Sample-by-sample ratio-of-ratios (ROR) for determination of peripheral oxygen saturation level

Abstract

Technologies are provided for generating an estimate of peripheral oxygen saturation level of a subject using a digital signal processing (DSP) approach that includes determining sample- by-sample ratio-of-ratios (ROR) The estimate of the peripheral oxygen saturation level can be provided to the subject or a computing device associated with the subject. The DSP approach permits efficient use of memory resources and computing resources.

Description

SAMPLE-BY- SAMPLE RATIO- OF-RATIOS (ROR) FOR DETERMINATION OF PERIPHERAL OXYGEN SATURATION LEVELCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of and priority to U.S. Provisional Patent Application No. 63 / 696,292, filed September 18, 2024, U.S. Provisional Patent Application No. 63 / 696,297, and U.S. Provisional Patent Application No. 63 / 696,305, the contents of each of which applications are hereny incorporated herein by reference in their entireties.SUMMARY

[0002] The following presents a simplified summary of one or more aspects to provide a basic understanding of such aspects. This summary is not an extensive overview of all contemplated aspects and is intended to neither identify key or critical elements of all aspects nor delineate the scope of any or all aspects. Its sole purpose is to present some concepts of one or more aspects in a simplified form as a prelude to the more detailed description that is presented later.

[0003] In some aspects, the disclosure provides a method including: for a signal sampling event, performing digital signal processing (DSP) operations comprising, obtaining a first AC component of raw photoplethysmography (PPG) signal from a first color channel; obtaining a first DC component of the raw signal from the first color channel; obtaining a second AC component of raw signal from a second color channel; obtaining a second DC component of the raw signal from the second color channel; conditioning the first AC component by applying an adapative bandpass filter based on a heart rate of a subject, resulting on a first conditioned AC component; conditioning the second AC component by applying the adaptive bandpass filter based on the heart rate of the subject, resulting on a second conditioned AC component; determining a perfusion index for the first color channel as a first ratio of the first conditioned AC component and the first DC component; determining a perfusion index for the second color channel as a second ratio of the second conditioned AC component and the second DC component; determining a value of a ratio-of-ratios (ROR) for the signal sampling event as the ratio of the first ratio and the second ratio; repeating the DSP operations for each one of multiple next signal sampling events until multiple values of the ROR are determined; determining a single value of ROR by applying an aggregation filter to the multiple values of the ROR; anddetermining, using the single value of ROR, a value of peripheral oxygen saturation (SpO2) of the subject.

[0004] In other aspects, the disclosure provides another method including: generating one or more first datasets for respective one or more wearable devices, with each dataset of the one or more first datasets including, reference ratio-of-ratios (ROR) values, wherein a first reference ROR value of the reference ROR values is defined by a reference peripheral oxygen saturation (SpO2) value and a first calibration function of a first wearable device of the one or more wearable devices, with the first calibration function mapping ROR values to reference SpO2values, and multiple first sets of signal features of conditioned photoplethysmography (PPG) signal from a first color channel, and multiple second sets of signal features of conditioned PPG signal from a second color channel; generating a second dataset as the union of the one or more first datasets; training, using the second dataset, a predictor model to determine a reference ROR corresponding to first signal features associated with the first color channel and second signal features associated with the second color channel; receiving first input PPG signal from a particular first color channel of a particular wearable device; receiving second input PPG signal from a particular second color channel of the particular wearable device; determining, using the first input PPG signal, a first input set of signal features; determining, using the second input PPG signal, a second input set of signal features; generating a second reference ROR by applying the predictor model to the first input set of signal features and the second input set of signal features; determining a value of SpO2using the second reference ROR.

[0005] In yet other aspects, the disclosure provides yet another method including: receiving first photoplethysmography (PPG) signals corresponding to multiple first measurement channels, the first PPG signals caused by excitation with light of a first wavelength; receiving second PPG signals corresponding to the multiple first measurement channels, the second PPG signals caused by excitation with light of a second wavelength; receive respective confidence levels for the multiple measurement channels; training, using the first PPG signals, the second PPG signals, and the respective confidence levels, a predictor model to determine a signal quality index of a measurement channel, with the signal quality index being indicative of a confidence level that a value of peripheral oxygen saturation (SpO2) determined using the measurement channel is an accurate value; receiving third PPG signals of a subject, with the third PPG signals corresponding to multiple second measurement channels of a wearable device and caused by excitation of tissueof the subject with light of the first wavelength; receiving fourth PPG signals of the subject, with the fourth PPG signals corresponding to the multiple second measurement channels and caused by excitation of the tissue of the subject with light of the second wavelength; generating multiple signal quality indices for respective ones of the multiple second measurement channels by applying the trained predictor model to the third PPG signals and the fourth PPG signals; generating, using the multiple signal quality indices, a ranking of the multiple second measurement channels; selecting, using the ranking, one or more particular measurement channels of the second measurement channels, with the one or more particular measurement channels having a defined placement within the ranking; and determining, using the one or more particular measurements channels, a value of SpCh of the subject.BRIEF DESCRIPTION OF THE DRAWINGS

[0006] The accompanying drawings form part of the disclosure and are incorporated into the subject specification. The drawings illustrate example aspects of the disclosure and, in conjunction with the following detailed description, serve to explain at least in part various principles, features, or aspects of the disclosure. Some aspects of the disclosure are described more fully below with reference to the accompanying drawings. However, various aspects of the disclosure can be implemented in many different forms and should not be construed as limited to the implementations set forth herein. Like numbers refer to like elements throughout.

[0007] FIG. 1 is a block diagram of an example of a wrist wearable device for non-invasive monitoring of vital signs of a subject, in accordance with one or more aspects of this disclosure.

[0008] FIG. 2 illustrates examples of wrist wearable devices having respective arrangements of light source devices and optical sensor devices, in accordance with one or more aspects of this disclosure.

[0009] FIG. 3A illustrates schematically contributions to light absorption, within a portion of a subject, of excitation light used to measure peripheral oxygen saturation (SpCh), in accordance with one or more aspects of this disclosure.

[0010] FIG. 3B depicts a DC contribution, a non-pulsatile contribution, and an AC contribution to an optical signal obtained using excitation light of wavelength X, in accordance with one or more aspects of this disclosure.

[0011] FIG. 4 illustrates molar absorption coefficients of HbO2 and RHb.

[0012] FIG. 5 presents AC and DC components of photoplethysmography (PPG) signals obtained in response to illumination of tissue of a subject with light of wavelength X, in accordance with one or more aspects of this disclosure.

[0013] FIG. 6A presents various charts illustrating results for SpCh obtained by implementing the sample-by-sample approach described herein, in a scenario involving motion artifacts, in accordance with one or more aspects of this disclosure.

[0014] FIG. 6B presents various charts illustrating results for SpCh obtained by implementing the sample-by-sample approach described herein, in another scenario not involving motion artifacts, in accordance with one or more aspects of this disclosure.

[0015] FIG. 7 illustrates an example of a method for generating an estimate of SpCh using the sample-by-sample approach described herein, in accordance with one or more aspects of this disclosure.

[0016] FIG. 8 present a chart that illustrates reference SpCh values overt time for a test subject, and another chart that illustrates calculated R values over time corresponding to the reference SpCh values for the subject, in accordance with one or more aspects of this disclosure.

[0017] FIG. 9 present a chart of reference SpCh values and calculated R values of all test subjects, in accordance with one or more aspects of this disclosure.

[0018] FIG. 10 presents schematic line plots of respective calibration curves for two hardware configurations, in accordance with one or more aspects of this disclosure.

[0019] FIG. 11 illustrates an example of the ROR predictor model, in accordance with one or more aspects of this disclosure.

[0020] FIG. 12 illustrates an example of a system to generate an estimate of SpCh based on an ML approach, in accordance with aspects of this disclosure.

[0021] FIGS. 13-15 illustrate performance of a NN-based R calculator relative to the signal processing-based approach, in accordance with one or more aspects of this disclosure.

[0022] FIG. 16 is a diagram that summarizes the multi-task learning approach that is described herein, to generate and apply an ROR predictor model in accordance with one or more aspects of this disclosure.

[0023] FIG. 17 illustrates an example of a ROR predictor model that can be used to predict R and a confidence level for the predicted R, in accordance with one or more aspects of this disclosure.

[0024] FIG. 18 illustrates another example of a system to generate an estimate of SpO2based on an ML approach, in accordance with one or more aspects of this disclosure.

[0025] FIG. 19 illustrates performance of an ML approach to determination of SpO2, in accordance with one or more aspects of this disclosure.

[0026] FIG. 20 is a schematic diagram of an example of an NN architecture of an NN-based SpO2 predictor model, in accordance with one or more aspects of this disclosure.

[0027] FIG. 21 illustrates an example of a system to generate, based on an ML-based approach, an estimate of SpO2and a confidence level / attribute of the estimate, in accordance with aspects of this disclosure.

[0028] FIG. 22 presents a flowchart of an example of a multi-channel method for determining an SpO2value, in accordance with aspects of this disclosure.

[0029] FIG. 23 presents an example output of the multi-channel method, where singlechannel SpO2 output is selected according to confidence level of the SpO2output.

[0030] FIG. 24 illustrates an example of a system to generate, based on a multi-channel ML- based approach, an estimate of SpO2, in accordance with aspects of this disclosure.

[0031] FIG. 25 presents an example of an NN architecture of the select / sort module, in accordance with aspects of this disclosure.

[0032] FIG. 26 presents an example of an NN architecture of the select / sort module and an example of an NN architecture of the SpO2predictor model, in accordance with aspects of this disclosure.

[0033] FIG. 27 presents a flowchart of an example of method for generating a selection / sorting model and determining select measurement channel(s) using the selection / sorting model, in accordance with aspects of this disclosure.

[0034] FIGS. 28A-28B illustrate various types of errors associated with ML-based predictor models described herein, in accordance with aspects of this disclosure.

[0035] FIG. 29 illustrates other types of errors associated with ML-based predictor models described herein, in accordance with aspects of this disclosure.

[0036] FIG. 30 is a block diagram of an example of a computing system that forms, or includes, a training subsystem in accordance with aspects of this discosure.DETAILED DESCRIPTION

[0037] Embodiments of this disclosure may address the issue of non-invasive measurement of SpO2in a subject. Monitoring oxygen saturation is a valuable element in the assessment of respiratory health and overall well-being the subject. One of the challenges included in pulse oximetry is the determination of ratio-of-ratios (ROR) used to determine an estimate of SpO2. Such a challenge may be more significant when photoplethysmography (PPG) signals are collected on wrist wearable devices because of low signal quality and potential large noise sources (e.g., wrist and finger motions, and respiration). Additionally, wrist PPG signals are obtained in reflective mode, resulting in subject dependent variations from population mean of the ROR to SpO2calibration curve. As an example, to address issues that may arise from motion, accelerometer signals may be used to determine if calculation of ROR is to be performed. As another example, to address potential measurement issues arising from respiration, carefully designed filtering may be used to isolate the heartbeat signal.

[0038] Embodiments of this disclosure provide devices, systems, and techniques for pulse oximetry. In accordance with aspects of this disclosure, pulse oximetry can be implemented using optical measurements and various SpO2measurement techniques that may accurately and non- invasively determine oxygen saturation levels in blood. The techniques / approaches described herein utilize optical sensor devices to obtain and analyze variations in light absorption, providing continuous or nearly continuous monitoring of SpO2levels in the bloodstream of a subject. The optical SpO2measurement techniques in accordance with this disclosure may provide a reliable and convenient solution for subjects, healthcare professionals, and clinicians. By accurately estimating oxygen saturation levels, embodiments of this disclosure may assist in detecting potential respiratory issues and / or facilitating timely medical intervention. As a result, embodiments of this disclosure may ultimately contribute to better patient outcomes and improved healthcare management. Indeed, the embodiments of this disclosure improve both wearable devices used for monitoring vital signs and the technical field of monitoring vital signs.

[0039] Embodiments of this disclosure use light source device that emit light through the skin of a subject and also use optical sensor device that can measure the amount of light absorbed by oxygenated and deoxygenated hemoglobin in the blood of the subject. The SpO2measurement techniques in accordance with aspects of this disclosure may analyze light absorption variations and may then determine SpO2levels accurately. Measurements of SpO2are non-invasive,painless, and may provide continuous or nearly continuous results. Accordingly, embodiments of this disclosure may provide a desirable solution for continuous monitoring in various healthcare settings. The SpCh measurement techniques in accordance with aspects of this disclosure may be readily integrated with wearable devices, and thus may provide a user-friendly and efficient solution for accurate SpO2 measurement.

[0040] In some aspects, to determine an estimate of SpCh level, RoR may be determined using sample-by-sample calculation corrected by a machine learning (ML) model. In addition, a ML- based confidence level may be used to quantify a degree of confidence / accuracy of an ROR determination and / or SpO2 determination. In cases where multiple measurement channels to measure PPG signals are available, ROR and / or SpO2 may be estimated using select RoRs that have a satisfactory confidence level and / or select PPG signals that have a satisfactory level of signal quality.

[0041] As is described in greater detail below, for excitation light of a defined wavelength, a PPG signal includes an AC component associated with arterial blood and a baseline component (or DC component) that is essentially stationary and is associated with non-arterial blood and other bodily tissues. Perfusion index (PI) at the defined wavelength may be defined as a ratio of the AC component to DC component of PPG signal. A conventional approach to obtaining an estimate of SpO2 includes generating two PPG signals at a location in the body of the subject. One PPG signal is generated using essentially monochromatic light in the red portion of the electromagnetic (EM) radiation spectrum, and the other PPG signal is generated using essentially monochromatic light in infrared (IR) portion of the EM radiation spectrum. The conventional approach also includes obtaining a first ratio of PPG signal intensity at peak and PPG signal intensity at trough for the red excitation light, and also obtaining a second ratio of PPG signal intensity at peak and PPG signal intensity at trough for the IR excitation light. The ratio of the first ratio and the second ration, what is referred to as the ratio of ratios (ROR), is then determined and used to generate the estimate of SpO2.

[0042] Such ROR can be approximated as the ratio of (a) a first ratio of amplitude of AC component and DC component for red excitation light and (b) a second ration of amplitude of AC component and DC component for IR excitation light. In scenarios where PPG signal have low signal-to-noise ratio (SNR) it may be challenging to identify a precise peak and trough at times. Embodiments of this disclosure may address such an issue by applying a sample-by-sampledetermination of ROR for PPG signals obtained with two monochromatic excitation lights. That is, at each signal sample (represented by the index i) a ratio-of-ratios Ri can be determined over a group of samples spanning a defined time interval. The many values of Ri obtained over the group of samples is then aggregated to determine a single ROR that is used to generated an estimate of SpO2. By aggregating the values of Ri, the estimate of SpO2becomes robust to noise artifacts or motion artifacts, or both. Such a sample-by-sample ROR approach can outperform the conventional peak-and-trough approach to ROR. Each Ri and / or an aggregated value of multiple individual values of Ri may be used as a feature in ML-based approaches in accordance with this disclosure.

[0043] Embodiments of this disclosure may use an ML-based approach that corrects and consolidates several different RORs determined based on sample-by-sample digital signal processing (DSP) as described herein. An ML-based approach may address the issue of one ROR being adequate in some cases (e.g., for some subjects) and another ROR being adequate in other cases (e.g., for other subjects). The ML-based approach permits generating an ML model that may determine, from acceptably large training datasets, the rules for when to use which ROR. Subjectdependent variation of calibration curves may be improved by correcting an estimated ROR.

[0044] A challenge presented in training a ML model as is described above is how to train a single model that can satisfactorily operate on datasets of different wearable devices having different hardware configurations and associate calibrations that are quite different. Embodiments of this disclosure may address such a challenge by determining ROR targets from the SpO2reference values for a calibration of a wearable device, and then training the ML model for a target in ROR rather than in SpO2value. It is noted that ROR targets are formal instruments, not observable quantities, that permit training an ML model in accordance with aspects of this disclosure. Training in this disclosure may be based on ROR features and / or other hand-crafted features that encode signal morphologies and signal qualities and beat consistencies. Examples of ROR features include signal-to-noise ratio (SNR), peak-to-peak amplitude changes, peak-to-peak intervals, trough-to-trough intervals, trough-to-through amplitude changes, and other signal attributes that quantify signal quality. Because a digital PPG signal is a series of sample values, other examples of ROR features include statistical features of at least a portion of the series of sample values, such as mean, variance, skewness, kurtosis, and the like. The statistical features can be determined for various quantities obtainable from the digital PPG signal, such as peakamplitude, trough amplitude, and the like. Such a portion of series of sample values can be defined by a rolling window of time. Other features can include include functions of statistical features, such as ratios of statistical features. For purposes of illustration, consistency of a beat refers to degree of regularity of heart beats over a defined time interval. Further, a handcrafted feature refers to a feature that has been manually designed a priori in order to characterize an aspect of a dataset. The disclosure is not limited to handcrafted features and, in some cases, non-handcrafted features may be used. An example of a non-handcrafted feature is raw signal. It is noted that for at least some features, different values of a feaure can be determined by calculating the feature in a particular time interval containing a particular number of signal samples. In some cases, a particular set of features can be configured or otherwise selected to train a ML model based on computational efficiency afforded by that particular set of features during the training of the ML model and / or application of the ML model. Such a flexibility contributes to the various improvements to the technical field of monitoring vital signs, as afforded by the ML approaches in accordance with aspects of this disclosure.

[0045] Various types of ML models may be used as an ROR predictor model or SpCh predictor model. Examples of ML models that can be used include artificial neural networks, such as multilayer perceptron (MLP) models, long short-term memory (LSTM) models, transformer models, State Space models, and convolution neural network (CNN) models. Regardless of type of ML model, the ML-based approach described herein can decrease errors made in determining RORs from digital signal processing. As such the ML-based approach can improve the technical field of monitoring of vital signs of a subject using a wearable device.

[0046] In the ML-based approach, RoR correction and confidence classification may be trained jointly, using a multi-task learning approach. A training loss function can be configured as a combination of a regression loss function and a classification loss function. Configuring the learning task in this way reduces the numbers of model weights in the predictor model being trained while maintaining good accuracy for predictions of RoR and confidence level. Further, in an aspect of the ML-based approach, a classification target may be set to true if an absolute difference between an RoR estimate and an RoR target is within a defined threshold value. Rather than generating training data for the confidence classification task before training the predictor model, the training data may be generated while the predictor model is being trained for the RoR correction task.

[0047] Multi-channel selection for different watch design and tissue diversity Due to tissue diversity on a location of the body of a subject and how a wearable device is worn, a probe positioned at different locations may yield optical signals that differ from one another in terms of signal quality. For example, tissue diversity in the wrist and how a wrist worn device is placed on the wrist, not all probe positions yield equally good signal qualities. Such a variation in signal quality across positions in a location of the body may be exacerbated in situations where the probe is positioned directly on top of a vein. Variations in signal quality across sensor positions may adversely affect the performance of a technique to estimate SpCh.

[0048] Embodiments of this disclosure may solve the issue of signal quality variation across probe positions by providing a wearable device having a hardware configuration that includes multiple measurement channels, and also providing and a ML-based MC selector module to select dynamically one or more measurement channels to use for SpCh estimation. Such a wearable device may be referred to as multiple-channel (MC) wearable device or multiple-channel hardware. The MC selector module may be a computationally lightweight version of a ML-based confidence classifier in accordance with aspects described herein, where the MC selector module may predict a confidence level without calculating all the DSP handcrafted features used by the ML-based confidence classifier and also without executing the full ML-based confidence classifier.

[0049] More specifically, embodiments of this disclosure provide a signal-quality predictor model that may be trained using a confidence level generated by an SpCh predictor model as a signal quality label for quality of an optical signal. The MC selector module can apply the signalquality predictor model to optical signal from each (or at least a subset) of the measurement channels present in a wearable device for measurement of SpCh. In response to applying the signal-quality predictor model, the MC selector module can rank the measurement channels according to signal quality, resulting in a quality ranking of the measurement channels. The MC selector module can then select, based on the quality ranking, one or more satisfactory measurement channels for SpCh measurement. In some cases, the MC selector module selects the best measurement channel for SpCh measurement. In other cases, the MC selector module selects both the best measurement channel and the second best measurement channel for SpCh measurement. The MC selector module may be executed at regular time intervals or according to a schedule in order to update a channel selection.

[0050] Having identified select measurement channel(s), embodiments of this disclosure may apply an SpO2predictor model to generate a respective SpO2estimate for each one of the select measurement channel(s). As a result, one or more SpO2estimates are predicted. Embodiments of the disclosure may then combine the SpO2estimate(s) to generate a single estimate of SpO2that represents a measurement of SpO2in the MC wearable device. The SpO2estimate(s) may be combined based on respective confidence levels of the SpO2estimate(s).

[0051] The use of an MC wearable device and MC selector module in accordance with this disclosure provides various improvements over commonplace devices and techniques to measure SpO2. An improvement includes increase in coverage. For purposes of illustration, coverage is the ratio of a time interval Atrwith SpO2estimation / report and a total measurement duration Atm. For example, for a total measurement duration of Atm = 200 s and an SpO2estimation process that yields 150 output values at 1 Hz, that is, Atr= 150 s, then the coverage is Atr / Atm= 150 s / 200 s = 0.75 or 75%. An additional, or alternative, improvement includes decreased SpO2estimation error.

[0052] Further, the MC selector module described herein differs from simple signal quality assessment modules. The MC selector module involves an algorithm-specific channel selector that applies a ML model (e.g., the signal quality predictor model) for distillation of DSP features. Having distilled the appropriate DSP features, embodiments of this disclosure may apply an SpO2predictor model in accordance with aspects of this disclosure. In sharp contrast to commonplace oximetry techniques, by combining the MC selector module and an SpO2evaluation module that applies an SpO2predictor model, embodiments of this disclosure provide an end-to-end ML-based approach to multiple-channel oximetry.

[0053] Although various aspects of the technologies disclosed herein are described with reference to a wrist wearable device, the technologies described herein are not limited in that respects. Indeed, the principles and practical applications of this disclosure can be implemented in other types of wearable devices. The wearable devices include patches, necklaces, neck strap devices, an arm strap device, chest strap devices, earbuds, rings, headbands, and similar devices.

[0054] FIG. 1 is a block diagram of an example of a wrist wearable device for non-invasive monitoring of vital signs of a subject, in accordance with one or more aspects of this disclosure. The vital signs include SpO2. The example wrist wearable device 110 may be affixed to a wrist of a subject 104. The example wrist wearable device 110 can perform optical measurements of SpO2. As such, the example wrist wearable device 110 includes multiple light source devices 114and multiple sensor devices 118 including multiple optical sensor devices 120. The multiple sensor device 118 also may include and one or more inertial sensor devices 122. An inertial sensor may be, or may include, a three-axis accelerometer that can generate acceleration signals corresponding to movement of the subject 104, such as movement of a hand or chest wall of the subject.

[0055] The multiple light source devices 114 and the multiple optical sensors 120 may be assembled in a section of the wrist wearable device 110 that faces the skin of the subject 104. The multiple light sources devices 114 and the multiple optical sensor 118 may be assembled according to a particular layout. In some cases, the multiple light source device 114 may include multiple light emitting diodes (LED). A first LED of the multiple LEDs is configured to emit light having a first wavelength within a first portion of the electromagnetic (EM) radiation spectrum. The first wavelength may be within the red portion of the EM radiation spectrum, for example. In addition, a second LED of the multiple LEDs is configured to emit light having a second wavelength within a second portion of the EM radiation spectrum. The second wavelength may be within the infrared (IR) portion of the EM radiation spectrum.

[0056] A pair formed by a light source device (LED) and an optical sensor device (also referred to as a photodetector) may be referred to as a “color channel.” A color channel may be characterized by the wavelength of light emitted by the light source device. As an example, a color channel that includes a light source device that emits in the red portion of the EM radiation spectrum may be referred to as a “red channel.” Similarly, also as an example, a channel that includes a light source device that emits in the IR portion of the EM radiation spectrum may be referred to as an “IR channel.”

[0057] Optical measurement of SpO2 may be conducted with optical signals corresponding to a first color channel and a second color channel. The first color channel and the second color channel form a measurement channel. For example, an optical measurement of SpO2 may be conducted with optical signals corresponding to an IR channel and a red channel. The IR channel and the red channel form a measurement channel.

[0058] The wrist wearable device 110 also includes one or more processors 124 functionally coupled with the sensor devices 118. The processor(s) 124, individually or in combination, can operate on optical signals received by a measurement channel. To that end, the processor(s) 124 can be functionally coupled with one or more memory devices 126 having stored thereonprocessor-accessible instructions 128. A processor accessible instruction is program code that may be accessed and / or executed by a processor. The processor-accessible instructions 128 may be arranged in modules in accordance with aspects of this disclosure. By executing the processor- accessible instructions 128, the processor(s) 124 can implement one or more of the various techniques / approaches described herein. The memory device(s) 126 also can include data 132 including various data and parameters in accordance with aspects of this disclosure.

[0059] The wrist wearable device 110 also can include I / O interfaces 130. An I / O interface of the I / O interfaces 130 can permit reporting data indictive of SpO2 value to another device that is external to the wrist wearable device 110. For example, the other device may be a display device. It some cases, the wrist wearable device 110 has a display device integrated thereon and another I / O interface may supply the data indictive of SpO2 values to the integrated display device.

[0060] The display device that is integrated into the wrist wearable device 110 is an example of a feedback device 140 that is may be part of the wrist wearable device 110 in some cases. The feeback device 140 can provide SpO2 values determined by the wrist wearable device 110 by implementing the techniques / approaches described herein. More specifically, to provide a SpO2 value, the wrist wearable device 110 can cause the feedback device 140 to present one or more types of indicia in response to a determination of an SpCh value of the subject 104. In addition, or in some cases, the wrist wearable device 110 can cause the feedback device 140 to present indicia indicative of an SpCh value of the subject 104 or a group of SpCh values of the subject satisfying an alert condition or another type of rule. In one example, the wrist wearable device 110 can cause the feedback device 140 to present indicia in response to the SpCh value of the subject 104 being less than a defined threshold value, where the indicia can be conspicuously presented and, thus, convey to the SpCh status to the subject. In another example, the wrist wearable device 110 can cause the feedback device 140 to present other indicia in response to a group of SpO2values satisfy or present a defined trend, e.g., downward trend, or another type of temporal dependence during a defined period of time. For example, the period of time may extend from a present time (or a recent time) to a defined past time. In such cases, the wearable device 110 of the subject 104 being less than a defined threshold value, where the indicia can be conspicuously presented and, thus, convey to the SpCh status to the subject.

[0061] The feedback device 140 may in some cases include other devices besides a display device. In other cases, the feedback device 140 may provide feedback on SpO2 status withoutreliance on a display device. In such cases, the wearble wrist device 110 may be used by, and benefit, a wider portion of a population monitoring SpO2. Accordingly, the wrist wearable device 110 can therefore provide value(s) of SpO2of the subject 104 by causing the feedback device 140 to present at least one of the value(s) of SpO2by one or more of displaying, via the feedbac device 140, visual indicia indicative of the value, outputting, via the feedback device, aural indicia indicative of the value, or providing, via the feedbacl device 140, a haptic stimulus representative of the value.

[0062] Regardless of the particular alarm condition and / or the type of indicia presented to provide SpO2 value(s) determined by the wrist wearable device 110, the indicia provide immediate access to a current SpO2status of the subject 104. Hence, improving both the monitoring of SpO2 relative to other types of commonplace devices that lack such a contemporaneous alerting and such commonplace devices.

[0063] The wrist wearable device 110 includes the I / O interfaces 130 and, in some cases, one or more network adapter(s) 134 that permit sending and / or receiving data and signaling with devices that are external to the wrist wearable device 110. The data and signaling may be sent and / or received wirelessly, via a radio unit included in the network adapter(s) 134. In addition, or as an alternative, the data and signaling may be sent and / or received via a wireline connection.

[0064] Thus, the wrist wearable device 110 can direct or otherwise cause a computing device or an apparatus that includes a computing device to perform an action in response to a value of SpO2or group of values of SpO2satisfying an alarm condition as described herein. For example, in some cases, the wrist wearable device 110 can be communicatively coupled (wirelessly or otherwise). The wrist wearable device 110 can then can cause or direct a mobile device (such as a smartphone; not depited in FIG. 1) to establish a call session with a mobile device or stationary device of a caregiver of the subject 104 in response to SpO2value(s) of the subject satisfying the alarm condition. In another example, the wirst wearable device 110 may be communicatively coupled to a wireless speaker device. The wrist wearable device 110 can then can cause or direct the wireless speaker device to emit a sound (steady or intermittent) or present another type of aural indicia (such as recorded utterance) in response to SpO2value(s) of the subject satisfying the alarm condition. The sound and the other aural indicia, individually or in combination, may convey an anomalous level (e.g., an unhealty level) of SpO2for the subject 104, for example.

[0065] In other scenarios, the wirst wearable device 110 may be communicatively coupled to a computing device (mobile or stationary (such as personal computer); not depicted in FIG. 1) via one or more networks (wireless and / or wireline) or another type of communication architecture. In one example, the computing device can be operated, owned, or leased, by a caregiver of the subject 104. The wrist wearable device 110 can then can cause or direct the computing device to present visual indicia, aural indicia, and / or haptic indicia in response to SpO2value(s) of the subject 104 satisfying the alarm condition. The sound and the other aural indicia, individually or in combination, may convey an anomalous level (e.g., an unhealty level) of SpO2for the subject 104, for example. In another example, the computing device can be operated, owned, or leased, by the subject 104. The wrist wearable device 110 canthen can cause or direct the computing device to compose an electronic message (e.g., an email, text message, or instant message) and then send the electronic message to another computing device of a caregiver of the subject 104.

[0066] The wrist wearable device 110 may be communicatively coupled with a computing device of a hospital or another type of healthcare facility, via one or more networks (wireless and / or wireline). The wrist wearable device 110 can then send data indicative of SpO2value(s) and an instruction to cause the computing device to update an electronic medical record of the subject 104. The data may be sent in nearly real-time, as the wrist wearable device 110 determines SpO2value(s). In addition, or an alternative, the data may be sent in batches at defined times (e.g., according to a schedule or periodically) or in response to a defined event. Hence, the wrist wearable device 110 and the techniques / approaches to determine SpO2can improve the monitoring of SpO2relative to other types of commonplace devices.

[0067] FIG. 2 illustrates examples of wrist wearable devices having respective layouts of light source devices and optical sensor devices (e.g., photodiode devices), in accordance with one or more aspects of this disclosure. The example wrist wearable device 200 has a first optical sensor device 210(1) and a second optical sensor device 210(2), a red LED 220, and an IRLED 230. The two optical sensor devices 210(1) and 220(2), the IR LED 230, and red LED 220 form two different measurement channels (also referred to as SpO2 channels). In some cases, each one of the optical sensor device 210(1) and the optical sensor device 210(2) is a photodiode. The first optical sensor device 210(1), the second optical sensor device 210(2), the red LED 220, and the IR LED 230 can be contained within a mesa 240 that protrudes from a surface 244 of the examplewrist wearable device 200. The surface 244 is a skin-facing surface upon placement of the example wrist wearable device 200 on a subject (e.g., subject 104 (FIG. 1). The two types of light source devices in the example wrist wearable device 200 are LEDs simply for the sake of illustration. The disclosure is not limited in that respect and other types of light souce devices may be used in the example wrist wearable device 200 or any other wearable devices in accordance with aspects of this disclosure.

[0068] The example wrist wearable device 250 has six optical sensor devices, three IRLEDs, and three red LEDs. The six optical sensor devices include a first optical sensor device 260(1), second optical sensor device 260(2), a third optical sensor device 260(3), a fourth optical sensor device 260(4), a fifth optical sensor device 260(5), and a sixth optical sensor device 260(6). The three red LEDs include a first red LED 270(1), a second red LED 270(2), and a third red LED 270(3). The three IR LEDs include a first IR LED 280(1), a second IR LED 280(2), and a third IR LED 280(3). The number of measurement channels is significantly greater than the number of measurement channels present in the example wrist wearable device 200. Indeed, there are 18 different measurement channels (also referred to as SpO2 channels) in the example wrist wearable device 250. In some cases, each one of the optical sensor devices 260(1 )-260(6) is a photodiode. The optical sensor devices 260(1 )-260(6) can be contained within a mesa 294 that protrudes from a surface 294 of the example wrist wearable device 250. The surface 294 is a skin-facing surface upon placement of the example wrist wearable device 250 on a subject (e.g., subject 104 (FIG. 1). The first red LED 270(1), the second red LED 270(2), the third red LED 270(3) also can be contained within the mesa 294. The first IR LED 280(1), the second IR LED 280(2), and the third IR LED 280(3) also can be contained within the mesa 294. The two types of light source devices in the example wrist wearable device 250 are LEDs simply for the sake of illustration. The disclosure is not limited in that respect and other types of light souce devices may be used in the example wrist wearable device 250 or any other wearable devices in accordance with aspects of this disclosure.

[0069] Each one of the example wrist wearable device 200 and the example wrist wearable device 250 can be embody or can include the wrist wearable device 110 (FIG. 1) and the elements included therein.

[0070] Regardless of the number of measurement channels in a wrist wearable device, SpCh can be measured using a single measurement channel. In some configurations, a measurement ofSpO2may be accomplished by digital signal processing of an IR photoplethysmography (PPG) signal and a red PPG signal. The IR PPG signal and the red PPG signal corresponding to a IR channel and a red channel, respectively.

[0071] More specifically, a PPG signal results from absorption and modulation of excitation light of a defined wavelength A by pulsatile arterial blood of a subject. Examples of the defined wavelength include 660 nm or 880 nm. The excitation light is absorbed by non-pulsatile arterial blood, non-arterial blood, other tissues, water, and other bodily elements within a section of a subject. Diagram 300 in FIG. 3A schematically illustrates contributions to absorption of excitation light Io within a portion of a subject. Diagram 350 in FIG. 3A illustrates PPG for two example wavelengths, a first wavelength within the red portion of the electromagnetic (EM) radiation spectrum (e.g., 660 nm, corresponding to trace 360) and a second wavelength within the infrared (IR) portion of the EM radiation spectrum (e.g., 880 nm, corresponding to trace 370.

[0072] Without intending to be bound by theory and / or modeling, the light that is detected by an optical sensor device after the excitation light has been absorbed within a portion of a subject causes a time-dependent signal IA(t) = BA+ 7V (t) +Here, V (t) is a time- dependent contribution arising from non-pulsatile arterial blood; BAis a baseline contribution arising from non-arterial blood, other tissues (including bone), water, and other bodily liquids; andP (t) is the pulsatile arterial contribution arising from cyclic or nearly cyclic flow of arterial blood. The baseline contribution and the pulsatile contribution may be referred to as DC component and AC component, respectively. Simply as an illustration, FIG. 3B depicts a DC component (panel (a)), a non-pulsatile contribution (panel (b)), and an AC component (panel (c)) to an optical signal obtained using excitation light of wavelength X. FIG. 3B also depicts the total time- dependent signals (panel (d)) and combination of DC and AC contributions (panel (e)).

[0073] The DC component and AC component of the received optical signals are different for different light wavelengths. This is due to the different absorption characteristics of HbO2, RHb, and other tissues for different light wavelengths. FIG. 4 illustrates molar absorption coefficients of HbO2(trace 410) and RHb (trace 420). For reference, FIG. 4 includes a line 430 at ordinate value corresponding to 660 nm (red light) and a line 440 corresponding to 880 nm (IR light).

[0074] The ratio of the AC component to the DC component is referred to as perfusion index (PI). The ratio of PI obtained for a first excitation wavelength and PI obtained for a secondexcitation wavelength may be used to determine a value of SpCh, as is outlined below, in accordance with this disclosure.

[0075] Hemoglobin (Hb) is an oxygen-transport protein in red blood cells (RBCs). The two main forms of Hb present in blood are oxygenated hemoglobin (oxy-hemoglobin, HbCL) and deoxygenated hemoglobin (deoxy-hemoglobin, RHb). Peripheral oxygen saturation (SpCh) is a measure of oxygen in arterial blood, which is described as a percentage of the amount of oxyhemoglobin to total hemoglobin. SpCh can be expressed as follows:where HbCL] and C[RHb] are the concentrations of HbCh and RHb, respectively.

[0076] Beer-Lambert law can be relied upon to model the optical absorption of light that occurs in response to transmitting light into a portion of the body of a subject. Beer-Lambert law describes the attenuation of light in terms of properties of the medium through which the light is propagating. According to the Beer-Lambert law,where I is the light intensity received after traversing the medium, Io is the intensity of the incident light (or excitation light), and A(A) is an attenuation defined as:Here, s(A) is the model extinction coefficient at wavelength A, C is the concentration of material, and d(A) is the optical path length for the light of wavelength A.

[0077] Considering the molecular compounds of tissue, Beer-Lambert law can be extended as follows:where the acronym NPBT stands for “non-pulsatile blood and tissue.”

[0078] To measure SpCh, two light source devices (e.g., LEDs) configured to emit light at different wavelengths may be used. These two wavelengths may be selected such that the molar absorption coefficients of HbCL and RHb are well spectrally separated. In some cases, a red LED that emits light at 660 nm and an infrared LED that emits at 880 nm are used in pulse oximetry to measure SpCh.

[0079] More specifically, with respect to FIG. 1, the wrist wearable device 110 can generate two separate PPG signals corresponding to excitation with light of respective wavelengths. A PPG optical signal that is generated is indicative of light absorbed by tissue of a subject in response to excitation light emitted by a first light source device (e.g., an LED) emitting light in a first portion of the EM radiation spectrum. A second PPG signal that is generated is indicative of light absorbed by the tissue of the subject in response to excitation light emitted by a second light source device (e.g., another LED) transmitting light in a second portion of the EM radiation spectrum. In some cases, the first portion and second portion of the EM radiation spectrum correspond to a red portion and an IR portion of the EM radiation spectrum, respectively.

[0080] For each one of the first and second PPG signals, the wrist wearable device 110 can detect peak and trough locations. The wrist wearable device 110 can determine the logarithm of a ratio of a PPG signal at peak and trough locations in order to remove the effect of absorbents different from arterial blood (e.g., non-pulsatile blood and tissue elements) as shown below. Specifically, in cases a red channel and an IR channel are used to measure SpCh, the intensity of PPG signals for the red channel can be expressed as follows:.(Red)TroughEqs. (5) and (6) result in the following attenuation for the red channel:

[0081] Similarly, for the IR channel,

[0082] A value of SpCh can be approximated as,SpO2« a x R2+ b x R + c (9) where R (also referred to herein as “ROR”) is a ratio-of-ratios that is defined as,and a, b, and c are calibration coefficients. These calibration coefficients are specific to the particular opto-mechanical arrangement of light source devices, optical sensor devices, and structure of a wearable device used to measure SpCh. For example, the calibration parameters a, b, and c depend on distance between light source device and optical sensor device (e.g., LED-PD distance), wavelengths of light emitted by respective light source devices, optical shield (or window) overlaying light source devices and optical sensor devices. Although a quadratic function on R is shown in Eq. (9), the disclosure is not limited to such a relationship between SpCh value and R. Indeed, other functions of R may be used to approximate or otherwise define SpCh values. Each one of those other functions may be referred to as a calibration curve and has at least one calibration parameter.

[0083] The approximation in Eq. (10) is used to compensate unknown wavelength- dependent path lengths, d(Red) and d(IR), and / or other non-idealities in a model for light-tissue interactions, such as geometry related optical scattering parameter(s) of tissue. An example of such a model includes Eq. (5) and Eq. (6) above.

[0084] Because of the cyclic nature of arterial blood, a single value of R can be obtained in a peak-and-trough approach to determining SpCh.

[0085] This disclosure, in one aspect, recognizes that a sample-by-sample determination of R also permits removing contributions from non-pulsatile contributions to light absorption. Thus, a determination of R, for a signal sample (represented by the index z) over a defined time interval spanning N samples may be used to determine an estimate of SpCh. Because multiples values {Ri} are used, the determination of SpO2 is more immune to noise and motion artifacts compared to a determination of SpCh based on peak-and-trough identification. Further, implementation of peak- and-trough identification may be more computationally intensive in terms of processor clock cycles and memory storage, than a determination of R, over a group of samples. Indeed, the amount of computing time involved in a determination of SpCh using a sample-by-sample approach may be about an order of magnitude less than the computing time involved in the determination of SpCh using peak-and-trough identification.

[0086] With reference to FIG. 1, in the sample-by-sample approach in accordance with this disclosure, the wrist wearable device 110 can generate an PPG signal sample <Ji (represented by the index z) in response to illumination of tissue of the subject 104 with light of wavelength X. The wrist wearable device 110 can generate a DC (or baseline) component DC^ and an AC component AC ) by applying, for example, a low-pass filter to the PPG signal sample Qi. Thus, for a PPG signal obtained with a light source device that emits in the IR portion of the EM radiation spectrum, the AC component and the DC (or baseline) component for sample Qi can be represented by AC^ and DC^, respectively. Similarly, for another PPG signal obtained using a light source device that emits in the red portion of the EM radiation spectrum, the AC component and the DC component can be represented by AC^dand DC^d, respectively.

[0087] Further, the wrist wearable device 110 can condition the AC component of the optical signal for the red channel by applying an adaptive bandpass filter based on heart rate (HR) frequency, resulting in a conditioned AC component The wrist wearable device 110 alsocan condition the AC component of the optical signal for the IR channel by applying the adaptive bandpass filter based on the HR frequency, resulting in a conditioned AC component AC^. For example, the adaptive bandpass filter may have a cutoff frequency that is equal to the HR frequency. The wrist wearable device 110 can determine the HR frequency. By applying the adaptive bandpass filter, non-pulsatile and tissue contributions to a PPG signal may be removed. Thus, the conditioned AC components described herein correspond to pulsatile contributions to the PPG signal.

[0088] The wrist wearable device 110 can then determine a sample-dependent perfusion index PIflgd f°rthe red channel as a first ratio of the conditioned AC component and the DC component of the red channel. Namely, PI^=Further, the wrist wearable device 110 can then determine a sample-dependent perfusion index PI^ for the IR channel. Namely,AC^ / DC^. The wrist wearable device 110 can determine a ratio-of-ratios R, for sample <Ji as the ratio of the first ratio and the second ratio: R[ = PIRBCI / ^R ■

[0089] Simply as an illustration, FIG. 5 presents AC and DC components of PPG signals obtained in response to illumination of tissue of a subject with light of wavelength A. Panel 510 in FIG. 5 depicts five values of R, Rt, Ri+i, Rt+2, Ri+s, andR;+4, for respective samples.

[0090] Back to referring to FIG. 1, wrist wearable device 110 can determine a single value of ROR by applying an aggregation filter to the set {Ri, R2, ... RN} of Rt for samples i = 1, 2, ... N. In some cases, the aggregation filter is a median filter with a defined length (10 seconds, for example). The number of A is then determined by the sampling rate (e.g., 25 Hz) used by the wrist wearable device 110 to generate an optical signal and the defined length. The disclosure is not limited to the aggregation filter being a median filter. Other types of aggregation filters can be used, such as an average filter.

[0091] The wrist wearable device 110 can determine a value of SpO2 via Eq. (9), using the single value of ROR obtained for N samples as the value of R, and calibration parameters a, b, and c for the wrist wearable device 110.

[0092] FIG. 6A presents various charts illustrating results for SpO2 obtained by implementing the sample-by-sample approach described herein, in a scenario involving motion artifacts. Double-headed arrows indicate time intervals where motion artifacts are presents. FIG. 6B 1presents various charts illustrating results for SpCh obtained by implementing the sample-by- sample approach described herein, in another scenario not involving motion artifacts.

[0093] In FIG. 6A, for red light, line plot 610a illustrates band filtered PPG signal (that is, AC signal), line plot 610b illustrates DC PPG signal, and line plot 610c illustrates raw PPG signal (which is the combination of AC and DC signals), each of the signal are illustrated as respective time series. For IR light, line plot 620a illustrates band filtered PPG signal, line plot 620b illustrates DC PPG signal, and line plot 620c illustrates raw PPG signal (which is the combination of AC and DC signals), each of the signals also are illustrated as respective time series.

[0094] In FIG. 6B, for red light, line plot 650a illustrates band filtered PPG signal (that is, AC signal), line plot 650b illustrates DC PPG signal, and line plot 650c illustrates raw PPG signal (which is the combination of AC and DC signals), each of the signal are illustrated as respective time series. For IR light, line plot 660a illustrates band filtered PPG signal, line plot 660b illustrates DC PPG signal, and line plot 660c illustrates raw PPG signal (which is the combination of AC and DC signals), each of the signals also are illustrated as respective time series.

[0095] FIG. 6A and FIG. 6B also include, respectively, a scatter plot 630 of AC / DC of red vs IR for the samples in the time series in the presence of motion artifacts and a scatter plot 670 of AC / DC of red vs IR for the samples in the time series in the absence of motion artifacts. It can be gleaned from the scatter plot 670 that, in the absence of motion artifacts, the samples form almost a line going through the origin and with a slope essentially equal to the ratio of ratios (ROR) calculated from a reference SpCh device. However, in the presence of motion artifacts (FIG. 6A), there are many samples outlying from the RoR slope (see line 634 in FIG. 6A).

[0096] FIG. 6A and FIG. 6B also include, respectively, (i) a histogram 640 of the slope of lines formed by each of the samples to origin, in the absence of motion artifacts, and (ii) a histogram 680 of the slope of lines formed by each of the samples to origin, in the absence of motion artifacts.

[0097] Comparing “with motion artifacts” and “without motion artifacts” scenarios, we see that both form a respective unimodal bell-shaped distribution. Determining the mode of the distribution is similar to determining the median of the samples used to generate the histogram. The mode is consistent with the ratio of ratios (ROR (also referred to as / ?)) calculated from a reference SpO2 device; represented by a solid circle near the peak of the distributions shown in the historgrams. Thus, embodiments of this disclosure provide a sample-by-sample approach tocalculate the RoR that is robust to small motions. Such an approach, and associated robustness, provide a clear improvement to the field of monitoring SpCh and other vital signs of a subject. Further it is noted that the sample-by-sample approach in accordance with aspects of this disclosure also can create efficiencies in the memory footprint of the wearable device that implements the sample-by-sample approach. Use of computational resources also may be improved. In other words, the sample-by-sample approach and the wearable devices in accordance with this disclosure improve both commonplace wearable devices (and also non-wearable devices) for monitoring SpCh and the technical field of monitoring vital signs.

[0098] FIG. 7 illustrates an example of a method for generating an estimate of SpCh using the sample-by-sample approach described herein. The example method 700 is presented in pseudocode, simply for the sake of illustration and clarity. A wearable device can implement the example method 700. To that end, the wearable device has one or more processors, one or more memory devices, and / or other computing resources. The wearable device may be a wrist wearable device or another type of wearable device, such as a patch, a necklace, a neck strap device, an arm strap device, a chest strap device, an earbud, a ring, a headband, or similar device. Regardless of type and / or form factor, the wearable device includes the elements of the example wrist wearable device 110. Indeed, in some cases, the wearable device that implements the example method 700 is the wrist wearable device 110 (FIG. 1).

[0099] As is described herein, in cases where the aggregation filter in the example method 700 yields the median of the set of sample, the implementation of the example method 700 can be interpreted in terms of the descriptin hereinbefore of FIG. 6A and FIG. 6B. Specifically, the wearable device (e.g., wrist wearable device 110 (FIG. 1)) that implements the example method 700 can determine, for each sample within a set of signal samples in the time series of PPG signals, the AC / DC (the perfusion index) of red channel vs IR channel. As noted in FIG. 7, the set of signal samples can have N samples. The wearable device, for each sample within the set of signal samples, can then determine a slope of a line formed by the sample to an origin (as discussed above in connection with FIG. 6A and FIG. 6B). In other words, as is described in FIG. 7, the wearable device can determine the / ? / for each sample i. Such a determination can yield A slopes. The wearable device can then determine a distribution (or histogram) of the slopes (see FIG. 6B, for example). Because determining the median of the slopes (or R, samples used to determine the histogram) is similar to determining the median of the slopes, the application of the aggregationfilter that applies the median can yield a single value of ROR that is robust against motion artifacts. As is described in FIG. 7, the wearable device can then determine an estimate (or value) of SpO2. In cases the wearable device is mounted to the subject 104, the estimate of SpO2 corresponds to the subject 104.

[0100] As is described herein, after obtaining a value of ROR, a wearable device can generate an estimate of SpO2using the value of ROR and a group of calibration parameters corresponding to the wearable device. To determine the group of calibration parameters (e.g., a, b, and c in Eq. (9)) a dataset of R values and reference SpO2values may be collected in a hypoxia laboratory. In the hypoxia laboratory, SpO2level of the test subjects are varied in a controlled manner and PPG signals from the test subjects are measured and recorded by the wearable device for which a calibration is being determined. During data collection, the test subjects use a gas mask to control their SpO2level. Through the gas mask, the blood oxygen content is reduced incrementally by changing the oxygen level of the test subject from 100% SpO2and lowering to 70% SpO2. The measured and recorded PPG signals may be used to determine R values for the reference SpO2values. An R value for a reference SpO2level may be determined using a peak-and-trough approach, for example. FIG. 8 presents a chart 800 that illustrates reference SpO2values over time for a test subject. FIG. 8 also presents a chart 850 that illustrates calculated R values over time corresponding to the reference SpO2values for the subject.

[0101] Reference SpO2values and calculated R values of all test subjects utilized during calibration may be aggregated as is shown in the scatter plot in FIG. 9. A polynomial curve, such as second- order polynomial curve or a first- order polynomial curve is fitted to the aggregated data in order to obtain the calibration parameters (e.g., a, b, and c in Eq. (9)) for use in a determination of SpO2based on observed values of R when the wearable device is in use. A line 910 and a line 920 corresponding, respectively, to a quadratic fit and a linear fit are also shown in FIG. 9.

[0102] Although signal processing-based R determinations described herein provide satisfactory performance, there may be situations where accuracy decreases due one or more factors, such as low optical signal quality. Because SpO2measurements may be performed using various types of wearable devices that obtain PPG signals from various body locations on a subject, such as wrist, chest, neck, upper arm, in-ear, forehead, and the like, PPG signals may have low signal quality due to low blood perfusion in some of those locations, unlike fingertips and ear lobes.

[0103] Embodiments of this disclosure may address the issue of low signal quality and, as a result, may improve SpCh measurement accuracy for a wide range of body locations. To that end, embodiments of this disclosure use machine learning (ML) techniques in addition to the signal processing-based approach described herein. Numerous optical signal features may be used so as to generate a ML model that can learn not to make errors that may be present in signal processingbased R determinations. As such, the ML model may be viewed as a corrector for signal processing-based R determinations.

[0104] One challenge posed by an ML training technique is the identification of a suitable reference / label for a ML model. More specifically, it is noted that in situations where the same PPG signals are obtained from two different wearable devices (or hardware configurations thereof) so that their respective calculated R values and other features are the same, respective estimates of SpO2 for those devices are different for the same inputs because the respective groups of calibration parameters for the devices are different. Such a situation may cause confusion in the learning process of ML model if reference SpCh values are used as labels. Calibration parameters could be added as input features so that the ML model can distinguish those two cases. However, the addition of calibration features may cause to ML model to be retrained for every hardware configuration to be used for measurement of SpCh, which may add significant resource overhead.

[0105] Embodiments of this disclosure may solve such a challenge in several ways. In some cases, embodiments of this disclosure solve the challenge by providing an ROR predictor model that is hardware agnostic and based on an ML approach, where the ROR predictor model need not be retrained when a new hardware configuration is used to measure SpCh. Examples of the ML- based predictor model include artificial neural network (ANN) models, such as MLP models, LSTM models, transformer models, State Space models, and convolution neural network (CNN) models.

[0106] Training of the ROR predictor model includes generating training data that includes reference R values for one or more hardware configurations. To generate reference R values for a hardware configuration (or hardware design), the calibration curve for the hardware configuration can be used to convert a reference SpO2 value to a reference R value RREF. Simply as an illustration, FIG. 10 presents schematic line plots of respective calibration curves for twohardware configurations. Although the calibration curves shown in FIG. 10 are straight lines, a reference RREF value can be obtained with any type of calibration curve.

[0107] The training data also includes multiple sets of values of signal features evaluated using PPG signal from a first measurement channel corresponding to excitation light of a first wavelength XA. The training data also includes second values of the signal features evaluated using second PPG signal from a second measurement channel corresponding to excitation light of a second wavelength XB. The signal features {fi,f2,fs . . . Q\, with Q > 1, include multiple handcrafted statistical and signal-processing-based features as described hereinbefore.

[0108] Embodiments of this disclosure can generate training data for one or more wearable devices (or hardware configurations) including respective one or more datasets. Each dataset including data defining (RREF;{A.B,C;... ^)nwk for each hardware configuration HWk.The data may thus be obtained from different body locations: wrist, fingertip, in-ear, forehead, upper arm, chest, and neck; different optical configurations and different source-photodetector distances (e.g., from 3 mm to 11 mm); different analog front ends (AFEs) and different mechanical designs; different subjects having skin color ranging from 1 to 6 in Fitzpatrick scale; different protocols, such as hypoxia laboratory, at rest, during sleep, and the like.

[0109] FIG. 11 illustrates an example of a ROR predictor model, in accordance with one or more aspects of this disclosure. The example ROR predictor model 1110 is an ANN having the neural network architecture that includes a 5-layer multi-layer perceptron (MLP) 1120 with continuously differentiable exponential linear unit (CELU) activation functions. In FIG. 11, each layer of the 5-layer MLP 1120 is represented by a rectangle. Dropout may be used for the first two layers of the 5-layer MLP with 0.2 probability, for example. The example ROR predictor model 1110 also has one fully connected layer for a single output. The output is a predicted R Value (Rpred).

[0110] It is noted that the ANN is not limited to the NN architecture depicted in FIG. 11, and other architectures, such as other number of layers and / or activation functions, may be implemented. Further, other ROR predictor models may be configured as an LSTM model.

[0111] FIG. 12 illustrates an example of a system to generate an estimate of SpO2 based on an ML approach, in accordance with aspects of this disclosure. The example system 1200 includes a training subsystem 1230 and an inference / application subsystem 1260. The training subsystem 1230 includes an ROR evaluation module 1234 (denoted by “reverse R calculator” in FIG. 12)and a feature extraction module 1238 (denoted by “feature extractor” in FIG. 12). For PPG signals corresponding to respective color channels present in a measurement channel, the feature extraction module 1238 can generate multiples features, such as a combination of the ROR features described hereinbefore.

[0112] The training subsystem 1230 also includes a constructor unit 1242 including one or more modules (not depicted in FIG. 12) that permit training an ROR predictor model in accordance with aspects of this disclosure and / or testing a trained ROR predictor module.

[0113] The inference / application subsystem 1260 includes the feature extraction module 1238. The inference / application subsystem 1260 also includes an evaluation module 1264 that can receive data defining values of features for a measurement channel. The evaluation module 1264 can determine a value of Rpred by applying a trained ROR predictor model 1268. The ROR evaluation module 1266 can supply the value of Rpred to an SpO2 evaluation module 1272 also included in the inference / application subsystem 1260. As is described herein, the trained ROR predictor model 1268 is hardware agnostic. Thus, the SpO2 evaluation module 1272 can determine an estimate 1276 of SpO2 using a group of calibration parameters for a wearable device (e.g., the wearable device 110 (FIG. 1)) that generates the IR PPG signal and the red PPG signal. As is depicted in FIG. 12, the group of calibration parameters may include three parameters: A, B, and C. As a result, the SpO2 evaluation module 1272 can supply the SpO2 estimate 1276. The inference / application subsystem 1260 can be implemented (e.g., configured or otherwise instealled in the werable device that measures the IR PPG signal and red PPG signal during inference, as part of probing SpO2 of a subject. The wearable device can be the wearable device 110 (FIG. 1) and the subject can be the subject 104 (FIG. 1). In some cases, the processor- accessible instructions 128 (FIG. 1) can include the feature extractor module 1238, the R evaluation module 1264, and the SpO2 evaluation module 1272. Data 132 (FIG. 1) can include parameters and other data defining the trained ROR model 1268. Data 132 also can include calibration data, including the parameters A, B, and C described herein.

[0114] Performance of a signal processing-based approach to determining R can be compared with performance of a ML-based approach to determining R as is described herein. Simply as an illustration, performance of an NN-based R calculator presents a significant improvement relative to the signal processing-based approach, as is shown in FIGS. 13-15. The overall root mean square error (RMSE) of test set, which includes a number of test subjects in hypoxia laboratory,was significantly improved in the ML-based approach relative to the signal processing-based approach. With respect to FIG. 13, the RMSE of a test set of 16 subject in hypoxia laboratory was improved from 3.28 % in the signal processing-based approach to 2.35 % in the ML-based approach. With respect to FIG. 14, the RMSE of a test set of 20 subject in hypoxia laboratory was improved from 3.42 % in the signal processing-based approach to 2.02 % in the ML-based approach. With respect to FIG. 15, the RMSE of a test set of 15 subject in hypoxia laboratory was improved from 5.20 % in the signal processing-based approach to 3.24 % in the ML-based approach. In other cases, the RMSE was reduced from 4.17% to 3.08%, which also is a substantive performance improvement. Such an improvement may permit wearable devices that use the ML- based techniques described herein to perform clinical-grade SpCh measurements. Such a performance provides an improvement over commonplace wearable devices that rely on typical signal processing-based approaches. Thus, ML-based approaches and the wearable devices in accordance with aspects of this disclosure clearly improve wearable devices for monitoring vital signs and the field of monitoring vital signs.

[0115] In some cases, an ROR predictor model can be configured to predict both R and a confidence level / attribute of the predicted R. To this end, a multi-task learning approach can be used, where the ROR predictor model is trained to predict R and the confidence level / attribute of R by solving an optimization problem (e.g., minimization problem) with respect to a total loss function that is equal to regression loss function plus a classification loss function. The regression loss function and the classification loss functions are trained simultaneously. Examples of regression losses are LI loss function or mean square error (MSE) loss function. An example of classification loss is binary cross-entropy (BCE) loss function. The disclosure is not limited to any particular type of loss function.

[0116] Confidence level / attribute may be a value between 0 and 1. Instead of relying on labeled data indicative of value of R and expert assignments of respective confidence levels to the values of R, a training subsystem, or a component thereof (such as the model constructor unit 1242 (FIG. 12)), generates a confidence label as the ROR predictor model is trained. Specifically, in some implementations, the confidence level is represented by a confidence label defined as the magnitude of the difference between predicted R and a realization of a confidence label of the predicted R. The confidence label is then assigned a value of 1 in response to the magnitude beinggreater than a defined threshold. The confidence label is assigned a value of 0 in response to the magnitude being less than or equal to the defined threshold value.

[0117] At a particular iteration during training, the regression loss function is evaluated as a function of predicted R and a confidence label adopted at that iteration particular iteration. In addition, the classification loss also is evaluated as a function of a predicted confidence label (e.g., a value in the interval [0,1]) at the confidence label determined at the particular iteration as the magnitude of the difference between the predicted R and the confidence level adopted at that particular iteration. Accordingly, as training progresses, the confidence label adopted at an iteration improves as the total loss function evolves towards an optimal (or otherwise satisfactory value).

[0118] FIG. 16 is a diagram 1600 that summarizes the multi-task learning approach that is described herein, to generate and apply an ROR predictor model in accordance with aspects of this disclosure. In the diagram 1600, the regression loss is an LI loss function and the classification loss is a BCE loss function. The total loss is the superposition of the regression loss and the classification loss. It is noted that this disclosure is not limited to an LI loss function and / or BCE loss function, and other types of loss functions can be used in the multi-task learning approach. For example, a mean squared error (MSE) loss and / or Gaussian negative log likelihood loss also may be used.

[0119] FIG. 17 illustrates an example of a ROR predictor model that can be used to predict R and a confidence level for the predicted R, in accordance with one or more aspects of this disclosure. The example ROR predictor model 1710 is an ANN having a neural network architecture that includes a 5-layer multi-layer perceptron (MLP) 1720 with CELU activation. In FIG. 17, each layer of the 5-layer MLP 1720 is represented by a rectangle. Dropout may be used for the first two layers of the 5-layer MLP with 0.2 probability, for example. The example ROR predictor model 1710 also has one fully connected layer for a first output corresponding to a predicted R value (Rpred) and a second output that is supplied to a sigmoid activation. The sigmoid activation applies a sigmoid function to the second output, and, in response, outputs the confidence level for Rpred.

[0120] It is noted that the ANN that forms the ROR predictor model 1710 is not limited to the NN architecture depicted in FIG. 17, and other architectures, such as other number of layersand / or activation functions, may be implemented. Further, other ROR predictor models can be configured, such as an LSTM model.

[0121] FIG. 18 illustrates another example of a system to generate an estimate of SpCh based on an ML approach, in accordance with aspects of this disclosure. The ML approach involves predicting R and a confidence level for R, as is described herein. The example system 1800 can be integrated into a wearable device for measurement of SpCh. The wearable device may be the wrist wearable device 110 (FIG. 1), for example. Indeed, example system 1800 can be implemented (e.g., configured or otherwise installed in the werable device that measures IRPPG signal and red PPG signal, and in some cases acceleration signals, during as part of probing SpCh of a subject. The wearable device can be the wearable device 110 (FIG. 1) and the subject can be the subject 104 (FIG. 1). In some cases, the processor-accessible instructions 128 (FIG. 1) can include the motion detection module 1810, the module 1820, the module 1830, the module 1840, the module 1850, and the module 1860 described herein. Data 132 (FIG. 1) can include parameters and other data defining the various models included in the example system 1800, in accordance with aspects of this disclosure. Data 132 also can include calibration data, including the parameters A, B, and C described herein.

[0122] The example system 1800 includes a motion detector module 1810 that can receive measurement signals from an inertial sensor device present in the wearable device. For example, the inertial device may be one of the inertial sensor devices 122. The inertial sensor device may be an accelerometer device and the measurement signals can correspond to signals from respective measurement channels in the accelerometer device. The motion detector module 1810 can generate a flag, and attribute, or another datum indicative of a degree of motion (e.g., in motion, stationary, etc.) of the wearable device.

[0123] The example system 1800 also includes a module 1820 that can operate on PPG signals for a first color channel (e.g., red channel) and a second color channel (e.g., IR channel). The module 1820 can condition the received by operating on the signal to remove artifacts or otherwise transform the signal to a more suitable form subsequent operations. The module 1820 also can perform AC / DC separation by applying low pass filtering to the received signal. The output of the low pass filtering yields the DC value. Subtracting the output of the low pass filtering from the original / raw signal gives the AC component. In cases the PPG signals that are receivedcorrespond to a red channel and IR channel, the module 1820 can generate a red AC signal, a red DC signal, an IR AC signal, and an IR DC signal.

[0124] The module 1820 also can generate a signal quality flag (or attribute or datum) representative of signal quality of the received PPG signals. To that end, the module 1830 can evaluate or otherwise check the perfusion index that may be obtained with the original / raw PPG signals in the frist color channel and the second color channel. In case the perfusion index is less than a defined threshold value, the module 1820 assigns low signal quality to the signal quality flag for the received PPG signals. In case the pefusion index is equal to or greater than the defined threshold value, the module 1820 can evaluate or otherwise check consistency of beat (or hearthrate-like periodicity) of the PPG signals. A consistency of beat that is less than another defined threshold value may result in the module 1820 assigning a low signal quality to the signl quality flag. Otherwise a consistenty of beat that is equal to or greater than the other defined threshold value may in the module 1820 assigning a satisfactory signal quality to the signal quality flag. The module 1820 also can analyze or otherwise check the DC signal for absence of fluctuations. In the absence of such fluctuations, the module 1820 can assign a satisfactory signal quality to the signal quality flag.

[0125] The example system 1800 also includes a signal processing-based evaluation module 1830. The signal processing-based evaluaton module 1830 that can operate on the AC and DC signals for the channels (e.g., red channel and IR channel) that supplied the raw signals. The module 1820 can supply such signals to the signal processing-based evaluation module 1830. By operating on those signals, the signal processing-based evaluation module 1830 can generate one or more estimates of ROR (R). In some cases, the signal processing-based evaluation module 1830 can generate a first R and a second R, which may be referred to as Main R and Auxiliary R, respectively, simply for the sake of nomenclature.

[0126] The example system 1800 also includes a feature extractor module 1840 that can operate in accordance with aspects described herein in connection with other feature extractor modules of this disclosure. The module 1820 also can supply, to the feature extractor module 1840, the AC and DC signals for the channels (e.g., red channel and IR channel) that supplied the raw signals. The feature extractor module 1840 can genereate one or more features 1844. The feature(s) 1844 may include one or more of the features described herein.

[0127] The example system 1800 also includes an ML- based R evaluation module 1850. Such a module 1850 can include an ROR predictor model (not depicted in FIG. 18) in accordance with this disclosure. Such ROR predictor model can be the ML model described in connection with FIG. 17 or FIG. 20 herein. As such, in accordance with aspects of this disclosure, the ML-based R evaluation module 1850 can generate an ROR prediction (Rpred) and a confidence level for the ROR prediction.

[0128] The ML-based R evaluation module 1850 can supply the ROR prediction to an SpO2 evaluation module 1860 that may be part of the example system and can operate in accordance with aspects of this disclosure. The ML-based R evaluation module 1850 can thus generate an estimate / value of SpO2. For example, such module 1860 can receive Rpred 1854 and can evaluate a calibration function using Rpred 1854 as an argument. In some cases, the calibration function is a quadratic function of R and has three calibration parameters a, b, and c (depicted as A, B, and C in FIG. 18). The module 1850 can supply the estimate / value of SpCh.

[0129] Simply as an illustration, performance of an ML approach to determination of SpCh is shown in FIG. 19. The ML approach may be implemented by the example system 1800 in some cases. More specifically, the results presented in FIG. 19 using the ML-based SpCh predictor model is an MLP as is described herein in connection with FIG. 17 or FIG. 20. Similar to other performance results described herein, the overall RMSE of test set, which includes a number of test subjects in hypoxia laboratory, was significantly improved in the ML-based approach relative to the signal processing-based approach.

[0130] As is described herein, embodiments of this disclosure also may implement an ML approach to predict an SpCh value and a confidence level that quantifies confidence / uncertainty of a predicted SpCh value. To that end, embodiments of the disclosure can train an SpCh predictor model. Similar to ROR predictor models described herein, in some cases, the SpO2 predictor model is an ANN having the NN architecture 2010 shown in FIG. 20. The NN architecture 2010 defines, at least partially, the SpO2 predictor model. Indeed, the NN architecture 2010 is an example of the ROR predictor model 1710 (FIG. 17) described herein. The NN architecture 2010 includes a 5-layer MPL with CELU activation. Dropout with a defined probability may be implemented for the first two layers, in some configurations. The NN architecture 2010 is configured to receive multiple hand-crafted statistical and signal-processing-based features. As is illustrated in FIG. 20, the NN architecture 2010 may be configured to receive 37 such features.The NN architecture 2010 is configured to yield two outputs: a prediction of SpCh and a confidence level for the prediction. Although not shown in FIG. 20, the prediction of SpCh is based on a defined calibration curve that maps an ROR predicted by the SpO2 predictor model (which model is an ML model and is represented by the NN architecture 2010) to a value of SpCh. As is described herein, the calibration curve may be a quadratic function of R (see Eq. (9)).

[0131] It is noted that the ANN that may form a SpCh predictor model is not limited to the NN architecture depicted in FIG. 20, and other architectures, such as a different number of layers and / or different activation function, may be implemented. Further, an SpCh predictor model is not limited to being or including an ANN. Indeed, an SpCh predictor model may be another type of ML-based models, such as an LSTM model, a transformer model, a State Space model, and a convolution neural network (CNN) model. It also noted that an SpCh predictor model can be an ROR predictor model in accordance with aspects described herein, where a prediction of SpO2 is obtained from a prediction of ROR (e.g., Rpred) and subsequent transformation, via a calibration function, of the prediction of ROR to a value of SpO2.

[0132] An SpO2predictor model in accordance with aspects of this disclosure can be trained in similar fashion to the training of ROR predictor models that are described herein. Specifically, the SpO2predictor model can be trained using multi-task learning, to output both SpO2 and confidence level. Reference SpO2 values obtained in various manners can be used in training. For example, the reference SpO2 values can be obtained in controlled experiments in a hypoxia laboratory. Labels for confidence level output, however, may be unavailable or may be impractical to produce. Thus, similar to other confidence levels / attributes described herein, confidence level / attribute outputs for respective SpCh predictions may be obtained dynamically during training of the ML model that forms the SpCh predictor model, at each iteration of the training.

[0133] For each group (or batch) of examples in a training set, a training iteration (via application of the ML model in training) yields a vector of SpCh predictions and a vector of probabilities that indicates whether or not SpCh predictions are within a defined distance with respect to reference SpCh values. As mentioned, an SpCh prediction is obtained from a defined calibration curve that maps an ROR predicted by the ML model to a value of SpO2. Eq. (11) below represents ther relationship between the ML modle in training and the SpO2 predictions (denoted by “SpO2_pred”), such a vector of probabilities (denoted by “In Band Probability”), and the defined distance (denoted by “Threshold”). A reference SpO2 vaue is denoted by “SpO2_label”in Eq. (11). The labels for In Band Probability are contained in a vector (denoted by “In Band Label”) that may be obtained as is shown in Eq. (12).[SpO2_pred, In Band Probability] = ML model(features) (11)In Band Label = |SpO2_pred - SpO2_label| < Threshold (12)

[0134] A total loss function L to be optimized during training is defined as a combination of two loss functions corresponding to respective learning tasks: (i) Regression of SpCh values and (ii) classification of SpO2predictions. In some cases, for the regression and classification optimization problems, LI loss and BCE loss are used, respectively. In such cases, L can be defined as follows:L = Ll_Loss(SpO2_pred, SpO2_label) + BCE_Loss(In_Band_Probability,In Band Label)

[0135] Relative to some commonplace ML techniques, an improvement of the foregoing ML approach to an SpO2predictor model is that labels for confidence level need not be available prior to training. Indeed, as is described herein, such labels are generated dynamically during training, at each training iteration.

[0136] As is described herein, and illustrated in FIGS. 13-15 and FIG. 19, performance of ML-based SpO2predictor models may be superior to performance of signal processing-based methods. In some scenarios (referred to as corner cases, for the sake of nomenclature) predictions of SpO2values may be improved by implementing multiple-channel techniques. Examples of corner cases include loose sensor-skin coupling, low perfusion, and similar cases.

[0137] FIG. 21 illustrates an example of a system to generate, based on an ML-based approach, an estimate of SpO2and a confidence level / attribute of the estimate, in accordance with aspects of this disclosure. The example system 2100 includes the training subsystem 1230 and an inference / application subsystem 2130. The training subsystem 1230 may be implemented in one or multiple computing devices. The inference / application subsystem 2130 may be implemented in a wearable device, such as the wearable device 110 (FIG. 1), for example. The inference / application subsystem 2130 includes the feature extraction module 1238. Theinference / application subsystem 2130 also includes the evaluation module 1264 that can receive data defining values of features for a measurement channel. As is described herein, the evaluation module 1264 can determine a value of Rpred by applying a trained ROR predictor model 1268. The ROR evaluation module 1264 can supply the value of Rpred to an SpO2evaluation module 2134 also included in the inference / application subsystem 2130. As is described herein, the trained ROR predictor model 1268 is hardware agnostic. Thus, the SpO2 evaluation module 2134 can determine an estimate 2142 of SpO2 and a confidence level 2144. The confidence level 2144 is indictive of an accuracy of the estimate 2142 of SpO2. The confidence level 2144 is defined as the confidence level of the value Rpred. To that end, the SpO2 evaluation module 2134 can determine a value of SpO2 using the predicted ROR (Rred) and a calibration curve, as is described herein. Such a value of SpO2 is the estimate 2142. The SpO2 evaluation module 2134 can supply the estimate 2142 and the confidence level 2144.

[0138] In some cases, the processor-accessible instructions 128 (FIG. 1) can include the feature extractor module 1238, the R evaluation module 1264, and the SpO2 evaluation module 2134. Data 132 (FIG. 1) can include parameters and other data defining the trained ROR model 1268 and the SpO2 evaluation module 2134. Data 132 also can include calibration data, including the parameters A, B, and C described herein.

[0139] FIG. 22 presents a flowchart of an example of a multi-channel method for determining an SpO2level, in accordance with aspects of this disclosure. The example multi-channel method 2200 may be implemented by a wearable device used to measure SpO2 levels. To that end, the wearable device has one or more processors, one or more memory devices, and / or other computing resources. In some cases, the wearable device is the wearable device 110 (FIG. 1). Other types of computing devices can implement the example method 2200.

[0140] At block 2210, the wearable device can determine single-channel output of a measurement channel. The single-channel output is indicative of the SpO2 level. For example, the single-channel output may include data quantifying the SpO2 level. The wearable device can host the SpO2evaluation module 2134 (FIG. 21) and can execute the SpO2 evaluation module 2134 to determine the single-channel output.

[0141] At block 2220, the wearable device can determine a confidence level of the measurement channel. To that end, the wearable device can host the SpO2evaluation module 2134 (FIG. 21) and can execute the SpO2 evaluation module 2134 to determine the confidence level.The wearable device can determine the single-channel output and the confidence level concurrently, by executing the SpCh evaluation module 2134 (FIG. 21).

[0142] At block 2220, the wearable device can determine if a next measurement channel is to be evaluated. In some instance, the wearable device determines that a next measurement channel is to be evaluated. In response to such a positive determination (“Yes” branch), the flow of the example method 2200 returns to block 2210, and the wearable device can implement block 2210 and block 2220 for the next measurement channel.

[0143] In other instances, the wearable device determines that there is no next measurement channel to be evaluated. For example, the wearable device may determine that single-channel output and confidence level have been determined for a satisfactory number of measurement channels (e.g., all measurement channels available in the wearable device have been evaluated). In response to such a negative determination (“No” branch), the flow of the example method can continue to block 2240 where the wearable device can increase a cumulative score of a second measurement channel having the highest confidence level.

[0144] Quality of PPG signals corresponding to respective color channels may change over time due to various factors, such as changes in the opto-mechanical coupling between the wearable device and a subject being monitored. Thus, cumulative scores of respective channels also may change over time. Accordingly, cumulative scores of measurements channels available in the wearable device may be updated over a defined update period (e.g., 15 second, 30 seconds, 60 seconds, 120 seconds, or 150 seconds).

[0145] Hence, at block 2250 the wearable device can determine if the update period has elapsed. In some instances, the wearable device determines that the update period has not elapsed. In response to such a negative determination (“No” branch), the flow of the example method 2200 can return to block 2210, and the wearable device can update the same or another cumulative score by implementing repeating the implementation of blocks 2210 to 2240. In other instances, the wearable device can determine that the update period has elapsed. In response to such a positive determination (“Yes” branch), the flow of the example method 2200 can continue to block 2260 where the wearable device can identify a particular measurement channel having a highest cumulative score.

[0146] At block 2270, the wearable device can report / supply the single-channel output corresponding to the particular measurement channel. For example, the wearable device can report / supply data quantifying an SpO2level.

[0147] FIG. 23 presents an example output of the multi-channel method 2200, where singlechannel SpO2output is selected according to confidence level of the SpO2output. The example output is obtained in a hardware configuration having two measurement channels, simply for purposes of illustration.

[0148] As the number of light source devices and optical sensor devices increases, the number of measurement channels increases rapidly. As a result, an ML- based approach similar to example multi-channel method 2200 may be impractical to implement due to a large number of measurements channels for which an SpO2prediction model is to be applied. Accordingly, embodiments of this disclosure may include an ML-based multi-channel approach where a subset of the available measurement channels is selected based on signal quality. That is, rather than processing all available measurement channels, PPG signals from a subset of select measurement channels are processed using ML-based techniques to determine an estimate of SpO2.

[0149] FIG. 24 illustrates an example of a system to generate, based on a multi-channel ML- based approach, an estimate of SpO2, in accordance with aspects of this disclosure. The example system 2400 shown includes a channel select / sort module 2410. The channel select / sort module 2410 can be configured as a ML-based predictor model that ingests raw channel data 2404 and outputs a signal quality index (SQI). An SQI of a measurement channel indicates, quantitiatively or otherwise, a confidence level that a value of SpO2determined using the measurement channel is an accurate value. As mentioned, the predictor model may be a ML-based predictor model. Examples of the ML-based predictor model include an ANN model and an LSTM model. The raw channel data includes, in some cases, infrared channel data, red channel data, and x, y, and z channel data from an accelerometer device that is present in a wearable device to measure SpO2. The AC components and DC components of the infrared channel and the red channel, which components are the main variables in an SpO2determination, are included as inputs to the channel select / sort module 2410.

[0150] As mentioned, manual expert labeling may be impractical, and, even if feasible, may limit size of a training set. Thus, the target values of the ML-based predictor model’s confidence output may be assigned on-the-fly according to error in SpO2estimates, as is described herein.These confidence targets also can be used as the target values of the channel select / sort module 2410.

[0151] Based on an assumption that signal quality conditions the uncertainty of the estimates, a Bayesian uncertainty measure of the ML-based predictor model may be used as target values of the channel select / sort module 2410. In some cases, a Bayesian uncertainty measure may be implemented by using dropout layers at the inference / application phase. Multiple inferences with different random dropout configurations give multiple SpCh estimates. The mean and variance of the estimates can be considered as the estimate and the uncertainty of the estimate, respectively. By this Bayesian approach, the ML-based predictor model can be used to obtain signal quality target values for any data independent of having reference SpCh values.

[0152] Because the example system 2400 may be implemented in a wearable device, it may be desirable that the NN architecture of the channel select / sort module 2410 be light in computation and memory footprint. FIG. 25 and FIG. 26 present an example of an NN architecture of the channel select / sort module 2410, in accordance with aspects of this disclosure. The NN architecture includes fully connected layer, skip connection, swish activation function, and layer normalization as building blocks of the selection / sorting model associated with the channel select / sort module 2410 (FIG. 24) The disclosure is, of course, not limited in these respects. For example, other activation functions could be used and / or layer normalization may be avoided. The input signal length may be configured to 128 samples to roughly cover a 5-second interval at 25 Hz sampling rate. The dimension of hidden features may be configured to 256.

[0153] FIG. 27 presents a flowchart of an example of method for generating a selection / sorting model and determining select measurement channel(s) using the selection / sorting model, in accordance with aspects of this disclosure. The example method 2700 may be implemented by a computing device or a system of computing devices. To that end, each of the computing devices includes one or more processors, one or more memory devices, and / or other computing resources.

[0154] The example method 2700 includes a training stage and an inference / application phase. The training phase may be implemented by a system of computing devices. The inference / application phase may be implemented by a wearable device used to measure vital signs, where the wearable device (e.g., wearable device 110 (FIG. 1)) hosts the channel select / sort module 2410 described herein.

[0155] The training phase includes block 2710 to block 2740. At block 2710, the system of computing devices (referred to as a computing system) can receive first input signals for multiple measurement channels. Each one of the first signals is caused by excitation with light of a first wavelength (e.g., red light with X = 660 nm).

[0156] At block 2720, the computing system can receive second input signals for the multiple measurement channels. Each one of the second input signals is caused by excitation with light of a second wavelength (e.g., IR light with X = 880 nm).

[0157] At block 2730, the computing system can receive confidence levels for multiple channels. A respective confidence level may be received for each one of the multiple measurement channels. A confidence level serves as a signal quality index (or label) for a measurement channel. The confidence level may correspond to the confidence level that is output by an SpCh predictor model when applied to signal features of a first PPG signal and a second PPG signal. The SpCh predictor model may be an ML-based model in accordance with aspects described herein.

[0158] At block 2740, the computing system can train a signal quality index (SQI) predictor model using the first input signals, the second input signals, and the confidence levels.

[0159] The inference / application phase includes block 2750 to block 2790. At block 2750, the wearable device that hosts the select / sort module 2410 can receive a third PPG signal and a fourth PPG signal. The third PPG signal and the fourth PPG signal are caused, respectively, by excitation with light of the first wavelength and second wavelength. As mentioned, in one example, the first wavelength is 660 nm, and the second wavelength is 880 nm.

[0160] At block 2760, the wearable device can generate a SQI for the measurement channel by applying the SQI predictor model to the third PPG signal and the fourth PPG signal.

[0161] At block 2770, the wearable device can determine if SQI for another channel of the multiple measurement channels is to be generated. In the affirmative case (“Yes” branch), flow of the example method 2700 returns to block 2750 and flow continues as described above. Block 2750 and block 2760 may be repeated as many times as the number Mcof measurement channels present in the multiple channels. As a result, example method 2700 yields Mcpredicted SQIs. For example, as described herein, in the hardware configuration 250 (FIG. 2), there as many as 18 measurement channels. Accordingly, block 2750 and block 2760 may be implemented 18 times, yielding 18 predicted SQIs.

[0162] In case the outcome of implementing block 2770 is negative (“No” branch), the flow of the example method continues to block 2780, where the wearable device can generate a ranking of the respective SQIs for the multiple measurement channels.

[0163] At block 2790, the wearable device can select one or more measurement channels having a particular placement within the ranking. As a result, the wearable device can identify a desired number mcof measurement of satisfactory channels that may be used to generate a multichannel estimate of SpCh. Here, mcis configurable and is less than of Moin some cases. In one example, the wearable device selects the measurement channel having the top ranked SQL That is, mc= 1. In another example, the wearable device selects the top-two ranked measurement channels. That is, mc= 2. In yet another example, the wearable device selects the top-three ranked measurement channels. That is, mc= 3.

[0164] Because mcis configurable, the value of mcmay be configured based on the amount of computing resources that are present in the wearable device that implements the inference / application stage of the example method 2700.

[0165] Channel selection can be updated / refreshed periodically, after a defined a refresh time interval elapses. In addition, or as an alternative, the channel selection also can be updated / refreshed according to a defined schedule and / or or in response to a defined event. As an example, a ranking of measurement channels can be triggered upon or after a change in the placement of the wearable device detected by a motion sensor device (e.g., an accelerometer device) present in the wearable device. As another example, the ranking can be updated in response to the longer of a “no report” duration of the algorithm and a defined time interval elapsing. Regardless of how the channel selection is updated / refreshed, such updates may permit the wearable device to use satisfactory measurement channels to determine a value of SpCh when monitoring SpCh level over time.

[0166] With further reference to FIG. 24, the example system 2400 also includes processing modules 2420 that can process, individually or in combination, PPG signals 2422 selected by the select / sort module 2410. The number of processing modules 2420 is configurable and may be different across wearable devices for measurement of SpCh, depending on computing resources present in the wearable devices. That is, a wearable device with greater amount of computing resources may have a greater number of processing modules 2420 than another wearable device that has a lesser amount of computing resources. Each one of the processing modules 2420 mayinclude a feature extractor module 2426 that can generate one or multiple signal processing-based features 2425 in accordance with aspects described herein. SpO2evaluation module 2428. Each one of the processing modules 2420 also may include an SpO2evaluation module 2428. The SpO2evaluation module 2428 can determine SpO2values and respective confidence levels by applying an ROR predictor model and transforming a predicted value of ROR Rpred) to a value of SpO2via a calbration curve, as is described herein. Such respective confidence levels are defined by respective confidence levels of predicted RORs or SpO2values, as is described herein. In accordance with aspects of this disclosure. The ROR predictor model is a machine- learned model (also referred to as machine- learning model). FIG. 26 presents an example of an NN architecture of the SpO2predictor model 2610, and the manner of training and predicting RORs or SpO2 values and confidence levels. The training is similar to the training described herein (see, e.g., FIG. 16) for n ROR predictor models in accordance with this disclosure.

[0167] Back to FIG. 24, the example system 2400 also includes a multi-channel fusion module 2430 that can generate a multi-channel estimate 2440 of SpO2based on one or more predictions of SpO2and respective confidence levels generated by the SpO2evaluation module 2428 for the selected channels corresponding to the signals 2422. To that end, the multi-channel fusion module 2430 can evaluate an aggregation function / expression that combines the one or more predictions of SpO2and respective confidence levels. FIG. 26 presents an example of the aggregation function / expression 2630 (denoted by “multi-channel fusion function”). The aggregation function / expression 2630 determines a multi-channel estima of SpO2as a weighted average of SpO2 predicted estimates, with weights provided by respective confidence levels of selected channels. The disclosure is not limited in that respect and other aggregation functions / expressions may be utilized.

[0168] In some cases, the processor-accessible instructions 128 (FIG. 1) can include the module 2410, the module 2426, and the SpO2 evaluation module 2428, and the fusion module 2430. Data 132 (FIG. 1) can include parameters and other data defining the trained ML models included in the module 2410 and the module 2428, respectively. Data 132 also can include calibration data, including the parameters A, B, and C described herein.

[0169] It is noted that red PPG signals and IR PPG signals alone may not contain sufficient information for characterizing subject dependent variations on a calibration curve. Consequently, in some cases, an ML model confidence might misclassify due to data (aleatoric) uncertainty.Embodiments of this disclosure may address such an issue by changing the training target for ML confidence from classifying overall error to classifying only the fluctuation error. Fluctuation error is defined as the error the ML model makes to a ROR target using individual calibration rather than population calibration. Here, for purposes of illustration, an individual calibration corresponds to a calibration curve corresponding to a single subject, where the calibration curve (referred to as “individual calibration curve”) maps a value of ROR to a value of SpO2. In addition, also for purposes of illustration, a population calibration refers to another calibration curve corresponding to a group of subjects, here that other calibration curve (referred to as “population calibration curve”) also maps a value of ROR to a value of SpO2. Without intending to be bound by modeling, changin the training target to fluctuation error may be based the understanding that the overall error that the ML model makes has to error contributions: (I) a fluctuation error and (II) a calibration error. Fluctuation error refers to random fluctuations around the individual calibration curve. Calibration error refers to deviation of the individual calibration curve from the population calibration line.

[0170] FIGS. 28A-28C illustrate various types of errors associated with ML-based predictor models described herein, in accordance with aspects of this disclosure. FIG. 28A and FIG. 28B are simulated as the ideal behavior if the ML model is able to classify for errors without misclassification. FIG. 28A illustrates SpCh regression error as a function of coverage based on ideal overall error confidence. FIG. 28B illustrates SpCh regression error as a function of coverage based on ideal fluctuation error confidence. The overall regression error (labeled “rmse all” and shown as blue circles) in both FIG. 28A and FIG. 28B trade off with coverage efficiently in the high coverage region (above 0.8 (or 80%) coverage, for example). However, the ideal fluctuation error confidence ceases to be effective in lower coverage regions, as the calibration error dominates. Because operation of a werable device for determination of SpCh in accordance with aspects of this disclosure may operate around the high coverage region only, the wearable device can yield satisfactory quality of data that the data that might be disqualified with the confidence metric are few, this shows the applicability of the fluctuation error confidence approach.

[0171] Results of fluctuation error confidence can be obtained. To that end, compared to the overall error confidence defined hereinbefore, where In Band Label = |SpO2_pred_population_calibration - SpO2_label | < Threshold is the ground truth, theIn Band Label is redefined as In Band Label = In_Band_Labelfe. = |SpO2_pred_individual_calibration - SpO2_label | < Threshold and is taken as the ground truth during training. Here, “fe” denotes fluctuation error and is used to convey that the training is based on fluctuation error confidence. The training of a ML model still may be based on a loss that is similar to the loss shown in Eq. (13) but with the In_Band_Labelfe; namely, the taning can be based on the following loss in some cases:£fe= Ll_Loss(SpO2_pred, SpO2_label) + BCE_Loss(In_Band_Probability,In Band Labelfe)This disclosure is not limited in that respect and the loss L may be defined in terms of combinations of other types of regression loss and / or classification loss.

[0172] By training the ML model for this new confidence target, the ML model classifies the fluctuation error accurately in some cases. As a result, a more efficient tradeoff between error and coverage may be accomplished when using the fluctuation error confidence compared to using the overall error confidence. Such an efficient tradeoff may be more prevalent in a higher coverage region corresponding to some product use cases. FIG. 29 present a comparison of the ideal fluctuation error confidence (keyed as “ideal” and shown as green circles), the trained overall error confidence (keyed as “v9” and shown as orange circles), and the trained fluctuation error confidence (keyed as “ind err conf’ and shown in blue circles). From FIG. 29 it can be readily gleaned that there is a gained efficiency in tradeoff of SpCh regression error versus coverage between 50 % and 90 % coverage regions.

[0173] FIG. 30 is a block diagram of an example of a computing system that forms, or includes, a training subsystem in accordance with aspects of this discosure. Accordingly, the example computing system 3000 can provide at least some of the functionality described herein in connection with training of ML-based predictor in accordance with aspects described herein.

[0174] The example computing system 3000 forms, or is part of, a cloud computing system that can permit training a ML-based predictor model. The example computing system 3000 represents a cloud computing system. However, the disclosre is not limited in that respect and other types of computing systems including a network of server devices or computing devices can be used for training of an ML-based predictor model. The example computing system 3000includes two types of server devices: Compute server devices 3020 and storage server devices 3030. A subset of the compute server devices 3020, individually or collectively, can host various modules that permit implementing the training of ML-based predictor model, in accordance with aspects described herein. The modules include the modules 3050. The architecture of each of the compute server devices 3020, including the compute server device 3022, includes multiple input / output (I / O) interfaces 3024, one or more processors 3026, one or more memory devices 3028, and a bus architecture 3025 that functionally couples the processor(s) and the memory device(s). In some cases, the compute server device 3022 can store the modules in at least one of such memory device(s) 3028. At least the subset of the compute server devices 3020 can be functionally coupled to one or multiple ones of the storage server devices 3030. The coupling can be direct or can be mediated by at least one of the gateway devices 3010. The storage server devices 3030 include data and / or metadata that can be used to implement the functionality described herein in connection with monitoring status changes of a cardiopulmonary condition using multi-modal measurement signals.

[0175] Each one of the gateway devices 3010 can include one or multiple processors functionally coupled to one or multiple memory devices that can retain application programming interfaces (APIs) and / or other types of program code for access to the compute server devices 3020 and storage server devices 3030. Such access can be programmatic, via a defined function call, for example. The subset of the compute server devices 3020 that host one or a combination of modules that can use API(s) supplied by the gateway devices 3010 in order to provide results of implementing the functionalities described herein in connection with monitoring a respiratory condition using variability of respiratory parameters in accordance with aspects described herein.

[0176] Numerous example embodiments emerge from the foregoing detailed description and annexed drawings. The example embodiments are represented by the following clauses:

[0177] Clause 1A. A method, comprising: for a signal sampling event, performing digital signal processing (DSP) operations comprising, obtaining a first AC component of raw photoplethysmography (PPG) signal from a first color channel; obtaining a first DC component of the raw signal from the first color channel; obtaining a second AC component of raw signal from a second color channel; obtaining a second DC component of the raw signal from the second color channel; conditioning the first AC component by applying an adapative bandpass filter based on a heart rate of a subject, resulting on a first conditioned AC component; conditioning thesecond AC component by applying the adaptive bandpass filter based on the heart rate of the subject, resulting on a second conditioned AC component; determining a perfusion index for the first color channel as a first ratio of the first conditioned AC component and the first DC component; determining a perfusion index for the second color channel as a second ratio of the second conditioned AC component and the second DC component; determining a value of a ratio- of-ratios (ROR) for the signal sampling event as the ratio of the first ratio and the second ratio; repeating the DSP operations for each one of multiple next signal sampling events until multiple values of the ROR are determined; determining a single value of ROR by applying an aggregation filter to the multiple values of the ROR; and determining, using the single value of ROR, a value of peripheral oxygen saturation (SpO2) of the subject.

[0178] Clause 2A. The method of clause 1 A, further comprising providing the value of SpO2 of the subject.

[0179] Clause 3 A. The method of any of clause 1A or claim 2A, wherein the providing comprises causing a computing device to present the value of SpO2 of the subject, and wherein presenting the value of SpO2 of the subject comprises one or more of displaying visual indicia indicative of the value, presenting aural indicia indicative of the value, or presenting a haptic stimulus representative of the value.

[0180] Clause 4A. The method of any of the preceding clauses, wherein the first color channel comprises a first light source device configured to emit light of a first light wavelength, and wherein the second color channel comprises a second light source device configured to emit light a second light wavelength.

[0181] Clause 5A. The method of any of the preceding clauses wherein the first light wavelength lies in the red portion of the electromagnetic (EM) radiation spectrum, and wherein the second light wavelength lies in the infrared (IR) portion of the EM radiation spectrum.

[0182] Clause 6A. Awearable device, comprising: at least one processor; at least one memory device coupled with the at least one processor, the at least one memory device retaining processorexecutable instructions that, in response to execution by the at least one processor, individually or in combination, cause the wearable device to: for a signal sampling event, perform digital signal processing (DSP) operations comprising, obtaining a first AC component of raw photoplethysmography (PPG) signal from a first color channel; obtaining a first DC component of the raw signal from the first color channel; obtaining a second AC component of raw signalfrom a second color channel; obtaining a second DC component of the raw signal from the second color channel; conditioning the first AC component by applying an adapative bandpass filter based on a heart rate of a subject, resulting on a first conditioned AC component; conditioning the second AC component by applying the adaptive bandpass filter based on the heart rate of the subject, resulting on a second conditioned AC component; determining a perfusion index for the first color channel as a first ratio of the first conditioned AC component and the first DC component; determining a perfusion index for the second color channel as a second ratio of the second conditioned AC component and the second DC component; determining a value of a ratio- of-ratios (ROR) for the signal sampling event as the ratio of the first ratio and the second ratio; repeat the DSP operations for each one of multiple next signal sampling events until multiple values of the ROR are determined; determine a single value of ROR by applying an aggregation filter to the multiple values of the ROR; and determine, using the single value of ROR, a value of peripheral oxygen saturation (SpO2) of the subject.

[0183] Clause 7A. The wearable device of clause 6A, wherein the first color channel comprises a first light source device configured to emit light of a first light wavelength, and wherein the second color channel comprises a second light source device configured to emit light a second light wavelength.

[0184] Clause 8A. The wearable device of clause 7A, wherein the first light wavelength lies in the red portion of the electromagnetic (EM) radiation spectrum, and wherein the second light wavelength lies in the infrared (IR) portion of the EM radiation spectrum.

[0185] Clause 9A. The wearable device of claim any of the preceding clauses, further comprising multiple optical sensor devices and at least two light source devices.

[0186] Clause 10A. The wearable device of any of the preceding clauses, further comprising a feedback device, wherein the at least one memory device retains further processor-executable instructions that, in response to execution by the at least one processor, individually or in combination, further cause the wearable device to provide, via the feedback device, the value of SpO2of the subject.

[0187] Clause 11 A. The wearable device of clause 10A, wherein providing the value of SpCh comprises causing the feedback device to present the value of SpCh of the subject by one or more of displaying visual indicia indicative of the value, outputting aural indicia indicative of the value, or providing a haptic stimulus representative of the value.

[0188] Clause 12A. At least one non-transitory computer-readable storage medium having processor-executable instructions encode thereon that, in response to execution by at least one processor, individually or in combination, cause a wearable device to: for a signal sampling event, perform digital signal processing (DSP) operations comprising, obtaining a first AC component of raw photoplethysmography (PPG) signal from a first color channel; obtaining a first DC component of the raw signal from the first color channel; obtaining a second AC component of raw signal from a second color channel; obtaining a second DC component of the raw signal from the second color channel; conditioning the first AC component by applying an adapative bandpass filter based on a heart rate of a subject, resulting on a first conditioned AC component; conditioning the second AC component by applying the adaptive bandpass filter based on the heart rate of the subject, resulting on a second conditioned AC component; determining a perfusion index for the first color channel as a first ratio of the first conditioned AC component and the first DC component; determining a perfusion index for the second color channel as a second ratio of the second conditioned AC component and the second DC component; determining a value of a ratio-of-ratios (ROR) for the signal sampling event as the ratio of the first ratio and the second ratio; repeat the DSP operations for each one of multiple next signal sampling events until multiple values of the ROR are determined; determine a single value of ROR by applying an aggregation filter to the multiple values of the ROR; and determine, using the single value of ROR, a value of peripheral oxygen saturation (SpO2) of the subject.

[0189] Clause 13 A. The at least one non-transitory computer- readable storage medium of clause 12A, wherein the wearable device comprises a feedback device, at least one non-transitory computer-readable storage medium having further processor-executable instructions encoded thereon that, in response to execution by the at least one processor, individually or in combination, further cause the wearable device to provide, via the feedback device, the value of SpCh of the subject.

[0190] Clause 14 A. The at least one non-transitory computer- readable storage medium of clause 13 A, wherein providing the value of SpCh comprises causing the feedback device to present the value of SpCh of the subject by one or more of displaying visual indicia indicative of the value, outputting aural indicia indicative of the value, or providing a haptic stimulus representative of the value.

[0191] Clause IB. A method, comprising: generating one or more first datasets for respective one or more wearable devices, with each dataset of the one or more first datasets including, reference ratio- of-ratios (ROR) values, wherein a first reference ROR value of the reference ROR values is defined by a reference peripheral oxygen saturation (SpCh) value and a first calibration function of a first wearable device of the one or more wearable devices, with the first calibration function mapping ROR values to reference SpO2 values, and multiple first sets of signal features of conditioned photoplethysmography (PPG) signal from a first color channel, and multiple second sets of signal features of conditioned PPG signal from a second color channel; generating a second dataset as the union of the one or more first datasets; training, using the second dataset, a predictor model to determine a reference ROR corresponding to first signal features associated with the first color channel and second signal features associated with the second color channel; receiving first input PPG signal from a particular first color channel of a particular wearable device; receiving second input PPG signal from a particular second color channel of the particular wearable device; determining, using the first input PPG signal, a first input set of signal features; determining, using the second input PPG signal, a second input set of signal features; generating a second reference ROR by applying the predictor model to the first input set of signal features and the second input set of signal features; determining a value of SpO2 using the second reference ROR

[0192] Clause 2B. The method of clause IB, further comprising providing the SpO2 value. Clause 3B. The method of any of clauses IB or 2B, wherein the determining the value of SpO2comprises evaluating, using the second reference ROR as an argument, a calibration function of the particular wearable device, wherein a result of the evaluating defines the value of SpO2.

[0193] Clause 4B. The method of any of the preceding clauses, wherein the training further comprises training, using the second dataset, the predictor model to determine a confidence level associated with the reference ROR.

[0194] Clause 5B. The method of any of the preceding clauses, wherein the predictor model is trained to determine the reference ROR and the confidence level jointly.

[0195] Clause 6B. The method of any of the preceding clauses, wherein the training the predictor model to determine the reference ROR and the confidence level jointly comprisessolving an optimization problem with respect to a multi-task loss function including a regression loss function and a classification loss function.

[0196] Clause 7B. The method of any of the preceding clauses, wherein the solving comprises evaluating a confidence label at each iteration during the training of the predictor model.

[0197] Clause 8B. The method of any of the preceding clauses, wherein the confidence label is based on a fluctuation error label.

[0198] Clause 9B. The method of any of the preceding clauses, wherein the predictor model includes one of an artificial neural network, a long short-term memory (LSTM) model, a transformer model, a State Space model, or a convolution neural network (CNN) model.

[0199] Clause 10B. A system, comprising: at least one processor; at least one memory device coupled with the at least one processor, the at least one memory device retaining processorexecutable instructions that, in response to execution by the at least one processor, individually or in combination, cause the system to: generate one or more first datasets for respective one or more wearable devices, with each dataset of the one or more first datasets including, reference ratio-of- ratios (ROR) values, wherein a first reference ROR value of the reference ROR values is defined by a reference peripheral oxygen saturation (SpCh) value and a first calibration function of a first wearable device of the one or more wearable devices, with the first calibration function mapping ROR values to reference SpO2 values, and multiple first sets of signal features of conditioned photoplethysmography (PPG) signal from a first color channel, and multiple second sets of signal features of conditioned PPG signal from a second color channel; generate a second dataset as the union of the one or more first datasets; and train, using the second dataset, a predictor model to determine a reference ROR corresponding to first signal features associated with the first color channel and second signal features associated with the second color channel.

[0200] Clause 1 IB. The system of clause 10B, wherein training the predictor model comprises training, using the second dataset, the predictor model to determine a confidence level associated with the reference ROR.

[0201] Clause 12B. The system of any of clause 10B or 11B, wherein training the predictor model comprises training the predictor model to determine the reference ROR and the confidence level jointly.

[0202] Clause 13B. The system of any of the preceding clauses, wherein the training the predictor model to determine the reference ROR and the confidence level jointly comprisessolving an optimization problem with respect to a multi-task loss function including a regression loss function and a classification loss function.

[0203] Clause 14B. The system of any of the preceding clauses, wherein the solving comprises evaluating a confidence label at each iteration during the training of the predictor model.

[0204] Clause 15B. The system of any of the preceding clauses, wherein the confidence label is based on a fluctuation error label.

[0205] Clause 16B. A wearable device, comprising: at least one processor; at least one memory device coupled with the at least one processor, the at least one memory device retaining processor-executable instructions that, in response to execution by the at least one processor, individually or in combination, cause the wearable device to: receive first input PPG signal from a first color channel of the wearable device; receive second input PPG signal from a second color channel of the wearable device; determine, using the first input PPG signal, a first input set of signal features; determine, using the second input PPG signal, a second input set of signal features; generate a reference ROR by applying the predictor model to the first input set of signal features and the second input set of signal features; determine a value of SpCh using the second reference ROR.

[0206] Clause 17B. The wearable device of clause 16B, further comprising a feedback device, wherein the at least one memory device retains further processor-executable instructions that, in response to execution by the at least one processor, individually or in combination, further cause the feedback device to provide the SpO2 value.

[0207] Clause 18B. The wearable device of clause 17B, wherein providing the value of SpO2 comprises causing the feedback device to present the value of SpCh of the subject by one or more of displaying visual indicia indicative of the value, outputting aural indicia indicative of the value, or providing a haptic stimulus representative of the value.

[0208] Clause 19B. The wearable device of any of the preceding clauses, wherein the determining the value of SpCh comprises evaluating, using the second reference ROR as an argument, a calibration function of the particular wearable device, wherein a result of the evaluating defines the value of SpO2.

[0209] Clause 20B. The wearable device of any of the preceding clauses, wherein the predictor model includes a as multilayer perceptron (MLP) model.

[0210] Clause 1C. A method, comprising: receiving first photoplethysmography (PPG) signals corresponding to multiple first measurement channels, the first PPG signals caused by excitation with light of a first wavelength; receiving second PPG signals corresponding to the multiple first measurement channels, the second PPG signals caused by excitation with light of a second wavelength; receive respective confidence levels for the multiple measurement channels; training, using the first PPG signals, the second PPG signals, and the respective confidence levels, a predictor model to determine a signal quality index of a measurement channel, with the signal quality index being indicative of a confidence level that a value of peripheral oxygen saturation (SpO2) determined using the measurement channel is an accurate value; receiving third PPG signals of a subject, with the third PPG signals corresponding to multiple second measurement channels of a wearable device and caused by excitation of tissue of the subject with light of the first wavelength; receiving fourth PPG signals of the subject, with the fourth PPG signals corresponding to the multiple second measurement channels and caused by excitation of the tissue of the subject with light of the second wavelength; generating multiple signal quality indices for respective ones of the multiple second measurement channels by applying the trained predictor model to the third PPG signals and the fourth PPG signals; generating, using the multiple signal quality indices, a ranking of the multiple second measurement channels; selecting, using the ranking, one or more particular measurement channels of the second measurement channels, with the one or more particular measurement channels having a defined placement within the ranking; determining, using the one or more particular measurements channels, a value of SpCh of the subject.

[0211] Clause 2C. The method of clause 1C, wherein the determining the value of SpCh of the subject comprises, determining an aggregate value of one or more values of SpCh , with each of the one or more values of SpCh determined using respective ones of the one or more particular measurement channels; and configuring the aggregate value as the value of SpCh of the subject.

[0212] Clause 3C. The method of any of clause 1C or 2C, further comprising updating the ranking of the multiple second measurement channels periodically, according to a defined schedule or in response to a defined event.

[0213] Clause 4C. The method of any of the preceding clauses, further comprising providing the second value of SpCh of the subject.

[0214] Clause 5C. The method of any of the preceding clauses, wherein the providing comprises causing a computing device to present the value of SpO2of the subject, and wherein presenting the value of SpO2of the subject comprises one or more of displaying visual indicia indicative of the value, presenting aural indicia indicative of the value, or presenting a haptic stimulus representative of the value.

[0215] Clause 6C. A method, comprising: receiving first photoplethysmography (PPG) signals of a subject, with the first PPG signals corresponding to multiple measurement channels of a wearable device and caused by excitation of tissue of a subject with light of a first wavelength; receiving second PPG signals of the subject, with the second PPG signals corresponding to the multiple measurement channels and caused by excitation with light of a second wavelength; generating respective signal quality indices for the multiple measurement channels by applying a machine-learned predictor model to the first PPG signals and the second PPG signals; generating, using the respective signal quality indices, a ranking of the multiple measurement channels; selecting, using the ranking, one or more particular measurement channels of the measurement channels, with the one or more particular measurement channels having a defined placement within the ranking; determining, using the one or more particular measurements channels, a value of SpO2of the subject.

[0216] Clause 7C. The method of clause 6C, wherein the determining the value of SpO2of the subject comprises, determining an aggregate value of one or more values of SpO2, with each of the one or more values of SpO2determined using respective ones of the one or more particular measurement channels; and configuring the aggregate value as the value of SpO2of the subject.

[0217] Clause 8C. The method of any of clause 6 or 7, further comprising updating the ranking of the multiple second measurement channels periodically, according to a defined schedule, or in response to a defined event.

[0218] Clause 9C. The method of any of the preceding clauses, further comprising providing the value of SpO2of the subject.

[0219] Clause 10C. The method of any of the preceding clauses, wherein the providing comprises causing a computing device to present the value of SpO2of the subject, and wherein presenting the value of SpO2of the subject comprises one or more of displaying visual indicia indicative of the value, presenting aural indicia indicative of the value, or presenting a haptic stimulus representative of the value.

[0220] Clause 11C. A wearable device, comprising: at least one processor; at least one memory device coupled with the at least one processor, the at least one memory device retaining processor-executable instructions that, in response to execution by the at least one processor, individually or in combination, cause the wearable device to: receive first photoplethysmography (PPG) signals of a subject, with the first PPG signals corresponding to multiple measurement channels of a wearable device and caused by excitation of tissue of a subject with light of a first wavelength; receive second PPG signals of the subject, with the second PPG signals corresponding to the multiple measurement channels and caused by excitation with light of a second wavelength; generate respective signal quality indices for the multiple measurement channels by applying a machine-learned predictor model to the first PPG signals and the second PPG signals; generate, using the respective signal quality indices, a ranking of the multiple measurement channels; select, using the ranking, one or more particular measurement channels of the measurement channels, with the one or more particular measurement channels having a defined placement within the ranking; determine, using the one or more particular measurements channels, a value of SpCh of the subject.

[0221] Clause 12C. The wearable device of clause 11C, the at least one memory device retaining further processor-executable instructions that, in response to execution by the at least one processor, individually or in combination, cause the wearable device to update the ranking of the multiple second measurement channels periodically, according to a defined schedule, or in response to a defined event.

[0222] Clause 13C. The wearable device of any of clauses 11C or 12C, further comprising a feedback device, wherein the at least one memory device retains further processor-executable instructions that, in response to execution by the at least one processor, individually or in combination, further cause the feedback device to provide the SpCh value.

[0223] Clause 14C. The wearable device of clause 12C, wherein providing the value of SpCh comprises causing the feedback device to present the value of SpCh of the subject by one or more of displaying visual indicia indicative of the value, outputting aural indicia indicative of the value, or providing a haptic stimulus representative of the value.

[0224] Various aspects of the disclosure may take the form of an entirely or partially hardware aspect, an entirely or partially software aspect, or a combination of software and hardware. Furthermore, as described herein, various aspects of the disclosure (e.g., systems and methods)may take the form of a computer program product comprising a computer-readable non-transitory storage medium having processor-accessible instructions (e.g., computer-readable and / or computer-executable instructions) such as computer software, encoded or otherwise embodied in such storage medium. Those instructions can be read or otherwise accessed and executed by one or more processors, individually or in combination, to perform or permit the performance of the operations described herein. The instructions can be provided in any suitable form, such as source code, compiled code, interpreted code, executable code, static code, dynamic code, assembler code, combinations of the foregoing, and the like. Any suitable computer-readable non-transitory storage medium may be utilized to form the computer program product. For instance, the computer-readable medium may include any tangible non-transitory medium for storing information in a form readable or otherwise accessible by one or more computers or processor(s) functionally coupled thereto. Non-transitory storage media can include read-only memory (ROM); random access memory (RAM); magnetic disk storage media; optical storage media; flash memory, and so forth.

[0225] Aspects of this disclosure are described herein with reference to block diagrams and flowchart illustrations of processor-implemented methods, systems, devices, apparatuses, and computer program products. It can be understood that each block of the block diagrams and flowchart illustrations, and combinations of blocks in the block diagrams and flowchart illustrations, respectively, can be implemented by processor-accessible instructions. Such instructions may include, for example, computer program instructions (e.g., processor- readable and / or processor-executable instructions). The processor-accessible instructions may be built (e.g., linked and compiled) and retained in processor-executable form in one or multiple memory devices or one or many other processor-accessible non-transitory storage media. These computer program instructions also can be stored in a processor-readable memory, where in response to execution by one or more processors, individually or in combination, the computer program instructions can direct a computer, a computing device, or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the processor- readable memory produce an article of manufacture including processor-accessible instructions (e.g., processor-readable instructions and / or processor-executable instructions) to implement the function specified in the flowchart blocks (individually or in a particular combination) or blocks in block diagrams (individually or in a particular combination). The computer programinstructions can be loaded onto a computer, a computing device, or other programmable data processing apparatus to cause a series of operations to be performed on the computer or other programmable apparatus to produce a computer-implemented process. The series of operations may be performed in response to execution by one or more processor or other types of processing circuitry. Thus, such instructions that execute on the computer or other programmable apparatus provide operations for implementing the functions specified in the flowchart blocks (individually or in a particular combination) or blocks in block diagrams (individually or in a particular combination).

[0226] In some implementations, the processor-accessible instructions may be loaded or otherwise incorporated into a general purpose computer, a special purpose computer, or another programmable information processing apparatus to produce a particular machine, such that the operations or functions specified in the flowchart block or blocks can be implemented in response to execution at the computer or processing apparatus. More specifically, the loaded processor- accessible instructions may be accessed and executed by one or multiple processors, individually or in combination, or other types of processing circuitry. In response to execution, the loaded processor-accessible instructions provide the functionality described in connection with flowchart blocks (individually or in a particular combination) or blocks in block diagrams (individually or in a particular combination). Thus, such instructions which execute on a computer, a computing device, or other programmable data processing apparatus can create a means for implementing the functions specified in the flowchart blocks (individually or in a particular combination) or blocks in block diagrams (individually or in a particular combination).

[0227] Unless otherwise expressly stated, it is in no way intended that any protocol, procedure, process, or method set forth herein be construed as requiring that its acts or steps be performed in a specific order. Accordingly, where a process or method claim does not actually recite an order to be followed by its acts or steps or it is not otherwise specifically recited in the claims or descriptions of the subject disclosure that the steps are to be limited to a specific order, it is in no way intended that an order be inferred, in any respect. This holds for any possible nonexpress basis for interpretation, including: matters of logic with respect to the arrangement of steps or operational flow; plain meaning derived from grammatical organization or punctuation; the number or type of aspects described in the specification or annexed drawings; or the like.

[0228] As used in this disclosure, including the annexed drawings, the terms “component,” “module,” “interface,” “system,” and the like are intended to refer to a computer-related entity or an entity related to an apparatus with one or more specific functionalities. The entity can be either hardware, a combination of hardware and software, software, or software in execution. One or more of such entities are also referred to as “functional elements.” As an example, a component can be a process running on a processor, a processor, an object, an executable, a thread of execution, a program, and / or a computer. For example, both an application running on a server or network controller, and the server or network controller can be a component. One or more components can reside within a process and / or thread of execution and a component can be localized on one computer and / or distributed between two or more computers. Also, these components can execute from various computer readable media having various data structures stored thereon. The components can communicate via local and / or remote processes such as in accordance with a signal having one or more data packets (e.g., data from one component interacting with another component in a local system, distributed system, and / or across a network such as the Internet with other systems via the signal). As another example, a component can be an apparatus with specific functionality provided by mechanical parts operated by electric or electronic circuitry, which parts can be controlled or otherwise operated by program code executed by a processor. As yet another example, a component can be an apparatus that provides specific functionality through electronic components without mechanical parts, the electronic components can include a processor to execute program code that provides, at least partially, the functionality of the electronic components. As still another example, interface(s) can include I / O components or Application Programming Interface (API) components. While the foregoing examples are directed to aspects of a component, the exemplified aspects or features also apply to a system, module, and similar.

[0229] In addition, the term “or” is intended to mean an inclusive “or” rather than an exclusive “or.” That is, unless specified otherwise, or clear from context, “X employs A or B” is intended to mean any of the natural inclusive permutations. That is, if X employs A; X employs B; or X employs both A and B, then “X employs A or B” is satisfied under any of the foregoing instances. Moreover, articles “a” and “an” as used in this specification and annexed drawings should be construed to mean “one or more” unless specified otherwise or clear from context to be directed to a singular form.

[0230] In addition, the terms “example” and “such as” are utilized herein to mean serving as an instance or illustration. Any aspect or design described herein as an “example” or referred to in connection with a “such as” clause is not necessarily to be construed as preferred or advantageous over other aspects or designs described herein. Rather, use of the terms “example” or “such as” is intended to present concepts in a concrete fashion. The terms “first,” “second,” “third,” and so forth, as used in the claims and description, unless otherwise clear by context, is for clarity only and doesn’t necessarily indicate or imply any order in time or space.

[0231] The term “processor,” as utilized in this disclosure, can refer to any computing processing unit or device comprising processing circuitry that can operate on data and / or signaling. A computing processing unit or device can include, for example, single-core processors; single-processors with software multithread execution capability; multi-core processors; multicore processors with software multithread execution capability; multi-core processors with hardware multithread technology; parallel platforms; and parallel platforms with distributed shared memory. Additionally, a processor can include an integrated circuit, an application specific integrated circuit (ASIC), a digital signal processor (DSP), a field programmable gate array (FPGA), a programmable logic controller (PLC), a complex programmable logic device (CPLD), a discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. In some cases, processors can exploit nanoscale architectures, such as molecular and quantum-dot based transistors, switches and gates, in order to optimize space usage or enhance performance of user equipment. A processor may also be implemented as a combination of computing processing units. In addition, or as an alternative, a processor may be implemented as a virtual machine (or virtual processor) where a host device can provide a software environment in which the virtual processor shares computing resources of the host device with other virtual processors and / or components of the host device. The software environment may be referred to as a virtualized environment and permits the virtual processor to perform operations by executing processor-executable instructions retained in a portion of the underlying computing resources. As is described herein, the computing resources may include, for example, an operating system (O / S), CPUs, memory, disk space, incoming bandwidth, and / or outgoing bandwidth.

[0232] In addition, terms such as “store,” “data store,” data storage,” “database,” and substantially any other information storage component relevant to operation and functionality ofa component, refer to “memory components,” or entities embodied in a “memory” or components comprising the memory. It will be appreciated that the memory components described herein can be either volatile memory or nonvolatile memory, or can include both volatile and nonvolatile memory. Moreover, a memory component can be removable or affixed to a functional element (e.g., device, server).

[0233] Simply as an illustration, nonvolatile memory can include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM), which acts as external cache memory. By way of illustration and not limitation, RAM is available in many forms such as synchronous RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), and direct Rambus RAM (DRRAM). Additionally, the disclosed memory components of systems or methods herein are intended to comprise, without being limited to comprising, these and any other suitable types of memory.

[0234] Various aspects described herein can be implemented as a method, system, device, apparatus, or article of manufacture using standard programming and / or engineering techniques. In addition, various of the aspects disclosed herein also can be implemented by means of program modules or other types of computer program instructions stored in a memory device and executed by a processor, or other combination of hardware and software, or hardware and firmware. Such program modules or computer program instructions can be loaded onto a general purpose computer, a special purpose computer, or another type of programmable data processing apparatus to produce a machine, such that the instructions which execute on the computer or other programmable data processing apparatus create a means for implementing the functionality of disclosed herein.

[0235] The term “article of manufacture” as used herein is intended to encompass a computer program accessible from any computer- readable device, carrier, or media. For example, computer readable media can include but are not limited to magnetic storage devices (e.g., hard drive disk, floppy disk, magnetic strips, or similar), optical discs (e.g., compact disc (CD), digital versatile disc (DVD), blu-ray disc (BD), or similar), smart cards, and flash memory devices (e.g., card, stick, key drive, or similar).

[0236] What has been described above includes examples of one or more aspects of the disclosure. It is, of course, not possible to describe every conceivable combination of components or methodologies for purposes of describing these examples, and it can be recognized that many further combinations and permutations of the present aspects are possible. Accordingly, the aspects disclosed and / or claimed herein are intended to embrace all such alterations, modifications and variations that fall within the spirit and scope of the detailed description and the appended claims. Furthermore, to the extent that one or more of the terms “includes,” “including,” “has,” “have,” or “having” is used in either the detailed description or the claims, such term is intended to be inclusive in a manner similar to the term “comprising” as “comprising” is interpreted when employed as a transitional word in a claim.

Claims

1. CLAIMS:

1. A method, comprising: for a signal sampling event, performing digital signal processing (DSP) operations comprising, obtaining a first AC component of raw photoplethysmography (PPG) signal from a first color channel; obtaining a first DC component of the raw signal from the first color channel; obtaining a second AC component of raw signal from a second color channel; obtaining a second DC component of the raw signal from the second color channel; conditioning the first AC component by applying an adapative bandpass filter based on a heart rate of a subject, resulting on a first conditioned AC component; conditioning the second AC component by applying the adaptive bandpass filter based on the heart rate of the subject, resulting on a second conditioned AC component; determining a perfusion index for the first color channel as a first ratio of the first conditioned AC component and the first DC component; determining a perfusion index for the second color channel as a second ratio of the second conditioned AC component and the second DC component; determining a value of a ratio-of-ratios (ROR) for the signal sampling event as the ratio of the first ratio and the second ratio; repeating the DSP operations for each one of multiple next signal sampling events until multiple values of the ROR are determined; determining a single value of ROR by applying an aggregation filter to the multiple values of the ROR; and determining, using the single value of ROR, a value of peripheral oxygen saturation (SpO2) of the subject.

2. The method of claim 1, further comprising providing the value of SpO2of the subject.

3. The method of any of claim 1 or claim 2, wherein the providing comprises causing a computing device to present the value of SpO2of the subject, and wherein presenting the valueof SpO2of the subject comprises one or more of displaying visual indicia indicative of the value, presenting aural indicia indicative of the value, or presenting a haptic stimulus representative of the value.

4. The method of any of the preceding claims, wherein the first color channel comprises a first light source device configured to emit light of a first light wavelength, and wherein the second color channel comprises a second light source device configured to emit light a second light wavelength.

5. The method of any of the preceding claims wherein the first light wavelength lies in the red portion of the electromagnetic (EM) radiation spectrum, and wherein the second light wavelength lies in the infrared (IR) portion of the EM radiation spectrum.

6. A wearable device, comprising: at least one processor; at least one memory device coupled with the at least one processor, the at least one memory device retaining processor-executable instructions that, in response to execution by the at least one processor, individually or in combination, cause the wearable device to: for a signal sampling event, perform digital signal processing (DSP) operations comprising, obtaining a first AC component of raw photoplethysmography (PPG) signal from a first color channel; obtaining a first DC component of the raw signal from the first color channel; obtaining a second AC component of raw signal from a second color channel; obtaining a second DC component of the raw signal from the second color channel; conditioning the first AC component by applying an adapative bandpass filter based on a heart rate of a subject, resulting on a first conditioned AC component;conditioning the second AC component by applying the adaptive bandpass filter based on the heart rate of the subject, resulting on a second conditioned AC component; determining a perfusion index for the first color channel as a first ratio of the first conditioned AC component and the first DC component; determining a perfusion index for the second color channel as a second ratio of the second conditioned AC component and the second DC component; determining a value of a ratio-of-ratios (ROR) for the signal sampling event as the ratio of the first ratio and the second ratio; repeat the DSP operations for each one of multiple next signal sampling events until multiple values of the ROR are determined; determine a single value of ROR by applying an aggregation filter to the multiple values of the ROR; and determine, using the single value of ROR, a value of peripheral oxygen saturation (SpO2) of the subject.

7. The wearable device of claim 6, wherein the first color channel comprises a first light source device configured to emit light of a first light wavelength, and wherein the second color channel comprises a second light source device configured to emit light a second light wavelength.

8. The wearable device of claim 7, wherein the first light wavelength lies in the red portion of the electromagnetic (EM) radiation spectrum, and wherein the second light wavelength lies in the infrared (IR) portion of the EM radiation spectrum.

9. The wearable device of claim any of the preceding claims, further comprising multiple optical sensor devices and at least two light source devices.

10. The wearable device of any of the preceding claims, further comprising a feedback device, wherein the at least one memory device retains further processor-executable instructionsthat, in response to execution by the at least one processor, individually or in combination, further cause the wearable device to provide, via the feedback device, the value of SpCh of the subject.

11. The wearable device of claim 10, wherein providing the value of SpCh comprises causing the feedback device to present the value of SpCh of the subject by one or more of displaying visual indicia indicative of the value, outputting aural indicia indicative of the value, or providing a haptic stimulus representative of the value.

12. At least one non- transitory computer- readable storage medium having processorexecutable instructions encode thereon that, in response to execution by at least one processor, individually or in combination, cause a wearable device to: for a signal sampling event, perform digital signal processing (DSP) operations comprising, obtaining a first AC component of raw photoplethysmography (PPG) signal from a first color channel; obtaining a first DC component of the raw signal from the first color channel; obtaining a second AC component of raw signal from a second color channel; obtaining a second DC component of the raw signal from the second color channel; conditioning the first AC component by applying an adapative bandpass filter based on a heart rate of a subject, resulting on a first conditioned AC component; conditioning the second AC component by applying the adaptive bandpass filter based on the heart rate of the subject, resulting on a second conditioned AC component; determining a perfusion index for the first color channel as a first ratio of the first conditioned AC component and the first DC component; determining a perfusion index for the second color channel as a second ratio of the second conditioned AC component and the second DC component; determining a value of a ratio-of-ratios (ROR) for the signal sampling event as the ratio of the first ratio and the second ratio;repeat the DSP operations for each one of multiple next signal sampling events until multiple values of the ROR are determined; determine a single value of ROR by applying an aggregation filter to the multiple values of the ROR; and determine, using the single value of ROR, a value of peripheral oxygen saturation (SpO2) of the subject.

13. The at least one non- transitory computer-readable storage medium of claim 12, wherein the wearable device comprises a feedback device, at least one non-transitory computer-readable storage medium having further processor-executable instructions encoded thereon that, in response to execution by the at least one processor, individually or in combination, further cause the wearable device to provide, via the feedback device, the value of SpCh of the subject.

14. The at least one non-transitory computer-readable storage medium of claim 13, wherein providing the value of SpCh comprises causing the feedback device to present the value of SpCh of the subject by one or more of displaying visual indicia indicative of the value, outputting aural indicia indicative of the value, or providing a haptic stimulus representative of the value.

Images