Sensing systems for measuring and determining physiological parameters of a patient, and devices and methods thereof

Abstract

Devices, systems, and methods herein relate to determining physiological information. In some embodiments, a method includes receiving, from at least one of a sensing device or a database, at least one signal associated with a physiological variable of a patient. The method includes generating an output based on the at least one signal, the output representing a prediction of a physiological variable. In some embodiments, an image model and a language model can be used to generate a prediction of a physiological variable.

Description

Attorney Docket No.: CRDS-008 / 01WO 348698-2034SENSING SYSTEMS FOR MEASURING AND DETERMINING PHYSIOLOGICAL PARAMETERS OF A PATIENT, AND DEVICES AND METHODS THEREOFCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority to and the benefit of U. S. Provisional Patent Application No. 63 / 727,638, filed December 3, 2024, titled “ SYSTEMS, DEVICES, AND METHODS FOR ANALYZING AND PRESEN TING PHYSIOLOGICAL INFORMATION,” U. S. Provisional Patent Application No. 63 / 727,647, filed December 3, 2024, titled “SYSTEMS, DEVICES, AND METHODS FOR ANALYZING AND PRESENTING PHYSIOLOGICAL INFORMATION,” and U. S. Provisional Patent Application No. 63 / 831,073, filed June 26, 2025, titled “ SYSTEMS, DEVICES, AND METHODS FOR ANALYZING AND PRESENTING PHYSIOLOGICAL INFORMATION,” the disclosures of each of which are incorporated herein by reference.TECHNICAL FIELD

[0002] Devices, systems, and methods disclosed herein relate to analyzing and presenting physiological information, including physiological information based on signals measured using one or more wearable devices on a patient’s body.BACKGROUND

[0003] Wearable devices may provide a convenient way to capture physiological signals or characteristics of a user. Signals from sensors used to measure physiological characteristics, such as photoplethysmography (PPG) sensors, electrocardiogram (ECG) sensors, and / or seismocardiogram (SCG) sensors, can often include noise associated with the sensor (e.g., high frequency noise, noise from movement, noise due to improperly positioned sensors, etc.). Because of the noise, physicians, when using sensor measurements to support decisions, often are not confident in the measurements from the sensors.

[0004] However, current methods for providing decisions support using wearable technology provide limited information and may be unreliable. Current methods may also be difficult to generate across multiple types of data and / or sensors because small differences in data collectionAttorney Docket No.: CRDS-008 / 01WO 348698-2034approach, dataset demographic differences, and ground truth collection methods can impact the datasets and therefore outputs or inferences derived from the datasets.

[0005] Thus, there is a need for devices and methods that can process and analyze physiological information, provide insights into reliability of such information or analysis, and / or present more robust or comprehensive information associated with such information or analysis.

[0006] In the United States, nearly 6 million adults have been diagnosed and live with the burden of heart failure (HF). HF is a chronic, progressive condition in which the heart is unable to pump enough blood to meet the body’s demands. HF costs currently exceed over $30 billion annually largely due to high rates of hospitalization and readmission. Unfortunately, management of these patients is difficult as the current standard of care techniques are ineffective and reactive.

[0007] A method for detecting a decline in cardiac function and reducing hospitalization for patients with HF that has held up in large randomized controlled trials involves filling pressure guided therapy aimed at optimizing volume status, HF is fundamentally a condition where the body’s normal compensatory mechanisms for addressing insufficient cardiac performance -namely increasing blood volume through fluid retention -- backfire and lead to a worsening of health status and, ultimately, acute decompensation and hospitalization. Monitoring filling pressures provides an early indication of the fluid retention resulting from such compensatory mechanisms, and thus allows for early intervention through increasing diuretic dosages to prevent fluid overload and acute hospitalization events. In clinic, these filling pressures are measured via invasive catheterization, which is unsuitable for remote monitoring. Meanwhile, the studies that demonstrated improved outcomes at home used FDA-approved implantable pressure sensors to measure pulmonary artery pressure (PAI’). Unfortunately, such implantable sensors require an invasive surgical procedure for placement, face reimbursement challenges, and are quite expensive (e.g., $25,000 per patient), which limits their usability to a small fraction of the overall patient population with HF that could benefit from this solution.

[0008] Recently, researchers have leveraged the advent of wearable technology and artificial intelligence to assess volume status towards enabling convenient remote monitoring without the need for an invasive procedure. Currently, FDA-cleared solutions for remotely evaluating volume status using noninvasive methods exist. However, these solutions rely on obtrusive measurements from blood pressure cuffs or systems with multiple cables and fail to capture trends due to their intrinsic binary classification of congestion status. Additionally, these methods typically provide7Attorney Docket No.: CRDS-008 / 01WO 348698-2034binary outputs (e.g., fluid overload versus not), and thus are not directly actionable using existing flowcharts for directing care as are available for implantable PAP sensors. Researchers have demonstrated the feasibility' of unobtrusively estimating changes m filling pressure for patients with HF using electrocardiogram (ECG) and seismocardiogram (SCG) signals acquired from a wearable device. However, this initial feasibility work was performed on a small dataset and did not provide an absolute pressure value, thus limiting its generalizability to a larger, more diverse population.SUMMARY

[0009] Described here are systems, devices, and methods for processing and / or analyzing physiological information, and presenting such information in a usable format for a user or physician.

[0010] In some embodiments, a method includes: receiving, from at least one of a sensing device, at least one signal associated with a physiological characteristic of a patient. The method includes processing at least one signal using a language model to output a clinically relevant physiological signal and a confidence score indicative of a reliability of the physiological signal. In some embodiments, the method includes transforming the at least one sensor signal into a format suitable for processing using a language model (e.g., a text-based format). In some embodiments, the method includes processes for training the model, e.g., training using auxiliary' tasks, dictionary learning, etc. In some embodiments, the model can be a multi-output model that can be trained to generate multiple different clinical signal outputs.

[0011] In some embodiments, a method includes: receiving a plurality of signals associated with observable physiological characteristics of a patient; transforming a first set of one or more signals from the plurality of signals into a first input format associated with a language model; generating, by inputting the first set of signals in the first input format into the language model, a first output associated with a physiological variable; transforming a second set of one or more signals from the plurality of signals into a second input format associated with an image model; generating, by inputting the second set of signals in the second input format into the image model, a second output associated with the physiological variable; and determining a predicted value for the physiological variable based on the first output and the second output.Attorney Docket No.: CRDS-008 / 01WO 348698-2034

[0012] In some embodiments, an apparatus includes: a plurality of sensors configured to measure a plurality of signals associated with observable physiological characteristics of a patient; a display configured to present information to a user; a processor operatively coupled to the plurality of sensors and the display, the processor configured to: transform a first set of one or more signals from the plurality of signals into a first input format associated with a language model; generate, by inputting the first set of signals m the first input format into the language model, a first output associated with a physiological variable; transform a second set of one or more signals from the plurality' of signals into a second input format associated with an image model; generate, by inputting the second set of signals in the second input format into the image model, a second output associated with the physiological variable; determine a predicted value for the physiological variable based on the first output and the second output; and present, via the display, the predicted value for the physiological variable,

[0013] In some embodiments, a method includes: receiving a plurality of signals associated with observable physiological characteristics of a patient; converting, using discrete latent codes, each signal of the plurality of signals into discrete latent representations of each signal of the plurality of signals; generating, by inputting the discrete latent representations into a language model, an output associated with a physiological variable; and converting, using the discrete latent codes, the output into an output waveform including a series of predicted values for the physiological variable,

[0014] In some embodiments, an apparatus includes: a plurality of sensors configured to measure a plurality of signals associated with observable physiological characteristics of a patient; a display configured to present information to a user; a processor operatively coupled to the plurality of sensors and the display, the processor configured to: convert, using discrete latent codes, each signal of the plurality of signals into discrete latent representations of each signal of the plurality of signals; generate, by inputting the discrete latent representations into a language model, an output associated with a physiological variable; convert, using the discrete latent codes, the output into an output waveform including a series of predicted values for the physiological variable; and present, via the display, the predicted value for the physiological variable.Attorney Docket No.: CRDS-008 / 01WO 348698-2034BRIEF DESCRIPTION OF THE DRAWINGS

[0015] FIG. 1A is a block diagram of a system for capturing and processing and / or presenting physiological data, according to an embodiment.

[0016] FIG. IB is a block diagram of a compute device of the system of FIG. 1A, according to an embodiment.

[0017] FIG. 2 is a block diagram of a network of devices for processing and / or presenting physiological data, according to an embodiment.

[0018] FIG. 3 schematically depicts the flow of inputs into and outputs from systems and devices for processing physiological data, according to an embodiment.

[0019] FIG. 4 is a flow chart illustrating the flow of information through systems and devices for processing physiological data, according to an embodiment,

[0020] FIG, 5 is a flow chart illustrating a method for processing physiological data, according to an embodiment.

[0021] FIG, 6 is a flow chart illustrating a method for training a language model to generate one or more physiological signals, according to an embodiment.

[0022] FIG, 7 is a high-level overview of a pulmonary capillary wedge pressure (PCWP) analysis algorithm, according to an embodiment.

[0023] FIG. 8 / X depicts data collection of noninvasive measurements from the wearable sensing device and an invasive measurement during right heart catheterization (RHC), according to an embodiment.

[0024] FIG. 8B depicts annotation of the desired physiological variable, according to embodiments.

[0025] FIG. 8C depicts dataset splitting randomly into training, validation, and test sets, according to an embodiment.

[0026] FIG. 9A depicts a flowchart for using an image model and a language model to determine PCWP, according to an embodiment.

[0027] FIG. 9B depicts a flowchart for training machine learning models in stages, according to an embodiment.

[0028] FIG. 10A depicts a scatter plot comparing noninvasively estimated PCWP to PCWP measured through a Swan-Ganz catheter during right heart catheterization (RHC).Attorney Docket No.: CRDS-008 / 01WO 348698-2034

[0029] FIG. 1 OB depicts a Bland Altman plot comparing noninvasively estimated PCWP to PCWP measured through a Swan-Ganz catheter during right heart catheterization (RHC).

[0030] FIG. 11 depicts an encoder network and a decoder network, according to embodiments.

[0031] FIG. 12 depicts a high level modeling plan, according to an embodiment.

[0032] FIG. 13A depicts an example of an ABP signal along with respective ABP generated codebook indices.

[0033] FIG. 13B depicts an example of an ECG signal along with respective ECG generated codebook indices.

[0034] FIG. 14 depict a model input and output structure, according to an embodiment.

[0035] FIG, 15A depicts an original ECG signal along with a reconstructed ECG signal using the VQ-VAE decoder, according to an embodiment.

[0036] FIG, 15B depicts an original PPG signal along with a reconstructed PPG signal using the VQ-VAE decoder, according to an embodiment.

[0037] FIG, 15C. depicts an original ABP signal along with a reconstructed ABP signal using the VQ-VAE decoder, according to an embodiment.

[0038] FIG. 16 depicts text generation for GPT Neo with scaling factors model, according to an embodiment.

[0039] FIG. 17 depicts a flow chart illustrating a method for using an image model and a language model for processing physiological information, according to an embodiment.

[0040] FIG. 18 depicts a flow chart illustrating a method for training machine learning models, according to an embodiment.DETAILED DESCRIPTION

[0041] Systems, apparatuses, and methods for processing and analyzing observable physiological signals and outputting clinically relevant physiological signals are described herein. In some embodiments, systems and devices described herein can be configured to receive observable patient information (e.g., physiological signals collected by sensing device, such as a pulse oximeter, an ECG device, a PPG device, a SCG device, an arterial line, etc.). The observable patient information may be captured by a wearable device (e.g., a smart w7atch, a patch, etc.), and / or may have been pre-collected in a home or clinical setting and stored in a database. The observable patient information may be associated with a physiological variable or characteristic of a patient.Attorney Docket No.: CRDS-008 / 01WO 348698-2034The systems and devices described herein may process and / or analyze the observable patient information, e.g., using pre-processing algorithms and / or trained models, and output a clinically relevant physiological signal. In some embodiments, other information may also be outputted with the clinically relevant physiological signal, e.g., including a confidence score indicative of the reliability of the output and / or information indicative of one or more health conditions of a user. In some embodiments, the clinically relevant physiological signal, the confidence score, and / or other information associated therewith can be used to support clinical decisions.

[0042] Conventional systems, devices, and / or methods for using wearable technology to provide decision support includes certain limitations. Many of these systems rely on a dedicated dataset, which includes the wearable signals together with ground truth (e.g., clinical variables such as blood pressure including diastolic blood pressure, systolic blood pressure, etc.). The dataset is typically split into training and testing groups, from which a model is formed using the training set to relate the signals measured with the wearable to low-dimensional variables of clinical interest. Such systems fail to provide information on the reliability of the variable that is estimated. Moreover, any trained models or other algorithms used in such systems are typically difficult to generalize across multiple datasets, since small differences in the collection approach, dataset demographics, and ground truth collection methods can have a significant influence on the performance of the models.

[0043] Systems, devices, and methods described herein are designed to overcome these limitations by providing more comprehensive information associated with a physiological variable of interest. In particular, systems, devices, and methods described herein use a trained model to output a waveform associated with a clinically relevant physiological variable, instead of a single value output. The full output waveform can be configured to provide more meaningful information to a clinician over single values associated with a variable. For example, for a model that is designed to assist clinicians in deciding howto control a person’s hemodynamic state, rather than outputting a limited number of values for blood pressure (e.g., 120 / 80 mmHg), systems, devices, and methods described herein are configured to output the full blood pressure waveform (e.g., multiple time points per section of blood pressure values, including, for example, hundreds of time points per second of blood pressure values). This has multiple advantages - first, the clinician is able to see the entire waveform rather than the two values alone, and thus there is more insight into the details of the predicted output and ability to verify the data that is presented; second, the fullAttorney Docket No.: CRDS-008 / 01WO 348698-2034waveform output can be used by the models described herein to further train the models; third, by predicting the output waveform and not low dimensional values (e.g., 120 / 80 mmHg), the model is configured to learn inherent information that is deeply embedded in the waveform output and not necessarily obvious at the surface of the waveform itself, which allows the model to be more accurate and more generalizable across different datasets and conditions. In some embodiments, systems, devices, and methods described herein can also provide a confidence score or other information indicative of a reliability of the waveform output.Overview of Systems and Methods

[0044] FIG. 1A is a block diagram of a system 100 for capturing and / or analyzing physiological information of a user, according to embodiments. The system includes a sensing device 110 operably coupled to a compute device 120. In some embodiments, the sensing device 110 and the compute device 120 can be implemented as separate or different devices, which can be operatively coupled to one another. In some embodiments, the sensing device 110 and the compute device 120 can be implemented on the same device.

[0045] The sensing device 110 can be configured to collect information about a user. For example, the sensing device 110 can be configured to capture observable physiological information of a user. The captured data can be in the form of signals, e.g., associated with one or more sensor] s) 116. In some embodiments, the sensing device 110 can be, or included in, a wearable device, such as a smart watch, a sleeve, a band, a patch, or the like. In some embodiments, the sensing device 110 can be configured to measure signals associated with one or more of volume changes, electrical activity, cardiac vibrations, ECG, heart rate, pulse rate, PPG, blood pressure, blood flow, SCG, muscle electrical potential, nerve electrical potential, temperature, brain waves, motion, measures of activity, number of steps taken, location, acceleration, pace, distance, altitude, direction, velocity, speed, time elapsed, time left, and / or the like. In some embodiments, the sensing device 110 can be configured to collect data of the user at predetermined times and / or time intervals. In some embodiments, the sensing device 110 can be configured to collect data of the user during predetermined activities (e.g., during rest and / or sleep, or other times when a user may be less likely to be moving). The sensing device 110 includes a processor 112, a memory 114, a sensor(s) 116, an input / output (I / O) device 118, and a communications interface 119 (or a multiplicity of such components), each operatively coupled to one another (e.g., via a system bus, a network, etc.). Suitable examples of a sensing device 110 are described in U. S. PatentAttorney Docket No.: CRDS-008 / 01WO 348698-2034Application No. 18 / 751,868, filed June 24, 2024, titled “SYSTEMS AND METHODS FOR MEASURING HEMODYNAMIC PARAMETERS WITH ARABLE CARDIOVASCULAR SENSING,” U. S. Patent Application No. 18 / 722,416, filed June 20, 2024, titled “MULTIWAVELENGTH PHOTOPLETHYSMOGRAM SYSTEM AND METHOD WITH MOTION ARTIFACT DETECTION,” and PCT Patent Application No. PCT / US25 / 44710, filed September 3, 2025, titled “SYSTEMS, DEVICES, AND METHODS FOR DETERMINING PHYSIOLOGICAL INFORMATION, INCLUDING HANDHELD AND IMPLANTABLE DEVICES,” the disclosures of each of which are incorporated herein by reference.

[0046] The processor 112 can be, for example, a hardware based integrated circuit (IC), or any other suitable processing device configured to run and / or execute a set of instructions or code. For example, the processor 112 can be a general-purpose processor, a central processing unit (CPU), an accelerated processing unit (APU), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a programmable logic array (PLA), a complex programmable logic device (CPLD), a programmable logic controller (PLC) and / or the like. The processor 112 can be operatively coupled to the memory 114, the I / O device, and / or the communications interface 119, e.g., through a system bus (for example, address bus, data bus and / or control bus).

[0047] The memory 114 can be, for example, a random-access memory (RAM), a memory buffer, a hard drive, a flash memory, a read-only memory (ROM), an erasable programmable read-only memory (EPROM), and / or the like. In some instances, the memory 114 can store, for example, one or more software programs and / or code that can include instructions to cause the processor 112 to perform one or more processes, functions, and / or the like. In some implementations, the memory 114 can be a portable memory (for example, a flash drive, a portable hard disk, and / or the like) that can be operatively coupled to the processor 112. In some instances, the memory 114 can be operatively coupled to the sensing device 110 and / or another compute device (e.g., compute device 120, database, etc. ). For example, in some embodiments, the memory 114 can be coupled to a remote server or database, e.g., for sending and / or receiving information therefrom. In some embodiments, the memory 114 and processor 112 may be implemented on a single chip. In other embodiments, the memory 114 and processor 112 may be implemented on separate chips.

[0048] The sensor(s) 116 can include one or more sensor(s) configured to measure an observable or measurable characteristic of a user (e.g., ECG, PPG, SCG, electrodermal activity (EDA), blood pressure, heart rate, skin temperature, etc.). The sensor(s) 116 can send a signal indicative of theAttorney Docket No.: CRDS-008 / 01WO 348698-2034measured characteristic to the processor 112, memory 114, and / or other components of the sensing device 110. For example, the sensor(s) 116 can measure and output one or more of a SCG waveform, a PPG waveform, an ECG waveform, etc. In some embodiments, the data from the sensor(s) 116 is stored in the memory 114. In some embodiments, the processor 112 can be configured to control the operation of the sensor(s) 116. For example, the processor 112 can be configured to activate the sensor(s) 116 and / or change one or more operational parameters (e.g., light wavelengths, length intensity, sampling frequency, etc.) of the sensor(s) 116. The sensor(s) 116 can be configured to operate continuously, sporadically, and / or periodically.

[0049] The I / O device 118 can include an input device and / or an output device, such as, for example, a display (e.g,, Cathode Ray tube (CRT) display, Liquid Crystal Display (LCD), Light Emitting Diode (LED) display, Organic Light Emitting Diode (OLED) display, and / or the like), mouse, keyboard, microphone, touch screen, speaker, scanner, headset, printer, camera, and / or the like. For example, the I / O device 118 may include an input device for a user to input information or instructions and / or an output device for a user to receive an output (e.g., SCG readings, ECG readings, PPG readings, etc.). In some embodiments, the I / O device 118 can be used to provide alerts to a user, e.g., to indicate to a user that there is too much movement for sensor data capture, to indicate to the user a possible issue with the sensor (e.g., sensor placement due to a wearable being worn too loosely, or sensor defect), etc. In some embodiments, the I / O device 118 can instruct a user to perform certain activities (e.g., to lay down or to minimize movement), e.g., to facilitate cleaner data capture by sensor(s) 116. In some embodiments, the I / O device 118 can display information received from the compute device 120. This information can include, for example, physiological information derived using a model, a confidence score or another reliability indication, etc., as further described herein.

[0050] The communications interface 119 of the sensing device 110 can be configured to receive information and / or send information to other devices (e.g., compute device 120). The communications interface 119 can be a wired or wireless communications interface. The communications interface 119 can, for example, be configured to send information captured by the sensor(s) 116 to the compute device 120. In some embodiments, the communications interface 119 can receive data, signals, and / or instructions from the compute device 120.

[0051] The compute device 120 can be configured to process and / or analyze sensor data, e.g., received from the sensing device 110, and / or other data, e.g., received from a user, a database, orAttorney Docket No.: CRDS-008 / 01WO 348698-2034other source. For example, the compute device 120 can be configured to filter, rectify, differentiate, integrate, enhance, pre-process, and / or combine the sensor data. In some embodiments, the compute device 120 can be configured to receive sensor data from more than one sensing device 110. In some embodiments, the compute device 120 can be nearby the sensing device 110, such as, for example, a local computer, laptop, mobile device, tablet, etc. In some embodiments, the compute device 120 can be a server that is remote from the sensing device 110 but can communicate with the sensing device 110, e.g., via a network (as depicted in FIG. 2). In some embodiments, the sensing device 110 can be configured to transmit sensor data to a nearby device (e.g,, a user device such as a mobile device) via a wireless network (e.g,, Wi-Fi, Bluetooth, etc.), and then that device can be configured to transmit the sensor data to the compute device 120 for further processing and / or analysis. In some embodiments, the compute device 120 is implemented as or includes a user device.

[0052] The compute device 120 can include a processor 122, a memory 124, an I / O device 128, and a communications interface 129 (or a multiplicity of such components). The memory 124 can be, for example, a random access memory (RAM), a memory buffer, a hard drive, a flash memory, a database, an erasable programmable read-only memory (EPROM), an electrically erasable read¬ only memory (EEPROM), a read-only memory (ROM), and / or so forth. In some embodiments, the memory 124 stores instructions that cause processor 122 to execute modules, processes, and / or functions associated with processing and / or analyzing sensor data from sensing device 110. In some instances, the memory 124 can be operatively coupled to other compute devices (e.g., as depicted in FIG. 2). In some embodiments, the memory 124 stores information associated with more than one user. For example, compute device 120 can be a household account, a medical provider account, and / or the like, and the memory 124 can be configured to store information associated with one or more users associated with that account. The administrator account can be utilized to allow one or more users (e.g., healthcare professionals, caretakers, etc.) to access information during operation.

[0053] The processor 122 of compute device 120 can be any suitable processing device configured to run and / or execute functions associated with processing and / or analyzing sensor data from the sensing device 110. The processor 122 can be a general purpose processor, microcontroller, a Field Programmable Gate Array (FPGA), an Application Specific Integrated Circuit (ASIC), a Digital Signal Processor (DSP), and / or the like. In some embodiments where the sensing device 110 andAttorney Docket No.: CRDS-008 / 01WO 348698-2034the compute device 120 are implemented as one device, the processor 122 and the processor 112 can be the same processor.

[0054] In some embodiments, the compute device 120 can be configured to process a signal from the sensor(s) 116 and / or other sensed information, e.g., to determine physiological information and / or a confidence score associated with the physiological information. The sensed information can include, for example, one or more of raw sensor signal information, processed sensor signal information, timestamp information, time window information, contextual information, and / or the like. In some embodiments, the sensed information can indicate a time period during which the sensed information was obtained and / or collected. In some embodiments, the compute device 120 can be configured to send instructions to the sensing device 110 to cause the sensing device 110 to operate according to one or more parameters. For example, the processor 122 can be configured to send instructions to the sensing device 110 to take measurements at predetermined times and / or intervals.

[0055] Generally, to generate physiological information and / or a confidence score, the compute device 120 is configured to process the sensor data received from the sensing device 110 using a model trained to predict or infer the physiological information based on the sensor data. For example, the sensor data, before or after pre-processing, can be input into a trained model, which in response can output clinically relevant physiological information. In some embodiments, the compute device 120 can be configured to use multinomial sampling to determine a confidence score associated with the predicted physiological information. For example, the sensor data may be input into the model multiple times to obtain a distribution of outputs. The compute device 120 can then analyze the distribution of outputs (e.g., by analyzing the variability of the outputs) to determine a confidence score associated with the predicted physiological information. In some implementations, the compute device 120 can be configured to generate an output. In some embodiments, the output can include a waveform of a physiological variable of interest. The output waveform can be used by a clinician or other user to gam insight into the patient’s health. In some embodiments, the output can include a single value or summary value of a physiological variable, e.g., an average (e.g., mean, median, etc.).

[0056] In some embodiments, the compute device 120 is configured to transform the signals or data received from the sensing device 110, such that the signals are suitable for processing by a model or other algorithm. For example, the compute device can transform the signals from theAttorney Docket No.: CRDS-008 / 01WO 348698-2034sensing device 110 from a first format (e.g., time-domain signal) to a second format (e.g., text¬ based signal or frequency- domain signal). In some embodiments, the model used by the compute device 120 can be a language model, such as, for example, a large language model (LLM). In such embodiments, the compute device 120 can be configured to transform a signal or waveform from the sensing device 110 into language or text-based data (e.g., a string, text, etc.). The compute device 120 can then process the transformed data using the model to generate an output. The compute device 120 can then transform the output into a suitable format for presentation to a user. For example, the compute device 120 can transform the output back into a time-domain signal waveform. In some embodiments, the compute device 120 is configured to implement an encoder and decoder architecture to transform or convert the input data into a suitable format for the model, and to revert the output data back to an initial format. Moreover, because language models and other types of models are transformer-based models, the compute device 120 can use multinomial sampling to generate a distribution of outputs based on the input data. As such, in some embodiments, the compute device 120 can be configured to generate a distribution of outputs based on the input data and to determine a confidence score or reliability of the outputs, e.g., based on a variability of the outputs. The outputs from the model can be associated with at least one physiological variable (e.g., heart rate, blood pressure, etc.). In some implementations, the outputs can be predictions of the at least one physiological variable generated by the model. For example, the outputs from the model can include or be used to provide predictions of one or more cardiopulmonary parameters of the user, such as, for example, changes in filling characteristics of the heart of a user (e.g., changes in hemodynamics, filling pressure, pulmonary artery (PA) pressure, PCWP, and the like, or any combination thereof), changes in pulmonary characteristics of the user (e.g., pleural effusion), and / or a clinical likelihood of heart failure of the patient.

[0057] FIG. IB provides a more detailed view of the compute device 120. The memory 124 can store processor-executable instructions that, when executed by a processor (e.g., processor 122, causes the processor to implement data / waveform processing 124a, data transformation or encoding-decoding 124b, an inference model 124c, and / or a physiological variable determination 124e. Optionally, in some embodiments, the memory 124 can store instructions that cause the processor to implement a confidence score determination 124d. In some implementations, the memory 124 can include additional instructions for operating the compute device 120 and / or instructions for operating the sensing device 110.Attorney Docket No.: CRDS-008 / 01WO 348698-2034

[0058] The data / waveform processing 124a includes pre-processing of data received from the sensing device 110 and / or stored in a database (e.g., functionally and / or structurally similar to the database(s) 260 of FIG. 2). The data can include data associated with physiological signals measured using the sensing device 110. The pre-processing can include filtering, augmenting, normalization, cleaning, cropping, averaging, and / or combining the data, and / or the like. In some embodiments, the pre-processing can implement one or more rules to assess the quality of an input signal and discard input signals that have a signal-to-noise ratio that is above a predetermined threshold.

[0059] The data transformation 124b includes transforming the data from a first format (e.g., type, domain, etc.) into a second format that is different from the first format. In some embodiments, the data transformation 124b includes implementing an encoder-decoder architecture. In particular, the data transformation 124b can include encoding the input data (e.g., data from the sensor(s)) into a format for processing by a model (e.g., inference model 124c), and the data transformation can include decoding the output data (e.g., data generated by the inference model 124c) back into an original format. The decoded data can include clinically relevant physiological information (e.g., waveforms of physiological variable(s) of interest), which can be presented to a clinical to evaluate and / or assess a user’s health. In some implementations, the data transformation 124b include transforming sensor data into a format suitable for processing by a language model (e.g., a string, text, etc.). In some implementations, the data transformation 124b can be configured to transform multiple signal types (e.g., different types of sensor data) into an input suitable for processing by a language model.

[0060] The inference model 124c includes generating one or more outputs associated with one or more physiological variables. In some implementations, such as when the inference model 124c is a multi-output model, the one or more outputs can include multiple different clinical signal outputs. As such, in some implementations, the inference model 124c can be configured to generate multiple signal outputs at the same time. The inference model 124c can be configured to learn shared, embedded characteristics of the different outputs that may not be obvious at a surface level that is observable but is shared at a level that the model can access. By learning multiple outputs at the same time, the model can focus on the underlying key characteristics of the output that are shared, as well as the nuances of each individual output signal that defines them. This type ofAttorney Docket No.: CRDS-008 / 01WO 348698-2034learning can be analogous to language models being trained to learn different dialects of the same language and thus better understand the salient characteristics of the language itself.

[0061] In some implementations, the inference model 124c can be a machine learning (ML) model. In some implementations, the inference model 124c includes a language model (e.g., natural language model, etc ). The output of such a model can be text-based. In some implementations, the inference model 124c includes a convolutional neural network. The model can be trained to identify and extract key features that can be used to determine or predict values associated with a physiological parameter of interest.

[0062] In some embodiments, the inference model 124c can be a model that uses self-attention to generate or predict outputs. Self-attention involves using already predicted values to predict the next values in a sequence. For example, in the case of waveform prediction, if the model has predicted a first portion of a waveform (e.g., a first second of a waveform), then the first portion can be used as an input to predict the next portion(s) of the waveform. The inference model 124c, by implementing self-attention, can learn from previous predictions and more accurately predict future values. This is different from traditional models that predict values based on fixed input sizes and / or are static. In some embodiments, the model is configured to predict sequences of output data points that are inherently related rather than individual data points or low-level variables. This is analogous to a language model predicting words rather than individual letters, or sentences rather than letter and / or words. The use of such language models for physiological signals provides more accurate predictions. This approach is further leveraged in the self-attention methodology, by which the model is configured to correct itself dynamically based on the outputs it generates by assessing the viability of the sequences of these outputs as being realistic based on types of signals it has previously observed or predicted.

[0063] In some embodiments, the inference model 124c is trained using one or more datasets, including datasets with observable physiological signals, clinically relevant signals, ground truth signals, and / or other types of signals. In some embodiments, when the inference model 124c is a language model, the model can be trained using dictionary learning on physiological signals to capture one or more essential or key characteristics of such signals, which may be non-obvious at the surface level but are deeply embedded information within the signals. Such characteristics may be shared across different signals and different datasets, allowing for training based on and learning of different signals and datasets. In some embodiments, the inference model 124c can also beAttorney Docket No.: CRDS-008 / 01WO 348698-2034trained using auxiliary tasks, e.g., tasks which do not directly relate to the main task at hand. While these tasks in and of themselves are less relevant from a clinical standpoint, they are easier to administer and require little to no ground truth domain expertise to conduct. Further details on the training of an inference model 124c, as described herein, are provided with reference to FIG. 6.

[0064] In some embodiments, the inference model 124c can be used to generate a distribution of outputs that are associated with a physiological variable of interest. The output can vary, depending on the reliability of the input data. Multinomial sampling allows the model to generate multiple predictions from the same input to give distributions of outputs. This differs from conventional models with fixed weights, which produce the same output from the same input data, and thus the reliability of the underlying data cannot be derived from and / or quantified based on the output of the model.

[0065] Multinomial sampling may be used to provide insight on the accuracy of the physiological variable prediction. For example, the more the distribution of outputs differ, the potential error in the input data may be higher. Optionally, in some embodiments, confidence score determination 124d can be implemented. The confidence score determination 124d can include determining a confidence score(s) associated with the output(s) of the inference model 124c. Taking the distribution of outputs generated by the inference model 124c, the confidence score determination 124d can analyze the variability of the distribution and / or other characteristic of the distribution and output a confidence score(s). The confidence score(s) can quantify for a user the reliability or accuracy of output(s) of the inference model 124c. Such reliability or accuracy may be affected by signal quality, sensor placement, movement, and / or the like. In some implementations, the confidence score can be determined by determining the standard deviation of the output distribution. If the standard deviation is higher (indicating greater variability), then the confidence score may indicate a lower confidence or reliability in the data. Conversely, if the standard deviation is lower (indicating lower variability), the confidence score may indicate a higher confidence or reliability in the data. In contrast to traditional approaches that may only use a pre¬ processing step to assess a signal quality of inputs based on rules, such as, for example, signal-to-noise ratio, and discard ones that are beyond a certain threshold, the confidence score determination 124d evaluates the inference model 124c’s understanding and familiarity with the input data and provides a confidence score indicative of the model’s ability to produce an accurate output based on its understanding of the inputs.Attorney Docket No.: CRDS-008 / 01WO 348698-2034

[0066] In some embodiments, the inference model 124c may include at least one of a language model or an image model. In some embodiments, a language model (as further described herein) can be used to generate an output associated with an output waveform. For example, input data (e.g., input signals of one or more observable physiological parameters) can be encoded (e.g., transformed into a format associated with the language model) and input into the language model, and the language model can generate one or more outputs based on the input data. The language model would have been previously trained based on similar input data to generate predictions associated with associated with one or more physiological variables of interest. The one or more outputs from the language model can then be decoded (e.g,, converted back into a time-series format or other format associated with the input data) and presented to a user. In some embodiments, the output can be an output waveform or a series of predictions. In some embodiments, the output can include or be used to determine a single value or prediction for a physiological variable of interest. In some embodiments, an image model (as further described herein) can be used to generate predictions of a physiological variable. For example, input data (e.g., input signals of one or more observable physiological parameters) can be encoded (e.g., transformed into a format associated with the image model, such as an image) and input into the image model, and the image model can generate one or more outputs based on the input data. The image model would have been previously trained based on similar input data to generate predictions associated with associated with one or more physiological variables of interest. The one or more outputs from the image model can include prediction(s) of the one or more physiological parameters of interest. In some embodiments, the output from the language model can be used together with the output from the image model to generate a final prediction of a physiological parameter of interest. For example, the final output or prediction can be generated by averaging a first prediction from the language model (e.g., a first output) and a second prediction from the image model (e.g., a second output). In some embodiments the prediction of the physiological variable includes assigning weights to the first and second predictions from the two models, e.g., based on the confidence score associated with the first and second predictions, the quality of the input data, etc. In some embodiments, the prediction of the physiological variable is based on which of the first or second prediction has a higher corresponding confidence score.

[0067] While two models (e g., a language model and an image model) are described herein, it can be appreciated that the models herein can include other types of machine learning models (e.g.,Attorney Docket No.: CRDS-008 / 01WO 348698-2034deep learning models, neural networks, etc.). In some embodiments, an ensemble model that combines two more more models to produce a prediction can be used. The ensemble model can use any suitable type of aggregation, including, for example, averaging, weighted averaging, etc..

[0068] The physiological variable determination 124e includes determining at least one physiological variable based on the output of the inference model 124c. In some implementations, the physiological variable determination 124e includes transforming the output generated by the inference model 124c to a waveform, e.g., a clinically relevant physiological waveform such as a blood pressure waveform. In some implementations, transforming can include using a decoder (as described above) to decode the output from the inference model 124c to a waveform data type. In some embodiments, the decode decodes the output back to an original format. In some implementations, the physiological variable determination 124e can include taking an average, median, mode, maximum, minimum, and / or the like of the output to determine the physiological variable. For example, if a distribution of outputs was generated for each second (or time period), the physiological variable determination 124e can involve averaging or otherwise combining the values from the distribution to arrive at a combined value for that second (or time period). The values over time can then be presented together to a user, e.g., as a waveform.

[0069] The I / O device 128 of the compute device 120 can be similar to the I / O device 118 of the sensing device 110. For example, the I / O device 128 can include an input device for receiving one or more inputs and / or commands from a user and / or an output device for presenting information to a user. The I / O device 128 can include any type of peripherals, such as an input device, an output device, a mouse, keyboard, microphone, touch screen, speaker, scanner, headset, printer, camera, and / or the like. In some embodiments, the I / O device 128 can be used by a user to view the processed data. For example, if the system 100 processes PPG, ECG, or SCG data, the I / O device 128 can be utilized to display physiological data metrics (e.g., heart metrics) and / or a confidence score derived from the data to the user.

[0070] The communications interface 129 of the compute device 120 can be configured to receive information and / or send information to other devices (e.g., sensing device 110, and / or other compute devices as depicted in FIG. 2). The communications interface 129 can be a wired or wireless communications interface. In some embodiments, the communications interface 129 can be configured to receive data from the sensing device 110, including the data associated with the sensor(s) 116.Attorney Docket No.: CRDS-008 / 01WO 348698-2034

[0071] FIG. 2 is a block diagram of a network of devices, including systems and devices for determining clinically relevant physiological information and / or a confidence level associated with such information, according to an embodiment. Such systems and devices can be configured to process signals to generate a physiological data waveform and / or a confidence score associated with the physiological data outputs. In some embodiments, the systems and devices can include a sensing device 210 (e.g., functionally and / or structurally similar to the sensing device 110 of FIG.1A) and / or a compute device 220 (e.g., functionally and / or structurally similar to the compute device 120 of FIGS. 1A-1B). The sensing device 210 and / or the compute device 220 can be operatively coupled to one or more other compute devices, including, for example, a server 250, a database 260, and / or optionally, one or more other device(s) 290, via one or more network(s) 202, The device(s) 290 can include, for example, additional sensing device(s) (e.g., functionally and / or structurally similar to the sensing device 110 of FIG, 1A) and / or additional compute device(s) (e.g,, functionally and / or structurally similar to the compute device 120 of FIGS, 1 A-1B). In some embodiments, the other device(s) 290 can include compute devices that are associated with one or more third parties, such as, for example, an administrator, a physician or healthcare provider, a hospital, a caretaker, etc.

[0072] The network 202 can be any type of network implemented as a wired network and / or wireless network and used to operatively couple the sensing device 210, the compute device 220, the server 250, the database 260, and / or other device(s) 290 to one another. The communication may or may not be encrypted. A wireless network may refer to any type of digital network that is not connected by cables of any kind. Examples of wireless communication in a wireless network include, but are not limited to cellular, near-field communication, radio, satellite, and microwave communication. However, a wireless network may connect to a wired network in order to interface with the Internet, other earner voice and data networks, business networks, and personal networks. A wired network is typically carried over copper twisted pair, coaxial cable and / or fiber optic cables. There are many different types of wired networks including wide area networks (WAN), metropolitan area networks (MAN), local area networks (LAN), Internet area networks (IAN), campus area networks (CAN), global area networks (GAN), like the Internet, and virtual private networks (VPN).

[0073] The network 202 may include or be coupled to the server 250 and the database 260 for processing and / or storage. The database 260 can be any device configured to store data, e.g.,Attorney Docket No.: CRDS-008 / 01WO 348698-2034received from other devices. For example, the database 260 can include instructions for storing signal data (e.g., signals captured by sensor(s) 116 of a sensing device 110), processed signal data, signal repositories, and / or the like. In some embodiments, the database 260 can store the final outputs from processing the signals, such as, for example, the clinically relevant physiological data waveforms and / or confidence scores associated therewith. In some embodiments, the database 260 can be configured to store other patient information, e.g., historical physiological characteristic information, patient demographic information, patient health history,, etc. The server 250 can be any device configured to process signals and / or data received from the sensing device 210 and / or the compute device 220. In some embodiments, the server 250 can be configured to execute some or all of the processes of the sensing device 210 and / or the compute device 220, as described above with reference to FIGS. 1 A and IB.

[0074] Similar to other sensing devices described above, the sensing device 210 can be operatively coupled to the compute device 220. For example, the sensing device 210 can be operatively coupled to the compute device 220 via near-field communication, a wireless connection (e.g,, WiFi, Bluetooth, etc.), and / or a wired connection. Optionally, the sensing device 210 can be coupled to the network(s) 202 and / or other compute devices (e.g., server 250, database 260, other device(s) 290). The sensing device 210 can be operatively coupled to the compute device 220 and / or one or more other compute devices such that the sensing device 210 can send information (e.g., sensor signals) to and / or receive information (e.g., instructions for monitoring a patient or subject, parameters for operation, etc.) from one or more such devices.

[0075] FIG. 3 is a flow 300 of information being inputted into and outputted by a pre-trained model 302, according to embodiments. The flow 300 can be implemented, for example, by a compute device such as the compute device 120 of FIGS. 1A-1B and / or the compute device 220 of FIG. 2.

[0076] The pre-trained model 302 can be structurally and / or functionally similar to the inference model 124c, as described above with reference to FIG. IB. In some embodiments, the pre-trained model 302 may have been trained using dataset(s) that can be from database(s), such as, for example, the database(s) 260 of FIG. 2. The dataset(s) can include observable signals (e.g., those measured by a sensing device such as a sensing device 110 of FIG. 1A and / or the sensing device 210 of FIG. 2) and / or other information collected of a patient. In some implementations, the observable signals may be associated with ground truth data, e.g., data collected of one or moreAttorney Docket No.: CRDS-008 / 01WO 348698-2034clinically relevant physiological variables using established measurement methods. For example, ground truth data for blood pressure may be captured using tonometry, a blood pressure cuff, etc. In some implementations, the pre-trained model 302 may have been trained using auxiliary tasks and data associated therewith.

[0077] The pre-trained model 302 can receive observable physiological signals 304 that are associated with a patient. The observable physiological signals may be measured by a sensing device (e.g., structurally and / or functionally similar to the sensing device 110 of FIG. 1A and / or the sensing device 210 of FIG. 2). The pre-trained model 302, by processing the observable physiological signals 304, can generate an output 306, The output 306 can include clinically relevant signals (e.g., physiological signals) which can include signals, waveforms, and / or the like related to information desired by a clinician to make a decision associated with the health of the patient. Additionally, the output may include a confidence score associated with the output. The confidence score indicates how likely the clinically relevant signals are representative of the actual characteristics of the patient. In some embodiments, a clinician 308 can receive the output 306 and decide on care based on the clinically relevant signals and the confidence score. The output 306 allows the clinician 308 to more confidently make a decision for a patient that may bring about a desired outcome as the signals may be directly related to the care and the confidence score 306 can indicate if the signals are reliable and should be considered by the clinician 308. For example, if the confidence score is below a predetermined threshold (or a threshold set by the clinician), the clinician may capture additional data from the user, e.g., after adjusting or reorienting the sensing device, or use other information to decide on care.

[0078] An example of a flow of inputting information into a model and obtaining an output is shown in FIG. 7. FIG. 7 depicts a high-level flow of determining or predicting PCWP.

[0079] As shown in FIG. 7, observable physiological signals 732 such as ECG, SCG, and PPG can be input into a model 736 (e.g., an analysis engine or machine learning model, structurally and / or functionally similar to the inference model 124c of FIG. 1 B) for determining a physiological variable 740 (structurally and / or functionally similar to the physiological variable determination 124e of FIG. IB) such as PCWP. The model 736 for determining PCWP can be a ML model developed to noninvasively estimate absolute PCWP using cardiac waveforms. The ML model can be implemented using a compute device that is functionally and / or structurally similar to those described in any of the systems described herein (e.g., functionally and / or structurally similar toAttorney Docket No.: CRDS-008 / 01WO 348698-2034the system 100 of FIGS. 1A-1B, the system 200 of FIG. 2, the system 300 of FIG. 3, etc.). The system can acquire ECG signals, SCG signals, and PPG signals, e.g., to provide an assessment of volume status through the combination of electrical, mechanical, and hemodynamic elements of cardiac function. Traditional heuristic-based feature engineering methods may be time-consuming and typically have limited scalability. Meanwhile, complex deep learning approaches require large datasets and are prone to overfitting. Therefore, systems, devices and methods described herein are configured to implement a deep learning approach that integrates image and language modeling with domain-specific pretraining techniques on larger datasets to effectively balance these trade¬ offs and create a scalable, robust model. While described herein in reference to PCWP, the image and language modeling with domain-specific pretraining techni ques can be used to determine other physiological variables, in some embodiments,

[0080] Initially, the model can be pretrained on multiple open-sourced datasets containing cardiovascular waveform data. To fine-tune the model to output the target PCWP label, data can be collected from a wearable device (e.g., a sensing device as described herein) during a right heart catheterization procedure as part of a multi -site, observational study. During the procedure, ECG, SCG, and PPG waveforms measured from the sensing device and pressure tracings from a Swan- Ganz catheter, can be collected from patients with heart failure reduced ejection fraction (HFrEF). The model can then be fine-tuned using other and / or additional patient data, according to methods as described herein. The outputs can represent accurate noninvasive absolute PCWP estimation using cardiac waveforms.Models and Applications

[0081] FIG. 4 is a flow 400 illustrating a process of analyzing patient data (e.g., observable physiological data, patient records, etc.) and generating physiological waveform data, according to an embodiment. The flow chart 400 includes a sensing device 410 (e.g., structurally and / or functionally similar to the sensing device 110 of FIG. 1 A and / or the sensing device 210 of FIG. 2) and a compute device 420 (e.g., structurally and / or functionally similar to the compute device 120 of FIG. IB and / or the compute device 220 of FIG. 2), and optionally database(s) 460 (e.g., functionally and / or structurally similar to the database(s) 260 of FIG. 2) and / or other device(s) 490 (e.g. functionally and / or structurally similar to the other device(s) 290 of FIG. 2).

[0082] The compute device 420 includes data processing 424a (e.g., structurally and / or functionally similar to the data / waveform processing 124a of FIG. IB), a data transformation 424bAttorney Docket No.: CRDS-008 / 01WO 348698-2034(e.g., structurally and / or functionally similar to the data transformation 124b of FIG. IB), an inference model 424c (e.g., functionally and / or structurally similar to the inference model 124c of FIG. IB and the pre-trained model of FIG. 3), a confidence score determination 424d (e g., functionally and / or structurally similar to the confidence score determination 124d of FIG. IB), and a physiological variable determination 424e (e.g., structurally and / or functionally similar to the physiological variable determination 124e of FIG. IB). In some implementations, the data processing 424a and the confidence score determination 424d are optional.

[0083] The data processing 424a receives inputs from the sensing device 410. Optionally, the data processing 424a receives inputs from the database(s) 460. The inputs can include observable physiological data measured by the sensing device 410 and / or stored in the database(s) 460. In some embodiments, the inputs can include historical patient data (e.g,, height, age, weight, physical state, mental state, activity level, etc.), patient demographic information, and / or other information associated with a patient. In some embodiments, where data was collected using a sensing device implemented as a wearable device (e.g,, a PPG sensor, a SCG sensor, an ECG sensor, etc ), the inputs can include information associated with the wearable and / or sensor(s) thereof (e.g., type of device and / or operational parameters associated with the device). The data processing 424a can be configured to pre-process the inputs. For example, the data processing 424a can include one or more of filtering, normalizing, cropping, segmenting, etc.

[0084] The data transformation 424b receives the observable physiological data and / or other patient data, e.g., directly from sensor(s) or after the data has been pre-processed via data processing 424a. The data transformation 424b transforms the waveforms from a first format to a second format, where the second format is suitable for processing using the inference model 424c. For example, the observable physiological data and / or other patient data can be transformed into a text-based or language-based format including text, phrases, and / or the like. In some implementations, the data transformation 424b can include implementing encoding.

[0085] The inference model 424c receives as inputs the transformed or encoded data. In some embodiments, the inference model 424c is a natural language model. The inference model 424c processes the inputs and generates outputs, including, for example, generated predictions of a physiological variable. In some implementations, the output can include a distribution of outputs, e.g., generated using multinomial sampling. In some implementations, the inference model 424cAttorney Docket No.: CRDS-008 / 01WO 348698-2034can be configured to use self-attention, whereby the model uses already predicted outputs to predict the next outputs in a sequence.

[0086] The physiological variable determination 424e receives the output from the inference model 424c and generates an output waveform (or multiple output waveforms), e.g., by decoding or transforming the output back into its original format. In some implementations, where a distribution of outputs was generated, the output for each time period (e.g., each millisecond, second, etc.) can be determined based on an average, median, mode, maximum, minimum, and / or the like of the output for that time period. In some implementations, the physiological variable determination 424e receives additional information associated with a patient from the database(s) 460. The additional patient information can be used when determining the physiological variables or output waveform and / or to associate the physiological variables or output waveform with a particular patient. In some implementations, the physiological variable determination 424e can send the output waveform to other device(s) 490, e.g., for review by one or more users (e.g,, a clinician).

[0087] Optionally, a confidence score determination 424d receives a distribution of outputs from the inference model 424c and determines a confidence score based on the distribution of outputs. In some implementations, the confidence score is determined based on a variability, standard deviation, and / or the like of the distribution of outputs. For example, if the output is found to have higher variability or a greater standard deviation, the confidence score can indicate that there is lower confidence in the determined value for a physiological variable. In some implementations, the confidence score determination 424d can receive the values for the physiological variable from the physiological variable determination 424e and use those values to determine the confidence score. In some implementations, the confidence score can be sent to the other device(s) 490, e.g., for review by one or more users.

[0088] FIG. 5 is a flow chart illustrating a method 500 for determining physiological data, according to an embodiment. The method can be executed by any of the systems and devices described herein, for example, any of the compute devices or sensing devices described in FIGS.1A-3.

[0089] At 502, the method 500 includes receiving, from a sensing device and / or database, at least one signal associated with an observable physiological characteristic of a patient and / or other patient information. At least one signal can be associated with a SCG waveform, PPG waveform,Attorney Docket No.: CRDS-008 / 01WO 348698-2034ECG waveform, heart rate, blood pressure, and / or the like. In some implementations, the other patient information can include information such as age, weight, expected blood pressure, resting heart rate, activity level, and / or the like.

[0090] At 508, the method 500 includes transforming the at least one signal into an input format for an inference model such as, for example, a language model. In some implementations, the input format can be a text-based format, a string format, and / or the like. In some implementations, transforming to the input format associated with the language model can include using an encoder to encode at least one signal. In some implementations, other patient information can also be transformed into a format associated with the language model.

[0091] At 510, the method 500 includes generating, using the language model, an output. In some implementations, the output can include at least one prediction of a physiological variable. The output can include distributions of outputs for multiple periods of time (e.g., millisecond, seconds, etc.). Each distribution can include a set of predicted physiological variables. In some implementations, generating the output includes self-attention. Self-attention can include using generated outputs as inputs to generating additional outputs, thus increasing the accuracy of future outputs. In some implementations, generating the output can be based on the other patient information (e.g., age, height, weight, demographic, physical state, mental state, activity level, etc.). For example, a patient’s age can indicate an expected range of a physiological variable for that patient.

[0092] At 512, the method 500 optionally includes determining a confidence score based on the output. Determining the confidence score can be based on the variability of the output, a standard deviation of the output, and / or the like. The confidence score can indicate a level of potential inaccuracy and / or reliability in the outputs generated by the model at 510.

[0093] At 514, the method 500 includes determining one or more measures or values associated with a physiological variable based on the output. In some implementations, 514 can include transforming the output from the language model into an output waveform. Where a distribution of outputs was generated for each time period in an output waveform, the individual values of the distribution can be determined by determining an average, a median, a mode, a maximum, and / or the like of the distribution. At 516, the method 500 optionally includes sending, to the display of a user device, the measure and / or the confidence score. A user can then review the displayed measure and / or the confidence score to aid in determining care for a patient.Attorney Docket No.: CRDS-008 / 01WO 348698-2034

[0094] FIG. 17 depicts a method 1700 for processing physiological information using an image model and a language model, according to an embodiment. The method can be executed by any of the systems and devices described herein, for example, the system 100 of FIG. 1A, the network environment 200 of FIG. 2, and / or the one or more compute devices described in FIGS. 1 A-2.

[0095] At 1702, the method 1700 includes receiving, from a sensing device and / or database, at least one signal associated with a physiological variable of a patient and / or other patient information. At least one signal can be associated with a SCG waveform, PPG waveform, ECG waveform, heart rate, blood pressure, and / or the like. In some implementations, the other patient information can include information such as age, weight, expected blood pressure, resting heart rate, activity level, and / or the like.

[0096] At 1704, the method 1700 includes transforming the at least one signal to an input format associated with a language model. In some implementations, the input format can be a text-based format, a string format, and / or the like. In some implementations, transforming to the input format associated with the language model can include using an encoder to encode at least one signal. In some implementations, other patient information can also be transformed into a format associated with the language model.

[0097] At 1706, the method 1700 includes generating, using the language model, a first output. In some implementations, the output can include at least one prediction of a physiological variable. The output can include distributions of outputs for multiple periods of time (e.g., millisecond, seconds, etc.). Each distribution can include a set of predicted physiological variables. In some implementations, generating the output includes self-attention. Self-attention can include using generated outputs as inputs to generate additional outputs, thus increasing the accuracy of future outputs. In some implementations, generating the output can be based on the other patient information. For example, a patient’s age can indicate an expected range of a physiological variable for that patient.

[0098] At 1708, the method 1700 includes processing the at least one signal into an input format associated with an image model. In some embodiments, processing can include converting the at least one signal into image data. In some embodiments, the image data can be resized, up-sampled, scaled, and / or the like. In some embodiments, the image data can be generated for a predetermined period of time.Attorney Docket No.: CRDS-008 / 01WO 348698-2034

[0099] At 1710, the method 1700 includes generating, using an image model, a second output. In some embodiments, the image model can be configured to receive and / or generate a waveform image associated with the at least one signal and to process the image. In some embodiments, processing the image can include up-sampling and / or resizing the image. To generate the second output, the image model may be configured to use regression to determine at least one prediction of the physiological variable (e.g., the second output).

[0100] At 1712, the method 1700 includes determining, based on the first output and the second output, a final output. In some embodiments, the first output and the second output are averaged to determine the final output. In some embodiments, weights associated with the first output and the second output can be used to determine a weighted average to determine the final output. In some embodiments, the weights can be associated with confidence scores of the first output and / or the second output. In some embodiments, the final output can be determined based choosing one of the first output or the second output based on which of the first output or the second output has a higher corresponding confidence score. In some embodiments, the final output can include one or more of an output waveform, an output value(s), and / or the like.

[0101] FIG. 9A depicts a flowchart for determining PCWP 940 (structurally and / or functionally similar to physiological variable determination 124e of FIG. IB) using an image model 935 (structurally and / or functionally similar to the inference model 124c of FIG. IB) and a language model 936 (structurally and / or functionally similar to the inference model 124c of FIG IB). While a language model and an image model are described herein, it can be appreciated that other suitable types of machine learning models or additional machine learning models can be used. In some embodiments, a device can be used to measure input signals 932 associated with a patient such as ECG, SCG, and PPG. In some embodiments, additional signals can also be used. In some embodiments, the input signals 932 can be transformed into a format associated with at least one of the language model 936 via signal to text conversion 934 (structurally and / or functionally similar to data transformation 124b of FIG. IB). In some embodiments, the input signals 932 may be transformed into a format associated with the image model 935 via signal to image conversion 933 (structurally and / or functionally similar to data transformation 124b of FIG. IB). In some embodiments, the image model 935 (e.g., ResNet-50, etc.) and the language model 936 (e.g., LLM, GPT, GPT-Neo 1.3B, etc.) are independent models. In some embodiments, the final output determining PCWP 940 is determined by a first prediction for the physiological variable based onAttorney Docket No.: CRDS-008 / 01WO 348698-2034the first output of the language model 936 and a second prediction for the physiological variable based on the second output from the image model 935. In some embodiments, the final output determining PCWP 940 is generated by averaging the first prediction and the second prediction 939. In some embodiments, a confidence score is generated for at least one of the first or second output based on a first or second distribution of outputs. In some embodiments the prediction of the physiological variable includes assigning weights to the first and second output based on the confidence score associated with the first and second output. In some embodiments, the prediction of the physiological variable is based on which of the first or second output has a higher corresponding confidence score. The flow described in reference to FIG. 9 A can be executed by any of the systems and devices described herein, for example, the system 100 of FIG. 1A, the network environment 200 of FIG, 2, and / or the one or more compute devices described in FIGS, 1 A-2, While FIG 9A and the below description discuss determining an output PCWP from ECG, SCG, and PPG input signals, other combinations of input signals and output(s) can be possible using the systems, methods, and devices described herein.

[0102] The model architecture (e.g,, the image model and the language model) was designed to increase (e.g., optimize) the amount of information that can be determined from the input cardiac signals. An LLM was chosen to be a part of the model pipeline due to LLMs being able to interpret complex patterns from data. Due to the complexity of LLMs, sufficient training data is generally needed to fine-tune the model. To compensate, an image-based model can be utilized in parallel to the LLM Using the complementary strengths of LLMs and image models, one can mitigate the risk of overfitting that is often associated with highly complex models. The model architecture can include two parallel model pipelines that are averaged together to produce a final PCWP estimation. Both pipelines are structured to input a period of time (e.g., seven to ten seconds) of waveform data from ECG, SCG, and PPG. Specifically, separate models were designed to enhance the robustness and accuracy of the model. Ensemble methods have demonstrated to generalize better to unseen data by aggregating multiple models to provide more reliable and trustworthy predictions. In some embodiments, the multiple models can have assigned weights for determining the final output (e.g., weighted average). For example, the models can have predetermined weights, weights associated with a confidence score, and / or the like.

[0103] Using this model architecture allows for providing a robust generalization to the unseen dataset. The performance of the ensemble averaging approach demonstrated an improvement ofAttorney Docket No.: CRDS-008 / 01WO 348698-20343% compared to a language model-only pipeline and 9% compared to a image model only pipeline. Therefore, the model architecture described above is capable of noninvasively estimating PCWP using cardiac bio-signals with comparable accuracy to existing implantable methods. This model architecture can be used adjunctively to the current standard of care to evaluate patients with HF. Accurate noninvasive estimation of PCWP using data collected from a wearable device represents a significant advancement in enabling convenient monitoring of intracardiac filling pressures for patients with HF. Widespread remote monitoring of volume status can aid both in identifying the risk of hospitalization and in post-discharge management, potentially reducing the length of stay and rate of readmission.

[0104] The processes (e.g,, models) described above can provide clinicians with the ability to obtain noninvasive estimates of intracardiac filling pressure in both inpatient and outpatient settings. In some embodiments, such as in an inpatient setting, clinicians can use this tool alongside physical exams to quantitatively track a patient’s progress from decompensated to compensated status, increasing confidence in discharge decision-making. Meanwhile in the outpatient setting, as witnessed in the improved patient outcomes associated with implantable devices, clinicians can identify post-discharge volume status, titrate medication dosage, and potentially detect early signs of congestion, facilitating earlier rehospitalization. As a result, the model architecture offers a convenient and affordable method to help facilitate improved HF management resulting in reduced costs, hospitalization, and improved quality of life for patients.

[0105] Referring back to FIG. 9A, in some embodiments, the model architecture can include using a ResNet-50 image model and a GPT-Neo 1.3B language model, however, other combinations of models can be used. The models are used to estimate PCWP. The inputs (e.g., ECG, SCG, and PPG) are transformed into a 3x224x224 waveform images through up-sampling and resizing of the signals. The waveform images are then fed or input into a ResNet-50 model structured as a regressor that will output a single value. Using the image model includes using convolutional layers to extract spatial features from the signals and thus allows for the recurrent layers to capture temporal dependencies. This model uses residual connections to mitigate the vanishing gradient problem, enabling improved feature extraction and model learning. The model training was improved (e.g., optimized) using mean squared error (MSE) as the loss function and with early stopping criteria included to prevent overfitting.Attorney Docket No.: CRDS-008 / 01WO 348698-2034

[0106] The model architecture further uses a large language model (LLM) and the GPT-Neo 1 3B model and baseline weights. This architecture includes an encoder-decoder framework with vector quantization consisting of a codebook. The input of the pipeline is fed into Vector Quantized Variation Autoencoders (VQ-VAE) to encode continuous waveforms into discrete latent representations, tokens that can be treated as individual words, as described further below. A unique prefix token is then applied to each set of tokens to denote tokens that belong to a specific waveform. For example, an “SCG” token is positioned before the token representative of the SCG waveform, and this process is repeated for each unique waveform. The prefix token prompts the model for the upcoming senes of tokens. GPT-Neo model is used in causal language model (CLM) training and is designed to generate human-like text based on a given input. During inferencmg, both the input signals and a prefix token representing the desired output are provided as a prompt.

[0107] VQ-VAE is a generative model based on a variational autoencoder (VAE). VAE architecture aims to make the latent space discrete using VQ techniques. VAE consist of 3 parts: (1) an encoder network: this network parametrizes the posterior q(z|x) over latents; (2) a prior distribution p(z); and (3) a decoder network: this network outputs the distribution of p(x|z) over the input data. FIG 11 depicts an encoder and decoder network relevant to VQ-VAE. VQ-VAE is used for encoding continuous waveforms into discrete latent representations. The discrete latent representations may be associated with discrete phonemes, text, or words. VQ-VAE is a type of autoencoder that generates a discrete latent representation using vector quantization. Since waveforms are continuous in nature, it is harder for any neural network to learn a learn a continuous distribution via gradient descent. VQ-VAE uses an encoder-decoder framework where it maintains a discrete codebook to generate reconstructed signals using the codebook. The encoder 1134 (structurally and / or functionally similar to data transformation 124b of FIG. IB), transforms input waveforms into continuous latent variables. Vector quantization maps continuous latent variables to discrete codes learned from the codebook. The decoder 1138 reconstructs the continuous waveforms using discrete latent codes. The discrete latent space serves as a codebook that can eventually be treated as individual words, phonemes, text, or words. This converts a sample waveform into a set of words or a sentence that can be fed in a language model and be trained to learn various patterns representing inter-waveform characteristics.

[0108] Once outputted by the models, the predictions from both ResNet-50 and GPT-Neo 1.3B can be used to obtain the final predicted PCWP (e.g., physiological variable of interest). In someAttorney Docket No.: CRDS-008 / 01WO 348698-2034embodiments, the output from the GPT Neo 1.3B (e g., language model) can be an output waveform, and systems and devices described herein can determine based on the output waveform a single value or prediction for PCWP. In some embodiments, the output from the GPT Neo 1.3B can be a single value or prediction for PCWP. This single value obtained from the output of the language model can be averaged with the output from ResNet-50 (e.g., image model). This averaging ensemble approach can allow the model pipeline to learn different aspects of the signals and smooth out the anomalies, leading to a more generalized and robust prediction. In some embodiments, averaging can include a weighted average using predetermined or calculated weights. For example, the weights can be calculated based on confidence scores associated with the outputs and / or the like. Herein, additional details on the model architecture and the sequential pretraining steps used to tram the model are provided.Training of Models

[0109] FIG. 6 is a flow chart illustrating a method 600 for training at least one machine learning model, according to an embodiment. The method can be executed by any of the systems and devices described herein, for example, the system 100 of FIG. 1 A, the network environment 200 of FIG. 2, and / or the one or more compute devices described in FIGS. 1A-2.

[0110] At 602, the method 600 includes receiving, from a database, data associated with at least one physiological signal type or physiological variable. The at least one physiological signal type can include signals from various sources of various types (e.g., waveform, values, array, etc.). The physiological variables can include heart rate, PPG, ECG, blood pressure, blood flow, and / or the like. In some implementations, the data can include additional health information such as weight, age, and / or the like associated with the data associated with at least one physiological signal type or physiological variable. In some embodiments, an output may be associated with inputting one or more characteristics or additional health information of the patient into a language model. For example, the one or more characteristics or additional health information being at least one of patient demographics, height, age, weight, physical state, mental state, or activity level.

[0111] At 604, the method 600 includes training at least one machine learning model based on the data. The at least one machine learning model can, in some implementations, include a language model (e.g., natural language model, etc.). At least one machine learning model is trained to include dictionary learning on physiological signals to determine characteristics which may be embedded within the signals. In some implementations, the machine learning model can be trainedAttorney Docket No.: CRDS-008 / 01WO 348698-2034to recognize the characteristics across different signals and different datasets. At least one machine learning model is trained to recognize and utilize various types of inputs. In some implementations, at least one machine learning model includes a multi-output model that can learn multiple different physiological outputs at the same time, by learning multiple outputs at the same time, the at least one machine learning model can determine characteristics of the outputs that are shared. The at least one machine learning model can be trained to generate outputs that are clinically desirable (e.g., desired by a clinic) and / or confidence scores.

[0112] At 606, the method 600 optionally includes training at least one machine learning model using auxiliary tasks. The auxiliary tasks can be tasks that are not directly related to the primary function of the at least one machine learning model, however, can still aid in generated output signals that are clinically relevant outputs and / or confidence scores. The auxiliary tasks can aid m a more accurate output of the machine learning model without increasing complexity as the auxiliary tasks may not need ground truth domain expertise to conduct. In some implementations, the auxiliary tasks can include pretraining the at least one machine learning model to predict ambulatory blood pressure (ABP) and / or pulmonary artery pressure (PAP). For example, the auxiliary tasks can include predicting ABP using ECG and PPG, classifying ECG arrhythmias, predicting ECG, PPG and / or SCG using signals such as ECG and PPG, SCG, PPG, and SCG, ECG, etc., and / or the like. In some implementations, the auxiliary tasks can include unmasking portion of signal.

[0113] At 608, the method 600 includes generating, using at least one machine learning model, an output. The output is generated as described in reference to 510 of FIG. 5. The output can be a distribution based off of an input from a sensing device. The output can include a prediction, distribution of prediction, and / or the like. At 610, the method 600 includes training the at least one machine learning model based on the output. Training the at least one machine learning model based on the output allows the at least one machine learning model to learn from previous predictions and more accurately predict future values.

[0114] In some embodiments, 610 can include calculating similarity scores between the input data (and / or the output) and reference data consisting of different combinations of physiological waveforms (e.g., morphology) and variables with varying signal quality. For example, a similarity score may be calculated between a plurality of input signals and pluralities of reference signals based on at least one of signal quality or morphology. A subset of the reference data with theAttorney Docket No.: CRDS-008 / 01WO 348698-2034highest similarity scores (e.g., indicating highest similarity) is then selected. Based on the selected subset, the model can be trained and / or updated. In some embodiments, updating the model can include updating algorithm stages and / or model weights. In some embodiments, one or more weights of at least one of a language model or an image model may be based on the similarity score. In some embodiments, the model can be updated based on a signal quality associated with the data. For example, the algorithm stages and / or model weights can be tuned and / or updated based on input data with similar signal quality to a subset of reference data. In some embodiments, the algorithm stages and / or model weights can be updated based on the confidence score. In some embodiments, updating and / or training the model can be repeated with different sets of input data (and / or outputs) until the confidence score(s) meets or exceeds a predetermined confidence threshold. For example, the first and second output can be iteratively generated by inputting different sets of one or more input signals (e.g., ECG, PPG, SCG, etc.) into at least one of the language model or the image model until the confidence score of the first or second output is greater than a predetermined threshold.

[0115] At 612, the method 600 includes a decision to continue training. If it is still desirable for the at least one machine learning model to generate additional outputs, the method 600 can continue down the “YES” path to return to 602 to continue training the at least one machine learning model. If it is not desirable for the at least one machine learning model to generate additional outputs, the method 600 can continue down the “NO” path to 614. At 614, the method 600 includes terminating training of the at least one machine learning model. In some implementations, the at least one machine learning model can be used for generating outputs. In some implementations, training can restart when the at least one machine learning model generates additional models.

[0116] FIG. 18 depicts a flow chart illustrating a method 1800 for training machine learning models, according to an embodiment. The method 1800 can be executed by any of the systems and devices described herein, for example, the system 100 of FIG. 1 A, the network environment 200 of FIG. 2, and / or the one or more compute devices described in FIGS. 1A-2.

[0117] Referring generally to the method 1800, the method 1800 includes sequential pretraining steps can use, in some embodiments, multiple datasets with various sets of physiological signals (e.g., ECG, SCG, PPG, and / or any other signal described herein) obtained via different devices (e.g., wearable device, implantable device, internal device, clinical device, and / or any other deviceAttorney Docket No.: CRDS-008 / 01WO 348698-2034described herein) to improve an ML model’s performance in learning the underlying relationships within the signals. This sequential training can increase the speed of convergence of the training process but also can lead to improved model generalizability. The improved model generalizability allows for the models to be fine-tuned for specific tasks for a limited dataset. As a further benefit, the method 1800 allows for the models to be trained for prompt fine-tuning where the prompts can include the signal type and specific task that the model needs to be fine-tuned on. Pretraining, combined with the inherent flexibility' associated with prompting, allows for a single model that is capable of outputting other cardiac parameters of relevance (e.g., cardiac output, PAP, etc.) that are intrinsically related to the physiological signals with minor adjustments. While the above is described in reference to the method 1800, pretraining can be used with any of the systems, devices, and methods described herein,

[0118] At 1802, the method 1800 includes receiving, from a database, a first set of data associated with a plurality of data types. In some embodiments, the plurality of data types can include different types of associated data. For example, the first set of data can include a plural ity of signals (e.g., ECG, SCG, PPG, ABP, PPG, PAP, PCWP, and / or the like), patient information (e g., demographic, height, weight, age, physical state, mental state, activity level, etc.), relationship between data (e.g., time stamps, etc.), data from various devices (e.g., wearable device, implantable device, internal device, clinical device, and / or any other device described herein), data from studies, and / or the like. In some embodiments, the data can be general (e.g., from a variety of patients) and / or targeted (e.g., from patients similar to the target patient).

[0119] At 1804, the method 1800 includes training at least one machine learning model based on the first set of data. The at last one machine learning model can include at least one of an image model, a language model, and / or any of the models described herein. In some embodiments, training can be supervised and / or unsupervised. Training on the first set of data allows for the at least one machine learning model to be generalized to recognizing patterns in the plurality of data types and allows for the at least one machine learning model to be fine-tuned to more specific inputs and desired outputs. In some embodiments, training during 1804 is a pretraining step.

[0120] At 1806, the method 1800 optionally includes receiving, from a user device, a prompt associated with at least one desired input type and at least one desired output type. The at least one desired input signal may be an input type that is expected and the at least one desired output type may be an output type that is expected. For example, the prompt can be associated with using aAttorney Docket No.: CRDS-008 / 01WO 348698-2034specific device configured to measure the desired input type and desired output type can be the specific type of output that is desired for observation, study, and / or the like. Including the prompt can allow for fine-tuning (e.g., customization) of the at least one machine learning model based on the prompt. At 1808, the method 1800 includes receiving a second set of data, the second set of data associated with the prompt. For example, the second set of data can include input data that is of the desired input type and output data that is of the desired output type.

[0121] At 1810, the method 1800 includes training the at least one machine learning model based on the second set of data. The second set of data is used to fine-tune the at least one machine learning model so that the at least one machine learning model can be customized to an expected input and a desired output, while keeping the context of the data of the first set of data to increase the accuracy of the predictions generated by the at least one machine learning model. At 1812, the method 1800 optionally includes generating, using the machine learning model, an output. In some embodiments, the output can be based on an input that is in the desired input type. In some embodiments, the output is in the at least one desired output type.

[0122] Wlnle the systems, methods, and devices described herein can be used with a variety of inputs and to generate a variety of output, the systems, methods, and devices described herein can be used to specifically train and use model that generates a PCWP model as described below. However, the below descriptions can be similarly used for different inputs and / or to generate different outputs than described.

[0123] FIG. 9B depicts a flowchart for training machine learning models in stages, according to an embodiment. The flowchart depicts training the machine learning model that use ECG, SCG, and / or PPG signals to generate a PCWP estimate. The model is initially pretrained using sequential multi-task pretraining steps 952 on both open-source datasets and studies followed by task-specific fine-tuning 954 on the SCG signal in cardiovascular monitoring for the study dataset.

[0124] As described above, a study was conducted to gather data for model development. In the study, a wearable sensing device (e.g., such as the sensing devices described herein) was placed on a subject’s sternum to record time- synchronized continuous ECG, PPG, and SCG signals. In this study, 500 subjects wore the wearable device in a seating or supine position while multiple blood pressure cuff readings were taken. Additional studies were also conducted where wearable data was similarly gathered, which included simultaneous ECG and SCG data collected on 220 subjects.Attorney Docket No.: CRDS-008 / 01WO 348698-2034

[0125] Data for model fine-tuning and evaluation was collected during the study — a prospective, multi-site, observational study to gather training data for model development. During the study, patients scheduled to undergo RHC wore a wearable device. Approximately 1000 subjects were recruited as part of the study, of which 266 subjects have HFrEF. As described in embodiments herein, the wearable device is a chest-worn wearable that records single-lead ECG, multi wavelength PPG, and triaxial SCG signals. The device adheres to the patient via two disposable ECG electrodes and is placed mid-sternum, two fingers below the suprasternal notch. The wearable device initiates a recording by interfacing to a computer via a communications interface and using a custom C# app, after which the device continuously captures the physiological data. The recording is stored locally on an internal memory and can be later downloaded using the aforementioned app. While the wearable device was worn during this study, other devices, including other wearable devices, can be used to collect data for similar studies.

[0126] FIGS. 8A-C shows a detailed overview of the study design. For example, FIGS. 8A-C depicts a diagram showing data collection, annotation, and splitting for the study. FIG, 8A depicts concurrent data capture of simultaneous noninvasive ECG, SCG, and PPG waveforms 832 from the wearable sensing device and pressure tracings 842 from the Swan-Ganz catheter during the RHC procedure.

[0127] To obtain the labels for the model, the one-minute Swan-Ganz pressure tracings 842 during the PCWP measurement were reviewed by subject matter experts. The tracings were divided between three different annotators (e.g., subject matter experts) who were asked to interpret the waveform and annotate the mean PCWP values. The annotators were also asked to provide feedback if unable to provide reliable PCWP measurements due to poor waveform quality. FIG.8B depicts PCWP value annotation 844 from core lab cardiologists and model label generation using an average value of closest annotations. As shown, in some embodiments, annotation can include using an annotation tiebreaker 846 to determine a desired annotation.

[0128] After annotation, the dataset (including the wearable device signals 832, the RHC pressure tracing 842, and the annotated PCWP value 844) was randomly separated into a training and holdout set using an -80 / 20 ratio, leading to a test set of 57 recordings from subjects with HFrEF. The hold-out set was siloed during model development and the input data was only exposed to the model for final evaluation. The labels of the hold-out set were not used for training or fine-tuningAttorney Docket No.: CRDS-008 / 01WO 348698-2034the model. FIG. 8C depicts dataset splitting 848 randomly into training, validation, and a completely held-out test set.

[0129] Once the data is prepared for training, the machine learning models can be trained on the data for determining the PCWP. Different pretraining strategies were employed for image-based and text-based causal language models (CLM). These pretraining strategies included sequential pretraining on multiple datasets enabling the model to progressively learn various waveform signals across multiple devices. The image-based model was pretrained on certain datasets to learn PAP and PCWP before fine-tuning on the main study dataset. Using a ResNet-50 model, which is trained on a large dataset such as ImageNet, pretraining on a first dataset was performed to learn mean PAP values. The ResNet-50 model was adapted to process ECG, PPG, and ABP waveforms as inputs. Further pretraining was performed on a second dataset using ECG, PPG, and SCG signals to learn PCWP.

[0130] Similarly, the causal-based LLM GPT-Neo was first pretrained on datasets to learn ECG, PPG, and ABP signals. The pretraining was continued on additional datasets to learn ECG, PPG, and the unintroduced SCG signals. Finally, the model was also pretrained on the main study dataset by first pretraining it to learn PAP signals followed by PCWP signals.

[0131] A causal language modeling task is used to pretrain the model on the waveform data. The waveforms are converted into VQ-VAE tokens and a custom tokenizer is generated that maps codebook indices to a fixed token value. A GPT-Neo model is used for the CSM training. GPT- Neo is a CSM for generating human-like text based on a given input. It operates by predicting the next word in a sequence, allowing it to produce coherent and contextually relevant responses. Sequential pretraining on multiple datasets is used to further train the machine learning models. In causal training, each token is generated while attending to previously generated tokens. During inferencing, the input signals are provided as a prompt to allow the model to start generating after the prompted signals.

[0132] Signals from both a reference device (e.g., during RHC) and the wearable sensing device are used for VQ-VAE training (i.e., dictionary learning). During training, a 32 sample length waveform (e.g., ~ 0.256 seconds) belonging to any of the wearable device signals (e.g., ECG, PPG, SCG, ABP) are fed into the encoder-decoder framework to generate a discrete latent variable of 8 code indices. The codebook size was 1024 tokens, though codebooks of other sizes can be used. During operation, any 32 sample length waveform can be encoded into 8 code indices from theAttorney Docket No.: CRDS-008 / 01WO 348698-2034codebook. While training for VQ-VAE, 32 sample worth of signals belonging to any of the ECG, PPG, ABP signal group were captured. Each token corresponds to 4 sample points. The codebook indices length is about 8 indices to encode a signal. FIGS. 13A-B depict some of the examples of signals along with their generated codebook indices.

[0133] The GPT-Neo model (e.g., the language model) was trained on ECG, PPG, ABP waveforms from datasets to understand the underlying hemodynamic relationships between cardiac bio-signals: ECG, PPG, and blood pressure. One of the first challenges is that these signals, being continuous in nature, cannot be utilized directly in the language models. As such, the high level modeling plan depicted in FIG. 12 is implemented.

[0134] In the example study, 1024 samples from each signal group were used which corresponds to approximately 8.192 seconds of data. The data was inputted into the causal language model 1236 (structurally and / or functionally similar to the inference model 124c of FIG. IB),. 864 samples were converted into 216 code indices using VQ-VAE encoder 1234 (structurally and / or functionally similar to data transformation 124b of FIG. IB). The ECG signal was hardcoded using a set of fixed parameters and both the PPG and the ABP signals were standard scaled. Along with the signals, two extra tokens for ABP scaling were also added in the model training: (1) a [MAP] token to represent the mean of the ABP waveform and (2) a [STD] token to represent the standard deviation of the ABP waveform. These predicted [MAP] and [STD] tokens were then utilized to inverse scale the? XBP waveform. Tokenizer mappings were updated to learn possible values of M / XP and ABP STD ranging from 0 to 250 along with VQ-VAE mappings. During mferencmg, ECG tokens 1434a (structurally and / or functionally similar to ECG tokens 1234a of FIG. 12) and PPG tokens 1434b (structurally and / or functionally similar to PPG tokens 1234b of FIG. 12) are provided as a prompt to the model to generate ABP tokens 1440 (structurally and / or functionally similar to ABP tokens 1240 of FIG. 12 and / or physiological variable determination 124e of FIG. IB). The model input and output structure are depicted in FIG. 14.

[0135] The GPT-Neo previously trained on open-source datasets to estimate ABP was further trained on ECG, SCG, and PPG signals from a wearable device dataset. The sampling rate for ECG and PPG was 125 Hz while the SCGz signal was upsampled to 250 Hz. Approximately 7 seconds of each signal were fed into the causal learning model that consisted of ECG, SCGz, PPG separator tokens followed by the respective signals: (["[ECG]"] + ECG + ["[SCGz]"] + SCG + ["[PPG]"] + PPG). During inferencing, ECG and SCG signal tokens were passed as a prompt toAttorney Docket No.: CRDS-008 / 01WO 348698-2034predict PPG tokens. This allowed the model to learn the newly introduced SCG waveform using similar tasks involving previously learned waveforms (i.e., ECG, PPG, and ABP) After pretraining the GPT-Neo model on ABP waveforms, pretraining was continued on additional datasets to estimate the PAP waveform and focus the model learning on intracardiac filling pressures indicative of pulmonary congestion. During model inferencing, ECG, PPG, ABP tokens are provided as a prompt to the model to generate PAP tokens.

[0136] Fine-tuning of the model was performed using the training set from the study. To ensure robust model performance, a five-fold cross validation (CV) was conducted on the training set. After identifying desirable (e.g., optimal) hyperparameters, the training set was further split into a new training and validation set. Forty subjects of the training set were separated into a validation set. The model was trained on this newtraimng subset and the model checkpoints with higher (e.g., the best) performance on the validation set were selected. These model checkpoints with higher performance were then used for inference on the held-out test set to estimate a PCWP value. This approach helped improve the model’s generalizability to the unseen dataset. Finally, statistical analysis was performed between the model derived PCWP and the Swan-Ganz PCWP on the holdout dataset.Model Evaluations

[0137] The machine learning models including a ResNet-50 and GPT-Neo mode, as described above, were evaluated for their efficacy using benchmark dataset. From open-source datasets, 3608 windows from 462 subjects were selected to be used an evaluation dataset. This dataset can be used for model evaluation for improved comparative analysis. The dataset include an average of 8 windows per subject and each window consisting of ECG, PPG and ABP signals. The signals are qualified to possess over 10 SNR.

[0138] The text generation of the language model was also evaluated. ABP generation was reviewed while evaluating Encoder-Decoder models or CLM models. The model decoding strategy was customized and evaluated using Beam search and Multinomial Sampling. Out of both of these strategies, multinomial sampling was performing as desired. Multinomial Sampling randomly selects the next token based on the probability distribution over the entre VQ-VAE vocabulary of 1024 tokens. The text is also generated N-times to retrieve various possible combinations generated by the model, providing multiple predictions using the same input data and therefore the standard deviation of the predicted MAP values can be used as a modelAttorney Docket No.: CRDS-008 / 01WQ 348698-2034confidence score. Upon analysis, it was discovered that the higher the standard deviation of predicted MAP, the higher the error.

[0139] The VQ-VAE model’s performance on the signals was evaluated. The following table shows the reconstructed signals MAE when evaluated on about 5000 random signals from the validation dataset:VQ-VAE Reconstruction Mean Absolute Error (MAE)ECG PPG ABPOpen-Source Dataset 1 0.02 0.03 1.26Open-Source Dataset 2 0.01 1.037 1.21VQ-VAE reconstruction MAE

[0140] Similarly to the table, FIGS. 15A-C depict the original signal along with the reconstructed signal using the VQ-VAE decoder. FIG. 15 A depicts an original ECG signal and a reconstructed ECG signal with a MAE of 0.016 and 0.020 respectively. FIG. 15B depicts an original PPG signal and a reconstructed PPG signal with a MAE of 0.030 and 0.027 respectively. FIG. 15C depicts an original ABP signal and a reconstructed ABP signal with a MAE of 1.152 and 1.225 respectively.

[0141] GPT-Neo was evaluated using an input prompt approach. When the GPT-Neo model was prompted with VQ-VAE tokens of ECG and PPG, the model generated the ABP tokens. The text generation for GPT Neo with scaling factors model is depicted in FIG. 16. As seen in FIG. 16, the GPT-NEO generates a prediction by receiving an input of ECG and PPG signals 1632, then decoding the predicted sequence into VQVAE tokens using a tokenizer 1634. Then, an inverse of the ABP VQ-VAE token is generated using the VQ-VAE decoder into a scaled ABP signal 1638. The scaled ABP is then descaled by using predicted scaling factors 1640. The model generates the scaling mean and standard deviation along with the standard scaled predicted ABP. These scaling factors are used to descale the ABP signal to the predicted ABP signal.

[0142] The pre-training of the model was also evaluated. As seen in the table below, details the patient demographics and clinical characteristics of the training and validation set and the test set.Demographics and Clinical CharacteristicsTraining Validation Sets Test Set Total(n = 169) (n = 40) (n = 57) (n = 266) GenderAttorney Docket No.: CRDS-008 / 01WO 348698-2034Male 106 (63%) 23 (58%) 37 (65%) 166 (62%) Female 63 (37%) 17 (43%) 20 (35%) 100 (38%) RaceWhite 76 (45%) 15 (38%) 28 (49%) 119 (45%) Black or African American 66 (39%) 19 (48%) 19 (33%) 100 (39%) Other 27 (16%) 6 (15%) 10 (18%) 43 (16%) NYHA Class1 5 (3%) 1 (3%) 2 (4%) 8 (3%) II 39 (23%) 8 (20%) 9 (16%) 56 (21%) III 92 (54%) 24 (60%) 38 (67%) 154 (58%) IV 30 (18%) 7 (18%) 7 (12%) 44 (16%) Not Reported 3 (2%) 0 (0%) 1 (2%) 4 (1%) Mitral Valve RegurgitationNone / Trivial / Mild 62 (37%) 18 (45%) 40 (70%) 120 (45%) Moderate / Severe 107 (63%) 22 (55%) 17 (30%) 146 (55%)

[0143] FIG. 10A depicts a Scatter plot and FIG. 10B depicts a Bland Altman plot comparing noninvasively estimated PCWP to PCWP measured through a Swan-Ganz catheter during right heart catheterization.

[0144] The performances of the individual models and the ensemble method (e.g., using both the image model and the language model) were evaluated using regression metrics including mean difference ± standard deviation, limits of agreement (LoA), and correlation coefficient (R). Metrics on the PCWP model and implantable devices (CardioMEMS HF remote monitoring system, Cordelia system) can be seen in the table below. Correlation and Bland-Altman plots comparing the estimated PCWP to PCWP measured through Swan-Ganz are shown in 10. For the 57 subjects in the test set, the average difference between the model derived PCWP and the annotated PCWP values was 2.19 ± 6.02 mmHg. Due to the low number of subjects classified as New York Heart Association (NYHA) Class I and those unreported, an analysis of the PCWP values was conducted after the removal of those subgroups.Model Performance on Hold-out SetAttorney Docket No.: CRDS-008 / 01WO 348698-2034n Mean Limits of Correlation Difference Agreement coefficient (mmHg) (mmHg) (R)PCWP Model 57 2.19 ± 6.02 [-9.6, 14.0] 0.61PCWP Model (NYHA Class II, III, IV) 54 2.49 ± 5.96 [-9.2, 14.2] 0.63 CardioMEMS dPAP443 4.40 ± 5.95 [-7.4, 16.3] N / ACordelia dPAP348 3.00 ± 6.50 [-10.1, 16.1] 0.60

[0145] The table below shows a comparison between the model derived PCWP and the reference annotated PCWP for each subpopulation. Specifically, the mean annotated PCWP, mean model estimated PCWP, mean differences of the two values, and the mean absolute difference (MAD) are presented for gender, NYHA class, and mitral valve regurgitation subgroups.Subpopulation Metricsn Mean Swan- Mean Model Mean MADGanz PCWP PCWP Difference (mmHg) (mmHg) (mmHg)(mmHg)All Population 57 17.75 ± 8.07 19.12 ± 5.64 2.19 ± 6.02 5.2GenderMale 37 16.36 ± 7.91 18.92 ± 5.97 2.56 ± 5.85 4.95Female 20 17.98 ± 8.52 19.48 ± 5.12 1.51 ± 6.41 5.65RaceWhite 28 15.3 ± 5.88 18.68 ± 5.35 3.37 ± 5.12 5.01Black or African 19 17.39 ± 10.09 19.15 ± 5.67 1.77 ± 7.08 5.78 AmericanOther 10 20.60 ± 8.77 20.29 ± 6.78 -0.31 ± 5.89 4.61NYHA Class1 2 18 ± 4.24 14.17 ± 2.81 -3.83 ± 7.05 4.98II 9 15.22 ± 7.11 20.17 ± 6.81 4.95 ± 4.24 5.19III 38 17.22 ± 8.48 18.75 ± 5.53 1.52 ± 6.08 5.06IV 7 17.14 ± 9.61 21.73 ± 4.95 4.59 ± 6.43 6.46Attorney Docket No.: CRDS-008 / 01WO 348698-2034Not Reported 1 17.33 15.36 -1.98 1.98Mitral ValveRegurgitationNone / Trivial / Mild 40 16.4 ± 8.31 18.81 ± 5.76 2.41 ± 6.42 5.83 Moderate / Severe 17 18.17 ± 7.66 19.85 ± 5.44 1.68 ± 5.08 4.83

[0146] In groups where the sample size was sufficient for analysis (n > 10), the subpopulation metrics presented in the table above show minor differences, and no two subpopulations of sufficient size had a MAD greater than 1 mmHg. The greatest discrepancy in performance was between the “White” and “African American” subgroups. Similar discrepancies are commonly observed between these subgroups in wearable devices, particularly those with optical sensors. However, while there is a 0.77 mmHg difference between the two subgroups, the cause is likely more related to outliers in the data rather than subgroup bias. After removing a single subject that is considered to have low signal quality due to the high amount of respiratory noise in the input signals, the difference between the subgroups drops from 0.77 mmHg to 0.24 mmHg. A bias-free ML model capable of estimating PCWP may help address existing health disparities in HF management.

[0147] The above demonstrates the ability of an AIL model to accurately predict PCWP from cardiac waveforms measured using noninvasive sensors. This model enables convenient hemodynamic assessment by using a wearable device attached to the sternum and feeding the cardiac signals through the ML model. This model is capable of alleviating the burden of invasive procedures or implantable devices to obtain an intracardiac filling pressure measurement.

[0148] Using Swan-Ganz derived PCWP as a reference, the PCWP model shows comparable performance to that of CardioMEMS (an implanted, more invasive device) and significant improvement to that of Cordelia (another implanted device) (p < 0.05). Furthermore, the PCWP model maximum limits of agreement have an improvement of 2.32 mmHg compared to CardioMEMS and 2.12 mmHg compared to Cordelia. To note, while these implantable devices are measuring diastolic PAP instead of PCWP, the two pressures are highly correlated and often diastolic PAP is used as a surrogate for PCWP. By showing similar or improved performance, the PCWP model not only can have a comparable impact in remote care as implantable devices, whichAttorney Docket No.: CRDS-008 / 01WO 348698-2034have been shown to reduce hospitalization and improve outcomes for patients with HF, but also with fewer limitations.

[0149] As described herein, a ML model that uses noninvasively acquired cardiac signals to estimate PCWP has advantages over implantable devices. These implantable devices use pressure transducers placed in the pulmonary' artery to measure PAP and require an invasive procedure for sensor placement and calibration. Sensor drift is also common for implantable devices and additional procedures are needed periodically to re-calibrate the transducer. The PCWP ML model, on the other hand, provides an absolute reading without the need for sensor calibration. As such, no costly procedures or equipment are necessary' for remote monitoring of volume status.

[0150] " While the RHC procedure is a standard for inpatient PCWP measurements, a noninvasive device presents certain advantages. For the approximately 1000 subjects collected during the mam study, no adverse effects were observed while patients underwent the RHC procedure. Although adverse events associated with RHC were not recorded as part of the main study, RHC, generally a low-risk procedure, can still result in serious complications, particularly during venous access and catheter placement. In addition, the systems, methods, and devices described herein enable clinicians to obtain more frequent pressure measurements using general healthcare practitioners, bypassing the need for the specialized equipment and highly trained professionals required for RHC procedures. Furthermore, improving patient outcomes requires not only careful interpretation of hemodynamic parameters but also repeated measurements to identify trends m congestion status. Augmenting the current standard of care with noninvasive PCWP estimates can help decide the point within the diagnostic pathway to perform an RHC. Decision-making with a noninvasive ML model can serve as guidance in treatment decisions and assist in ensuring diagnostic accuracy.

[0151] It should be understood that the disclosed embodiments are not intended to be exhaustive, and functional, logical, operational, organizational, structural and / or topological modifications can be made without departing from the scope of the disclosure. As such, all examples and / or embodiments are deemed to be non-limiting throughout this disclosure.

[0152] All definitions, as defined and used herein, should be understood to control over dictionary definitions, definitions in documents incorporated by reference, and / or ordinary' meanings of the defined terms.

[0153] Examples of computer code include, but are not limited to, micro-code or micro¬ instructions, machine instructions, such as produced by' a compiler, code used to produce a webAttorney Docket No.: CRDS-008 / 01WO 348698-2034service, and files containing higher-level instructions that are executed by a computer using an interpreter. For example, embodiments can be implemented using Python, Java, JavaScript, C++, and / or other programming languages and development tools. Additional examples of computer code include, but are not limited to, control signals, encrypted code, and compressed code.

[0154] The drawings primarily are for illustrative purposes and are not intended to limit the scope of the subject matter described herein. The drawings are not necessarily to scale; in some instances, various aspects of the subject matter disclosed herein can be shown exaggerated or enlarged in the drawings to facilitate an understanding of different features. In the drawings, like reference characters generally refer to like features (e.g., functionally similar and / or structurally similar elements).

[0155] The acts performed as part of a disclosed method(s) can be ordered m any suitable way. Accordingly, embodiments can be constructed in which processes or steps are executed in an order different than illustrated, which can include performing some steps or processes simultaneously, even though shown as sequential acts in illustrative embodiments. Put differently, it is to be understood that such features can not necessarily be limited to a particular order of execution, but rather, any number of threads, processes, services, servers, and / or the like that can execute serially, asynchronously, concurrently, in parallel, simultaneously, synchronously, and / or the like in a manner consistent with the disclosure. As such, some of these features can be mutually contradictory, in that they cannot be simultaneously present in a single embodiment. Similarly, some features are applicable to one aspect of the innovations, and inapplicable to others.

[0156] Where a range of values is provided, it is understood that each intervening value, to the tenth of the unit of the lower limit unless the context clearly dictates otherwise, between the upper and lower limit of that range and any other stated or intervening value in that stated range is encompassed within the disclosure. That the upper and lower limits of these smaller ranges can independently be included in the smaller ranges is also encompassed within the disclosure, subject to any specifically excluded limit in the stated range. Where the stated range includes one or both of the limits, ranges excluding either or both of those included limits are also included in the disclosure.

[0157] The phrase “and / or,” as used herein in the specification and in the embodiments, should be understood to mean “either or both” of the elements so conjoined, i.e., elements that are conjunctively present in some cases and disjunctively present in other cases. Multiple elementsAttorney Docket No.: CRDS-008 / 01WO 348698-2034listed with “and / or” should be construed in the same fashion, i.e., “one or more” of the elements so conjoined. Other elements can optionally be present other than the elements specifically identified by the “and / or” clause, whether related or unrelated to those elements specifically identified. Thus, as a non-limiting example, a reference to “A and / or B”, when used in conjunction with open-ended language such as “comprising” can refer, in one embodiment, to A only (optionally including elements other than B); in another embodiment, to B only (optionally including elements other than A); in yet another embodiment, to both A and B (optionally including other elements); etc.

[0158] As used herein in the specification and in the embodiments, “or” should be understood to have the same meaning as “and / or” as defined above. For example, when separating items in a list, “or” or “and / or” shall be interpreted as being inclusive, i.e,, the inclusion of at least one, but also including more than one of a number or list of elements, and, optionally, additional unlisted items. Only terms clearly indicated to the contrary, such as “only one of’ or “exactly one of,” or, when used in the embodiments, “consisting of,” will refer to the inclusion of exactly one element of a number or list of elements. In general, the term “or” as used herein shall only be interpreted as indicating exclusive alternatives (i.e., “one or the other but not both”) when preceded by terms of exclusivity, such as “either,” “one of,” “only one of,” or “exactly one of.” “Consisting essentially of,” when used in the embodiments, shall have its ordinary meaning as used in the field of patent law.

[0159] As used herein in the specification and in the embodiments, the phrase “at least one,” in reference to a list of one or more elements, should be understood to mean at least one element selected from any one or more of the elements in the list of elements, but not necessarily including at least one of each and every element specifically listed within the list of elements and not excluding any combinations of elements in the list of elements. This definition also allows that elements can optionally be present other than the elements specifically identified within the list of elements to which the phrase “at least one” refers, whether related or unrelated to those elements specifically identified. Thus, as a non-limiting example, “at least one of A and B” (or, equivalently, “at least one of A or B,” or, equivalently “at least one of A and / or B”) can refer, in one embodiment, to at least one, optionally including more than one, A, with no B present (and optionally including elements other than B); in another embodiment, to at least one, optionally including more than one, B, with no A present (and optionally including elements other than A); in yet anotherAttorney Docket No.: CRDS-008 / 01WO 348698-2034embodiment, to at least one, optionally including more than one, A, and at least one, optionally including more than one, B (and optionally including other elements); etc.

[0160] In the embodiments, as well as in the specification above, all transitional phrases such as “comprising,” “including,” “carrying,” “having,” “containing,” “involving,” “holding,” “composed of,” and the like are to be understood to be open-ended, i.e., to mean including but not limited to. Only the transitional phrases “consisting of’ and “consisting essentially of’ shall be closed or semi-closed transitional phrases, respectively, as set forth in the United States Patent Office Manual of Patent Examining Procedures, Section 2111.03.

[0161] Some embodiments described herein relate to a computer storage product with a non-transitory / computer- readable medium (also can be referred to as a non-transitory processor-readable medium) having instructions or computer code thereon for performing various computer-implemented operations. The computer-readable medium (or processor-readable medium) is non-transitory in the sense that it does not include transitory propagating signals per se (e.g., a propagating electromagnetic wave carrying information on a transmission medium such as space or a cable). The media and computer code (also can be referred to as code) can be designed and constructed for the specific purpose or purposes. Examples of non-transitory computer-readable media include, but are not limited to, magnetic storage media such as hard disks, floppy disks, and magnetic tape; optical storage media such as Compact Disc / Digital Video Discs (CD / DVDs), Compact Disc-Read Only Memories (CD-ROMs), and holographic devices; magneto-optical storage media such as optical disks; carrier wave signal processing modules; and hardware devices that are specially configured to store and execute program code, such as Application- Specific Integrated Circuits (ASICs), Programmable Logic Devices (PLDs), Read-Only Memory (ROM) and Random-Access Memory (RAM) devices. Other embodiments described herein relate to a computer program product, which can include, for example, the instructions and / or computer code discussed herein.

[0162] Some embodiments and / or methods described herein can be performed by software (executed on hardware), hardware, or a combination thereof. Hardware modules can include, for example, a processor, a field programmable gate array (FPGA), and / or an application specific integrated circuit (ASIC). Software modules (executed on hardware) can include instructions stored m a memory' that is operably coupled to a processor and can be expressed in a variety of software languages (e.g., computer code), including C, C++, Java™, Ruby, Visual Basic™, and / orAttorney Docket No.: CRDS-008 / 01WO 348698-2034other object-oriented, procedural, or other programming language and development tools. Examples of computer code include, but are not limited to, micro-code or micro-instructions, machine instructions, such as produced by a compiler, code used to produce a web service, and files containing higher-level instructions that are executed by a computer using an interpreter. For example, embodiments can be implemented using imperative programming languages (e.g., C, Fortran, etc.), functional programming languages (Haskell, Erlang, etc.), logical programming languages (e.g., Prolog), object-oriented programming languages (e.g., Java, C++, etc.) or other suitable programming languages and / or development tools. Additional examples of computer code include, but are not limited to, control signals, encrypted code, and compressed code.

Claims

Attorney Docket No.: CRDS-008 / 01WO 348698-2034CLAIMSWe claim:

1. A method, comprising:receiving a plurality of signals associated with observable physiological characteristics of a patient;transforming a first set of one or more signals from the plurality of signals into a first input format associated with a language model;generating, by inputting the first set of signals in the first input format into the language model, a first output associated with a physiological variable;transforming a second set of one or more signals from the plurality of signals into a second input format associated with an image model;generating, by inputting the second set of signals in the second input format into the image model, a second output associated with the physiological variable; anddetermining a predicted value for the physiological variable based on the first output and the second output.

2. The method of claim 1, wherein the first set of signals includes one or more continuous waveforms, andtransforming the first set of signals into the first input format associated with the language model includes encoding, using an encoder, the one or more continuous waveforms into one or more discrete latent representations.

3. The method of claim 2, further comprising:decoding, using a decoder, the first output generated by the language model to generate a continuous waveform associated with the physiological variable.

4. The method of claim 1, wherein the plurality of signals includes at least one of: a seismocardiogram (SCG) signal, a photothermography (PPG) signal, or an electrocardiogram (ECG) signal.Attorney Docket No.: CRDS-008 / 01WO 348698-20345. The method of claim 1, wherein determining the predicted value for the physiological variable is determined by:determining a first prediction for the physiological variable based on the first output of the language model;determining a second prediction for the physiological variable based on the second output of the image model; andaveraging the first prediction and the second prediction to determine the predicted value for the physiological variable.

6. The method of claim 1, wherein the plurality of signals includes a plurality of continuous waveforms, andgenerating the first output further includes inputting one or more characteristics of the patient into the language model, the one or more characteristics including: a demographic of the patient, a height of the patient, an age of the patient, a weight of the patient, a physical state of the patient, an activity level of the patient, or a mental state of the patient.

7. The method of claim 1, further comprising:generating a distribution of first outputs; andgenerating a confidence score associated with the first output based on the distribution of first outputs.

8. The method of claim 7, further comprising:generating a distribution of second outputs; andgenerating a confidence score associated with the second output based on the distribution of second outputs.

9. The method of claim 8, wherein determining the predicted value for the physiological variable includes assigning weights to the first output and the second output based on the confidence score associated with the first output and the confidence score associated with the second output.Attorney Docket No.: CRDS-008 / 01WO 348698-203410. The method of claim 8, wherein determining the predicted value for the physiological variable is based on which of the first output or the second output has a higher corresponding confidence score.

11. The method of claim 1, further comprising:determining a similarity' score between the plurality of signals and pluralities of reference signals based on at least one of signal quality or morphology; andsetting one or more weights of at least one of the language model or the image model based on the similarity score,12. The method of claim 1, further comprising:determining a confidence score of the first output or the second output; and iteratively generating the first output or the second output by inputting different sets of one or more signals from the plurality of signals into the language model or the image model until the confidence score of the first output or the second output is greater than a predetermined threshold.

13. The method of claim 1, wherein the physiological variable is a pulmonary capillary wedge pressure (PCWP) of the patient.

14. An apparatus, comprising:a plurality of sensors configured to measure a plurality of signals associated with observable physiological characteristics of a patient;a display configured to present information to a user;a processor operatively coupled to the plurality of sensors and the display, the processor configured to:transform a first set of one or more signals from the plurality of signals into a first input format associated with a language model;generate, by inputting the first set of signals in the first input format into the language model, a first output associated with a physiological variable;Attorney Docket No.: CRDS-008 / 01WO 348698-2034transform a second set of one or more signals from the plurality of signals into a second input format associated with an image model;generate, by inputting the second set of signals in the second input format into the image model, a second output associated with the physiological variable;determine a predicted value for the physiological variable based on the first output and the second output; andpresent, via the display, the predicted value for the physiological variable.

15. The apparatus of claim 14, wherein the first set of signals includes one or more continuous waveforms, andthe processor is configured to transform the first set of signals into the first input format associated with the language model includes encoding, using an encoder, the one or more continuous waveforms into one or more discrete latent representations.

16. The apparatus of claim 15, wherein the processor is further configured to:decode, using a decoder, the first output generated by the language model to generate a continuous waveform associated with the physiological variable.

17. The apparatus of claim 14, wherein the plurality of signals includes at least one of: a seismocardiogram (SCG) signal, a photothermography (PPG) signal, or an electrocardiogram (ECG) signal.

18. The apparatus of claim 14, wherein the processor is configured to determine the predicted value for the physiological variable is determined by:determining a first prediction for the physiological variable based on the first output of the language model;determining a second prediction for the physiological variable based on the second output of the image model; andaveraging the first prediction and the second prediction to determine the predicted value for the physiological variable.Attorney Docket No.: CRDS-008 / 01WO 348698-203419. The apparatus of claim 14, wherein the plurality of signals includes a plurality of continuous waveforms, andthe processor is configured to generate the first output by inputting one or more characteristics of the patient into the language model, the one or more characteristics including: a demographic of the patient, a height of the patient, an age of the patient, a weight of the patient, a physical state of the patient, an activity level of the patient, or a mental state of the patient.

20. The apparatus of claim 14, wherein the processor is further configured to:generate a distribution of first outputs;generate a confidence score associated with the first output based on the distribution of first outputs; andpresent, via the display, the confidence score associated with the first output.

21. The apparatus of claim 20, wherein the processor is further configured to:generate a distribution of second outputs;generate a confidence score associated with the second output based on the distribution of second outputs; andpresent, via the display, the confidence score associated with the second output.

22. The apparatus of claim 21, wherein the processor is configured to determine the predicted value for the physiological variable by assigning weights to the first output and the second output based on the confidence score associated with the first output and the confidence score associated with the second output.

23. The apparatus of claim 21, wherein the processor is configured to determine the predicted value for the physiological variable by determining the predicted value for the physiological variable is based on which of the first output or the second output has a higher corresponding confidence score.

24. The apparatus of claim 14, wherein the physiological variable is a pulmonary capillary wedge pressure (PCWP) of the patient.Attorney Docket No.: CRDS-008 / 01WO 348698-203425. A method, comprising:receiving a plurality of signals associated with observable physiological characteristics of a patient;converting, using discrete latent codes, each signal of the plurality of signals into discrete latent representations of each signal of the plurality of signals;generating, by inputting the discrete latent representations into a language model, an output associated with a physiological variable; andconverting, using the discrete latent codes, the output into an output waveform including a series of predicted values for the physiological variable.

26. The method of claim 25, wherein each signal of the plurality of signals is a continuous waveform of an observable physiological characteristic of the patient, andconverting each signal of the plural ity of signal s into discrete latent representations includes using vector quantization to map the continuous waveforms to the discrete latent codes.

27. The method of claim 25, wherein the discrete latent representations are associated with discrete phonemes, text, or words.

28. The method of claim 25, wherein the physiological variable is a pulmonary capillary wedge pressure (PCWP) of the patient.

29. The method of claim 25, wherein generating the output includes generating a distribution of outputs, and the method further comprising:generating a confidence score based on the distribution of outputs.

30. The method of claim 25, wherein the plurality of signals includes at least one of: a seismocardiogram (SCG) signal, a photothermography (PPG) signal, or an electrocardiogram (ECG) signal.Attorney Docket No.: CRDS-008 / 01WO 348698-203431. The method of claim 25, wherein each signal of the plurality of signals is a continuous waveform of an observable physiological characteristic of the patient, andgenerating the output further includes inputting one or more characteristics of the patient into the language model, the one or more characteristics including: a demographic of the patient, a height of the patient, an age of the patient, a weight of the patient, a physical state of the patient, an activity level of the patient, or a mental state of the patient.

32. An apparatus, comprising:a plurality of sensors configured to measure a plurality of signals associated with observable physiological characteristics of a patient;a display configured to present information to a user;a processor operatively coupled to the plurality of sensors and the display, the processor configured to:convert, using discrete latent codes, each signal of the plurality of signal s into discrete latent representations of each signal of the plurality of signals;generate, by inputting the discrete latent representations into a language model, an output associated with a physiological variable;convert, using the discrete latent codes, the output into an output waveform including a series of predicted values for the physiological variable; andpresent, via the display, the predicted value for the physiological variable.

33. The apparatus of claim 32, wherein each signal of the plurality of signals is a continuous waveform of an observable physiological characteristic of the patient, andthe processor is configured to convert each signal of the plurality of signals into discrete latent representations by using vector quantization to map the continuous waveforms to the discrete latent codes.

34. The apparatus of claim 32, wherein the discrete latent representations are associated with discrete phonemes, text, or words.Attorney Docket No.: CRDS-008 / 01WO 348698-203435. The apparatus of claim 32, wherein the physiological variable is a pulmonary capillary wedge pressure (PCWP) of the patient.

36. The apparatus of claim 32, wherein the processor is configured to generate the output by generating a distribution of outputs, and the processor being further configured to:generating a confidence score based on the distribution of outputs; andpresent, via the display, the confidence score.

37. The apparatus of claim 32, wherein the plurality of signals includes at least one of: a seismocardiogram (SCG) signal, a photothermography (PPG) signal, or an electrocardiogram (ECG) signal,38. The apparatus of claim 32, wherein each signal of the plurality of signals is a continuous waveform of an observable physiological characteristic of the patient, andthe processor is configured to generate the output further by inputting one or more characteristics of the patient into the language model, the one or more characteristics including: a demographic of the patient, a height of the patient, an age of the patient, a weight of the patient, a physical state of the patient, an activity level of the patient, or a mental state of the patient.

Images