Machine-learned models for denoising biometric signals collected using mobile biometric sensors

Abstract

Training data is generated for training a machine-learned biometric noise reduction model. Generating the training data comprises obtaining information descriptive of a plurality of clean biometric signals measured using a corresponding medical device. For each clean biometric signals, noise is sampled from noise source(s) of a plurality of candidate noise sources. The noise is applied to at least a portion of the clean biometric signal. The training data is processed with the machine-learned biometric noise reduction model to obtain a plurality of denoised biometric signals. The machine-learned biometric noise reduction model is trained based on a loss function that evaluates a difference between the plurality of denoised biometric signals and the plurality of clean biometric signals.

Description

MACHINE-LEARNED MODELS FOR DENOISING BIOMETRIC SIGNALSCOLLECTED USING MOBILE BIOMETRIC SENSORSPRIORITY CLAIM

[0001] The present application is based on and claims priority to United States Provisional Application Number 63 / 573,239 having a filing date of April 2, 2024. Application claims priority to and the benefit of each of such applications and incorporates all such applications herein by reference in their entirety.FIELD

[0002] The present disclosure relates generally to denoising biometric signals. More particularly, the present disclosure relates to machine-learned models for removing noise from biometric signals collected using mobile biometric sensors.BACKGROUND

[0003] In recent years, biometric sensors have become an increasingly important area of competition for developers of mobile computing devices (e.g., smartphones, smart watches, wearable devices, etc.). Some mobile computing devices are capable of collecting biometric signals that were previously only collected using purpose-built medical devices. For example, blood oxygen levels have historically been measured using a purpose-built medical device such as a pulse oximeter that attaches to a user’s finger. However, some recently developed mobile computing devices include sensors capable of regularly measuring blood oxygen levels while the device is being worn by the user. For example, a wearable smart watch device can include a pulse oximeter sensor on the bottom surface of the watch such that the sensor directly contacts the user’s wrist. In such fashion, biometric information can be regularly collected from a user to detect health conditions, provide statistical information to the user, etc.SUMMARY

[0004] Aspects and advantages of embodiments of the present disclosure will be set forth in part in the following description, or can be learned from the description, or can be learned through practice of the embodiments.

[0005] One example aspect of the present disclosure is directed to a computer- implemented method. The method includes generating, by a computing system comprisingone or more processor devices, training data for training a machine-learned biometric noise reduction model, wherein generating the training data comprises obtaining, by the computing system, information descriptive of a plurality of clean biometric signals measured using a corresponding medical device. The method includes, for each clean biometric signal of the plurality of clean biometric signals, sampling, by the computing system, noise from one or more noise sources of a plurality of candidate noise sources. The method includes, for each clean biometric signal of the plurality of clean biometric signals, applying, by the computing system, the noise to at least a portion of the clean biometric signal. The method includes processing, by the computing system, the training data with the machine-learned biometric noise reduction model to obtain a plurality of denoised biometric signals. The method includes training, by the computing system, the machine-learned biometric noise reduction model based on a loss function that evaluates a difference between the plurality of denoised biometric signals and the plurality of clean biometric signals.

[0006] Another example aspect of the present disclosure is directed to a user computing device, comprising one or more processor devices and one or more tangible, non-transitory computer readable media storing computer-readable instructions that when executed by the one or more processor devices cause the user computing device to perform operations. The operations include obtaining a noisy biometric signal via a mobile biometric measurement device associated with the user computing device, wherein the noisy biometric signal comprises a first quantity of noise. The operations include causing the noisy biometric signal to be processed with a machine-learned biometric noise reduction model to obtain a denoised biometric signal comprising a second quantity of noise less than the first quantity of noise, wherein the machine-learned biometric noise reduction model is trained based on a loss function that evaluates a difference between a plurality of noisy biometric training signals and a plurality of clean biometric ground-truth signals, and wherein the plurality of noisy biometric training signals are generated by applying noise from randomly selected noise sources to randomly selected portions of the clean biometric signals. The operations include providing biometric measurements for display to a user of the user computing device, wherein the biometric measurements are based at least in part on the denoised biometric signal.

[0007] Another example aspect of the present disclosure is directed to a One or more tangible, non-transitory computer readable media storing computer-readable instructions that when executed by one or more processor devices cause the one or more processor devices to perform operations. The operations include generating training data for training a machine-learned biometric noise reduction model, wherein generating the training data comprises obtaining information descriptive of a plurality of clean biometric signals measured using a corresponding medical device. The operations include, for each clean biometric signal of the plurality of clean biometric signals, selecting at least a portion of the clean biometric signal. The operations include, for each clean biometric signal of the plurality of clean biometric signals, sampling noise from one or more noise sources of a plurality of candidate noise sources. The operations include, for each clean biometric signal of the plurality of clean biometric signals, applying the noise to the at least the portion of the clean biometric signal. The operations include processing the training data with the machine-learned biometric noise reduction model to obtain a plurality of denoised biometric signals. The operations include training the machine-learned biometric noise reduction model based on a loss function that evaluates a difference between the plurality of denoised biometric signals and the plurality of clean biometric signals.

[0008] Other aspects of the present disclosure are directed to various systems, apparatuses, non-transitory computer-readable media, user interfaces, and electronic devices.

[0009] These and other features, aspects, and advantages of various embodiments of the present disclosure will become better understood with reference to the following description and appended claims. The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate example embodiments of the present disclosure and, together with the description, serve to explain the related principles.BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Detailed discussion of embodiments directed to one of ordinary skill in the art is set forth in the specification, which makes reference to the appended figures, in which:

[0011] Figures 1 A, IB, and 1C each illustrate a perspective view of an example, nonlimiting computing device (e.g., wearable computing device) according to one or more example embodiments of the present disclosure.

[0012] Figure 2 is a block diagram of an environment suitable for implementing machine-learned models for denoising biometric signals collected using mobile biometric sensors according to some implementations of the present disclosure.

[0013] Figure 3 is a block diagram for the training data generator of the model training module of Figure 2 according to some implementations of the present disclosure.

[0014] Figure 4 illustrates a block diagram of an example, non-limiting wearable device according to one or more example embodiments of the present disclosure.

[0015] Figure 5 illustrates a diagram of an example, non-limiting user assessment management system according to one or more example embodiments of the present disclosure.

[0016] Figure 6 depicts a flow chart diagram of an example method to perform training of machine-learned models for denoising biometric signals collected using mobile biometric sensors according to example embodiments of the present disclosure.

[0017] Figure 7 depicts a flow chart diagram of an example method to utilize machine- learned models for denoising biometric signals collected using mobile biometric sensors according to example embodiments of the present disclosure according to example embodiments of the present disclosure.

[0018] Reference numerals that are repeated across plural figures are intended to identify the same features in various implementations.DETAILED DESCRIPTIONOverview

[0019] Generally, the present disclosure is directed to denoising biometric signals. More particularly, the present disclosure relates to removing noise from biometric signals collected using mobile biometric sensors. In recent years, accurate biometric sensors have become an increasingly competitive area of development for wearable computing devices. Some wearable computing devices are capable of collecting biometric signals that were previously only collected using purpose-built medical devices. For example, blood oxygen levels have historically been measured using a purpose-built medical device such as a pulse oximeter that attaches to a user’s finger. However, recent wearable computing devices include sensors capable of measuring blood oxygen levels while the device is being worn by the user. For example, a wearable smart watch device can include a pulse oximeter sensor on the bottom surface of the watch such that the sensor directly contacts the user’s wrist.

[0020] In some instances, conventional medical sensors and / or signal collection techniques can be incorporated into wearable computing devices without requiring substantial modifications. To follow the previous example, both the wearable smart watch attached to the user’s wrist and the pulse oximeter attached to the user’s finger can measure blood oxygen by projecting infrared light into the user’s skin. However, other types of medical sensors must bemodified substantially before they can be incorporated within certain wearable computing devices. For example, electrocardiogram (ECG) signals are collected in medical contexts by placing a large quantity of electrodes (e.g., 8-12) on different areas of a user’s body, and then measuring each electrode to obtain the ECG signal. Conversely, a wearable computing device, such as a smartwatch, is usually limited to contacting a single area of the user’s body (e.g., an electrode on the bottom surface of the watch such that the electrode contacts the user’s wrist, etc.).

[0021] This problem is further exacerbated by additional limitations that are inherent to wearable computing devices. Specifically, wearable computing devices are often constructed with a minimal form factor, and as such, suffer from the constraints typical to such form factors (e.g., reduced cost, size, power, storage, memory, etc.). Because of these limitations and constraints, biometric signals collected with wearable computing devices are generally noisy and difficult to interpret.

[0022] Similarly, biometric signals collected using purpose-built medical devices generally include a relatively small quantity of noise. In most instances, this noise is sufficiently small for conventional denoising techniques to rectify. For example, noise typically found in ECG signals collected with a purpose-built medical ECG device can be reduced using conventional techniques such as filters, etc. However, the noise present in biometric signals collected with wearable computing devices is much higher, and conventional noise reduction techniques are often insufficient.

[0023] Machine learning techniques have been successfully applied across a variety of technologies to reduce noise that could not be sufficiently reduced using conventional techniques. In particular, machine-learned models trained to reduce noise via supervised learning have proven to be particularly effective at reducing signal noise. However, because biometric sensors incorporated into wearable computing devices are a recent development, there is no existing training data sufficient to train a model to reduce the noise present in biometric signals collected with wearable computing devices. As such, training data sufficient to train such a model is greatly desired.

[0024] Accordingly, implementations described herein propose machine-learned models for denoising biometric signals collected using mobile biometric sensors. Specifically, a computing system can obtain a plurality of clean biometric signals measured using a corresponding medical device (e.g., ECG signals collected with a medical ECG device, etc.). The computing system can generate training data by applying noise to the clean biometric signals. Specifically, for each clean biometric signal, the computing system can sample noisefrom a random selection of noise sources. The computing system can select a random portion of the signal, and can apply the sampled noise to the portion of the signal. The computing system can process the training data with the model to obtain denoised biometric signals. The computing system can then train the model based on a loss function that evaluates a difference between the plurality of denoised biometric signals and the plurality of clean biometric signals.

[0025] Aspects of the present disclosure provide a number of technical effects and benefits. As one example technical effect and benefit, implementations described herein sufficiently optimize biometric signals such that the biometric signals can be analyzed to improve user health and user safety. Specifically, regular analysis of some biometric signals, such as ECG and PPGs signals, can offer a variety of benefits (e.g., earlier detection of health complications, optimizing sleep or exercise routines, etc.). Noisy biometric signals must be sufficiently de-noised before they are analyzed. However, when collected using mobile biometric sensors, conventional noise reduction techniques have proven insufficient. Accordingly, implementations described herein provide for machine-learned models that are capable of sufficiently denoising biometric signals collected using mobile biometric sensors, thus enabling analysis of noisy biometric signals and the variety of benefits provided therefrom.

[0026] With reference now to the Figures, example embodiments of the present disclosure will be discussed in further detail.Example Devices and Systems

[0027] Figures 1 A, IB, and 1C each illustrate a perspective view of an example, nonlimiting computing device (e.g., wearable computing device) 100 according to one or more example embodiments of the present disclosure. In example embodiments described herein, wearable device 100 can constitute and / or include a wearable computing device. For instance, in these or other example embodiments, wearable device 100 can constitute and / or include a wearable computing device such as, for example, a wearable physiological monitoring device that can be worn by a user (also referred to herein as a “wearer”) and / or capture one or more types of physiological data of the user (e.g., heart rate (HR) data, motion data (e.g., accelerometer data), body temperature data, respiration rate data, blood pressure data, blood oxygenation level data, deoxyribonucleic acid (DNA) data, electrodermal activity (EDA) data, stress related data, etc.).

[0028] Wearable device 100 according to example embodiments of the present disclosure can include a display 102, an attachment component 104, a securement component 106, and a button 108 that can be located on a side of wearable device 100. In at least one embodiment, two sides of display 102 can be coupled (e.g., mechanically, operatively) to attachment component 104. In some embodiments, securement component 106 can be located on, coupled to (e.g., mechanically, operatively), and / or integrated with attachment component 104. In these or other embodiments, securement component 106 can be positioned opposite display 102 on an opposing end of attachment component 104. In some embodiments, button 108 can be located on a side of wearable device 100, underneath display 102.

[0029] Display 102 according to example embodiments described herein can constitute and / or include any type of electronic display or screen known in the art. For example, in some embodiments, display 102 can constitute and / or include a liquid crystal display (LCD) or organic light emitting diode (OLED) display such as, for instance, a transmissive LCD display or a transmissive OLED display. Display 102 according to example embodiments can be configured to provide brightness, contrast, and / or color saturation features according to display settings that can be maintained by control circuitry and / or other internal components and / or circuitry of wearable device 100. In some embodiments, display 102 can constitute and / or include a touchscreen such as, for instance, a capacitive touchscreen. For example, in these embodiments, display 102 can constitute and / or include a surface capacitive touchscreen or a projective capacitive touch screen that can be configured to respond to contact with electrical charge-holding members or tools, such as a human finger.

[0030] In some embodiments, display 102 can be configured to provide (e.g., render) a variety of information such as, for example, the time, the date, body signals (e.g., physiological data of a user wearing wearable device 100), readings based upon user input, and / or other information. In one embodiment, such body signals can include, but are not limited to, heart rate data (e.g., heart beats per minute), motion data (e.g., movement data, accelerometer data), blood pressure data, body temperature data, respiration rate data, blood oxygenation level data, deoxyribonucleic acid (DNA) data, electrodermal activity (EDA) data, stress related data and / or any other body signal that one of ordinary skill in the art would understand that can be measured by a wearable device such as, for instance, wearable device 100. In some embodiments, the readings based upon user input can include, but are not limited to, the number of steps a user has taken, the distance traveled by the user, the sleep schedule of the user, travel routes of the user, elevation climbed by the user, and / or anyother metric that one of ordinary skill in the art would understand that can be input by a user into a wearable device such as, for instance, wearable device 100.

[0031] In some embodiments, the display 102 can be configured to display a low- power operating domain interface or a high-performance operating domain interface, or to switch between the two. In some implementations, the low-power operating domain interface can be similar, or identical to, the high-performance operating domain interface. For example, both the low-power operating domain interface and the high-performance operating domain interface can display the same interface elements, but the high-performance operating domain interface can utilize a higher brightness parameter value (e.g., making the display more luminous) and can update real-time data elements at a higher frequency than the update frequency of the low-power operating domain interface.

[0032] Additionally, or alternatively, in some implementations, the low-power operating domain interface can be different than the high-performance operating domain interface. For example, to minimize power consumption, the low-power operating domain interface can include interface elements that can be fully operated via a low-power processor device of the wearable device 100, such as a time indicator, a weather indicator, a date indicator, etc. Conversely, the high-performance operating domain interface can include a number of real-time data elements that depict real-time data, such as biometric data (e.g., a heartbeat of a user, motion information, etc.), stock quotes, calendar information (e.g., scheduled meetings, etc.), etc.

[0033] The low-power operating domain interface can be displayed with the display 102 when the wearable device 100 is operating in a low-power operating domain. The low- power operating domain can be implemented when a user is not interacting with the wearable device 100, or is not predicted to imminently begin interacting with the wearable device 100. As such, the low-power operating domain interface can generally be displayed via the display 102 when the wearable device 100 is not directly observed by the user. In other words, when the user is not actively looking at the wearable device 100, the low-power operating domain interface can be displayed.

[0034] In at least one embodiment of the present disclosure, the above-described body signals and / or readings based upon user input can be used to calculate further analytics to provide a user with data such as, for instance, a fitness score, a sleep quality score, a number of calories burned by the user, and / or other data. In some embodiments, wearable device 100 can take in (e.g., capture, collect, receive, measure) outside data irrespective of the user such as, for example: an ambient temperature of an environment surrounding and / or external towearable device 100; an amount of sun exposure wearable device 100 is subjected to; an atmospheric pressure of the environment surrounding and / or external to wearable device 100; an air quality of the environment surrounding and / or external to wearable device 100; the location of wearable device 100 based on, for instance, a global positioning system (GPS); and / or other outside factors that one of ordinary skill in the art would understand a wearable device such as, for instance, wearable device 100 can take in (e.g., capture, collect, receive, measure).

[0035] Attachment component 104 according to example embodiments described herein can be used to attach (e.g., affix, fasten) wearable device 100 to a user of wearable device 100. In some embodiments, attachment component 104 can take the form of, for example, a strap, an elastic band, a rope, and / or any other form of attachment one of ordinary skill in the art would understand can be used to attach a wearable device such as, for instance, wearable device 100 to a user.

[0036] Securement component 106 according to example embodiments of the present disclosure can facilitate attachment of attachment component 104 upon a user of wearable device 100. In some embodiments, securement component 106 can include, but is not limited to, a pin and hole locking mechanism (e.g., a buckle), a magnet system, a lock, a clip, and / or any other type of securement that one of ordinary skill would understand can be used to facilitate attachment of a wearable device such as, for instance, wearable device 100 to a user. In one embodiment, wearable device 100 does not include securement component 106. For example, in this or another embodiment, wearable device 100 can be secured to a user with a strap that can be tied around the user’s wrist and / or another suitable appendage.

[0037] Button 108 according to example embodiments described herein can allow for a user to interact with wearable device 100 and / or allow for the user to provide a form of input into wearable device 100. In the example embodiment depicted in Figures 1 A, IB, and 1C, one button 108 is shown on wearable device 100. However, it should be appreciated that wearable device 100 is not so limiting. For example, in some embodiments, wearable device 100 can include any number of buttons that allow a user to further interact with wearable device 100 and / or to provide alternative inputs. In at least one embodiment, wearable device 100 does not include button 108. For instance, as described above, in example embodiments, wearable device 100 can include a screen such as, for example, a touch screen that can receive inputs through (e.g., by way of) the touch of the user. In additional or alternative embodiments, wearable device 100 can include a microphone that can receive inputs through (e.g., by way of) voice commands of a user.

[0038] Some implementations of the present disclosure are described in the context of a user intending a particular type of interaction with the wearable device 100 that requires the wearable device 100 to operate in the high-performance operating domain. In some implementations, this particular type of interaction can include interacting with the button 108. Alternatively, in some implementations, interacting with the button 108 is considered to be the particular type of interaction depending on which operations (if any) are performed in response to the button 108 being pressed. For example, if a gaming application is selected and the user interacts with the button 108 (and thus executes the application), the interaction can be considered to be of the particular type of interaction. For another example, if a “sleep” feature is highlighted and the user interacts with the button 108, the user interaction with the button 108 may not be considered to be of the particular type of interaction.

[0039] Additionally, or alternatively, in some implementations, the device can transition to a high-performance operating domain (e.g., increasing screen brightness, modifying operating parameters of processor devices of the device (e.g., clock speed, available cores, virtualized instances, etc.), etc.) independent of a user input. For example, if biometric signals (e.g., an ECG signal, a PPG signal, etc.) obtained via the device are denoised and analyzed to identify a health condition, the device can transition to the high- performance operating domain to alert the user.

[0040] In some embodiments, wearable device 100 can constitute a portable computing device that can be designed so that it can be inserted into a wearable case (e.g., as illustrated in the example embodiments depicted in Figures 1A, IB, and 1C). In some embodiments, wearable device 100 can constitute a portable computing device that can be designed so that it can be inserted into one or more of multiple different wearable cases (e.g., a wristband case, a belt-clip case, a pendant case, a case configured to be attached to a piece of exercise equipment such as a bicycle). Wearable device 100 according to embodiments described herein can be formed into one or more shapes and / or sizes to allow for coupling to (e.g., secured to, worn, borne by) the body or clothing of a user. In some embodiments, wearable device 100 can constitute a portable computing device that can be designed to be worn in limited manners such as, for instance, a computing device that is integrated into a wristband in a non-removable manner and / or can be intended to be worn specifically on a person's wrist (or perhaps ankle).

[0041] Irrespective of configuration, wearable device 100 according to example embodiments of the present disclosure can include one or more physiological and / or environmental sensors (e.g., internal physiological sensor(s) 143, external physiologicalsensor(s) 145, and / or environmental sensor(s) 155) that can be configured to collect physiological and / or environmental data in accordance with various embodiments disclosed herein. In some embodiments, wearable device 100 can be configured to analyze and / or interpret collected physiological and / or environmental data to perform one or more health, wellness, and / or well-being assessments (e.g., physical, mental, emotional, behavioral, and / or sleep quality assessment(s)) of a user (e.g., a wearer) of wearable device 100 according to one or more embodiments described herein. In additional and / or alternative embodiments, wearable device 100 can be configured to communicate with another computing device or server that can perform such one or more health, wellness, and / or well-being assessments (e.g., physical, mental, emotional, behavioral, and / or sleep quality assessment(s)) of a user (e.g., a wearer) of wearable device 100 according to one or more embodiments described herein.

[0042] Wearable device 100 in accordance with one or more example embodiments of the present disclosure can include one or more physiological and / or environmental components and / or modules that can be designed to determine one or more physiological and / or environmental metrics associated with a user (e.g., a wearer) of wearable device 100. In at least one embodiment, such physiological and / or environmental component(s) and / or module(s) can constitute and / or include one or more physiological and / or environmental sensors. For instance, although not depicted in the example embodiments illustrated in Figures 1 A, IB, and 1C, in some embodiments, wearable device 100 can include one or more physiological and / or environmental sensors such as, for example, an accelerometer, a heart rate sensor (e.g., photoplethysmography (PPG) sensor), an electrodermal activity (EDA) sensor, an Electrocardiogram (ECG) sensor, a body temperature sensor, an environment temperature sensor, and / or another physiological and / or environmental sensor. In these or other embodiments, such physiological and / or environmental sensor(s) can be disposed on, coupled to, and / or otherwise be associated with an underside and / or a backside (e.g., back 134) of wearable device 100.

[0043] In some embodiments, the above-described physiological and / or environmental sensor(s) can be disposed on, coupled to, and / or otherwise be associated with wearable device 100 such that the sensor(s) can be in contact with or substantially in contact with human skin when wearable device 100 is worn by a user. For example, in embodiments where wearable device 100 can be worn on a user’s wrist, the physiological and / or environmental sensor(s) can be disposed on, coupled to, and / or otherwise be associated with back 134 that can be substantially opposite display 102 and touching an arm of the user. Inone embodiment, the above-described physiological and / or environmental sensor(s) can be disposed on, coupled to, and / or otherwise be associated with an interior or skin-side of wearable device 100 (e.g., a side of wearable device 100 that contacts, touches, and / or faces the skin of the user such as, for instance, back 134 and / or bottom 142). In another embodiment, the physiological and / or environmental sensors can be disposed on one or more sides of wearable device 100, including the skin-side (e.g., back 134, bottom 142) and one or more sides (e.g., first side 136, second side 138, top 140, display 102) of wearable device 100 that face and / or are exposed to the ambient environment (e.g., the external environment surrounding wearable device 100).

[0044] Figure 2 is a block diagram of an environment suitable for implementing machine-learned models for denoising biometric signals collected using mobile biometric sensors according to some implementations of the present disclosure. In some implementations, a computing environment 200 includes a computing system 202. The computing system 202 can be a system for training, providing, or otherwise implementing various machine-learned models for reducing noise in biometric signals, and can include a processor device(s) 204 and memory 206. In some implementations, the computing system 202 may be a computing system that includes multiple computing devices. Alternatively, in some implementations, the computing system 202 may be one or more computing devices within the computing environment 200 that includes multiple distributed devices and / or systems. Similarly, the processor device(s) 204 may include any computing or electronic device capable of executing software instructions to implement the functionality described herein.

[0045] The memory 206 can be or otherwise include any device(s) capable of storing data, including, but not limited to, volatile memory (random access memory, etc.), nonvolatile memory, storage device(s) (e.g., hard drive(s), solid state drive(s), etc.). In particular, the memory 206 can include a containerized unit of software instructions (i.e., a “packaged container”). The containerized unit of software instructions can collectively form a container that has been packaged using any type or manner of containerization technique.

[0046] The containerized unit of software instructions can include one or more applications, and can further implement any software or hardware necessary for execution of the containerized unit of software instructions within any type or manner of computing environment. For example, the containerized unit of software instructions can include software instructions that contain or otherwise implement all components necessary forprocess isolation in any environment (e.g., the application, dependencies, configuration files, libraries, relevant binaries, etc.).

[0047] In some implementations, the memory 206 of the computing system 202 can include a machine learning module 208. The machine learning module 208 can train, instantiate, de-instantiate, optimize, store, or otherwise implement various machine-learned model(s). To do so, the machine learning module 208 can include a model repository 210. The model repository 210 can store untrained, partially trained, or fully trained machine- learned models. The model repository 210 can also catalog information related to the machine-learned models (e.g., versioning information, performance metrics, predicted computing resource costs, etc.).

[0048] The model repository 210 can include machine-learned models 212 for reducing noise in biometric signals. For example, the model repository 210 may include a machine-learned ECG noise reduction model 212A trained to reduce noise in ECG signals collected with mobile biometric sensors. For another example, the model repository 210 may include a machine-learned PPG noise reduction model 212B trained to reduce noise in PPG signals collected with mobile biometric sensors. For yet another example, the model repository 210 may include a machine-learned multimodal noise reduction model 212C trained to reduce noise in multiple types of biometric signals collected with mobile biometric sensors.

[0049] The machine learning module 208 can include a model training module 214. The model training module 214 can be configured to implement training of machine learned models for utilization by the computing system 202 or other computing device(s) within the computing environment 200. For example, the model training module 214 can utilize training data to train a model for provision to a wearable computing device, mobile computing device, etc.

[0050] The model training module 214 can include training data 216. The training data 216 can include, or describe, training examples for supervised learning, data for unsupervised learning, etc. Specifically, in some implementations, the training data 216 can include training examples 218 for supervised training of the machine-learned models 212. The training examples 218 can include noisy biometric signals 220 (e.g., ECG signals, PPG signals, etc.) and corresponding clean biometric signals 222 to serve as ground truth signals. The model training module 214 can train the machine-learned models 212 using the training data 216.

[0051] Specifically, model training module 214 can train the machine-learned models to transform the noisy biometric signals 220 from the training examples 218 to their corresponding clean biometric signals 222 (e.g., “ground-truth” signals). For example, the model training module 214 can process a noisy ECG signal from the noisy biometric signals 220 with the machine-learned ECG noise reduction model 212A to generate a de-noised biometric signal. The model training module 214 can train the machine-learned ECG noise reduction model 212A based on a difference between the de-noised biometric signal and the clean ground-truth biometric signal of the training example.

[0052] To generate the training data 216, the model training module 214 can include a training data generator 224. The training data generator 224 can include clean biometric signal information 226. The clean biometric signal information 226 can include, or otherwise describe, some or all of the clean biometric signals 222 of the training examples 218. Each of the clean biometric signals 222 can include a relatively small quantity of noise (e.g., noise that is less than a threshold degree of noise, etc.) or lack noise entirely. Additionally, or alternatively, in some implementations, the clean biometric signal information 226 can include, or otherwise describe, clean biometric signals captured using a medical device (e.g., a medical biometric sensor, etc.). For example, the clean biometric signal information 226 may describe an ECG signal captured using a medical ECG sensing device that measures an ECG signal from multiple electrodes placed at different locations of a patient’s body. For another example, the clean biometric signal information 226 may describe a PPG signal captured using a medical PPG sensing device that measures a PPG signal from sensor(s) placed on the tips of a patient’s finger(s). By using biometric signals known to be “clean,” such as those captured using purpose-built medical devices, implementations described herein can train denoising models to denoise noisy biometric signals such that the de-noised signals emulate the “clean” signals with regards to accuracy, noise inclusion, etc.

[0053] As such, the clean biometric signals 222 described by the clean biometric signal information 226 can serve as “ground truth” examples in the training examples 218 for the models 212 to emulate. Additionally, the clean biometric signals 222 can serve as the source for the corresponding noisy biometric signals 220 of the training examples 218 by adding noise to the clean biometric signals 222. More specifically, training machine-learned de-noising models, such as the machine-learned models 212, can be accomplished accurately and efficiently using supervised learning techniques. Supervised learning generally refers to a training process in which a machine-learned model generates an output, and the output is compared to a “ground-truth” output that is known to be correct. Model performance can beimproved over time by modifying parameters of the models to reduce differences between the model outputs and the ground-truth outputs. However, the large majority of biometric signals have conventionally been collected using purpose-built medical devices (e.g., ECG machines, PPG machines, etc.), which generally not capture significant quantities of noise. Further, the noise that is present in such signals can generally be de-noised using conventional techniques. As such, there currently exists a lack of training data sufficient to train machine-learned models for denoising biometric signals collected using mobile biometric sensors.

[0054] Accordingly, the training data generator 224 can generate synthetic “noisy” training examples for the training examples 218 by applying sampled noise to the clean biometric signal information 226. To do so, the training data generator 224 can include a noise source selector 228. The noise source selector 228 can include, or otherwise access, a plurality of noise sources 230A - 230N (generally, noise sources 230). As described herein, the “noise” within a signal generally refers to unwanted, unintended, irrelevant, etc. disturbances or fluctuations that obscure or distort the original information carried by the signal. Noise can arise from various sources, including electrical interference, environmental factors, imperfections in equipment, and inherent limitations of the signal transmission medium.

[0055] The noise sources 230 can include any type or manner of noise. In some implementations, the noise sources 230 can include white noise (e.g., broad-spectrum noise), pink noise (e.g., noise concentrated on a lower-frequency range of the spectrum), brown noise (e.g., noise even more strongly concentrated on the lower-frequency range of the spectrum), etc. For example, the noise source 230 can be an algorithm configured to replicate white noise in the form typically found within various biometric signals, or may include a number of algorithms each specifically configured to replicate white noise found within a particular type of biometric signal (e.g., an algorithm that replicates white noise found in ECG readings, another algorithm that replicates white noise found in PPG readings, etc.).

[0056] Additionally, or alternatively, in some implementations, the noise sources 230 can include a signal adjustment process. A signal adjustment process can refer to a process that modifies characteristics of the biometric signal to emulate noise that may be captured by a mobile biometric sensor. For example, the noise sources 230 may include a signal compression process that compresses a portion of the clean biometric signal. For another example, the noise sources 230 may include a signal expansion process that expands a portion of the clean biometric signal. For yet another example, the noise sources 230 mayinclude a signal cropping process that crops (i.e., removes) a portion of the clean biometric signal.

[0057] Additionally, or alternatively, in some implementations, the noise sources 230 can include noise sampled from noisy biometric signals collected using medical devices. Specifically, in some implementations, in addition to the clean biometric signal information 222, the training data generator 224 can obtain noisy biometric signal information captured using the same (or similar) medical devices used to capture the clean biometric signal information 222. For example, the noisy biometric signal information may be captured by identifying extracting the relatively small quantity of existing noise present in the clean biometric signal information 222 (if present). For another example, the noisy biometric signal information may be noisy signals caused by incorrect usage of the medical devices, usage of older or outdated medical devices, etc.

[0058] The noise source selector 228 can select one (or more) of the noise sources 230 to for application to some (or all) of the clean biometric signals 222 described by the clean biometric signal information 226. To do so, the noise source selector 228 can perform a noise selection process. In some implementations, the noise selection process can be partially or fully random. For example, the noise selection process can be a randomized noise selection process that first identifies a random quantity of noise sources to select, and then randomly selects that quantity of noise sources from the noise sources 230. Alternatively, in some implementations, the noise selection process can be a non-random or deterministic noise selection process. For example, the noise selection process can evaluate the clean biometric signals 222 described by the clean biometric signal information 226 and then, based on the evaluation, select particular noise source(s) from the noise sources 230.

[0059] In some implementations, the noise source selector 228 can select a quantity and / or type(s) of the noise sources 230 on a per-signal basis for each of the clean biometric signals 222 described by the clean biometric signal information 226 (e.g., repeating the noise selection process once for each of the noisy biometric signals 220). Additionally, or alternatively, in some implementations, the noise source selector 228 can select a quantity and / or type(s) of the noise sources 230 for multiple portions of each of the clean biometric signals 222 (e.g., repeating the noise selection process multiple times to apply noise to multiple portions of each of the noisy biometric signals 220). Alternatively, in some implementations, the noise source selector 228 can select a quantity and / or type(s) of the noise sources 230 for certain sets of the clean biometric signals 222, or may select a set of noise source(s) to be applied to all of the clean biometric signals 222 described by the cleanbiometric signal information 226 (e.g., repeating the noise selection process one or more times for one or more sets of the noisy biometric signals 220).

[0060] The training data generator 224 can include a noise sampler 232. The noise sampler 232 can sample noise from the noise sources 230 selected by the noise source selector 228. Specifically, the noise sampler 232 can generate sampled noise information 234 descriptive of the noise sampled from the noise sources 230. The noise sampler 232 can sample noise from the noise sources 230 using any type or manner of conventional noise sampling technique.

[0061] The training data generator 224 can include a signal portion selector 236. The signal portion selector 236 can select one of the clean biometric signals 222. The signal portion selector 236 can identify a portion of the signal (or all of the signal) for application of the sampled noise described by the sampled noise information 234. In some implementations, the signal portion selector 236 can select multiple portions of one signal of the clean biometric signals 222 for application of sampled noise (e.g., the same set of sampled noise, different sets of sampled noise, etc.). Additionally, or alternatively, in some implementations, the signal portion selector 236 can select the same portion, or type of portion, of multiple signals of the clean biometric signals 222. For example, if the clean biometric signals 222 include ECG signals that are a minimum of five seconds in length, the signal portion selector 222 can select the first two seconds of multiple signals from the clean biometric signals 222.

[0062] The training data generator 224 can include a noise applicator 238. The noise applicator 238 can apply the sampled noise to the identified portion of the clean biometric signals 222 to generate corresponding noisy biometric signal 220. In such fashion, the training data generator 224 can sample and apply noise to the clean biometric signals 222 to obtain the corresponding noisy biometric signals 220, thus creating the training examples 218 described by the training data 216 for supervised learning of the machine-learned models 212.

[0063] In some implementations, the training data generator 224 can generate a single training example from each of the clean biometric signals 222 described by the clean biometric signal information 226. In other words, the training data generator 224 can create a total quantity of training examples equal to or less than the quantity of the clean biometric signals 222 described by the clean biometric signal information 226. For example, assume that the clean biometric signals 222 includes three biometric signals. The noise source selector 228 can select some (or all) of the noise source(s) 230, and the noise sampler 232 can sample noise from the selected noise source(s) 230 to generate sampled noise information234. The signal portion selector 236 can select a portion of the first clean biometric signal, and the noise applicator 238 can apply the sampled noise to the selected portion of the clean biometric signal to generate a corresponding first noisy biometric signal of the noisy biometric signals 220. This process can be repeated for the second and third clean biometric signals to generate corresponding second and third noisy biometric signals.

[0064] Additionally, or alternatively, in some implementations, the training data generator 224 can generate multiple training examples from a single clean biometric signal described by the clean biometric signal information 226. To follow the previous example, rather than immediately repeating the noise application process for the second and third clean biometric signals, the training data generator can generate another noisy biometric signal from the first clean biometric signal by applying different noise to the signal and / or selecting a different portion of the clean biometric signal for noise application. For example, after generating the first noisy biometric signal by applying sampled noise to the first clean biometric signal, the training data generator can generate a second noisy biometric signal from the first clean biometric signal by selecting different noise (e.g., selecting different source(s), sampling the source(s) in a different way, etc.), and / or by selecting a different portion of the signal and then applying the sampled noise to the selected portion of the signal to generate the second noisy biometric signal. In such fashion, by generating multiple training examples from the same clean biometric signal, implementations described herein can generate rich training datasets from smaller quantities of ground-truth training examples.

[0065] The model training module 214 can include a model optimizer 240. The model optimizer 240 can perform various model training and / or optimization operations, such as evaluating loss functions, adjusting parameters or weights of the models 212, quantizing model weights, performing distillation training, etc. Specifically, the model optimizer 240 can include loss function(s) 242. The loss function(s) 242 can include loss functions or “optimization” functions that are configured to evaluate model outputs for the purposes of training the models.

[0066] In some implementations, the loss function(s) can include multiple loss functions for evaluating outputs from models trained to de-noise multiple corresponding types of biometric signals. For example, one of the loss function(s) 242 can be configured to train a model to de-noise ECG signals while another of the loss function(s) 242 can be configured to train a model to de-noise PPG signals. The loss function(s) 242 can be evaluated by the model optimizer 240, and based on the evaluation, the model optimizer 240 can adjust parameter(s) of the model(s) evaluated by the loss function 242. The modeloptimizer 240 and the loss function(s) 242 will be discussed in greater detail with regards to Figure 3.

[0067] In some implementations, the model optimizer 240 can perform various optimization tasks for trained models (e.g., distillation training, quantization, fine-tuning, storage, indexing, transmittal of models to other devices, versioning implementation, etc.). In some implementations, the model optimizer 240 may perform a quantization process for various model(s) of the machine-learned models 212 after training the models. For example, the model optimizer 240 may reduce the precision of parameters of the model (e.g., from single-precision floating point (FP32) to half-precision floating point (FP16.16) or 8-bit integer (INT8), etc.) and thus reduce the computing resources necessary to utilize the model (e.g., memory resources, compute cycles, etc.).

[0068] For another example, the model optimizer 240 can perform distillation training to generate machine-learned de-noising distillation models. As described herein, “machine- learned de-noising distillation models” can refer to models that have been trained to de-noise biometric signals based on “teacher” models. Specifically, the distillation models can be trained on the inputs and outputs of “teacher” models such as the machine-learned models 212. The distillation models can exhibit similar (or the same) de-noising performance of the machine-learned models 212 while utilizing fewer computing resources, thus making the distillation models ideal for utilization by mobile computing devices with access to fewer and less powerful computing resources.

[0069] In some implementations, the computing environment 200 can include a wearable computing device 244. The wearable computing device 244 can be any type or manner of wearable device that can implement biometric sensors, either directly (e.g., with biometric sensors physically built into the wearable computing device 244) or indirectly (e.g., via biometric sensors communicatively coupled to the wearable computing device 244. For example, the wearable computing device may be the wearable device 100 of Figures 1 A-1C. Other examples of the wearable computing device 244 can include a smartwatch, a biometric tracking device worn on the wrist or another appendage, a biometric sensor, a distributed collection of sensors and devices, an Augmented Reality (AR) or Virtual Reality (VR) device, glasses, clothing devices, etc.

[0070] The wearable computing device 244 can include processor device(s) 246 and memory 248. The processor device(s) 246 and memory 248 will be discussed in greater detail with regards to Figure 4. The wearable computing device 244 can include mobile biometric sensor(s) 250. As described herein, a “mobile” biometric sensor generally refers to any typeor manner of biometric sensor that is not constrained to specific, purpose-built medical devices or medical contexts. For example, a conventional ECG machine generally requires multiple electrodes to be attached to a patient to capture a biometric signal. Furthermore, conventional ECG machines are relatively immobile, and usage is generally restricted to medical service providers. Conversely, a mobile ECG sensor incorporated into the wearable computing device 244 may incorporate a single electrode that contacts the patient’s skin when worn, thus enabling the sensor to regularly capture biometric signals from the user.

[0071] In particular, the wearable computing device 244 can incorporate, or otherwise communicatively access, mobile biometric sensors 250. In some implementations, the mobile biometric sensors 250 can be incorporated in the wearable computing device 244 to contact the skin of a user that wears the wearable computing device 244. For example, if the wearable computing device 244 is a smart watch device, the mobile biometric sensor(s) 250 can be located opposite the face of the watch on a lower or “bottom” surface of the watch which contacts a user’s wrist. In such fashion, the mobile biometric sensor(s) 250 can measure various biometric signals from the user via a surface of the user that contacts the mobile biometric sensor(s) 250.

[0072] The memory 248 of the wearable computing device 244 can include a biometric analysis module 252. In some implementations, the biometric analysis module 252 can include a machine-learned biometric noise reduction model 254. The machine-learned biometric noise reduction model 254 can be a model trained to reduce noise present in biometric signals measured using the mobile biometric sensors 250. For example, the machine-learned biometric noise reduction model 254 can be one or more of the machine- learned models 212 trained at the computing system 202. For another example, the machine- learned biometric noise reduction model 254 can be one or more models created based on the machine-learned models 212 (e.g., via distillation training, quantization, etc.). In some implementations, the memory 248 can include a model training module 255 that can train the machine-learned biometric noise reduction model 254 as described with regards to the model training module 214 of the computing system 202.

[0073] The biometric analysis module 252 can obtain noisy biometric signals 256 via the mobile biometric sensor(s) 250. In some implementations, the biometric analysis module 252 can process the noisy biometric signals 256 with the machine-learned biometric noise reduction model 254 to obtain de-noised biometric signals 258. The de-noised biometric signals 258 can include less noise than the noisy biometric signals 256 while retaining most (or all) of the signal present within the noisy biometric signals 256. For example, if the noisybiometric signals 256 are ECG signals, the de-noised biometric signals 258 can be ECG signals that exhibit a reduced quantity of noise while retaining consistency between ECG signal characteristics (e.g., P-wave, etc.). Additionally, or alternatively, in some implementations, the biometric analysis module 252 can provide the noisy biometric signals 256 to the computing system 202 for de-noising (e.g., using the machine-learned models 212, etc.).

[0074] The biometric analysis module 252 can include a signal analyzer 260. In some implementations, the signal analyzer 260 can analyze the de-noised biometric signals 258 to obtain biometric measurements 262 (e.g., physiological values or measurements, etc.). Additionally, or alternatively, in some implementations, the signal analyzer 260 can provide the de-noised biometric signals 258 to the computing system 202 for analysis. In some implementations, the signal analyzer 260 can cause information to be displayed or otherwise notified to a user of the wearable computing device 244. Various interactions implemented by the signal analyzer 260 will be discussed in greater detail with regards to Figure 4.

[0075] In some implementations, the computing environment 200 can include a user computing device 264. The user computing device 264 can include processor device(s) 266 and memory 268 as described with regards to the computing system 202 and / or the wearable computing device 244. The user computing device 264 can be any type or manner of computing device utilized by a user of the wearable computing device, such as a smartphone, laptop, desktop computer, tablet, AR / VR device, cloud device, virtualized device, automobile computing device, etc.).

[0076] The memory 268 of the user computing device 264 can include a biometric analysis module 270. The biometric analysis module can perform some, or all, of the functionality described with regards to the biometric analysis module 252. For example, after obtaining noisy biometric signals 256, the wearable computing device 244 may partially or fully de-noise the noisy biometric signals 256 with the machine-learned biometric noise reduction model 254, and then provide the partially or fully de-noised biometric signals to the biometric analysis module 270 of the user computing device 264. The biometric analysis module 270 can obtain the de-noised biometric signals 258 (e.g., by fully de-noising partially de-noised signals if necessary). The biometric analysis module 270 can analyze the de-noised biometric signals 258 to obtain the biometric measurements 262, and can provide the biometric measurements 262 to the wearable computing device 244.

[0077] Figure 3 is a block diagram for the training data generator 224 of the model training module 214 of Figure 2 according to some implementations of the present disclosure.Specifically, the training data generator 224 can include the noise source selector 228. The noise source selector 228 can include the noise sources 230, a random selection function 302, and a deterministic selection function 304. The random selection function 302 can be configured to select a random number of source(s) from the noise sources 230. The deterministic selection function 304 can select particular source(s) from the noise sources 230 based on source selection information 306. The source selection information 306 can map certain noise sources 230 to corresponding noise exhibited by noisy signals collected using mobile biometric sensors. In other words, given a particular mobile biometric sensor (e.g., an ECG sensor), the source selection information 306 can identify certain source(s) of the noise sources 230 that most closely match the noise found in signals collected using that particular mobile sensor.

[0078] The training data generator 224 can generate selected noise source information 308 using the noise source selector 228. The selected noise source information 308 can indicate the noise source(s) selected by the noise source selector 224. Noise sampler 310 can generate sampled noise 312 by sampling noise from the noise sources 230 indicated by the selected noise source information 308. The training data generator can process the clean biometric signals 222 with the signal portion selector 236 to obtain selected signal portion information 314. The selected signal portion information 314 can indicate portion(s) of the clean biometric signals 222 for applying the sampled noise 312. The noise applicator 238 can apply the sampled noise 312 to the portion(s) of the clean biometric signals 222 indicated by the selected signal portion information 314.

[0079] Figure 4 illustrates a block diagram of the above-described example, nonlimiting wearable device 100 and / or 244 according to one or more example embodiments of the present disclosure. That is, for instance, Figure 4 illustrates a block diagram of one or more internal and / or external components of the above-described example, non-limiting wearable device(s) 100 and / or 244 according to one or more example embodiments of the present disclosure.

[0080] As described above with reference to the example embodiments depicted in Figures 1 A-1C, 2, and 3, wearable device 100 can constitute and / or include a wearable computing device such as, for instance, a wearable physiological monitoring device. For example, in the example embodiment depicted in Figure 4, wearable device 100 can constitute and / or include a wearable physiological monitoring device that can be worn by a user 10 (also referred to herein as a “wearer” or “wearer 10”) and / or can be configured to gather data regarding activities performed by user 10 and / or data regarding user's 10physiological state (e.g., temperature, blood volume, pulse, pulse oximetry, heart rate, heart rhythm, etc.). In this or another embodiment, such data can include data representative of the ambient environment around user 10 or user’s 10 interaction with the environment. For example, in some embodiments, the data can constitute and / or include motion data regarding user’s 10 movements, ambient light, ambient noise, air quality, and / or physiological data obtained by measuring various physiological characteristics of user 10 (e.g., heart rate, respiratory data, body temperature, blood oxygen levels, perspiration levels, movement data).

[0081] Although certain embodiments are disclosed herein in the context of wearable physiological monitoring devices, it should be appreciated that the present disclosure is not so limiting. For example, it should be understood that one or more embodiments can by performed and / or implemented using any suitable or desirable type of computing device or combination of computing devices such as, for example, a client computing device, a laptop, a tablet, a wearable computing device (e.g., wearable device 100), a smartphone, and / or another computing device, whether wearable or not.

[0082] As illustrated in Figure 4, wearable device 100 according to example embodiments of the present disclosure can include one or more audio and / or visual feedback components 130 such as, for instance, electronic touchscreen display units, light-emitting diode (LED) display units, audio speakers, light-emitting diode (LED) lights, buzzers, and / or another type of audio and / or visual feedback module. In certain embodiments, one or more audio and / or visual feedback modules 130 can be located on and / or otherwise associated with a front side of wearable device 100 and / or display 102. For example, in wearable embodiments of wearable device 100, an electronic display such as, for instance, display 102 can be configured to be externally presented to user 10 viewing wearable device 100.

[0083] Wearable device 100 according to example embodiments of the present disclosure can include control circuitry 110. Although certain modules and / or components are illustrated as part of control circuitry 110 in the diagram of Figure 4, it should be understood that control circuitry 110 associated with wearable device 100 and / or other components or devices in accordance with example embodiments of the present disclosure can include additional components and / or circuitry such as, for instance, one or more additional components of the illustrated components depicted in Figure 4. Furthermore, in certain embodiments, one or more of the illustrated components of control circuitry 110 can be omitted and / or different than that shown in Figure 4 and described in association therewith.

[0084] The term “control circuitry” is used herein according to its broad and / ordinary meaning and can include any combination of software and / or hardware elements, devices,and / or features that can be implemented in connection with operation of wearable device 100. Furthermore, the term “control circuitry” can be used substantially interchangeably in certain contexts herein with one or more of the terms “controller,” “integrated circuit,” “IC,” “application-specific integrated circuit,” “ASIC,” “controller chip,” or the like.

[0085] Control circuitry 110 according to example embodiments of the present disclosure can constitute and / or include one or more processors, data storage devices, and / or electrical connections. In one embodiment, control circuitry 110 can be implemented on a system on a chip (SoC), however, those skilled in the art will recognize that other hardware and / or firmware implementations are possible.

[0086] In one or more embodiments of the present disclosure, control circuitry 110 can constitute and / or include one or more processors 181 that can be configured to execute computer-readable instructions that, when executed, cause wearable device 100 to perform one or more operations. In at least one embodiment, control circuitry 110 can constitute and / or include processor(s) 181 that can be configured to execute operational code (e.g., instructions, processing threads, software) for wearable device 100 such as, for instance, firmware or the like. Processor(s) 181 according to example embodiments described herein can each be a processing device. For instance, in the example embodiment depicted in Figure 4, processor(s) 181 can each be a central processing unit (CPU), microprocessor, microcontroller, integrated circuit (e.g., an application-specific integrated circuit (ASIC)), and / or another type of processing device. In this or another example embodiment, processor(s) 181 can be coupled to (e.g., electrically, communicatively, physically, operatively) to one or more components of control circuitry 110 and / or wearable device 100 such that processor(s) 181 can facilitate one or more operations in accordance with one or more example embodiments described herein.

[0087] In at least one embodiment of the present disclosure, the above-described computer-readable instructions and / or operational code that can be executed by processor(s) 181 can be stored in one or more data storage devices of wearable device 100. In the example embodiment depicted in Figure 4, such computer-readable instructions and / or operational code can be stored in memory 183 of wearable device 100. In this or another example embodiment, memory 183 can be coupled to (e.g., electrically, communicatively, physically, operatively) to one or more components of control circuitry 110 and / or wearable device 100 such that memory 183 can facilitate one or more operations in accordance with one or more example embodiments described herein.

[0088] Memory 183 according to example embodiments described herein can store computer-readable and / or computer executable entities (e.g., data, information, applications, models, algorithms) that can be created, modified, accessed, read, retrieved, and / or executed by each of processor(s) 181. In some embodiments, memory 183 can constitute, include, be coupled to (e.g., operatively), and / or otherwise be associated with a computing system and / or media such as, for example, one or more computer-readable media, volatile memory, nonvolatile memory, random-access memory (RAM), read only memory (ROM), hard drives, flash drives, and / or other memory devices. In these or other embodiments, such one or more computer-readable media can include, constitute, be coupled to (e.g., operatively), and / or otherwise be associated with one or more non-transitory computer-readable media. Although not depicted in the example embodiment illustrated in Figure 4, in some embodiments, memory 183 can include (e.g., store) a biometric analysis module 111, and / or other modules and / or data that can be used to facilitate one or more operations described herein.

[0089] Control circuitry 110 according to example embodiments of the present disclosure can constitute and / or include biometric analysis module 111. The biometric analysis module 111 according to example embodiments of the present disclosure can constitute and / or include one or more hardware and / or software components and / or features that can be configured to perform switching, or transitioning, from one operating domain to another in accordance with one or more embodiments described herein. For example, in some embodiments, the biometric analysis module 111 can constitute and / or include the biometric analysis module 252 of Figure 2.

[0090] In some embodiments, the biometric analysis module 111 can receive a noisy biometric signal (e.g., via the internal physiological sensor(s) 143, the external physiological sensor(s) 145, the environment sensor(s) 155, etc.), de-noise the received signal, and generate biometric measurements based on the de-noised signal. Based at least in part on (e.g., in response to) obtaining a particular biometric measurement, the wearable device 100 can perform one or more operations described herein to facilitate alteration (e.g., improvement) of user’s 10 health, wellness, and / or well-being (e.g., physical, mental, emotional, behavioral, and / or sleep quality).

[0091] For example, in at least one embodiment, wearable device 100 can perform operation(s) that can include, but not limited to: presenting the biometric measurement to user 10 and / or another computing device; providing user 10 and / or another computing device with an explanation of the biometric measurement, which can include a defined activity as described herein, suggesting one or more health improvement recommendations and / orengage another computing device to make such recommendation(s) based at least in part on (e.g., using) the biometric measurement (e.g., recommendation that user 10 seek medical attention, or seek epidemiological testing); implementing one or more wellness promoting features and / or engage another computing device to implement such feature(s) based at least in part on (e.g., using) the correlation or absence of correlation (e.g., vibrating in a particular manner to indicate to a user that they should reduce physical exertion, etc.); and / or another operation according to one or more example embodiments of the present disclosure.

[0092] In certain embodiments, the biometric analysis module 111 can be communicatively coupled with one or more internal physiological sensors 143 that can be embedded and / or integrated in wearable device 100. In certain embodiments, the biometric analysis module 111 can be optionally in communication with one or more external physiological sensors 145 not embedded and / or integrated in wearable device 100 (e.g., an electrode or sensor integrated in another electronic device). In some embodiments, examples of internal physiological sensors 143 and / or external physiological sensors 145 can constitute and / or include, but are not limited to, one or more sensors that can measure (e.g., capture, collect, receive) physiological data of user 10 such as, for instance, heart rate, blood oxygen level, movement, respiration, perspiration, stress data, and / or other physiological data of user 10.

[0093] In the example embodiment depicted in Figure 4, wearable device 100 can include one or more data storage components 151 (denoted as “data storage 151” in Figure 4). Data storage component(s) 151 according to example embodiments can constitute and / or include any suitable or desirable type of data storage such as, for instance, solid-state memory, which can be volatile or non-volatile. In some embodiments, such solid-state memory of wearable device 100 can constitute and / or include any of a wide variety of technologies such as, for instance, flash integrated circuits, phase change (PC) memory, phase change (PC) random-access memory (RAM), programmable metallization cell RAM (PMC -RAM or PMCm), ovonic unified memory (OUM), resistance RAM (RRAM), NAND memory, NOR memory, EEPROM, ferroelectric memory (FeRAM), MRAM, or other discrete NVM (non-volatile solid-state memory) chips. In some embodiments, data storage component(s) 151 can be used to store system data, such as operating system data and / or system configurations or parameters. In some embodiments, wearable device 100 can include data storage utilized as a buffer and / or cache memory for operational use by control circuitry 110.

[0094] Data storage component s) 151 according to example embodiments can include various sub-modules that can be implemented to facilitate the physiological monitoring and the health, wellness, and / or well-being assessment principles and features disclosed herein (e.g., temperature sensing) in accordance with one or more embodiments. For example, in at least one embodiment, data storage 151 can include one or more sub-modules that can include, but not limited to: an information collection module (e.g., biometric analysis module 111) that can manage the collection of physiological and / or environmental data relevant to any health, wellness, and / or well-being assessment described herein (e.g., body temperature sensing); a heart rate determination module that can determine values and / or patterns of one or more types of heart rates of user 10; a condition determination module that can determine a condition that may cause the temperature at the surface of the user 10 (e.g., the user’s skin), such as a disease, hyperthermia, hypothermia, exercise, etc.); a presentation module that can manage presentation of information to user 10 that can be associated with any health, wellness, and / or well-being assessment described herein (e.g., body temperature, heart rate, heart rhythm, etc.); a feedback management module for collecting and interpreting any input data and / or feedback received from user 10; and / or another sub-module.

[0095] Wearable device 100 according to example embodiments can further include a power storage module 153 (denoted as “power storage 153”), which can constitute and / or include a rechargeable battery, one or more capacitors, or other charge-holding device(s). In some embodiments, the power stored by power storage module 153 can be utilized by control circuitry 110 for operation of wearable device 100, such as for powering display 102. In some embodiments, power storage module 153 can receive power over a host interface of wearable device 100 (e.g., via one or more host interface circuitry and / or components 176 (denoted as “host interface 176” in Figure 4)) and / or through other means.

[0096] Wearable device 100 according to example embodiments can further include one or more environmental sensors 155. In at least one embodiment, examples of such environmental sensors 155 can include, but are not limited to, sensors that can determine and / or measure, for instance, ambient light, external (non-body) temperature, altitude, device location (e.g., global-positioning system (GPS)), and / or another environmental data.

[0097] Wearable device 100 according to example embodiments can further include one or more connectivity components 170, which can include, for example, a wireless transceiver 172. Wireless transceiver 172 according to example embodiments can be communicatively coupled to one or more antenna devices 195, which can be configured to wirelessly transmit and / or receive data and / or power signals to and / or from wearable device100 using, but not limited to, peer-to-peer, WLAN, and / or cellular communications. For example, wireless transceiver 172 can be utilized to communicate data and / or power between wearable device 100 and an external computing device (not illustrated in Figure 4) such as, for instance, an external client computing device (e.g., a smartphone, tablet, computer) and / or an external host system (e.g., a server), which can be configured to interface with wearable device 100. In certain embodiments, wearable device 100 can include one or more host interface circuitry and / or components 176 (denoted as “host interface 176” in Figure 4) such as, for instance, wired interface components that can communicatively couple wearable device 100 with the above-described external computing device (e.g., a smartphone, table, computer, server) to receive data and / or power therefrom and / or transmit data thereto.

[0098] Connectivity component(s) 170 according to example embodiments can further include one or more user interface components 174 (denoted as “user interface 174” in Figure 4) that can be used by wearable device 100 to receive input data from user 10 and / or provide output data to user 10. In some embodiments, user interface component(s) 174 can be coupled to (e.g., operatively, communicatively) and / or otherwise be associated with audio and / or visual feedback component s) 130. For instance, in these embodiments, display 102 of wearable device 100 can constitute and / or include a touchscreen display that can be configured to provide (e.g., render) output data to user 10 and / or to use audio and / or visual feedback component(s) 130 to receive user input through user contact with the touchscreen display. In some embodiments, user interface component(s) 174 can further constitute and / or include one or more buttons or other input components or features.

[0099] Connectivity component(s) 170 according to example embodiments can further include host interface circuitry and / or component(s) 176, which can be, for example, an interface that can be used by wearable device 100 to communicate with the above-described external computing device (e.g., a smartphone, table, computer, server) over a wired or wireless connection. Host interface circuitry and / or component(s) 176 according to example embodiments can utilize and / or otherwise be associated with any suitable or desirable communication protocol and / or physical connector such as, for instance, universal serial bus (USB), micro-USB, Wi-Fi, Bluetooth, FireWire, PCIe, or the like. For wireless connections, host interface circuitry and / or component(s) 176 according to example embodiments can be incorporated with wireless transceiver 172.

[0100] Although certain functional modules and components are illustrated and described herein, it should be understood that authentication management functionality in accordance with the present disclosure can be implemented using a number of differentapproaches. For example, in some embodiments, control circuitry 110 can constitute and / or include one or more processors (e.g., processor(s) 181) that can be controlled by computerexecutable instructions that can be stored in a memory (e.g., memory 183, data storage component(s) 151) so as to provide functionality such as is described herein. In other embodiments, such functionality can be provided in the form of one or more specially designed electrical circuits. In some embodiments, such functionality can be provided by one or more processors (e.g., processor(s) 181) that can be controlled by computer-executable instructions that can be stored in a memory (e.g., memory 183, data storage component(s) 151) that can be coupled to (e.g., communicatively, operatively, electrically) one or more specially designed electrical circuits. Various examples of hardware that can be used to implement the concepts outlined herein can include, but are not limited to, application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), and general- purpose microprocessors that can be coupled with memory that stores executable instructions for controlling the general-purpose microprocessors.

[0101] Figure 5 illustrates a diagram of an example, non-limiting user assessment management system 500 according to one or more example embodiments of the present disclosure. User assessment management system 500 depicted in Figure 5 illustrates an example, non-limiting networked relationship between wearable device 100, an external computing device 504, and / or one or more smart systems 512 in accordance with one or more embodiments. In particular, it should be noted that the user assessment management system 500, and some, or all, of the components or sub-systems of the user assessment management system 500, are illustrated to provide an example of operations that require the high- performance operating domain.

[0102] With reference to the example embodiment described above and depicted in Figure 4, wearable device 100 according to example embodiments of the present disclosure can perform one or more health, wellness, and / or well-being assessments (e.g., physical, mental, emotional, behavioral, and / or sleep quality assessment(s)) of user 10 and / or perform operation(s) to facilitate alteration (e.g., improvement) of user’s 10 health, wellness, and / or well-being based on such assessment s). As such, in certain embodiments described in the present disclosure, wearable device 100 can be capable of and / or configured to collect physiological sensor readings of user 10 and / or perform such assessment s) and / or operation(s) using such readings.

[0103] However, in additional and / or alternative embodiments, wearable device 100 and / or another electronic and / or computing device that can be used to detect physiologicalinformation of user 10, can be in communication with external computing device 504. In these and / or other embodiments, external computing device 504 can be configured to use such physiological information of user 10 to perform such one or more health, wellness, and / or well-being assessments (e.g., body temperature sensing) of user 10 according to one or more embodiments described herein. In these and / or other embodiments, based at least in part on (e.g., in response to) performing such assessment s), external computing device 504 can perform one or more operations described herein to facilitate alteration (e.g., improvement) of user’s 10 health, wellness, and / or well-being (e.g., physical, mental, emotional, behavioral, and / or sleep quality).

[0104] Wearable device 100 according to example embodiments can be configured to collect one or more types of physiological and / or environmental data using embedded sensors and / or external devices, as described throughout the present disclosure, and communicate or relay such information over one or more networks 506 to other devices. This includes, in some embodiments, relaying information to devices capable of serving as Internet-accessible data sources, thus permitting the collected data to be viewed, for example, using a web browser or network-based application at, for instance, external computing device 504. For example, while user 10 is wearing wearable device 100, wearable device 100 can capture, calculate, and / or store environment data and / or user’s 10 physiological data (e.g., heart rate, motion data, temperature, respiration, perspiration, EDA, stress data) using one or more environmental and / or physiological sensors. Wearable device 100 according to example embodiments can then transmit data representative of such environment data and / or user's 10 physiological data over network(s) 506 to an account on a web service, computer, mobile phone, and / or health station where the data can be stored, processed, and visualized by user 10 and / or another entity (e.g., a health care professional).

[0105] While wearable device 100 is shown in example embodiments of the present disclosure to have a display, it should be understood that, in some embodiments, wearable device 100 does not have any type of display unit. In some embodiments, wearable device 100 can have audio and / or visual feedback components such as, for instance, light-emitting diodes (LEDs), buzzers, speakers, and / or a display with limited functionality. Wearable device 100 according to example embodiments can be configured to be attached to user’s 10 body or clothing. For example, in these or other embodiments, wearable device 100 can be configured as a wrist bracelet, watch, ring, electrode, finger-clip, toe-clip, chest-strap, ankle strap, and / or a device placed in a pocket. In additional or alternative embodiments, wearable device 100 can be embedded in something in contact with user 10 such as, for instance,clothing, a mat that can be positioned under user 10, a blanket, a pillow, and / or another accessory.

[0106] In one or more embodiments of the present disclosure, the communication between wearable device 100 and external computing device 504 can be facilitated by network(s) 506. In some embodiments, network(s) 506 can constitute and / or include, for instance, one or more of an ad hoc network, a peer-to-peer communication link, an intranet, an extranet, a virtual private network (VPN), a local area network (LAN), a wireless LAN (WLAN), a wide area network (WAN), a wireless WAN (WWAN), a metropolitan area network (MAN), a portion of the Internet, a portion of the public switched telephone network (PSTN), a cellular telephone network, and / or any other type of network. In some embodiments, the communication between wearable device 100 and external computing device 504 can also be performed through a direct wired connection. In these or other embodiments, this direct-wired connection can be associated with any suitable or desirable communication protocol and / or physical connector such as, for instance, universal serial bus (USB), micro-USB, Wi-Fi, Bluetooth, FireWire, PCIe, or the like.

[0107] In example embodiments of the present disclosure, a variety of computing devices can be in communication with wearable device 100 to facilitate user’s 10 health, wellness, and / or well-being assessment and / or alteration (e.g., improvement). Although external computing device 504 is depicted as a smartphone in the example embodiment illustrated in Figure 5, it should be understood that the present disclosure is not so limiting. For instance, external computing device 504 according to example embodiments can constitute and / or include, for example, a smartphone with a display 508 as depicted in Figure 5, a personal digital assistant (PDA), a mobile phone, a tablet, a personal computer, a laptop computer, a smart television, a video game console, a server, and / or another computing device that can be external to wearable device 100.

[0108] The networked relationship depicted in the example embodiment illustrated in Figure 5 demonstrates how, in some embodiments, external computing device 504 can be implemented to perform one or more health, wellness, and / or well-being assessments (e.g., a body temperature assessment) of user 10 and / or perform operation(s) to facilitate alteration (e.g., improvement) of user’s 10 health, wellness, and / or well-being based on such assessment(s). For example, in one embodiment, user 10 can wear wearable device 100 that can be equipped as a bracelet with one or more physiological sensors but without a display. In this and / or another embodiment, while user 10 is wearing wearable device 100, wearable device 100 can capture, calculate, and / or store environment data and / or user’s 10physiological data (e.g., temperature) using the physiological sensors. Wearable device 100 according to example embodiments can then transmit data representative of such environment data and / or user's 10 physiological data over network(s) 506 to an account on a web service, computer, mobile phone, and / or health station where the data can be stored, processed, and visualized by user 10 and / or another entity (e.g., a health care professional). In some embodiments, wearable device 100 can periodically or continuously transmit such information to external computing device 504 over network(s) 506.

[0109] In additional and / or alternative embodiments, wearable device 100 can store the above-described collected physiological and / or environmental data and transmit this data to external computing device 504 in response to a trigger event such as, for instance, detection of an abnormal surface temperature of the skin of the user 10 (e.g., above or below a threshold range of standard human temperatures) after a period performing the defined activity. In some embodiments, wearable device 100 can transmit such data to external computing device 504 in response to detecting that a command has been performed by external computing device 504 such as, for instance, manual or automatic execution of an instruction to synchronize collected physiological and / or environmental data and perform one or more health, wellness, and / or well-being assessments (e.g., body temperature assessment) of user 10 as described herein.

[0110] In some embodiments, external computing device 504 can present (e.g., provide, render) a possible cause of the body temperature of user 10. For instance, in these or other embodiments, external computing device 504 can generate an intelligent notification 510 that can include such body temperature and / or one or more health improvement recommendations (e.g., a suggestion to reduce physical activity, a suggestion to see a healthcare provider, etc.) that, if and / or when implemented by user 10, can facilitate alteration (e.g., improvement) of user’s 10 health, wellness, and / or well-being (e.g., body temperature). In the example embodiment depicted in Figure 5, external computing device 504 can render intelligent notification 510 having such body temperature and the health improvement recommendation(s) on display 508 such that user 10 and / or another entity (e.g., health care professional, mental health care professional, sleep therapy provider, doctor, caregiver) can view such information.

[0111] Although not illustrated in the example embodiment depicted in Figure 5, in some embodiments, wearable device 100 can: sense a body temperature of the user 10; determine one or more health improvement recommendations based on (e.g., in response to) sensing the body temperature; generate intelligent notification 510 such that it includes thebody temperature and the health improvement recommendation(s); and render this information on display 102 of wearable device 100.

[0112] In one embodiment of the present disclosure, wearable device 100 and / or external computing device 504 can implement (e.g., initiate, run, operate) one or more wellness promoting features that can be included with wearable device 100 and / or external computing device 504 such as, for instance, a wellness promoting audio feature (e.g., by playing a sound that alerts the user to the occurrence of an abnormal body temperature), and / or another wellness promoting feature of wearable device 100 and / or external computing device 504.

[0113] In another embodiment of the present disclosure, wearable device 100 and / or external computing device 504 can facilitate implementation of one or more wellness promoting features of another computing device such as, for instance, a computing device of one or more smart systems 512. In this or another embodiment, smart system(s) 512 can constitute and / or include, but are not limited to, an audio system (e.g., a home audio system), a lighting system (e.g., a home lighting system), an HVAC system (e.g., a home HVAC system), an exercise system (e.g., an exercise machine), and / or another system that can be included in, coupled to, and / or operated by a computing device other than wearable device 100 and / or external computing device 504. For instance, in some embodiments, smart system(s) 512 can constitute and / or include a smart audio system, a smart lighting system, a smart HVAC system, and / or a smart exercise system (e.g., a smart exercise machine). In these or other embodiments, wearable device 100 and / or external computing device 504 can facilitate implementation of one or more wellness promoting features of smart system(s) 512 such as, for instance: a wellness promoting audio feature of a smart audio system; a wellness promoting lighting feature of a smart lighting system; a wellness promoting ambient temperature feature of a smart HVAC system; a wellness promoting exercise feature (e.g., a certain exercise mode or setting) of a smart exercise system; and / or another wellness promoting feature of smart system(s) 512.

[0114] In some embodiments described herein, wearable device 100 and / or external computing device 504 can send instructions to smart system(s) 512 that, when executed by such system(s) (e.g., via one or more processors), can cause the system(s) to perform operations to implement one or more wellness promoting features of such system(s). In one embodiment, wearable device 100 and / or external computing device 504 can send instructions to a smart audio system that, when executed by such a system (e.g., via one or more processors), can cause it to inform the user 10 that an abnormal body temperature hasbeen detected. In another embodiment, wearable device 100 and / or external computing device 504 can send instructions to a smart HVAC system that, when executed by such a system (e.g., via one or more processors), can cause it to output air at a certain wellness promoting temperature (e.g., a certain temperature that can be defined by user 10). In one embodiment of the present disclosure, wearable device 100 and / or external computing device 504 can send instructions to a smart exercise system that, when executed by such a system (e.g., via one or more processors), can cause it to operate in a certain mode or setting and / or to provide a recommendation to the user to select such a mode or setting.Example Methods

[0115] Figure 6 depicts a flow chart diagram of an example method 600 to perform training of machine-learned models for denoising biometric signals collected using mobile biometric sensors according to example embodiments of the present disclosure. Although Figure 6 depicts steps performed in a particular order for purposes of illustration and discussion, the methods of the present disclosure are not limited to the particularly illustrated order or arrangement. The various steps of the method 600 can be omitted, rearranged, combined, and / or adapted in various ways without deviating from the scope of the present disclosure.

[0116] At 602, a computing system can generate training data for training a machine- learned biometric noise reduction model.

[0117] At 604, to generate the training data, the computing system can obtain information descriptive of a plurality of clean biometric signals measured using a corresponding medical device.

[0118] At 606, in some implementations, to generate the training data, for each clean biometric signal of the plurality of clean biometric signals, the computing system can select at least a portion of the clean biometric signal.

[0119] At 608, to generate the training data, the computing system can sample noise from one or more noise sources of a plurality of candidate noise sources. In some implementations, prior to sampling the noise from the one or more noise sources of the plurality of candidate noise sources, the computing system can, for each clean biometric signal of the plurality of clean biometric signals, select the one or more noise sources from the plurality of candidate noise sources. In some implementations, selecting the one or more noise sources from the plurality of candidate noise sources can include performing a randomized noise selection process to select the one or more noise sources from the pluralityof candidate noise sources for the clean biometric signal. In some implementations, selecting the one or more noise sources from the plurality of candidate noise sources can include performing a noise selection process that evaluates the clean biometric signal to select the one or more noise sources from the plurality of candidate noise sources for the clean biometric signal.

[0120] At 610, to generate the training data, the computing system can apply the noise to the at least the portion of the clean biometric signal. In some implementations, the computing system can select at least a portion of the clean biometric signal to apply the noise to. In some implementations, the computing system can perform a randomized signal selection process to select the at least the portion of the clean biometric signal. Additionally, or alternatively, in some implementations, the computing system can perform a signal selection process that evaluates the clean biometric signal to select the at least the portion of the clean biometric signal.

[0121] At 612, the computing system can process the training data with the machine- learned biometric noise reduction model to obtain a plurality of denoised biometric signals.

[0122] At 614, the computing system can train the machine-learned biometric noise reduction model based on a loss function that evaluates a difference between the plurality of denoised biometric signals and the plurality of clean biometric signals. Specifically, in some implementations, the computing system can obtain the information descriptive of the plurality of clean biometric signals. The plurality of clean biometric signals can respectively include a plurality of clean electrocardiogram (ECG) signals measured using a medical ECG device.The computing system can process the training data with the machine-learned biometric noise reduction model to obtain a plurality of denoised ECG signals.

[0123] The training data can include a respective plurality of noisy ECG signals. The computing system can train the machine-learned biometric noise reduction model based on a loss function. Each of the plurality of noisy ECG signals can include information indicative of a P-wave, and the loss function can be configured to train the machine-learned biometric noise reduction model to preserve the information indicative of the P-wave in each the plurality of denoised ECG signals. Additionally, or alternatively, in some implementations, the computing system can obtain a plurality of clean PPG signals measured using a medical PPG device.

[0124] Specifically, in some implementations, the computing system can train the machine-learned biometric noise reduction model based on a loss function that includes one or more loss terms. Some of these loss term(s) can be configured to enforce certainconsistencies (or a lack thereof) between inputs and outputs of the model. For example, a P- wave is a wave found in ECG waveforms that carries biological significance. However, in some instances, training the machine-learned biometric noise reduction model with a default loss function can cause the model to recognize P-waves as noise, and then “de-noise” (i.e., remove) the P-wave from the ECG waveform. To discourage this behavior, the loss function can include a consistency loss term. The consistency loss term can enforce a consistency between P-waves in the inputs and outputs of the model. In other words, the consistency loss term can train the machine-learned biometric noise reduction model to not recognize P-waves as noise.

[0125] Additionally, or alternatively, in some implementations, the loss function can include a consistency loss term for dicrotic notch in PPG signals. A dicrotic notch refers to a feature of the PPG waveform that, when present in the waveform, can be more pronounced or less pronounced. Similarly to the P-wave, the presence / absence (and magnitude) of the dicrotic notch can carry some biological significance. As such, the consistency loss term can enforce a consistency between dicrotic notches in the inputs and outputs of the model. In other words, the consistency loss term can train the machine-learned biometric noise reduction model to not recognize dicrotic notches as noise.

[0126] In some implementations, the computing system can provide the machine-learned biometric noise reduction model to a user computing device for denoising biometric signals received at the user computing device.

[0127] In some implementations, the computing system can train a machine-learned biometric noise reduction distillation model based on the machine-learned biometric noise reduction model. The computing system can provide the machine-learned biometric noise reduction distillation model to a user computing device for denoising biometric signals received at the user computing device.

[0128] In some implementations, the computing system can receive, from a user computing device, a noisy biometric signal obtained by the user computing device. The computing system can process the noisy biometric signal with the machine-learned biometric noise reduction model to obtain a biometric signal with reduced noise. The computing system can send information descriptive of the biometric signal with reduced noise to the user computing device.

[0129] In some implementations, the computing system can perform a biometric analysis process to the biometric signal with reduced noise to obtain biometric measurements. The computing system can send the information descriptive of the biometric signal with reducednoise to the user computing device. The information descriptive of the biometric signal with reduced noise can include the biometric measurements.

[0130] Figure 7 depicts a flow chart diagram of an example method 700 to utilize machine-learned models for denoising biometric signals collected using mobile biometric sensors according to example embodiments of the present disclosure according to example embodiments of the present disclosure. Although Figure 7 depicts steps performed in a particular order for purposes of illustration and discussion, the methods of the present disclosure are not limited to the particularly illustrated order or arrangement. The various steps of the method 700 can be omitted, rearranged, combined, and / or adapted in various ways without deviating from the scope of the present disclosure.

[0131] At 702, a user computing device (e.g., a mobile computing device, a wearable computing device, etc.) can obtain a noisy biometric signal via a mobile biometric measurement device associated with the user computing device. The noisy biometric signal can include a first quantity of noise. In some implementations, the user computing device can be a wearable computing device that includes the mobile biometric measurement device. The mobile biometric measurement device can include a surface of the wearable computing device that contacts the user of the wearable computing device. Obtaining the noisy biometric signal via the mobile biometric measurement device can include measuring the noisy biometric signal from the user with the mobile biometric measurement device via the surface of the wearable computing device that contacts the user.

[0132] In some implementations, obtaining the noisy biometric signal via the mobile biometric measurement device associated with the user computing device can include receiving the noisy biometric signal from the wearable computing device worn by the user. The wearable computing device can include the mobile biometric measurement device. The wearable computing device can measure the noisy biometric signal from the user with the mobile biometric measurement device via the surface of the wearable computing device that contacts the user.

[0133] In some implementations, obtaining the noisy biometric signal via the mobile biometric measurement device associated with the user computing device can include obtaining the noisy biometric signal via the mobile biometric measurement device associated with the user computing device. The noisy biometric signal can include a noisy ECG signal. The mobile biometric measurement device can include a mobile ECG device. Additionally, or alternatively, in some implementations, the noisy biometric signal can include a noisy PPG signal, and the mobile biometric measurement device can include a mobile PPG device.

[0134] At 704, the user computing device can cause the noisy biometric signal to be processed with a machine-learned biometric noise reduction model to obtain a denoised biometric signal that includes a second quantity of noise less than the first quantity of noise. The machine-learned biometric noise reduction model can be trained based on a loss function that evaluates a difference between a plurality of noisy biometric training signals and a plurality of clean biometric ground-truth signals. The plurality of noisy biometric training signals can be generated by applying noise from randomly selected noise sources to randomly selected portions of the clean biometric signals.

[0135] In some implementations, causing the noisy biometric signal to be processed with the machine-learned biometric noise reduction model can include processing the noisy biometric signal with the machine-learned biometric noise reduction model to obtain the denoised biometric signal that includes the second quantity of noise less than the first quantity of noise.

[0136] In some implementations, causing the noisy biometric signal to be processed with the machine-learned biometric noise reduction model can include processing the noisy biometric signal with a machine-learned biometric noise reduction distillation model to obtain the denoised biometric signal. The machine-learned biometric noise reduction distillation model can be trained by a computing system via distillation from the machine- learned biometric noise reduction model. In some implementations, prior to processing the noisy biometric signal with the machine-learned biometric noise reduction distillation model, the user computing device can receive the machine-learned biometric noise reduction distillation model from the computing system.

[0137] At 706, the user computing device can provide biometric measurements for display to a user of the user computing device. The biometric measurements can be based at least in part on the denoised biometric signal. In some implementations, the information indicative of the denoised biometric signal from the computing system can include the biometric measurements. In some implementations, providing the biometric measurements for display to the user of the user computing device can include performing a biometric analysis process to the denoised biometric signal with reduced noise to obtain the biometric measurements.Additional Disclosure

[0138] The technology discussed herein makes reference to servers, databases, software applications, and other computer-based systems, as well as actions taken and information sentto and from such systems. The inherent flexibility of computer-based systems allows for a great variety of possible configurations, combinations, and divisions of tasks and functionality between and among components. For instance, processes discussed herein can be implemented using a single device or component or multiple devices or components working in combination. Databases and applications can be implemented on a single system or distributed across multiple systems. Distributed components can operate sequentially or in parallel.

[0139] While the present subject matter has been described in detail with respect to various specific example embodiments thereof, each example is provided by way of explanation, not limitation of the disclosure. Those skilled in the art, upon attaining an understanding of the foregoing, can readily produce alterations to, variations of, and equivalents to such embodiments. Accordingly, the subject disclosure does not preclude inclusion of such modifications, variations and / or additions to the present subject matter as would be readily apparent to one of ordinary skill in the art. For instance, features illustrated or described as part of one embodiment can be used with another embodiment to yield a still further embodiment. Thus, it is intended that the present disclosure cover such alterations, variations, and equivalents.

Claims

WHAT IS CLAIMED IS:

1. A computer-implemented method, comprising: generating, by a computing system comprising one or more processor devices, training data for training a machine-learned biometric noise reduction model, wherein generating the training data comprises: obtaining, by the computing system, information descriptive of a plurality of clean biometric signals measured using a corresponding medical device; for each clean biometric signal of the plurality of clean biometric signals: sampling, by the computing system, noise from one or more noise sources of a plurality of candidate noise sources; and applying, by the computing system, the noise to at least a portion of the clean biometric signal; processing, by the computing system, the training data with the machine-learned biometric noise reduction model to obtain a plurality of denoised biometric signals; and training, by the computing system, the machine-learned biometric noise reduction model based on a loss function that evaluates a difference between the plurality of denoised biometric signals and the plurality of clean biometric signals.

2. The computer-implemented method of claim 1, wherein the method further comprises: providing, by the computing system, the machine-learned biometric noise reduction model to a user computing device for denoising biometric signals received at the user computing device.

3. The computer-implemented method of claim 1, wherein the method further comprises: training, by the computing system, a machine-learned biometric noise reduction distillation model based on the machine-learned biometric noise reduction model; and providing, by the computing system, the machine-learned biometric noise reduction distillation model to a user computing device for denoising biometric signals received at the user computing device.

4. The computer-implemented method of claim 1, further comprising: receiving, by the computing system from a user computing device, a noisy biometric signal obtained by the user computing device;processing, by the user computing device, the noisy biometric signal with the machine- learned biometric noise reduction model to obtain a biometric signal with reduced noise; and sending, by the computing system, information descriptive of the biometric signal with reduced noise to the user computing device.

5. The computer-implemented method of claim 4, wherein sending the information descriptive of the biometric signal with reduced noise to the user computing device comprises: performing, by the computing system, a biometric analysis process to the biometric signal with reduced noise to obtain biometric measurements; and sending, by the computing system, the information descriptive of the biometric signal with reduced noise to the user computing device, wherein the information descriptive of the biometric signal with reduced noise comprises the biometric measurements.

6. The computer-implemented method of claim 1, wherein, prior to sampling the noise from the one or more noise sources of the plurality of candidate noise sources, the method comprises, for each clean biometric signal of the plurality of clean biometric signals: selecting, by the computing system, the one or more noise sources from the plurality of candidate noise sources.

7. The computer-implemented method of claim 6, wherein selecting the one or more noise sources from the plurality of candidate noise sources comprises: performing, by the computing system, a randomized noise selection process to select the one or more noise sources from the plurality of candidate noise sources for the clean biometric signal.

8. The computer-implemented method of claim 6, wherein selecting the one or more noise sources from the plurality of candidate noise sources comprises: performing, by the computing system, a noise selection process that evaluates the clean biometric signal to select the one or more noise sources from the plurality of candidate noise sources for the clean biometric signal.

9. The computer-implemented method of claim 1, wherein applying the noise to the at least the portion of the clean biometric signal comprises:performing, by the computing system, a randomized signal selection process to select the at least the portion of the clean biometric signal.

10. The computer-implemented method of claim 1, wherein applying the noise to the at least the portion of the clean biometric signal comprises: performing, by the computing system, a signal selection process that evaluates the clean biometric signal to select the at least the portion of the clean biometric signal.

11. The computer-implemented method of claim 1, wherein sampling the noise from the one or more noise sources of the plurality of candidate noise sources comprises: selecting, by the computing system, the one or more noise sources from the plurality of candidate noise sources, wherein the one or more noise sources comprises at least one of: noise extracted from noisy biometric signals measured using the corresponding medical device; one or more noise colors of a plurality of noise colors comprising brown noise and white noise; or one or more signal adjustment processes of a plurality of signal adjustment processes, comprising: a signal compression process that compresses a portion of the clean biometric signal; a signal cropping process that removes a portion of the clean biometric signal; or a signal expansion process that stretches a portion of the clean biometric signal.

12. The computer-implemented method of claim 1, wherein obtaining the information descriptive of the plurality of clean biometric signals measured using the corresponding medical device comprises: obtaining, by the computing system, the information descriptive of the plurality of clean biometric signals, wherein the plurality of clean biometric signals respectively comprise a plurality of clean electrocardiogram (ECG) signals measured using a medical ECG device.

13. The computer-implemented method of claim 12, wherein processing the training data comprises:processing, by the computing system, the training data with the machine-learned biometric noise reduction model to obtain a plurality of denoised ECG signals, wherein the training data comprises a respective plurality of noisy ECG signals; and wherein training the machine-learned biometric noise reduction model comprises: training, by the computing system, the machine-learned biometric noise reduction model based on a loss function, wherein each of the plurality of noisy ECG signals comprises information indicative of a P-wave, and wherein the loss function is configured to train the machine-learned biometric noise reduction model to preserve the information indicative of the P-wave in each the plurality of denoised ECG signals.

14. The computer-implemented method of claim 1, wherein obtaining the information descriptive of the plurality of clean biometric signals measured using the corresponding medical device comprises: obtaining, by the computing system, the information descriptive of the plurality of clean biometric signals, wherein the plurality of clean biometric signals respectively comprise a plurality of clean photoplethysmography (PPG) signals measured using a medical PPG device.

15. A user computing device, comprising: one or more processor devices; one or more tangible, non-transitory computer readable media storing computer- readable instructions that when executed by the one or more processor devices cause the user computing device to perform operations, the operations comprising: obtaining a noisy biometric signal via a mobile biometric measurement device associated with the user computing device, wherein the noisy biometric signal comprises a first quantity of noise; causing the noisy biometric signal to be processed with a machine-learned biometric noise reduction model to obtain a denoised biometric signal comprising a second quantity of noise less than the first quantity of noise, wherein the machine-learned biometric noise reduction model is trained based on a loss function that evaluates a difference between a plurality of noisy biometric training signals and a plurality of clean biometric ground-truth signals, and wherein the plurality of noisy biometric training signals are generated by applying noise from randomly selected noise sources to randomly selected portions of the clean biometric signals; andproviding biometric measurements for display to a user of the user computing device, wherein the biometric measurements are based at least in part on the denoised biometric signal.

16. The user computing device of claim 15, wherein the user computing device comprises a wearable computing device that comprises the mobile biometric measurement device, wherein the mobile biometric measurement device comprises a surface of the wearable computing device that contacts the user of the wearable computing device; and wherein obtaining the noisy biometric signal via the mobile biometric measurement device comprises: measuring the noisy biometric signal from the user with the mobile biometric measurement device via the surface of the wearable computing device that contacts the user.

17. The user computing device of claim 15, wherein obtaining the noisy biometric signal via the mobile biometric measurement device associated with the user computing device comprises: receiving the noisy biometric signal from a wearable computing device worn by the user, wherein the wearable computing device comprises the mobile biometric measurement device, and wherein the wearable computing device measures the noisy biometric signal from the user with the mobile biometric measurement device via the surface of the wearable computing device that contacts the user.

18. The user computing device of claim 15, wherein causing the noisy biometric signal to be processed with the machine-learned biometric noise reduction model to obtain the denoised biometric signal comprises: processing the noisy biometric signal with the machine-learned biometric noise reduction model to obtain the denoised biometric signal comprising the second quantity of noise less than the first quantity of noise.

19. The user computing device of claim 18, wherein processing the noisy biometric signal with the machine-learned biometric noise reduction model comprises: processing the noisy biometric signal with a machine-learned biometric noise reduction distillation model to obtain the denoised biometric signal, wherein the machine-learned biometric noise reduction distillation model is trained by a computing system via distillation from the machine-learned biometric noise reduction model; and wherein, prior to processing the noisy biometric signal with the machine-learned biometric noise reduction distillation model, the method comprises: receiving the machine-learned biometric noise reduction distillation model from the computing system.

20. The user computing device of claim 15, wherein causing the noisy biometric signal to be processed with the machine-learned biometric noise reduction model to obtain the denoised biometric signal comprises: sending the noisy biometric signal to a computing system to be processed at the computing system with the machine-learned biometric noise reduction model; and responsive to sending the noisy biometric signal, receiving information indicative of the denoised biometric signal from the computing system.

21. The user computing device of claim 20, wherein the information indicative of the denoised biometric signal from the computing system comprises the biometric measurements.

22. The user computing device of claim 15, wherein providing the biometric measurements for display to the user of the user computing device comprises: performing a biometric analysis process to the denoised biometric signal with reduced noise to obtain the biometric measurements.

23. The user computing device of claim 15, wherein obtaining the noisy biometric signal via the mobile biometric measurement device associated with the user computing device comprises: obtaining the noisy biometric signal via the mobile biometric measurement device associated with the user computing device, wherein the noisy biometric signal comprises a noisy ECG signal, and wherein the mobile biometric measurement device comprises a mobile ECG device.

24. The user computing device of claim 15, wherein obtaining the noisy biometric signal via the mobile biometric measurement device associated with the user computing device comprises:obtaining the noisy biometric signal via the mobile biometric measurement device associated with the user computing device, wherein the noisy biometric signal comprises a noisy PPG signal, and wherein the mobile biometric measurement device comprises a mobile PPG device.

25. One or more tangible, non-transitory computer readable media storing computer- readable instructions that when executed by one or more processor devices cause the one or more processor devices to perform operations, the operations comprising: generating training data for training a machine-learned biometric noise reduction model, wherein generating the training data comprises: obtaining information descriptive of a plurality of clean biometric signals measured using a corresponding medical device; and for each clean biometric signal of the plurality of clean biometric signals: selecting at least a portion of the clean biometric signal; sampling noise from one or more noise sources of a plurality of candidate noise sources; and applying the noise to the at least the portion of the clean biometric signal; processing the training data with the machine-learned biometric noise reduction model to obtain a plurality of denoised biometric signals; and training the machine-learned biometric noise reduction model based on a loss function that evaluates a difference between the plurality of denoised biometric signals and the plurality of clean biometric signals.

Images