WO2025178770A1

WO2025178770A1 - Methods and systems for photoplethysmographic data preparation in machine-learning training

Info

Publication number: WO2025178770A1
Application number: PCT/US2025/014908
Authority: WO
Inventors: David Mcnamee; Daniel VECH
Original assignee: Roche Diabetes Care Inc
Current assignee: Roche Diabetes Care Inc
Priority date: 2024-02-20
Filing date: 2025-02-07
Publication date: 2025-08-28
Anticipated expiration: 2026-08-20
Also published as: WO2025178609A1

Abstract

A method for training a machine-learning (ML) model using photoplethysmographic (PPG) data has been developed. The method includes receiving PPG training data including series of PPG data corresponding to first and second optical wavelengths, generating time segments in the at least one PPG data set, selecting only a portion of the time segments having high- frequency noise beneath a predetermined threshold for inclusion in training data, removing a rolling average value for the from each time segment in the training data, identifying peaks and valleys in the training data, generating normalized amplitudes of the peaks and valleys in the training data, identifying waves in the normalized amplitudes, each wave being identified based on a spike in the normalized amplitudes, and generating a trained ML model to classify data relevant to a biomarker using a training process with an untrained ML model and the identified waves.

Description

METHODS AND SYSTEMS FOR PHOTOPLETHYSMOGRAPHIC DATA PREPARATION IN MACHINE-LEARNING TRAINING

CLAIM OF PRIORITY

[0001] This application claims priority to International Patent Application No.

PCT/US2024/016460, filed February 20, 2024, which is hereby incorporated by reference herein in its entirety.

TECHNICAL FIELD

[0002] This disclosure is directed to systems and methods to improve the quality of signal data received from photoplethysmographic (PPG) sensors, which enables improvements to machine- learning classifiers for physiological data encoded in the PPG signals that rely on a training process.

BACKGROUND

[0003] In the health and fitness fields, photoplethysmographic (PPG) sensors are a common form of non-invasive sensor that monitor the physiological characteristics of humans or other animals. Examples of commercially available PPGs include one or more optical transceivers that are placed on the skin of a user. The optical transceivers transmit pulses of light that penetrate through at least a portion of the skin and blood vessels of the user and the optical transceivers detect return signals from the pulses, which are then processed to measure a biomarker in the user. In particular, a time series of PPG measurements can detect the expansion and contraction of blood vessels in a user, although the detected light signals in the PPG data may include other properties that can be used to detect biomarkers in the user. Some PPG sensors use multiple wavelengths of light including green, red, yellow, and near infrared light transceivers, which penetrate the skin to different degrees and deliver complementary signals. Common uses of PPG sensors in health and fitness devices include the measurement of heart rate and blood oxygenation (pulse oximetry) bio markers. [0004] In addition to traditional uses of PPG sensors, there is interest in the art in extracting other biomarkers from PPG sensor data. Many of these efforts include the use of machinelearning (ML) models to classify the biomarkers based on latent features in PPG signals. However, commercially available PPG sensors are known to generate large amounts of noise. While signal-processing techniques exist to mitigate the effects of noise for common uses such as heart rate and pulse oximetry, the noise may introduce artifacts that lead to errors in the training process for ML models. If the errors are introduced during training, then later inferencing use of the trained ML models to classify biomarkers of interest may produce undesirable inaccuracies when the ML model erroneously interprets noise as a feature of interest in the PPG data. Consequently, improvements to systems and methods for processing of PPG data to improve training of ML models and the quality of trained ML models in practical inferencing would be beneficial.

SUMMARY

[0005] In one embodiment, a method for training a machine -learning model using photoplethysmographic data has been developed. The method includes receiving, with a training system, sensor data comprising at least one photoplethysmographic (PPG) data set, the at least one PPG data set further comprising a first series of PPG data corresponding to a first optical wavelength and a second series of PPG corresponding to a second optical wavelength, generating, with the training system, a plurality of time segments in the at least one PPG data set, selecting, with the training system, only a portion of the plurality of time segments having high-frequency noise beneath a predetermined threshold for inclusion in a training data set, removing, with the training system, a rolling average value for the first series of PPG data and the second series of PPG data from one or more time segments in the training data set, identifying, with the training system, a plurality of peaks and valleys in each of the first series of PPG data and the second series of PPG data within one or more time segments in the training data set, generating, with the training system, a plurality of normalized amplitudes of the plurality of peaks and valleys corresponding to each of the first series of PPG data and the second series of PPG data within one or more time segments in the training data set, identifying, with the training system, a plurality of waves in the plurality of normalized amplitudes of the plurality of peaks and valleys corresponding to each of the first series of PPG data and the second series of PPG data within one or more time segments in the training data set, each wave being identified based on a spike in the plurality of normalized amplitudes, and generating, with the training system, a trained machine-learning model to classify data relevant to a biomarker using a training process with an untrained machine -learning model and the plurality of waves identified in the training data set.

[0006] In another embodiment, a system for training a machine -learning model using photoplethysmographic data has been developed. The system includes a memory and a processor operatively connected to the memory. The memory is configured to store program instructions, sensor data comprising at least one photoplethysmographic (PPG) data set, the at least one PPG data set further comprising a first series of PPG data corresponding to a first optical wavelength and a second series of PPG corresponding to a second optical wavelength, and data corresponding to a machine- learning model. The processor is operatively connected to the memory, the processor being configured to execute the stored program instructions to generate a plurality of time segments in the at least one PPG data set, select only a portion of the plurality of time segments having high-frequency noise beneath a predetermined threshold for inclusion in a training data set, remove a rolling average value for the first series of PPG data and the second series of PPG data from one or more time segments in the training data set, identify a plurality of peaks and valleys in each of the first series of PPG data and the second series of PPG data within one or more time segments in the training data set, generate a plurality of normalized amplitudes of the plurality of peaks and valleys corresponding to each of the first series of PPG data and the second series of PPG data within one or more time segments in the training data set, identify a plurality of waves in the plurality of normalized amplitudes of the plurality of peaks and valleys corresponding to each of the first series of PPG data and the second series of PPG data within one or more time segments in the training data set, each wave being identified based on a spike in the plurality of normalized amplitudes, and generate a trained machine-learning model stored in the memory to classify data relevant to a biomarker using a training process with an untrained machine-learning model and the plurality of waves identified in the training data set.

BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWINGS

[0007] To easily identify the discussion of any particular element or act, the most significant digit or digits in a reference number refer to the figure number in which that element is first introduced.

[0008] FIG. 1 is a schematic diagram of a system for performing training and inferencing with a machine-learning (ML) model based on photoplethysmographic (PPG) sensor data. [0009] FIG. 2 is a block diagram of a method for preparing PPG data for ML model training, training an ML model with the prepared training data, and using the trained ML model.

[0010] FIG. 3 is a depiction of two different time segments of a photoplethysmogram including waves from two different light wavelengths with a low and high levels of high- frequency noise.

[0011] FIG. 4 is a depiction of zero crossings in a time segment of a photoplethysmogram and a histogram of zero crossing distributions that depict low frequency noise and high-frequency noise.

[0012] FIG. 5 is a depiction of peaks and valleys in waves of a photoplethysmogram and normalized waves generated from the photoplethysmogram data.

[0013] FIG. 6 is a depiction of spike identification in the normalized waves of the photoplethysmogram data.

[0014] FIG. 7 is an example of an auto-encoder ML model that is trained using the method depicted in FIG. 2.

DETAILED DESCRIPTION

[0015] These and other advantages, effects, features and objects are better understood from the following description. In the description, reference is made to the accompanying drawings, which form a part hereof and in which there is shown by way of illustration, not limitation, embodiments of the inventive concept. Corresponding reference numbers indicate corresponding parts throughout the several views of the drawings.

[0016] While the inventive concept is susceptible to various modifications and alternative forms, exemplary embodiments thereof are shown by way of example in the drawings and are herein described in detail. It should be understood, however, that the description of exemplary embodiments that follows is not intended to limit the inventive concept to the particular forms disclosed, but on the contrary, the intention is to cover all advantages, effects, and features falling within the spirit and scope thereof as defined by the embodiments described herein and the embodiments below. Reference should therefore be made to the embodiments described herein and embodiments below for interpreting the scope of the inventive concept. As such, it should be noted that the embodiments described herein may have advantages, effects, and features useful in solving other problems. [0017] The devices, systems and methods now will be described more fully hereinafter with reference to the accompanying drawings, in which some, but not all embodiments of the inventive concept are shown. Indeed, the devices, systems and methods may be embodied in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will satisfy applicable legal requirements.

[0018] Likewise, many modifications and other embodiments of the devices, systems and methods described herein will come to mind to one of skill in the art to which the disclosure pertains having the benefit of the teachings presented in the foregoing descriptions and the associated drawings. Therefore, it is to be understood that the devices, systems and methods are not to be limited to the specific embodiments disclosed and that modifications and other embodiments are intended to be included within the scope of the embodiments. Although specific terms are employed herein, they are used in a generic and descriptive sense only and not for purposes of limitation.

[0019] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of skill in the art to which the disclosure pertains. Although any methods and materials similar to or equivalent to those described herein can be used in the practice or testing of the methods, the preferred methods and materials are described herein.

[0020] Moreover, reference to an element by the indefinite article “a” or “an” does not exclude the possibility that more than one element is present, unless the context clearly requires that there be one and only one element. The indefinite article “a” or “an” thus usually means “at least one.” Likewise, the terms “have,” “comprise” or “include” or any arbitrary grammatical variations thereof are used in a non-exclusive way. Thus, these terms may both refer to a situation in which, besides the feature introduced by these terms, no further features are present in the entity described in this context and to a situation in which one or more further features are present. For example, the expressions “A has B,” “A comprises B” and “A includes B” may refer both to a situation in which, besides B, no other element is present in A (z.e., a situation in which A solely and exclusively consists of B) or to a situation in which, besides B, one or more further elements are present in A, such as element C, elements C and D, or even further elements. [0021] As used herein, the term Person with Diabetes (PwD) refers to a patient who is diagnosed with or is at-risk for being diagnosed with one or more forms of diabetes including pre-diabetes, type 1 diabetes, type 2 diabetes, gestational diabetes, and, optionally, one or more comorbidities that are associated with diabetes.

[0022] As used herein, the term “biomarker” refers to any quantifiable aspect of a PwD's physiology that is either identified directly or calculated from one or more measurements generated by an analytical device, such a PPG sensor. For a PwD, a bio marker of interest may be directly linked to monitoring one or more of diabetes in the PwD, a diabetes -related comorbidity in the PwD, or to the general health and wellness of the PwD. Non-limiting examples of bio markers that may be encoded in PPG data include pulse oximetry, heart rate, respiration rate, blood pressure, as well as chemical analytes in the blood or other body fluids and tissues.

[0023] As used herein, the term “PPG data set” refers to at least one time series of data generated by a photoplethysmographic sensor that is placed on the epidermis of a human or animal user. In the examples described herein, each PPG data set includes a series of discrete PPG data points, which are also referred to as “samples” or “measurements” herein interchangeably. In certain embodiments described herein, the PPG data set further include PPG data from PPG sensors that operate at two different optical wavelengths concurrently. As such, these PPG data sets include first and second series of PPG data for the user that represent concurrent measurements taken at different optical wavelengths. Each series of PPG data is also referred to as a photoplethysmogram herein.

[0024] Fig. 1 depicts a system 104 for training a machine-learning model to classify a biomarker in photoplethysmogram sensor data received from a user-operated PPG sensor. Fig.

1 further depicts a plurality of training users 152 with corresponding training PPG sensors 154, a plurality of inferencing users 162 with corresponding inferencing PPG sensor device systems 164, and a plurality of healthcare providers (HCPs) 172 with corresponding HCP computing terminals 174.

[0025] In FIG. 1, the training PPG sensors 154 collect training PPG sensor data from the training users 152 during a training process for a machine-learning model. The inferencing PPG sensor device systems 164 collect PPG sensor data from the inferencing users 162 after the machine- learning model is trained, and the trained machine-learning model produces an output indicative of a biomarker in the PPG data. In some embodiments, the hardware configuration of the training PPG sensors 154 and the inferencing PPG sensor device systems 164 is identical. In other embodiments, the inferencing PPG sensor device systems 164 either incorporate or are communicatively coupled to additional hardware elements to perform inferencing operations. As such, references to the inferencing PPG sensor device systems 164 herein refer to both the PPG sensor hardware and additional hardware and software elements that interface with the PPG sensor including, for example, a smartphone, smartwatch, personal computer, or other electronic device used by the inferencing users 162. In the configurations described herein, both the training PPG sensors 154 and the inferencing PPG sensor device systems 164 include multi-wavelength transceivers that emit light and detect reflected or transmitted light, depending upon the sensor configuration, with wavelengths in at least two different wavelength ranges. For example, the PPG sensors 154 and 164 are configured to emit and detect light in a first optical wavelength range of approximately 500 nm to 600 nm corresponding to green light and a second optical wavelength range of approximately 750 nm to 1,400 nm corresponding to near infrared light. Other PPG configurations optionally use different selected optical wavelengths corresponding to other light ranges, which are typically in the visible-light or infrared spectra due to the optical absorption characteristics of the human epidermis and subcutaneous tissues. In most practical embodiments the inferencing users 162 represent a larger population that the training users 152, although it is envisioned that some or all of the training users 152 may also be inferencing users 162 upon completion of the training process.

[0026] In the configuration of FIG. 1, the training system 104 includes a processor 108, communication interface 112, and a memory 120. The processor 108 is formed from one or more digital processing devices such as central processing units (CPUs), graphics processing units (GPUs), field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), digital signal processors (DSPs) and any other suitable digital logic devices. In some embodiments, the processor 108 optionally includes one or more neural network accelerators (NNAs) that are configured to perform either or both of machine-learning model training and machine-learning model inferencing operations in an efficient manner, although these accelerators are not strictly required to implement the system 104 and to perform the methods described herein. The communication interface 112 is a device that enables the training system 104 to communicate with other computing systems, including the training PPG sensors 154, inferencing PPG sensor device systems 164, and the HCP terminals 174, via the network 140. Non- limiting examples of the communication interface 112 include wired and wireless network adapters such as Ethernet and IEEE 802.11 family wireless network adapters. The network 140 is typically a local area network (LAN) or wide area network (WAN) such as the Internet. While the network 140 enables an expedient remote exchange of data between the training/inferencing system 104 and the PPG sensors 154/164, in an alternative configuration the network 140 is a physical “sneakernet” that enables the exchange of data storage devices such as USB memory sticks or other data storage media to enable the training/inferencing system 104 to receive photoplethysmographic data from the PPG sensors 154/164 by directly accessing the data storage media.

[0027] In the training system 104, the memory 120 is a non-transitory memory system that includes one or more data storage devices including at least one of a non-volatile or volatile memory device configured to store program instructions 122, original PPG training data 124, machine- learning (ML) model 126, PPG inferencing data 132, and ML output data 136. The stored program instructions 122 include instructions for a data transformation process that receives the original PPG training data 124 and generates training PPG data 128 corresponding to waves that are identified in high-quality sections of PPG data while eliminating low-quality data. The original PPG training data 124 optionally include additional ground-truth data about a bio marker of interest for a supervised training process, which enables the training system 104 to verify the progress of training the ML model 126 from an initial untrained state to a trained state. The ground-truth data are typically generated using independent sensing devices (not shown) that the training users 152 operate concurrently with the operation of the PPG sensors 154, but after completion of the training process the inferencing users 162 operate the inferencing PPG sensor device systems 164 with the trained ML model 126 without the use of additional independent sensing devices.

[0028] The stored program instructions further include instructions for a training process to generate a trained ML model 126. In some embodiments, the trained ML model 126 further incorporates multiple ML models that are configured in series or in parallel to classify information about one or more biomarkers of interest in PPG inferencing data. An example of an ML model is an auto-encoder model, although other ML models may be trained and used for inferencing using the same techniques described herein. In particular, in one configuration an auto-encoder ML model is trained to reproduce high-quality representations of the PPG signal data to reduce or eliminate signal noise and other artifacts in the original PPG sensor data, and auto-encoder optionally generates intermediate training output data, inferencing output data, or both, for a second trained ML model that classifies information relevant to one or more bio markers. In a configuration in which the system 104 also acts as an inferencing system, the memory 120 stores PPG inferencing data 132, which include a similar PPG data type as the original training data 124, but these data are not part of a predetermined training data set. In FIG. 1, the inferencing system 104 receives PPG inferencing data 132 from the PPG sensor device systems 164 that are typically outside of the original training pool. During an inferencing operation, the inferencing system 104 applies the PPG inferencing data 132 to the trained ML model 126 to generate ML output data 136. In the context of FIG. 1, the ML output data 136 include data for an estimate of a bio marker of relevant interest to one of the inferencing users 162.

[0029] During operation, the training system 104 receives training data in the form of a plurality of photoplethysmograms from PPG sensors 154 that are operated by training users 152. After the data cleaning and training operations that are described in further detail below, the training system 104 trains a machine -learning (ML) model 126. The trained ML model 126 is then used to perform inferencing optionally using the training system 104 as a combined training/inferencing system 104 or by providing the trained ML model 126 to other computing devices that perform inferencing. While the training system 104 is depicted as a single device for illustrative purposes, those of skill in the art will recognize that the training system 104 may be implemented using a cluster of multiple computing systems that are communicatively coupled via the network 140 or other networked connections. As described above, the training system 104 optionally acts as an inferencing system 104, but other computing devices including a dedicated network-connected inferencing system or individual electronic devices such as the PPG sensor device systems 164 or other electronic devices operated by the inferencing users 162 may perform the inferencing operation using a previously trained machine-learning model. As such, any references herein to the training system 104 or the inferencing system 104 are understood to apply to any of these hardware and software configurations.

[0030] FIG. 2 is a block diagram of a process 200 for processing PPG data to reduce the effects of noise and other low-quality data in an ML model training process. In the description below, a reference to the process 200 performing a function or action refers to the operation of a processor to execute stored program instructions to perform the function or action in coordination with other components of a PPG sensing and machine -learning system. The process 200 is described in conjunction with the embodiment of FIG. 1 for illustrative purposes.

[0031] The process 200 begins as the training/inference system 104 receives PPG training data from the training PPG sensors 154 (block 204). In the configuration of FIG. 1, the PPG sensors 154 collect PPG data from the training users 152 and transmit the PPG data to the training system 104 via the network 140. The training system 104 stores the photoplethysmograms in the original PPG training data 124 in the memory 120. Each set of PPG data includes a photoplethysmogram for the first optical wavelength and the second optical wavelength that both correspond to a time series of measurements taken from a training user 152, who operates one of the training PPG sensors 154.

[0032] As described above, the original PPG training data includes both useful PPG sensor data as well as noise and other artifacts that the training PPG sensors 154 generate during operation, but that is not useful for training a machine-learning model. The process 200 continues as the training system 104 performs a series of operations to reduce the effects of noise and to select time segments that include high quality PPG signal data (training PPG wave data 128 stored in the memory 120), which the training system 104 subsequently uses to generate the trained ML model 126. During process 200, the training system generates a plurality of time segments from the PPG training data and selects only a portion of the time segments that include sufficiently low levels of high frequency noise to be useful for a machine- learning training process (block 208). In one embodiment, each time segment lasts for four (4) seconds and the training PPG sensors 154 generate 25 samples per second, yielding 100 PPG data samples in each time segment. In another embodiment, each time segment is 60 seconds in length and the training PPG sensors 154 also generate 25 samples per second, yielding 1,500 PPG data samples in each time segment, but alternative configurations may use different time segment lengths and sampling rates. The selection process includes a wavelength correlation process and a zero crossing measurement operation to identify if each time segment has a sufficiently low level of high-frequency noise to be selected for a training operation. In the wavelength correlation process, the training system 104 identifies a correlation between waves in the first series of PPG data at the first wavelength and the second series of PPG data at the second wavelength for each time segment. The correlation level is inversely related to the levels of high-frequency noise in both the first and second series of PPG data, and since the first and second series of PPG data are expected to have a strong positive correlation for acceptable levels of high-frequency noise. If the correlation score exceeds a predetermined threshold, then the training system 104 stores the time segments for both the first and second optical wavelengths for further processing to identify waves within the PPG data set, but if the score is too low then the time segments are excluded from the training PPG data 128. FIG. 3 depicts two different examples of PPG measurements with a first PPG time series graph 304 that includes a low level of high-frequency noise and a second PPG time series graph 316 that includes a high level of high-frequency noise. As depicted in the graph 304, the trace 308 represents the PPG data of the first optical wavelength corresponding to green light (e.g. approximately 500 nm to 600 nm) and the second trace 312 represents the PPG data of the second optical wavelength corresponding to near infrared light (e.g. approximately 750 nm to 1,400 nm). The first and second traces exhibit a high level of positive correlation, which indicates a low level of high-frequency noise. During operation, the training system 104 selects this time segment for further processing to identify normalized waves to be included with the training PPG data 128. By contrast, the graph 316 depicts PPG data traces 320 and 324, which correspond to the first optical wavelength of green light and the second optical wavelength of infrared light, respectively. The graph 316 depicts a low level of correlation between the traces 320 and 324, which indicates levels of high-frequency noise that exceed the predetermined threshold. During operation, the training system 104 does not select this time segment for inclusion with the training PPG data 128.

[0033] In addition to the correlation process that is described above, the training system 104 determines the number of zero crossing counts for signals in each time segment to assess low- frequency noise, high-frequency noise, or both low and high frequency noise. FIG. 4 depicts the zero-crossing count technique for filtering in more detail. The graph 404 depicts signals in a time segment crossing over a zero amplitude threshold at zero crossing points 408A - 408G. The training system 104 detects each zero crossing by identifying consecutive samples in the PPG data that transition from negative-to-positive values, positive-to-negative values, or when a single PPG data sample has a value of zero. In the example of FIG. 4, the time segment in graph 404 includes seven (7) zero-crossings at the references 408A - 408G. To determine the level of noise in the time window, the training system 104 compares the detected number of zero crossings to one or both of a minimum threshold and a maximum threshold of zero- crossings. The minimum threshold of zero-crossings represents a minimum number of expected zero-crossings to be identified in the time segment, and a value below the minimum threshold indicates low-frequency noise in the time segment. The maximum threshold of zero -crossings represents a maximum number of expected zero-crossings to be identified in the time segment, and a value above the minimum threshold indicates high-frequency noise in the time segment. The training system 104 rejects time segments that exhibit excessive low-frequency or high- frequency noise based on these thresholds and to include time segments that are within either or both of the thresholds in the training PPG data 128. To identify the minimum and maximum thresholds, the training system 104 refers to stored data corresponding to a histogram of zero crossings that are observed in a large number of time segments in the training data, such as the histogram 412 in FIG. 4. In the example of FIG. 4, the training system 104 identifies a minimum threshold of seven (7) zero crossings at reference 416 and identifies a maximum threshold of nine (9) zero crossings at reference 420 in the histogram 412, although the precise values of the minimum and maximum thresholds may vary in other configurations.

[0034] Referring again to FIG. 2, the process 200 continues as the training system 104 subtracts a rolling average of the PPG data in each selected time segment to mitigate direct current (DC) signal bias (block 212). In the embodiments described herein, the rolling average refers to a mean value of a window of ten (10) PPG data samples in the time series PPG data around each PPG data point, although larger or smaller rolling average windows may be used. As used herein, the term “average” includes weighted or non-weighted arithmetic, geometric, harmonic, and logarithmic means, or any other techniques that are recognized in the art for averaging time series data. The training system 100 subtracts the rolling average value from each data point to mitigate the effects of DC bias so that PPG data time segments generated at different times and between different training users 152 show reduced variance due to DC bias. Since the DC bias generally does not contain useful information for training an ML model, this process improves the consistency between different sets of training data. The training system 104 performs the subtraction of the rolling average values for both the first and second wavelengths in the PPG data.

[0035] The high frequency noise filtering and rolling average subtraction processes that are described above are performed prior to the remainder of the processing to generate training data described in the illustrative example of FIG. 2. However, in alternative configurations these operations may be performed at a later stage of processing, in particular after normalization of the PPG signal data. Additionally, the high frequency filtering and rolling average subtraction processes may be performed in any order or concurrently with parallelized computer processing systems.

[0036] The process 200 continues as the training system 104 generates normalized amplitudes in the training data based on identified peaks and valleys in the photoplethysmogram data of the selected time segments after the subtraction of the rolling average values described above (block 216). FIG. 5 depicts an example of standard and normalized PPG data in more detail. In FIG. 5, the graph 504 depicts a time series of PPG data with varying peaks and valleys that indicate changes the detected optical signal due to expansion and contraction of the volume of blood vessels in the body of one of the training users 152. During the process 200, the training system 104 divides the PPG data into predetermined time segments, such as a 60-second time segment depicted in FIG. 5, although the length of time segments may be shorter or longer. Because the blood vessels expand and contract cyclically, the PPG data also depicts cyclical peaks and valleys. As depicted in the graph 504, the absolute values of the PPG data vary over time even if the peak-to-valley amplitude of some signals in the series are quite similar to each other, such as similar peak-to-valley amplitudes at references 506 and 508 even though the absolute values of the PPG data vary between these references. The graph 516 depicts a normalized graph, which corresponds to a different segment of PPG data than is depicted in the graph 504, in which the valleys of each wave are set to a predetermined level, such as zero (0) in FIG. 5. The normalized data enables the identifications of waves in the PPG data with similar amplitudes and frequencies even if the absolute values of the waves vary in the original PPG data, and enable the identification of waves with amplitudes and frequencies that are substantially different from the other waves in a time segment, such as the wave 520. The training system 104 performs the normalization described above for the photoplethysmograms of both the first and second wavelengths in the PPG data. The process 200 utilizes the normalized PPG data for the remainder of the training process described herein. During later inferencing operations, the inferencing system 104, or another system that uses the trained ML model 126, also applies the normalization process described above to PPG inferencing data, such as the PPG inferencing data 132 depicted in FIG. 1.

[0037] Referring again to FIG. 2, the process 200 continues with the identification of waves in the normalized data based on continuous amplitude spikes (block 220). This process identifies waves in the PPG data that correspond to the aforementioned expansion/contraction cycles in blood vessels while reducing the likelihood of misidentifying signal noise from a corresponding training PPG sensor 154 as representing a physiological signal from a training user 152. Referring to FIG. 6, the graph 600 depicts the identification of wave peaks in the normalized PPG data in more detail. To identify the spike for each wave, the training system 104 processes the time-series PPG data to identify when the signal level first exceeds a predetermined level (dashed line 604) in the normalized waveform data, such as a threshold of 0.4 (on a scale of 0.0 to 1.0) in FIG. 6, although other threshold levels may be used in other embodiments. The training system 104 further identifies a peak in a series of waveform samples by observing monotonically increasing samples followed by monotonically decreasing samples, which indicate the rising and falling portions of a spike in the waveform. In the example of FIG. 6, a series of samples increases monotonically to a peak followed by a monotonic decrease toward the predetermined threshold indicated by the line 604 to identify spikes for waveform peaks 608A - 608D. The training system 104 does not identify a set of PPG data that do not include a spike matching these criteria as a wave to discard noise and false signals from the normalized PPG data. In one configuration, the training system 104 identifies at predetermined minimum number of consecutive waves, such as four (4) waves in the example of FIG. 6, in which there are no detections of false-spikes between the detected waves, and includes the normalized PPG data for the identified waves in the training data 128. In alternative configurations, the training system 104 identifies individual waves or a different number of waves in the training data. The normalization process described above enables an efficient method to identify waves in the training data set, which would be more complex and error-prone if processing the original PPG data directly. The training system 104 performs the identification of waves described above for the photoplethysmograms of both the first and second optical wavelengths in the PPG data.

[0038] Referring again to FIG. 2, the process 200 continues as the training system 104 trains the ML model 126 using the waves that are identified in the normalized training PPG data 128 (block 224). One non-limiting example of an ML model is an auto-encoder, and FIG. 7 depicts an example of an auto-encoder ML model 700 for illustrative purposes. The auto-encoder 700 includes sets of nodes that form an encoder 704, latent encoded space 712, and a decoder 716. In more detail, the encoder 704 includes an input layer 706, where each input is a onedimensional array of 100 amplitude values taken from the training PPG data 128 corresponding to at least one identified wave. The encoding layer 708 is a “hidden layer” that includes a smaller number of nodes that the input layer 706. Each node in the encoding layer 708 receives the input data and compresses the input data by applying weights to the combined inputs of the input layer 706 and uses a non-linear output activation function to generate an output value based on a combination of the inputs. While FIG. 7 depicts one intermediate layer 708 in the encoder 704, alternative configurations use multiple intermediate layers in the encoder. The latent encoded space layer 712 includes the fewest number of nodes in the auto-encoder 700 that compressed information about the most important features of the input data. Since there are fewer nodes in the latent space layer 712 compared to the input layer 706, the latent space layer 712 stores a lossy representation of the most important features in the original input data, where the training process adjusts the auto-encoder to recognize the most important features of the training data. After completion of the training process, in some configurations the compressed encoded data in the latent space layer 712 represent the output of the auto -encoder. The decoder 716 further includes at least one intermediate layer 718 that receives outputs from the nodes in the latent space layer 712 and generates multiple output values that decode the encoded information from the latent space layer 712 into an output with a greater amount of information, and the final output layer 720 of the decoder includes still more nodes that combine the outputs from the intermediate layer 718 to generate a final decoded output. As with the encoder 704, the nodes in the intermediate layer 718 and the output layer 720 of the decoder 716 include weight parameters that are generated during the training process. In the auto-encoder 700 the output layer 720 includes 100 output nodes, which matches the number of input nodes in the input layer 706, but in alternative configuration the output layer optionally includes a different number of nodes compared to the input layer.

[0039] During the training operation, the training system 104 uses a training process including a loss function and, optionally, a penalty function to provide sparsity in the nodes within the auto-encoder 700 to generate the specific numeric weight values in each of the internal layers of the encoder 704, latent space 712, and decoder 716. For many auto-encoders, the training process is unsupervised because the training goal is to reproduce an output that mimics an input to within a predetermined tolerance threshold so no additional a priori ground truth data are necessary, although supervised training techniques may be used with auto -encoders and other types of ML models. The process 200 enables the training system to provide high-quality PPG training data 128 to the ML model 126 during the training process, which in turn improves the quality of the encoded data in the latent space 712 and in the other layers of the encoder 704 and decoder 716. As such, during later inferencing operations, even if the input PPG inferencing data 132 include noise or other signal artifacts, the trained ML model 126 is properly configured to recognize the most important features in the PPG inferencing data 132 with a reduced likelihood of misinterpreting noise as important features in the PPG inferencing data 132. During inferencing operations after completion of the training process, in some configurations the auto-encoder 700 omits the use of the decoder 716 and generates an output based on the encoded data in the latent space layer 712. While an auto-encoder is one example of an ML model that may be trained with the techniques described herein, those of skill in the art will recognize that other ML models including, but not limited to, support vector machines, random forest classifiers, convolutional neural networks, recurrent neural networks, long -short term memory neural networks, transformer neural networks, Bayesian neural networks, and other forms of shallow and deep artificial neural networks that are known to the art are suitable for use with the process 200. Different forms of ML models may be employed to measure or otherwise characterize various biomarkers of interest based on inferencing PPG data. In embodiments of the training system 104 that are configured for training the ML model 126, the process 200 optionally concludes with the generation of the trained ML model 126.

[0040] Referring again to FIG. 2, the process 200 optionally continues after completion of the training process the trained ML model 126 is used to perform inferencing operations to generated information pertaining to one or more biomarkers of interest (block 228). In the configuration of FIG. 1, the system 104 acts as an inferencing system 104. During operation, the inferencing system 104 receives PPG data from the inferencing PPG sensor device systems 164 that record PPG measurements for the corresponding inferencing users 162. While of course it is possible for one or more of the training users 152 to also be amongst the inferencing users 162, in many practical situations the inferencing users 162 and the inferencing PPG sensor device systems 164 played no role in the earlier training process, but still rely on the trained ML model 126 to process the PPG inferencing data 132.

[0041] During operation, the inferencing system 104 applies the PPG inferencing data 132 to the trained ML model 126 to generate ML output data 136. In one configuration, the inferencing system 104 applies some or all of the processing steps described above with reference to blocks 208 - 220 to the inferencing data in substantially the same manner as applied to the training data in order to improve the quality of inferencing PPG data that are provided to the trained ML model 126. The ML output data 136 may be, for example, a classifier result or other output data generated by the trained ML model to indicate information about one or more biomarkers of interest. Using the auto-encoder 700 of FIG. 7 as an example, in one configuration the inferencing system 104 applies the PPG inferencing data 132 to the input layer 706 of the auto-encoder 700. The processor 108 extracts ML output data 136 from the nodes in the latent space layer 712 that encodes a compressed representation of the PPG inferencing data 132 to classify the biomarker. These inferencing output data can be compared to other encoded representations of the biomarker via, for example, a clustering algorithm or as input to another trained machine-learning model to either measure the biomarker directly or provide other analytically relevant information about the biomarker. In still another configuration, the full auto-encoder 700 produces output PPG waveforms based on the PPG inferencing data, where the output PPG waveforms are provided to a second trained ML model for either or both of further training of the second ML model or further classification using the second ML model, where the second trained ML model can be any of the ML models described above. The advantage of this configuration is that the trained auto-encoder 700 reproduces the most important features in the PPG inferencing data 132 while acting as a filter to remove signal noise and other artifacts that may be present in the original inferencing data generated by the inferencing PPG sensor device systems 164. In still other configurations, a different ML model may produce direct classifications of the biomarker in the ML output data 136. In some embodiments, the biomarker data are optionally transmitted to the HCP terminals 174, either directly or indirectly through an online electronic health record service, to enable the HCPs 172 to review the information as part of providing medical treatment to the inferencing users 162.

[0042] As described above, while the training/inferencing system 104 is capable of performing both the training and inferencing operations described herein, in alternative configurations the trained ML model 126 is transmitted to one or more additional computing devices to enable performance of the inferencing operations in a wide range of hardware and software configurations. In a first non-limiting alternative configuration, one or more separate networked server computing systems receive copies of the ML model 126 and perform the inferencing process in a similar manner to that described above. In a second non-limiting alternative configuration, the inferencing PPG sensor device systems 164 and associated computing devices receive copies of the ML model 126, which enables the individual inferencing PPG sensor device systems 164 to generate the ML output data 136.

[0043] While the embodiments described herein relate to improvements to the processing of PPG sensor data and machine learning for the classification of data relevant to one or more biomarkers, those of skill in the art will understand how these techniques are equally applicable to other forms of non-invasive sensing. For example, the techniques described herein are equally applicably to other non-invasive sensing technologies including radio frequency sensors, terahertz imaging sensors, ultrasonic sensors, and the like. In particular, techniques described herein are applicable to multi-spectral sensing technologies that produce measurement data for two or more frequencies in the electromagnetic band. Furthermore, the techniques described herein may be applied to frequency domain signals in addition to the time domain signals discussed above.

[0044] This disclosure is described in connection with what are considered to be the most practical and preferred embodiments. However, these embodiments are presented by way of illustration of the improvements to the art described herein, and these improvements are not strictly limited to the disclosed embodiments. Accordingly, one of skill in the art will realize that this disclosure encompasses all modifications and alternative arrangements within the spirit and scope of the disclosure and as set forth in the following claims.

Claims

CLAIMS What is claimed is:

1 . A method for training a machine-learning model using photoplethysmographic data, comprising: receiving, with a training system, sensor data comprising at least one photoplethysmographic (PPG) data set, the at least one PPG data set further comprising a first series of PPG data corresponding to a first optical wavelength and a second series of PPG corresponding to a second optical wavelength; generating, with the training system, a plurality of time segments in the at least one PPG data set; selecting, with the training system, only a portion of the plurality of time segments having high-frequency noise beneath a predetermined threshold for inclusion in a training data set; removing, with the training system, a rolling average value for the first series of PPG data and the second series of PPG data from one or more time segments in the training data set; identifying, with the training system, a plurality of peaks and valleys in each of the first series of PPG data and the second series of PPG data within one or more time segments in the training data set; generating, with the training system, a plurality of normalized amplitudes of the plurality of peaks and valleys corresponding to each of the first series of PPG data and the second series of PPG data within one or more time segments in the training data set; identifying, with the training system, a plurality of waves in the plurality of normalized amplitudes of the plurality of peaks and valleys corresponding to each of the first series of PPG data and the second series of PPG data within one or more time segments in the training data set, each wave being identified based on a spike in the plurality of normalized amplitudes; and generating, with the training system, a trained machine-learning model to classify data relevant to a biomarker using a training process with an untrained machine-learning model and the plurality of waves identified in the training data set.

2. The method of claim 1 wherein the selecting of the portion of the plurality of time segments for inclusion in the training data set further comprises: identifying the high-frequency noise based on a correlation between PPG data in the first series of PPG data and the second series of PPG data for each time segment, and based on a number of zero-crossings identified in the first series of PPG data and the second series of PPG data for each time segment.

3. The method of claim 2, wherein the selecting of the portion of the plurality of time segments for inclusion in the training data further comprises: identifying, with the training system, a number of zero crossings in each time segment; and selecting, with the training system, only time segments in the plurality of time segments having the number of zero crossings that is less than a first predetermined threshold to eliminate high-frequency noise and having the number of zero crossings that is greater than a second predetermined threshold to eliminate low frequency noise for inclusion in the training data set.

4. The method of claim 2, wherein the selecting of the portion of the plurality of time segments for inclusion in the training data further comprises: identifying, with the training system, a number of zero crossings in each time segment; and selecting, with the training system, only time segments in the plurality of time segments having the number of zero crossings that is greater than a first predetermined threshold to eliminate low frequency noise for inclusion in the training data set.

5. The method of claim 1 , wherein the identifying of each wave in the plurality of waves further comprises: identifying, with the training system, each spike based on a series of consecutive samples in the plurality of normalized amplitudes rising to a peak value monotonically and falling from the peak value monotonically while remaining above a predetermined threshold.

6. The method of claim 1 , wherein the first optical wavelength is in a range of approximately 500 nm to 600 nm corresponding to green light and the second optical wavelength is in a range of approximately 750 nm to 1 ,400 nm corresponding to near infrared light.

7. The method of claim 1 , wherein an auto-encoder receives the training data set to generate input training data for the training of the machine-learning model.

8. The method of claim 1 further comprising: receiving, with an inferencing system, sensor data comprising at least one inferencing PPG data set, the at least one PPG data set further comprising a first series of PPG data corresponding to the first optical wavelength and a second series of PPG data corresponding to the second optical wavelength; and performing, with the inferencing system, an inferencing process to classify the data relevant to the biomarker in the at least one inferencing PPG data set using the trained machine-learning model.

9. The method of claim 8, wherein the inferencing system is the training system.

10. The method of claim 1 further comprising: providing the trained machine-learning model to an inferencing system, the inferencing system being separate from the training system.

11. A system for training a machine-learning model using photoplethysmographic data, comprising: a memory configured to store: program instructions; sensor data comprising at least one photoplethysmographic (PPG) data set, the at least one PPG data set further comprising a first series of PPG data corresponding to a first optical wavelength and a second series of PPG corresponding to a second optical wavelength; and data corresponding to a machine-learning model; and a processor operatively connected to the memory, the processor being configured to execute the stored program instructions to: generate a plurality of time segments in the at least one PPG data set; select only a portion of the plurality of time segments having high- frequency noise beneath a predetermined threshold for inclusion in a training data set; remove a rolling average value for the first series of PPG data and the second series of PPG data from one or more time segments in the training data set; identify a plurality of peaks and valleys in each of the first series of PPG data and the second series of PPG data within one or more time segments in the training data set; generate a plurality of normalized amplitudes of the plurality of peaks and valleys corresponding to each of the first series of PPG data and the second series of PPG data within one or more time segments in the training data set; identify a plurality of waves in the plurality of normalized amplitudes of the plurality of peaks and valleys corresponding to each of the first series of PPG data and the second series of PPG data within one or more time segments in the training data set, each wave being identified based on a spike in the plurality of normalized amplitudes; and generate a trained machine-learning model stored in the memory to classify data relevant to a biomarker using a training process with an untrained machine-learning model and the plurality of waves identified in the training data set.

12. The system of claim 11 wherein the processor is further configured to: identify the high-frequency noise based on a correlation between PPG data in the first series of PPG data and the second series of PPG data for each time segment, and based on a number of zero-crossings identified in the first series of PPG data and the second series of PPG data for each time segment.

13. The system of claim 12 wherein the processor is further configured to: identify a number of zero crossings in each time segment; and select only time segments in the plurality of time segments having the number of zero crossings that is less than a first predetermined threshold to eliminate high- frequency noise and having the number of zero crossings that is greater than a second predetermined threshold to eliminate low frequency noise for inclusion in the training data set.

1 . The system of claim 12 wherein the processor is further configured to: identify a number of zero crossings in each time segment; and select only time segments in the plurality of time segments having the number of zero crossings that is greater than a first predetermined threshold to eliminate low frequency noise for inclusion in the training data set.

15. The system of claim 11 wherein the processor is further configured to: identify each spike based on a series of consecutive samples in the plurality of normalized amplitudes rising to a peak value monotonically and falling from the peak value monotonically while remaining above a predetermined threshold.

16. The system of claim 11 , wherein the first optical wavelength is in a range of approximately 500 nm to 600 nm corresponding to green light and the second optical wavelength is in a range of approximately 750 nm to 1 ,400 nm corresponding to near infrared light.

17. The system of claim 11 , wherein an auto-encoder receives the training data set to generate input training data for the training of the machine-learning model.

18. The system of claim 11 , wherein the memory is further configured to: store sensor data comprising at least one inferencing PPG data set, the at least one PPG data set further comprising a first series of PPG data corresponding to the first optical wavelength and a second series of PPG data corresponding to the second optical wavelength; and wherein the processor is further configured to execute the stored program instructions to: perform an inferencing process to classify the data relevant to the biomarker in the at least one inferencing PPG data set using the trained machine-learning model.

19. The system of claim 11 , wherein the system is further configured to: provide the trained machine-learning model to a separate inferencing system, the inferencing system being configured to use the trained machine-learning model to perform an inferencing process to classify the data relevant to the biomarker in at least one inferencing PPG data set using the trained machine-learning model.