlapping time segments of time-series hemodynamic data. Each of the overlapping time segments may respectively include a predefined duration of hemodynamic data. For example, a first time segment may include time-series data from minute 1 to minute 10 (e.g., following determination of the physiological baseline, or other surgical event such as insertion of a catheter). A second time segment may include time-series data from minute 2 to minute 11, a third time segment may include time-series data from minute 3 to minute 12, and so on. It should be understood that the aforementioned distribution of time segments is meant to be exemplary and not limiting. For example, the first time segment may include time-series data from minute 1 to minute 10, and the second time-segment may include time-series data from minute 1 and one second to minute 10 and one second. In addition to, or as an alternative to the raw time-series data included in the overlapping time segments, any of the calculated values described above (e.g., rate of change) associated with the data from one or more of the overlapping segments may be input into the first and / or second machine learning models.

[0118] At block 210, the system may predict least one adverse event occurring at a future time using a first machine learning model. In some examples, the system may determine a probability of an adverse event occurring at a future time using the first machine learning model. In some examples, the system may predict a type of adverse event. In some examples, the system may predict the occurrence of any adverse event. The future time may be during a surgical procedure, following a surgical procedure, or at any future time following receipt of the second physiological data. The prediction / determined probability of an adverse event may be based on the second physiological data and the physiological baseline of the patient. The first machine learning model may be trained to predict that one or more physiological parameters, including hemodynamic parameters, will fall below or exceed a threshold value (e.g., a percentage decline relative to the pertinent physiological baseline parameter, such as the heart rate baseline), that a rate of decline will exceed a threshold value, and / or that a rate of increase will exceed a thresholdvalue based on the time-series data and / or static data included in the second physiological data. The first machine learning model may predict various adverse events based on specific criteria, as described in further detail below. For instance, the primary adverse event of interest may be “heart rate deceleration,” which may include a more than 20% decrease in baseline heart rate and the administration of one of the following medications: epinephrine boluses, calcium, bicarbonate, atropine, ephedrine, phenylephrine, or dopamine. As noted above, a heart rate deceleration may include a drop in heart rate below a threshold and / or a rate of heart rate decline exceeding a threshold.

[0119] The at least one adverse event may include one or more of: a heart rate dropping below a first threshold, a blood pressure dropping below a second threshold, and an oxygen saturation dropping below a third threshold, etc. The at least one adverse event may additionally or alternatively include an oxygen saturation dropping below a third threshold along with an increase in the amount of oxygen (FIO2) a patient is receiving. The first threshold may include a predefined deviation from a heart rate baseline of the physiological baseline. For example, the at least one adverse event may include a decrease in heart rate of 20% to 30% relative to the baseline heart rate of the physiological baseline, a decrease of 30% to 50% relative to the baseline heart rate of the physiological baseline, and / or a decrease of 50% or more relative to the baseline heart rate of the physiological baseline. Predicting a decrease in heart rate of 20% to 30% relative to the baseline heart rate may include predicting a moderate heart rate deceleration. Predicting a decrease in heart rate of 30% to 50% relative to the baseline heart rate may include predicting a severe heart rate deceleration. Predicting a decrease in heart rate of 50% or more relative to the baseline heart rate may include predicting a cardiac arrest.

[0120] The second threshold may include a predefined deviation from a mean blood pressure baseline of the physiological baseline. For example, the at least one adverse event may include a decrease in mean blood pressure of 20% to 30% relative to the baseline mean blood pressure of the physiological baseline, a decrease in mean blood pressure of 30% to 50% relative to the baseline mean blood pressure of the physiological baseline, and / or a decrease in mean blood pressure of 50% or more relative to the baseline mean blood pressure of the physiological baseline. Predicting a decrease in mean blood pressure of 20% to 30% relative to the baseline mean blood pressure may include predicting a moderate blood pressure decline. Predicting a decrease in mean blood pressure of 30% to 50% relative to the baseline mean blood pressuremay include predicting a severe blood pressure decline. Predicting a decrease in mean blood pressure of 50% or more relative to the baseline mean blood pressure may include predicting a circulatory collapse.

[0121] The third threshold may include a predefined deviation from an oxygen saturation baseline of the physiological baseline. For example, the at least one adverse event may include a decrease in oxygen saturation of 20% to 30% relative to the baseline oxygen saturation of the physiological baseline, a decrease in oxygen saturation of 30% to 50% relative to the baseline oxygen saturation of the physiological baseline, and / or a decrease in oxygen saturation of 50% or more relative to the baseline oxygen saturation of the physiological baseline.

[0122] In some examples, the first, second, and third thresholds described above may correspond to a rate of increase or decrease for, e.g., a heart rate, mean blood pressure, and oxygen saturation. In some examples, thresholds may be set for a variety of other parameters (e.g., temperature, end-tidal CO2, hydration, and so on) and adverse events may be predicted based on a predicted decline below or rise above any of the respective thresholds.

[0123] The adverse event prediction of the first machine learning model may be influenced by static data in addition to the time-series hemodynamic data. For instance, the predicted adverse event(s) may be based at least in part on the pre-procedure cardiac status. As described above, the pre-procedure cardiac status may influence the model’s prediction by increasing or decreasing the likelihood of an adverse event (e.g., a relatively worse status may increase the probability of an adverse event).

[0124] The at least one adverse event predicted by the first machine learning model may also include a predicted pharmaceutical intervention, such as administration of at least one of epinephrine boluses, calcium, bicarbonate, atropine, ephedrine, dopamine, and phenylephrine during the surgical procedure. In other words, the first trained machine learning model may be trained to predict both a decrease in heart rate (or mean blood pressure, oxygen saturation, etc.) relative to baseline and the administration of any of epinephrine boluses, calcium, bicarbonate, atropine, ephedrine, dopamine, and / or phenylephrine during the surgical procedure.

[0125] The future time at which the adverse event is predicted may be a predefined time after a segment of the plurality of overlapping segments (e.g., at least one minute, at least two minutes, at least three minutes, at least four minutes, at least five minutes, at least six minutes, at least seven minutes, at least eight minutes, at least nine minutes, at least ten minutes, no morethan ten minutes, no more than nine minutes, no more than eight minutes, no more than seven minutes, no more than six minutes, no more than five minutes, no more than four minutes, no more than three minutes, no more than two minutes, and / or no more than one minute). The system may continuously receive sets of time series data and predict whether an adverse event will occur at a predefined time or set of times (e.g., an interval) following each set of time-series data. Accordingly, the future time may include a predefined time following each respective set of time-series hemodynamic data.

[0126] The first machine learning model may be a classifier model, such as a deep neural network model trained (and optionally iteratively retrained) for classification of adverse events based on both static and dynamic data, although other models may be configured to predict adverse events, for instance, the first machine learning model may instead be a transformer model, such as a temporal fusion transformer model described in additional detail below with reference to the second trained machine learning model, or may be any classifier model capable of handling both static data and time-series data. The first machine learning model may be trained (and optionally iteratively retrained) to predict adverse events based on static data and time-series hemodynamic data acquired from a plurality of subjects (e.g., patients). For instance, the first machine learning model may be trained using data acquired from the Boston Children’s Hospital (BCH) Anesthesia Information System (AIMS) or the BCH Sensis data source, which is a database of cardiac data. The first machine learning model may additionally or alternatively be trained using synthetic data, as described further below with reference to FIG. 6.

[0127] Using both static data (e.g., pre-procedure cardiac status, age, biological sex, or any of the other data points described above) and dynamic data (e.g., time-series hemodynamic data, drug administration data, clinician initiated event data, etc.) may enable the machine learning model to capture both the time-invariant characteristics and the temporal dynamics present in the dataset. As described above, both the static and dynamic data may be preprocessed by the system to, for instance, remove outliers, fill missing data, segment the time-series data into overlapping / cascading time segments (e.g., 10-minute segments), and / or ensure that all variables are in the correct format.

[0128] The first machine learning model may include separate pathways or channels for the dynamic and static data. For example, the static data may be processed using a first pathway including a feedforward network, dense layers, etc., and the dynamic data may be processedusing a second pathway including recurrent neural network(s) RNN(s) such as LSTM (Long Short-Term Memory) or Gated Recurrent Units (GRUs), which are configured to handle timeseries data and capture temporal dependencies. After the data is processed using one of the respective pathways, representations (e.g., higher dimensional embeddings or other representations of features and patterns recognized in the data) learned from both of the static and dynamic data may be combined, for instance, by concatenation (e.g., joining vector representations of static data to vector representations of dynamic data) or more complex merging strategies. The first machine learning model (e.g., the adverse event prediction model) may be trained using both the static and dynamic feature representations jointly, which may enable the model to learn relationships between the static features and temporal patterns / trends. The first machine learning model may be trained based on the static and dynamic data representations using backpropagation where the error is calculated based on the classification outcomes (e.g., via comparison to ground truth values) and propagated back through the model to adjust various weightings and / or other configurations during training and / or retraining.

[0129] In some embodiments, the first machine learning (ML) model is selected from a set of ML models configured (e.g., trained) to predict adverse events at future times (e.g., five minutes into the future). As described in further detail below, experimental data indicate that different ML models (e.g., models having different architectures, models having different features, models trained using different hyperparameter values and / or different training data, etc.) can perform better (e.g., provide more accurate, sensitive, specific, and / or precise predictions of adverse events) for different subsets of patients. For example, different ML models can perform better for subsets of patients having similar characteristics, such as the characteristics represented by the patients’ static data (e.g., the patient’s pre-procedural cardiac status, age, ventricular function, ejection fraction, ventricular size (Z score), valvular function, regurgitation or stenosis, pulmonary vascular resistance, etc., or any combination thereof).

[0130] In some examples, the first ML model is selected from the set of ML models based on characteristics of the patient. For example, the first ML model can be selected based on (1) the patient’s static data, (2) a subset of patients to which the patient is assigned (or with which the patient is associated) based on the patient’s static data, and / or (3) the performance of the first ML model on a validation dataset drawn from patients having static data similar to the static data of the patient or similar to the static data of the subset of patients to which the patient is assigned. Insome examples, the use of different ML models for different subsets of patients may be described as “ensemble modeling.”

[0131] Similarity of the patient’s static data to the static data of any other patient or subset of patients may be assessed using any suitable criteria or technique. In some examples, a patient’s static data is similar to the static data of a subset of patients if one or more specified features of the patients’ static data have values that satisfy one or more corresponding criteria (e.g., patient age within a specified range, patent ejection fraction within a specified range, etc.). In some examples, a patient’s static data is similar to the static data of a subset of patients if (1) a group of patients partitioned into subsets by applying a clustering algorithm to the patients’ static data, and (2) the patient is assigned to the same cluster as the subset of patients.

[0132] At block 212, the system may predict a value of at least one physiological parameter at one or more future times using a second machine learning model. The one or more future times may be during a surgical procedure. However, as noted above, physiological parameters may be predicted outside of a surgical procedure. In some embodiments, both blocks 210 and 212 may be performed. In some embodiments, only one of blocks 210 and 212 may be performed (e.g., the system may predict adverse events but not physiological parameter values, or the system may predict physiological parameter values and not adverse events). Blocks 210 and 212 may be performed in parallel or in sequence, in any order. Prediction of physiological parameter values (in addition to, or in place of, predicting adverse events) may be a desirable feature to enable physicians to more precisely tailor interventions relative to adverse event predictions. For instance, by providing projections of specific physiological parameter values, such as a specific heart rate, blood pressure, etc. at a future time during a surgical procedure, physicians can select a response / intervention specifically tailored to that predicted value.

[0133] The system may be configured to predict adverse events and / or physiological parameters at the time corresponding to the adverse event. For instance, the system may predict a heart rate deceleration adverse event and may also output a predicted heart rate at time corresponding to the predicted deceleration. The second machine learning model may predict the value of at least one physiological parameter based on the second physiological data. The prediction may additionally be based at least in part on the physiological baseline of the patient and / or on static data, such as the pre-procedure cardiac status determined based on preprocedural data at block 202. The static data may additionally or alternatively include a patient’sage, a duration of the surgical procedure, ventricular function of the patient, ejection fraction of the patient, ventricular size of the patient (Z score), valvular function of the patient, regurgitation or stenosis, pulmonary vascular resistance of the patient, or any combination thereof. The prediction may also be based at least in part on information associated with administration of one or more anesthetics during the procedure. For example, the individual effects of common anesthetic medications such as, but not limited to Isoflurane, Morphine, Fentanyl, and / or Propofol may be received as part of the second physiological data and provided to the model.

[0134] Accordingly, in some examples (e.g., where the techniques described herein are implemented) during a procedure, the second trained machine learning model may receive a patient’s determined physiological baseline determined based on the first physiological data described above; static data including pre-procedure cardiac status (i.e., precardiac score) and / or a patient’s age, a duration of the surgical procedure, ventricular function of the patient, ejection fraction of the patient, ventricular size of the patient (Z score), valvular function of the patient, regurgitation or stenosis, pulmonary vascular resistance of the patient, information associated with an anesthetic administered during the procedure, or any combination thereof; and dynamic time-series physiological data included in the second physiological data. Based on the received data, described above, the second machine learning model may predict at least one future physiological parameter of the patient.

[0135] The at least one physiological parameter predicted by the second machine learning model may include at least one of: a heart rate, an oxygen saturation, an end-tidal carbon dioxide, a mean blood pressure, a temperature, patient hydration, and / or other disease specific factors. For example, the second machine learning model may predict a future heart rate, future oxygen saturation, future end-tidal carbon dioxide, and / or a future mean blood pressure during a surgical procedure based on a time series of heart rate data, a time series of oxygen saturation data, a time series of end-tidal carbon dioxide data, and / or a time series of mean blood pressure data collected during the surgical procedure. The predicted physiological parameter(s) may include a time-series of predicted values. For instance, the second machine learning model may be configured to predict a plurality of future heart rates (e.g., the heart rate at each minute for a predefined time period following the end of a segment of received time-series physiological data) at a plurality of different times. In other words, the second machine learning model may beconfigured to predict a time series of physiological parameter(s) based on the received timeseries of physiological parameters.

[0136] The second machine learning model may be trained to predict future values of physiological parameters based on a time series of absolute physiological parameter values, a time series of changes in physiological parameter values compared to a physiological baseline value, and / or a time series of changes in physiological parameter values as a percentage of the baseline value. The second machine learning model may also be trained based on any of the static data described above, as well as based on data associated with intra-procedural events such as balloon dilation events and / or coronary angiography events, and / or drug administration indicators. The second machine learning model may be trained using data acquired from the Boston Children’s Hospital (BCH) Anesthesia Information System (AIMS) or the BCH Sensis data source, a database of cardiac data. The second machine learning model may additionally or alternatively be trained using synthetic data, as described further below with reference to FIG. 6.

[0137] Like the first trained machine learning model, the second trained machine learning model may be trained based on both static and dynamic data (e.g., time-series data). In some examples, the second trained machine learning model may include a different model architecture than the first trained machine learning model. For instance, the first trained machine learning model may be a deep neural network such as a dense neural network trained for classification tasks, and the second machine learning model may be a transformer model, such as a temporal fusion transformer model trained (and optionally iteratively retrained) for regression tasks (although, as noted above, the first machine learning model may also include the same temporal fusion model architecture as the second model, but may be trained (and optionally iteratively retrained) to predict adverse events).

[0138] Transformer models are a type of deep learning architecture introduced by Ashish Vaswani et al., Attention is All You Need, Neural Information Processing Systems (2017), which is incorporated herein by reference in its entirety. Transformer models offer a variety of technical advantages that make them powerful tools for various prediction tasks. For instance, transformer models utilize self-attention mechanisms, which allow them to weigh the importance of different points in the input data. Additionally, transformer models utilize positional encoding, meaning they process all data points simultaneously, unlike Recurrent Neural Networks (RNNs) or Long Short-Term Memory (LSTM) models that process data sequentially. Transformers have shownsuperior performance in various tasks, from language translation to time-series prediction. See Irwan Bellow, et al., Attention Augmented Convolutional Networks, Proceedings of the IEEE / CVF International Conference on Computer Vision 3286 (2019); Eli Simhayev, Kashif Rasul & Niels Rogge, Yes, Transformers are Effective for Time Series Forecasting (+ Autoformer), HUGGING FACE (June 16, 2023), https: / / huggingface.co / blog / autoformer. They capture complex temporal dependencies without the constraints of sequentially processed models. This may allow for more flexible attention over the data, leading to more nuanced predictions. For instance, transformer models handle long sequences and high dimensional data better relative to LSTM models and RNN models. Additionally, transformer models enable prediction of physiological parameters at least five minutes into the future, leveraging their superior pattern recognition and learning capabilities. Finally, transformers reduce the amount of data and time needed to train a predictive model, while improving performance.

[0139] As noted above, the second trained machine learning model may include a temporal fusion transformer model, for instance, as described in an article published by University of Oxford and Google Cloud Al in 2021. See Bryan Lim, et al., Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting, 31 International Journal of Forecasting 1748 (2021), https: / / doi.Org / 10.1016 / j.ijforecast.2021.03.012, which is incorporated herein by reference in its entirety. Such a temporal fusion transformer model may use gating mechanisms, variable selection networks, static covariate encoders, and temporal processing components to enable interpretable multi-horizon forecasting. Multi-horizon forecasting refers to the process of predicting multiple future values of a time series across different future time steps or horizons. Unlike single-horizon forecasting, which predicts a single future point, multi-horizon forecasting aims to forecast a sequence of future values, each corresponding to a different time horizon. Accordingly, as described above, the second machine learning model described herein may be configured to predict a sequence of physiological parameters at multiple future time points (the first machine learning model may also be configured to predict a sequence of probabilities of adverse events at multiple time points). For example, the second machine learning model may predict values for five minutes into the future at each future minute, wherein each future minute forms a horizon of the multi-horizon forecast.

[0140] The temporal fusion transformer model may be configured to receive static data, time- varying past inputs, and / or time- varying a priori known future inputs. The gatingmechanisms may be configured to control the flow and processing of information based on the received data within the model, for instance by determining the extent of non-linear processing required for various inputs using a Gated Residual Network (GRN) and suppressing parts of the architecture not needed for a given dataset using component gating layers based on Gated Linear Units (GLUs). The variable selection networks may enable the temporal fusion transformer model both to provide insight into which variables in the data contribute most to the model’s prediction and to remove noise from the data that could negatively impact performance. Variable selection may significantly improve model performance by allocating learning capacity to the most salient variables. All static (e.g., pre-procedure cardiac status and / or other static data described above), past (e.g., dynamic time series data acquired during the surgical procedure), and future inputs may utilize separate variable selection networks. The static covariate encoders may be configured to generate different context vectors based on static input data. The context vectors may include contexts for temporal variable selection, local processing of temporal features, and enriching of temporal features with static information. The context vectors may be incorporated at various different positions in the temporal fusion decoder, which is described in more detail below.

[0141] The temporal fusion decoder may be configured to implement a series of processing layers to learn temporal relationships present in the dataset. A sequence-to-sequence layer captures local context (e.g., anomalies, cyclical patterns, etc.) by feeding processed temporal features into the encoder and decoder, generating uniform temporal features as inputs into the temporal fusion decoder itself. Context vectors obtained using static covariate encoders may be used to enable static data to influence local processing. A static enrichment layer may enrich temporal features with static data. A temporal self-attention layer may apply self-attention to static-data-enriched temporal features. This layer employs interpretable multi-head attention to capture long-range dependencies and preserves causal information flow by applying interpretable multi-head attention to static enriched features grouped into a single matrix at each forecast time. Finally, a position- wise feed-forward layer applies additional non-linear processing to the outputs of the self-attention layer using GRNs.

[0142] The temporal fusion decoder may be configured to generate prediction intervals on top of point forecasts by simultaneously predicting various percentiles (e.g., 10th, 50th and 90th) at each time step. Quantile forecasts may be generated by applying linear transformation to theoutput of the temporal fusion decoder. Finally, the temporal fusion transformer may be trained by minimizing a loss function. For instance, the temporal fusion transformer may be trained by jointly minimizing the quantile loss summed across all quantile outputs.

[0143] In some embodiments, the second machine learning (ML) model is selected from a set of ML models configured (e.g., trained) to predict future values of one or more physiological parameters of a patient (e.g., up to five minutes into the future). As described in further detail below, experimental data indicate that different ML models can perform better (e.g., provide more accurate, sensitive, specific, and / or precise predictions of future values of a patient’s physiological parameters) for different subsets of patients. For example, different ML models can perform better for subsets of patients having similar characteristics (e.g., characteristics represented by the patients’ static data). In some examples, the second ML model is selected from the set of ML models based on characteristics of the patient. For example, the second ML model can be selected based on (1) the patient’s static data, (2) a subset of patients to which the patient is assigned (or with which the patient is associated) based on the patient’s static data, and / or (3) the performance of the second ML model on a validation dataset drawn from patients having static data similar to the static data of the patient or similar to the static data of the subset of patients to which the patient is assigned. Similarity of the patient’s static data to the static data of any other patient or subset of patients may be assessed using any suitable criteria or technique. In some examples, the use of different ML models for different subsets of patients may be described as “ensemble modeling.”

[0144] At block 214, the system may cause display of a visualization including at least one of the one or more predicted values of the at least one physiological parameter. In some examples, the visualization further includes an indication of the predicted adverse event and / or an indication of the probability of the at least one adverse event. The visualization may include a real-time feed depicting both received time-series physiological data and predicted physiological data and / or adverse events. An exemplary visualization is provided in FIG. 3 and described in more detail below. The system may, for instance, cause a visualization to be displayed on an interface in an operating room for utilization by physicians or other medical personnel during a surgical procedure. The visualization may enable physicians / medical personnel to continuously monitor the predicted future physiological parameters of a patient and to be alerted of predicted adverse events during the procedure. For instance, the system may cause display of an alertindicating a predicted adverse event (e.g., heart rate deceleration), which may alert a physician that an intervention should be taken to counter / pr event the predicted adverse event. In some embodiments, the system may predict and cause display of a probability of the adverse event (e.g., 80% probability of heart rate deceleration adverse event), which may enable the physician to more effectively determine the necessity of intervening. In some embodiments, the system may cause display of a physiological parameter value in addition to or in place of the predicted adverse event, which may alert the physician to various trends in physiological parameters, magnitudes of predicted drops or increases in those parameters, and so on, which may enable the physician to tailor any treatments / pharmacological interventions based on the predicted values.

[0145] FIG. 3 illustrates an exemplary visualization 300 that may be generated and displayed according to some embodiments. The visualization may be a real-time feed illustrating time-series data received from one or more sensors monitoring a patient’s physiological parameters during a surgical procedure, including, for instance, monitored and predicted values for a patient’s heart rate, oxygen saturation, end tidal CO2, noninvasive mean blood pressure, hydration, temperature, other disease specific factors, and / or any other suitable physiological parameters, including but not limited to physiological parameters described throughout. The visualization may include a heart rate visualization portion 302 that includes a monitored heart rate indicator 310 illustrating time-series heart rate data collected by the sensors during a surgical procedure and a predicted heart rate indicator 312 illustrating predicted future heart rate values predicted, for instance, by the second trained machine learning model, for a predefined duration following the end of the time-series heart rate data received. The visualization may include an oxygen saturation visualization portion 304 that includes a monitored oxygen saturation indicator 314 illustrating time-series oxygen saturation data collected by the sensors during a surgical procedure and a predicted oxygen saturation indicator 316 illustrating predicted future oxygen saturation values predicted, for instance, by the second trained machine learning model, for a predefined duration following the end of the time-series oxygen saturation data received. The visualization may include an end tidal CO2 visualization portion 306 that includes a monitored end tidal CO2 indicator 318 illustrating time-series end tidal CO2 data collected by the sensors during a surgical procedure and a predicted end tidal CO2 indicator 320 illustrating predicted future end tidal CO2 values predicted, for instance, by the second trained machine learning model, for a predefined duration following the end of the time-series end tidal CO2 data received.As shown the indicators 312, 316, and 320 provide predicted values for a five-minute window 308 following the end of a ten-minute segment of collected data. However, it should be understood that the system may be configured to predict a larger or smaller window (e.g., more or less than five minutes) based on the characteristics of the collected data. As indicated, the predicted heart rate during the future five-minute window is 128.00 BPM, the predicted oxygen saturation is 87.00%, the predicted end tidal CO2 is 43.00, and the predicted non-invasive mean blood pressure is 48.00 mm Hg. It should be understood that the system may predict discrete values at each time point during the future window (e.g., a discrete heart rate prediction may be predicted for each millisecond, second, minute, etc. during the future window, rather than a single value across the window as shown in the figure). Additionally, it should be understood that the predictions may be generated by the system based on more or less observation data than the ten-minute window illustrated in FIG. 3.

[0146] As described above, the one or more machine learning models described herein may be trained to predict adverse events and / or future values of physiological parameters. Both a first machine learning model (e.g., trained for a classification task) and a second machine learning model (e.g., trained for a regression task) may be trained based on observation data collected from subjects undergoing surgical procedures and / or synthetic data created based on the real data collected from subjects undergoing surgical procedures. An exemplary implementation of the systems and methods described herein was developed using observation (i.e., “real”) training data that included data obtained from an anesthesia record database containing minute-minute hemodynamic information during the anesthetic management of pediatric patients undergoing catheterization (155,568 data points from 1250 patients). The training data included preprocedural features, hemodynamic variables, medication administration, and adverse events. The data was segmented into training (70% of patients), validation (15% of patients) and testing (15% of patients) data sets. Adverse events in the training data were defined as hemodynamic changes leading to pharmacologic or other clinical interventions. Synthetic training data was also created by projecting 10 minutes of observation data to the 15th minute with added noise. FIGS. 4A-9B illustrate model performance of the classifier model (e.g., first machine learning model) based on observation training data and validation data and FIG. 6 illustrates model performance of the regression model (e.g., second machine learning model) based on synthetic training data.

[0147] FIGS. 4A and 4B illustrate the performance of an exemplary Dense Neural Network algorithm (e.g., an exemplary embodiment of the first machine learning model described herein) continuously predicting adverse events at the 15th minute based on 10 minutes of heart rate timeseries data, 10 minutes of blood pressure time-series data, 10 minutes of oxygen saturation time-series data, and 10 minutes of End Tidal CO2 time-series data, along with static data including balloon inflation and angiogram event information, and pre-procedure cardiac status. The accuracy of the model based on the training dataset measured by area under the ROC curve (AUC), shown in FIG. 4A, was 0.97, and the accuracy of the model based on the validation dataset measured by AUC, shown in FIG. 4B was 0.91. Changes in heart rate may be the leading hemodynamic variable predicting adverse events in real-time models. FIGS. 5A and 5B illustrate the performance of the same Dense Neural Network algorithm when all variables are removed except heart rate. As shown in FIG. 5A, the accuracy of the model based on the training dataset measured by AUC was 0.93, and as shown in FIG. 5B, the accuracy of the model based on the validation dataset measured by AUC was 0.94. As illustrated by the Dense Neural Network model performance in FIGS. 4A-5B, the systems and methods described herein are capable of accurately identifying hemodynamic patterns of change leading to an adverse event during cardiac catheterization. As described above, the models for adverse event prediction described herein can augment clinician perspective on evolving adverse events with clinically significant lead-time in which care teams can take actions to mitigate adverse event occurrence.

[0148] FIG. 6 illustrates model performance of the regression model (e.g., an exemplary embodiment of the second machine learning model described herein) based on synthetic training data. Specifically, FIG. 6 illustrates performance of a model configured to predict heart rate five minutes into the future based on 10 minutes of time-series heart rate data. The exemplary model was implemented with PyTorch’s transformer module, and the best model was trained in 32 epochs with a Root Mean Squared Error of approximately 10 heart beats.

[0149] As discussed throughout, different steps leading to the adverse event prediction and physiological parameter prediction objectives described herein may correspond to different events of a surgical procedure. FIG. 7 illustrates an exemplary timeline 700 of such events. The timeline 700 includes a pre-op step 702 and in lab step 704 in which static data may be collected. The static data may be input (e.g., by a user of the system and / or automatically by a computing system communicatively coupled to one or more sensors configured to collect various staticdata) directly to one or both of the first and second machine learning models for adverse event prediction and / or physiological parameter prediction, and / or may be used to determine a preprocedure cardiac status (e.g., a score of 1-5) that is input (e.g., by a user and / or automatically by a computing system) to one or both of the models. The timeline 700 includes an intubation step 706, which may be performed as part of a surgical procedure. Information about the intubation step (e.g., an indicator of its occurrence, a time stamp, etc.) may be input to one or both of the models. Timeline 700 includes a baseline calculation step 716 in which a physiological baseline of the patient is determined, for instance, by determining (e.g., using computing system 100 described above) an average value of various physiological parameters (e.g., heart rate, blood pressure) measured using one or more sensors prior to one or more surgical events. The physiological baseline may be determined based on a predefined duration (e.g., 10 minutes) of time-series physiological data obtained for the patient, such as heart rate data, noninvasive mean blood pressure data, etc., and may include an average heart rate, average noninvasive mean blood pressure, etc. as described above. Baseline calculation 716 may be performed at a predefined time prior to a catheter insertion step 710 (labeled “start access” in FIG. 7), such as 10 minutes before insertion. However, as described above, the physiological baseline may be determined at a variety of alternative times (e.g., during in- lab, pre-op, prior to intubation, etc.) or may be determined using a standardized baseline quantification method based on one or more characteristics of the patient (e.g., age, weight, biological sex, pre-existing conditions, etc.). As noted, step 710 corresponds to a catheter insertion step following determination of the physiological baseline. After catheter insertion, time-series physiological data may be collected for the patient and predictions of adverse events and future values of the physiological parameters may be predicted during the surgical procedure using the first and second machine learning models described herein based on the physiological data observed during the procedure and / or the static data received prior to catheter insertion. Step 712 marks the catheter removal step. The system may or may not continue to process physiological data and predict adverse events and future values of the physiological parameters after removal of the catheter at step 712 until the end of the procedure at step 714.

[0150] FIG. 8 depicts an exemplary computing device 800 for predicting adverse events and / or future values of physiological parameters (e.g., during a surgical procedure), in accordance with one or more examples of the disclosure, which can be computing system 100 ofFIG. 1 or a component thereof. Device 800 can be a host computer connected to a network. Device 800 can be a client computer or a server. As shown in FIG. 8, device 800 can be any suitable type of microprocessor-based device, such as a personal computer, laptop, workstation, server, or handheld computing device (portable electronic device) such as a phone or tablet. The device can include, for example, one or more of processors 802, input device 806, output device 808, storage 810, and communication device 804. Input device 806 and output device 808 can generally correspond to those described above and can either be connectable or integrated with the computer.

[0151] Input device 806 can be any suitable device that provides input, such as a touch screen, keyboard or keypad, mouse, or voice-recognition device. Output device 808 can be any suitable device that provides output, such as a touch screen, haptics device, or speaker.

[0152] Storage 810 can be any suitable device that provides storage, such as an electrical, magnetic, or optical memory, including a RAM, cache, hard drive, or removable storage disk. Communication device 804 can include any suitable device capable of transmitting and receiving signals over a network, such as a network interface chip or device. The components of the computer can be connected in any suitable manner, such as via a physical bus or wirelessly.

[0153] Software 812, which can be stored in storage 810 and executed by processor 802, can include, for example, source code, object code, and / or executable code that embodies the functionality of the present disclosure (e.g., as embodied in the devices as described above). For example, software 812 can include software for performing one or more steps of method 200 of FIG. 2.

[0154] Software 812 can also be stored and / or transported within any non-transitory computer-readable storage medium for use by or in connection with an instruction execution system, apparatus, or device, such as those described above, that can fetch instructions associated with the software from the instruction execution system, apparatus, or device and execute the instructions. In the context of this disclosure, a computer-readable storage medium can be any medium, such as storage 810, that can contain or store programming for use by or in connection with an instruction execution system, apparatus, or device.

[0155] Software 812 can also be propagated within any transport medium for use by or in connection with an instruction execution system, apparatus, or device, such as those described above, that can fetch instructions associated with the software from the instruction executionsystem, apparatus, or device and execute the instructions. In the context of this disclosure, a transport medium can be any medium that can communicate, propagate, or transport programming for use by or in connection with an instruction execution system, apparatus, or device. The transport readable medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, or infrared wired or wireless propagation medium.

[0156] Device 800 may be connected to a network, which can be any suitable type of interconnected communication system. The network can implement any suitable communications protocol and can be secured by any suitable security protocol. The network can comprise network links of any suitable arrangement that can implement the transmission and reception of network signals, such as wireless network connections, T1 or T3 lines, cable networks, DSL, or telephone lines.

[0157] Device 800 can implement any operating system suitable for operating on the network. Software 812 can be written in any suitable programming language, such as C, C++, Java, or Python. In various embodiments, application software embodying the functionality of the present disclosure can be deployed in different configurations, such as in a client / server arrangement or through a Web browser as a Web-based application or Web service, for example.Experimental Results

[0158] FIG. 9 illustrates deep neural network classification model performance results for different classification model configurations. As shown, three different models were evaluated. Each of the three models was trained to predict adverse events five minutes into the future. The first model (left column) labeled “Baseline Model + Prev AE” was configured to predict adverse events based on a subset of features (baseline features) including static feature(s) such as Precardiac Status Score (PCS) (for instance, the pre-procedure cardiac status described throughout), Hemodynamic Features including Heart Rate, Non-Invasive Blood Pressure Mean (NBPm), Oxygen Saturation. End Tidal CO2, Hemodynamic Perturbation Events, and Balloon Dilation, along with other critical events such as Coronary Angiography. The first model was further configured to consider whether a previous adverse event had occurred in predicting future adverse events (the “Prev AE” feature). The Prev AE feature was a binary flag that activated the moment that an AE occurred (e.g., 0 at the beginning of the procedure and turns to 1 the moment an AE occurs). It should be understood that other configurations for the Prev AE feature arewithin the scope of this disclosure (e.g., the Prev AE feature may indicate a number of adverse events, types of adverse events, and so on).

[0159] The second model was configured to consider the baseline features and Prev AE feature described above and was additionally configured to consider the administration of any of a plurality of drugs, including Dopamine, Epinephrine, Calcium Gluconate, Sodium Bicarbonate, Atropine, Ephedrine, Phenylephrine, Fentanyl, Hydromorphone, Morphine, Propofol, End Tidal Desflurane, End Tidal Isoflurane, End Tidal Nitrous Oxide, Inspired Sevoflurane. Similar to the Prev AE feature, each of the aforementioned drug features were configured to be assigned a binary value of either 0 or 1. That is, if the drug was not administered, it’s value was assigned to 0, and if it was administered, its value was assigned to 1. As with the Prev AE feature, alternative drug feature configurations are possible. For instance, the model may be configured to consider a dosage amount of a given drug type administered, an administration method (e.g., IV, oral, etc.), and so on. In one example, the second model attained AUC of 0.92, sensitivity of 0.82, specificity of 0.92, precision of 0.21, and Fl score of 0.33. The third model was configured to consider the Prev AE feature and drug features described above but was not configured to consider all of the baseline features. Instead, the third model was configured to consider a single baseline feature: heart rate. In one example (e.g., the third model having been trained with a first set of training data), the third model attained AUC of 0.94, sensitivity of 0.87, specificity of 0.92, precision of 0.21, and Fl score of 0.34. In another example (e.g., the third model having been trained a second set of training data), the third model attained AUC of 0.95, sensitivity of 0.90, and specificity of 0.92, with precision and Fl scores comparable to the prior example of the third model. The precision of the second and third models was lower than expected (and lower than the precision of the first model) due to class imbalances in the model’s defined adverse events (AE) outcome variable.

[0160] FIG. 10A illustrates an AUC-ROC curve indicative of a deep neural network adverse event classification model performance on training data according to some embodiments. The model of FIG. 10A was configured to consider only heart rate, Prev AE, and drug features similar to the third model of FIG. 9. FIG. 10B illustrates an AUC-ROC curve indicative of a deep neural network adverse event classification model performance on validation data according to some embodiments. The model of FIG. 10B was the same as the model of FIG.10A. As shown, the AUC for the training data was 0.99 and the AUC for the validation data was 0.94.

[0161] FIG. 11 A illustrates regression performance of a baseline temporal fusion transformer model, an updated temporal fusion transformer model, and a tuned temporal fusion transformer model each configured to predict physiological parameter values (e.g., to continuously predict physiological parameter values 5 minutes into the future) according to some embodiments. As shown, each of the models was evaluated according to three different metrics: symmetric mean absolute percentage error (SMAPE), root mean squared error (RMSE), and mean absolute error (MAE).

[0162] The baseline model of FIG. 11 A was trained with 1 attention head, a hidden size of 32, and a subset of features. The baseline model was trained with 1 attention head, a relative time index, an angiogram flag, a balloon inflation flag, a drug binary indicator, a category indicator (e.g., a pre-procedure cardiac status), an oxygen saturation, an end tidal CO2, a non-invasive mean blood pressure, a scaled heart rate change from baseline, a scaled oxygen saturation change from baseline, a scaled end tidal CO2 change from baseline, and a scaled non-invasive mean blood pressure change from baseline. The updated model included more features (e.g., heart rate, age, raw measurements of changes in heart rate, oxygen saturation, end tidal CO2, and noninvasive mean blood pressure from baseline, baseline measurements for each of heart rate, oxygen saturation, end tidal CO2, and noninvasive mean blood pressure) relative to the baseline model, along with updated hyperparameters, including: a hidden size increased from 32 to 64, attention heads 1 increased to 2, and max epochs increased from 15 to 30.

[0163] Finally, the tuned model was trained on all of the features from the baseline model and updated model. Hyperparameter tuning was performed to search for optimal hyperparameter values for gradient clip, hidden size, hidden continuous size, attention head size, dropout range, and learning rate. As shown, the tuned model demonstrated the least prediction error according to each of the three metrics (SMAPE, RMSE, and MAE). The error levels of the updated and tuned models were deemed actionable by a group of anesthesiologists.

[0164] Additionally, another example of a model trained to predict future values of a physiological parameter (“parameter forecasting model”) showed improved performance on a subset of patients. In general, the performance of the above-described parameter forecasting models varied based on patient characteristics, such that different parameter forecasting modelsperformed better on different subsets of patients. Thus, the parameter forecasting models are suitable for personalized healthcare applications, in which the most suitable model for a patient is selected based on a patient’s characteristics (e.g., characteristics known prior to the patient’s surgery).

[0165] FIG. 1 IB illustrates regression performance of a second baseline temporal fusion transformer model, a second updated temporal fusion transformer model, and a second tuned temporal fusion transformer model each configured to predict physiological parameter values (e.g., to continuously predict physiological parameter values 5 minutes into the future) according to some embodiments. The second baseline transformer model is trained to predict heart rate five minutes ahead using heart rate values from the previous minute. The second updated transformer model is similar to the first updated transformer model, but configured with different parameter values (post tuning). The second tuned transformer model is similar to the first tuned transformer model, but configured with different parameter values (post tuning). Each of the second transformer models was evaluated according to three different metrics: symmetric mean absolute percentage error (SMAPE), root mean squared error (RMSE), and mean absolute error (MAE). As FIG. 1 IB shows, the second updated and tuned models perform significantly better than the second baseline model. The second updated model performs comparably to the second tuned model, although the second tuned model performs slightly better. Furthermore, as described below, the updated and tuned models focus their attention on different aspects of the input data.

[0166] FIG. 12 illustrates a chart of heart rate predictions predicted by the baseline model of FIG. 11 A compared against ground truth values. As discussed below, the updated model and the tuned model each demonstrated improved performance over the baseline model. FIG. 13A illustrates heart rate predictions by the updated model of FIG. 11 A compared against ground truth values. FIG. 13B illustrates heart rate predictions by the tuned model of FIG. 11A compared against the ground truth values of FIG. 13 A. As show, both the updated and the tuned model track ground truth heart rate well, but the tuned model appeared more inclined to predict flat, unchanging heart rates.

[0167] FIG. 14A illustrates heart rate predictions by the updated model of FIG. 11 A compared against additional ground truth values. FIG. 14B illustrates heart rate predictions by the tuned model of FIG. 11A compared against the ground truth values of FIG. 14A. As shown, the tuned model of FIG. 14B slightly underperformed the updated model shown in FIG. 14A andpredicts a flat heart rate around 70 bpm in several instances. FIG. 15A illustrates heart rate predictions by the updated model of FIG. 11 A compared against additional ground truth values. FIG. 15B illustrates heart rate predictions by the tuned model of FIG. 11 A compared against the ground truth values of FIG. 15A according to some embodiments. As can be seen in FIGS. 15A and 15B, the updated and tuned models both generally track the reported true heart rate closely, but sometimes do not track sharp, brief peaks or dips in the reported true heart rate. Such peaks or dips in the reported true heart rate are, in some cases, erroneous values arising from sensing anomalies. After such peaks or dips, the updated and tuned models both quickly return to accurate tracking. FIG. 16A illustrates heart rate predictions by the updated model of FIG. 11A compared against additional ground truth values according to some embodiments. FIG. 16B illustrates heart rate predictions by the tuned model of FIG. 11 A compared against the ground truth values of FIG. 16A according to some embodiments.

[0168] FIG. 17A illustrates model attention by the updated model of FIG. 11 A to data at different times in a time series of input data according to some embodiments. As shown, the updated model of FIG. 11A paid significantly more attention to the latest (e.g., most recent) 1-4 minutes of the time-series input data (having time indexes -1 to -4) than to the earliest (e.g., least recent) 1-5 minutes (having time indexes -6 to -10). The 1st minute (e.g., the least recent minute of the 10-minute period, having time index -10) had slightly more attention than the following 2- 4 minutes (having time indexes -9 to -6). FIG. 17B illustrates model attention by the tuned model of FIG. 11 A to data at different times in a time series of input data according to some embodiments. The tuned model of FIG. 11 A paid significantly more attention to the earliest (e.g., least recent) 1-3 minutes of the 10 minute sequence of time-series data (having time indexes -10 to -8) than to the latest (e.g., most recent) 1-3 minutes (having time indexes -1 to -3). The middle minutes (e.g., the 4 minute period roughly equidistant from the beginning and end of the 10-minute period, having time indexes -4 to -7) had slightly more attention than the latest (e.g., most recent) 1-3 minutes. The 10th minute (e.g., the most recent minute of the 10-minute period, having time index -1) was more important than the preceding 1-2 minutes (having time indexes -2 to -3). Identifying the portions of the input data to which various models pay the most attention enhances the explainability of the models, thereby facilitating adoption of the models within the healthcare industry.

[0169] FIG. 18A illustrates static variable importance to the updated model of FIG. 11 A. As shown, the heart rate baseline was the most important feature, although each feature contributed significantly to model prediction. FIG. 18B illustrates static variable importance to the tuned model of FIG. 11 A. As shown, end tidal CO2 was the most important static feature for the tuned model, followed by heart rate baseline.

[0170] FIG. 19A illustrates dynamic variable importance to the updated model of FIG. 11 A according to some embodiments. The analysis of time variant (dynamic) features depicted in FIGS. 19A indicated that the previous heart rate was the most important for the updated model’s prediction. However, this does not mean that the other features do not contribute. In fact, together the other features describe the variance and nuance in the trends that the past heart rate alone does not. FIG. 19B illustrates dynamic variable importance to the tuned model of FIG. 11 A according to some embodiments. For the tuned model, the previous heart rate was by far the most important to the model’s prediction.

[0171] FIG. 20 illustrates heart rate predictions by two models compared against the ground truth values according to some embodiments. In FIG. 20, heart rate graph 2010 shows heart rate predictions by a first model (e.g., tuned model) compared against ground truth data for a set of patients, and heart rate graph 2020 shows heart rate predictions by a second model (e.g., updated model) compared against ground truth data for the same set of patients. More specifically, heart rate graphs 2010 and 2020 show heart rate predictions by the two models (e.g., five minutes into the future) and ground truth data scaled to the patient’s baseline average heart rate. For example, a predicted value of 110 in heart rate graph 2010 or 2020 indicates that the respective model predicts that the patient’s heart rate will be 110% of the patient’s baseline average heart rate at a future time (e.g., five minutes into the future). In addition, the x-axes of heart rate graphs 2010 and 2020 indicate the heart rate baselines of the patients, normalized on a scale from two standard deviations less than the mean baseline heart rate for the set of patients to two standard deviations greater than the mean baseline heart rate for the set of patients. As FIG. 20 shows, the second model is significantly more accurate than the first model for a subset of patients having normalized baseline heart rates roughly 0.25 to 0.40 standard deviations less than the mean baseline heart rate for the set of patients. In some examples, the second model may be selected to predict future values of the heart rate for the subset of patients having normalized baseline heart rates between 0.25 and 0.40 standard deviations less than the mean baseline heart rate for the setof patients, and the first model may be selected to predict future values of the heart rate for other patients. Thus, FIG. 20 illustrates in scenario in which ensemble modeling techniques can be used to improve model accuracy for one or more subsets of patients.

[0172] Some examples have been described in which heart rate is a feature used by a machine- learning model to predict future values of a patient’s physiological parameters and / or to predict the likelihood of an adverse event occurring. In some embodiments, the heart rate feature is a continuous feature with / a numeric value representing the heart rate (e.g., 60 bpm). In some embodiments, the heart rate feature is a categorical feature, with a value indicating a categorical characterization of the heart rate, such as a range which includes the heart rate (e.g., 55-59 bpm, 60-64 bpm, 65-69 bpm, etc.). Any suitable heart rate categories can be used.

[0173] Reference is made herein to “time-series data,” “time-series physiological data,” “time-series hemodynamic data,” etc. One of ordinary skill in the art will appreciate that the techniques described herein are applicable to time-series data (e.g., timestamped values, values sampled at regular or irregular time intervals, etc.) and / or to temporally varying data (e.g., any values that can change over time). All references to “time-series” data are also applicable to “temporally-varying” or “time-varying” data unless the context specifically indicates otherwise.

[0174] The term “approximately”, the phrase “approximately equal to”, and other similar phrases, as used in the specification and the claims (e.g., “X has a value of approximately Y” or “X is approximately equal to Y”), should be understood to mean that one value (X) is within a predetermined range of another value (Y). The predetermined range may be plus or minus 20%, 10%, 5%, 3%, 1%, 0.1%, or less than 0.1%, unless otherwise indicated.

[0175] Although the disclosure and examples have been fully described with reference to the accompanying figures, it is to be noted that various changes and modifications will become apparent to those skilled in the art. Such changes and modifications are to be understood as being included within the scope of the disclosure and examples as defined by the claims. Finally, the entire disclosure of the patents and publications referred to in this application are hereby incorporated herein by reference.

Claims

CLAIMSWhat is claimed is:

1. A medical monitoring system comprising: one or more processors; and one or more computer-readable media storing instructions that, when executed by the one or more processors, cause the system to perform operations including: obtaining physiological data of a patient, the physiological data indicating one or more sensed values of at least one physiological parameter of the patient, the one or more sensed values including a current value of the physiological parameter and / or one or more past values of the physiological parameter; generating, using at least one machine learning model, one or more future values of the physiological parameter, the one or more future values of the physiological parameter being predicted values of the physiological parameter at one or more future times, the generating being based on the physiological data and a physiological baseline of the patient; and providing visualization data to a display device configured to display a visualization of the physiological parameter, the visualization indicating at least a portion of the one or more sensed values of the physiological parameter and at least a portion of the one or more future values of the physiological parameter.

2. The medical monitoring system of claim 1, further including the display device.

3. The medical monitoring system of claim 1, wherein the one or more sensed values of the physiological parameter are obtained via one or more sensors.

4. The medical monitoring system of claim 3, further comprising the one or more sensors.

5. The medical monitoring system of claim 4, wherein the one or more sensors include a blood pressure monitor, an oxygen saturation monitor, a heart rate monitor, an end tidal CO2 monitor, a temperature monitor, and / or a hydration monitor.

6. The medical monitoring system of claim 1, wherein the visualization includes a timevarying graph of the portion of the one or more sensed values and the portion of the one or more future values of the physiological parameter.

7. The medical monitoring system of claim 1, wherein the visualization simultaneously indicates the portion of the one or more sensed values and the portion of the one or more future values of the physiological parameter.

8. The medical monitoring system of claim 1, wherein the machine learning model is a first machine learning model, and wherein the operations further include determining, by a second machine learning model, a probability of at least one adverse event occurring at one or more future times, wherein the determined probability is based on the physiological data and the physiological baseline of the patient.

9. The medical monitoring system of claim 8, wherein the operations further include: determining that the determined probability exceeds a threshold probability; and alerting a user to a potential occurrence of the adverse event.

10. The medical monitoring system of claim 8, wherein the operations further include obtaining the physiological baseline of a patient.

11. The medical monitoring system of claim 10, wherein the physiological data are second physiological data, and wherein obtaining the physiological baseline of the patient comprises: obtaining first physiological data of the patient prior to obtaining the second physiological data of the patient, wherein the first physiological data comprise one or more values of one or more physiological parameters of the patient; and determining the physiological baseline of the patient based on the first physiological data.

12. The medical monitoring system of claim 1, wherein the predicted values of the physiological parameter at one or more future times include a predicted value of the physiological parameter at a time approximately five minutes subsequent to a current time.

13. The medical monitoring system of claim 12, wherein the machine learning model includes a generative Al model.

14. The medical monitoring system of claim 13, wherein the generative Al model includes a transformer.

15. The medical monitoring system of claim 1, wherein the operations further include selecting the machine learning model from a plurality of machine learning models configured to generate future values of the physiological parameter, wherein the selecting the machine learning model is based on one or more characteristics of the patient.

16. A medical monitoring method comprising: obtaining physiological data of a patient, the physiological data indicating one or more sensed values of at least one physiological parameter of the patient, the one or more sensed values including a current value of the physiological parameter and / or one or more past values of the physiological parameter; generating, using at least one machine learning model, one or more future values of the physiological parameter, the one or more future values of the physiological parameter being predicted values of the physiological parameter at one or more future times, the generating being based on the physiological data and a physiological baseline of the patient; and providing visualization data to a display device configured to display a visualization of the physiological parameter, the visualization indicating at least a portion of the one or more sensed values of the physiological parameter and at least a portion of the one or more future values of the physiological parameter.

17. A non-transitory computer-readable storage medium storing computer-executable instructions that, when executed by one or more processors, cause the processors to perform medical monitoring operations including: obtaining physiological data of a patient, the physiological data indicating one or more sensed values of at least one physiological parameter of the patient, the one or more sensedvalues including a current value of the physiological parameter and / or one or more past values of the physiological parameter; generating, using at least one machine learning model, one or more future values of the physiological parameter, the one or more future values of the physiological parameter being predicted values of the physiological parameter at one or more future times, the generating being based on the physiological data and a physiological baseline of the patient; and providing visualization data to a display device configured to display a visualization of the physiological parameter, the visualization indicating at least a portion of the one or more sensed values of the physiological parameter and at least a portion of the one or more future values of the physiological parameter.

18. A method for determining a probability of an adverse event for a patient, the method comprising, at a computing system: receiving first physiological data of the patient, wherein the first physiological data comprises values of a plurality of physiological parameters of the patient; determining a physiological baseline of the patient based on the first physiological data; receiving second physiological data of the patient, the second physiological data comprising data obtained after the first physiological data; determining a probability of at least one adverse event occurring at one or more future times using a first trained machine learning model, wherein the determined probability is based on the second physiological data and the physiological baseline of the patient; and predicting one or more values of at least one physiological parameter at the one or more future times using a second trained machine learning model based on the second physiological data and the physiological baseline of the patient.

19. The method of claim 18, wherein the second physiological data are obtained during a surgical procedure, wherein the one or more future times are during the surgical procedure, and wherein the surgical procedure is a pediatric cardiac catheterization procedure.

20. The method of claim 18, wherein the first and second physiological data are obtained from a patient being monitored outside of a surgical procedure, and wherein the physiologicalbaseline is determined prior to at least one of insertion of a catheter into the patient, intubation of the patient, and administration of an anesthetic to the patient.

21. The method of claim 18, wherein the physiological baseline comprises a heart rate baseline, a mean blood pressure baseline, an oxygen saturation baseline, an end tidal CO2 baseline, or any combination thereof, and wherein the second physiological data of the patient comprises one or more sets of time-series hemodynamic data.

22. The method of claim 21, wherein the one or more sets of time-series hemodynamic data comprise at least one of: a time series of heart rate measurements, a time series of mean blood pressure measurements, a time series of oxygen saturation measurements, and a time series of end tidal CO2 measurements obtained from the patient.

23. The method of claim 21, wherein the one or more sets of time-series hemodynamic data comprise a plurality of overlapping time segments of time-series hemodynamic data, wherein each of the overlapping time segments respectively include a predefined duration of hemodynamic data.

24. The method of claim 18, wherein the at least one adverse event comprises at least one of: a heart rate dropping below a first threshold, a blood pressure dropping below a second threshold, and an oxygen saturation dropping below a third threshold, and wherein the first threshold comprises a predefined deviation from a heart rate baseline of the physiological baseline, the second threshold comprises a predefined deviation from a mean blood pressure baseline of the physiological baseline, and the third threshold comprises a predefined deviation from an oxygen saturation baseline of the physiological baseline.

25. The method of claim 18, comprising: determining a pre-procedure cardiac status, wherein the probability of the occurrence of the at least one adverse event is determined based on the preprocedure cardiac status, wherein the pre-procedure cardiac status is determined based on information associated with a surgical procedure and a health status of the patient, wherein the probability of the at least one adverse event occurring at one or more future times during thesurgical procedure is determined by the first trained machine learning model based on the preprocedure cardiac status, and wherein the predicted one or more values of at least one physiological parameter are predicted based on the pre-procedure cardiac status.

26. The method of claim 18, wherein the at least one adverse event is predicted by the first trained machine learning model based on a physician-initiated event, wherein the physician- initiated event comprises any of an administration of an anesthetic, a balloon inflation event, or a coronary angiography event, and wherein the at least one physiological parameter comprises at least one of a heart rate, an oxygen saturation, an end-tidal carbon dioxide, a mean blood pressure, a temperature, or patient levels of hydration.

Images