Physiological monitoring using low sample rate accelerometer data

Abstract

Techniques for physiological monitoring using low-sampling rate accelerometer data are described and can be implemented to generate analysis results related to a user's state during a long-wearing period. In an example, low-sampling rate accelerometer data is received from a wearable device mounted on a skin surface of a user's chest region during a wearing period, the low-sampling rate accelerometer data having a sampling rate that is below a sampling rate threshold. The low-sampling rate accelerometer data is processed to extract one or more motion-derived parameters that characterize temporal motion patterns of the user during the wearing period. Based on the one or more motion-derived parameters, analysis results related to the user's state during the wearing period are generated for presentation. The analysis results can include, but are not limited to, predictions of sleep states, activity states, and inactivity states, determinations of body angles and body postures, and detections of device-inversion events.

Description

[0001] Cross-reference to related applications

[0002] This application claims priority to U.S. Provisional Application No. 63 / 703,833, filed October 4, 2024, entitled “Prediction of Sleep and Activity under Constraints”; U.S. Provisional Application No. 63 / 703,710, filed October 4, 2024, entitled “Detection of Angles and Positions of Body and Devices”; and U.S. Non-Provisional Application No. 19 / 347,602, filed October 1, 2025, entitled “Physiological Monitoring Using Low Sampling Rate Accelerometer Data”, the entire contents of which are incorporated herein by reference. Background Technology

[0003] Wearable devices for monitoring physiological parameters and activity patterns have become increasingly prevalent in healthcare and health management applications, enabling healthcare providers and individuals to track health indicators without frequent clinical visits or disruptions to daily activities. These devices typically employ a variety of sensors and monitoring components to track physiological parameters, behavioral patterns, and / or activity levels over extended periods. However, wearable monitoring devices may face challenges that limit their effectiveness in continuous, long-term monitoring applications. For example, the computational overhead involved in real-time acquisition, processing, and analysis of sensor data can introduce power consumption challenges, potentially shortening battery life and reducing monitoring duration, thus offsetting the advantages offered by wearable devices. Attached Figure Description

[0004] Figure 1 This is a block diagram of a non-limiting example of an environment in which the system described herein can be used.

[0005] Figure 2 A non-limiting example of a monitoring device is shown.

[0006] Figure 3 A non-limiting system is shown in an example implementation for physiological monitoring using low-sampling-rate accelerometer data, wherein more details are shown. Figure 1 The operation of the prediction system.

[0007] Figure 4 A non-limiting example of physiological monitoring using low-sampling-rate accelerometer data is shown, in which analysis results including sleep / activity states are generated.

[0008] Figure 5 A non-limiting example of physiological monitoring using low-sampling-rate accelerometer data is shown, in which body angles and body postures are generated based on the accelerometer data.

[0009] Figure 6a , Figure 6b and Figure 6cA non-limiting example of physiological monitoring using low-sampling-rate accelerometer data is shown, in which a machine learning system is trained, configured, and implemented to generate analytical results based on various parameters.

[0010] Figure 7 A non-limiting example of physiological monitoring using low-sampling-rate accelerometer data is shown, where device inversion may lead to inaccurate measurements.

[0011] Figure 8 A non-limiting example of physiological monitoring using low-sampling-rate accelerometer data is shown, which includes the detection of device inversion events.

[0012] Figures 9a to 9e A non-limiting example of physiological monitoring using low-sampling-rate accelerometer data is shown, in which a report including various analytical results is output in a health reporting interface.

[0013] Figure 10 A flowchart is shown illustrating an algorithm as a step-by-step process in an example implementation, which can be executed by a processing device to generate analysis results related to the conditions during the wearing period based on low-sampling-rate accelerometer data from the wearable device.

[0014] Figure 11 A flowchart is shown illustrating an algorithm as a step-by-step process in an example implementation, which can be executed by a processing device to generate analysis results related to user state based on low-sampling-rate accelerometer data.

[0015] Figure 12 A flowchart is shown illustrating an algorithm as a step-by-step process in an example implementation, which can be executed by a processing device to determine body angles and body postures based on low-sampling-rate accelerometer data acquired during extended periods of wear.

[0016] Figure 13 A flowchart is shown illustrating an algorithm as a step-by-step process in an example implementation, which can be executed by a processing device to use machine learning to integrate ECG and accelerometer data, thereby enhancing sleep and activity prediction.

[0017] Figure 14 A flowchart is shown illustrating an algorithm as a step-by-step process in an example implementation, which can be executed by a processing device to detect device inversion events during a long monitoring period. Detailed Implementation

[0018] Traditional activity tracking devices typically include accelerometers operating at relatively high sampling frequencies (typically 25 Hz to 100 Hz) to capture motion data for activity analysis. However, this high-frequency sampling generates significant power consumption, limiting battery life and reducing monitoring duration for continuous wear applications. For example, power limitations can lead to frequent charging, disrupting monitoring continuity, limiting the amount of analyzable data, and reducing user compliance, particularly in clinical applications requiring prolonged wear for accurate health assessments. Furthermore, factors such as individual differences in motion patterns, device placement, and signal noise can affect the accuracy of such devices.

[0019] Therefore, this paper describes techniques, methods, and systems for physiological monitoring using low-sampling-rate accelerometer data. These techniques, methods, and systems overcome limitations by accurately generating various analytical insights, such as user status and device attributes, while operating under power and memory constraints. As an example, the predictive system utilizes a wearable device positioned on the user's chest region to acquire accelerometer data at a relatively low sampling rate (e.g., approximately 1.56 Hz) during extended wear periods, such as 1 to 14 days. Because the sampling rate is significantly lower than conventional techniques, the wearable device can acquire measurements for the duration of the wear period without requiring device charging or battery replacement.

[0020] In various examples, low-sampling-rate accelerometer data is processed to extract motion-derived parameters characterizing a user's temporal motion patterns. For instance, motion-derived parameters include calculated values ​​extracted from raw accelerometer data that quantify various aspects of the user's motion and posture over time. Motion-derived parameters may include metrics such as acceleration amplitude, standard deviation of acceleration amplitude over various time intervals, reference vectors and / or position vectors corresponding to body angles, and rolling averages of one or more components of the accelerometer data.

[0021] Based on motion-derived parameters, the prediction system can generate various analytical results related to the conditions during the wearing period and / or the user's state during the wearing period. For example, the prediction system can generate analytical results including predictions of sleep, activity, and inactivity states at discrete (e.g., minute-level) resolution throughout the wearing period. Therefore, the system can determine "when" the user is asleep, active, or awake but inactive throughout the wearing period.

[0022] The system can also determine a user's body angles and / or body posture based on low-sampling-rate accelerometer data. In the example, for this purpose, the system calculates a reference vector based on periods of high activity where the user is likely to be upright, and calculates a position vector corresponding to the user's posture throughout the entire wearing period. The system calculates the body angle as the polar angle between the reference vector and the position vector in a spherical coordinate system, where zero degrees represents an upright posture and ninety degrees represents a supine posture.

[0023] The system can also determine a user's body posture by classifying calculated body angles into posture categories using predefined angle thresholds. For example, a body angle close to zero degrees may indicate an upright or standing posture, a moderate angle (e.g., between 30 and 60 degrees) may correspond to a tilted posture, and an angle close to ninety degrees may indicate a lying or supine posture. In some implementations, the system uses the determined body angles and postures to inform sleep and activity predictions. For example, the system may determine that a supine posture during periods of relatively low movement indicates inactivity, while an upright posture with increased movement may correspond to an active state.

[0024] In various examples, the system can also determine the orientation of the wearable device, such as detecting device inversion events during wear. For instance, a wearable device may flip or rotate from its intended orientation during initial application or when reattached or repositioned mid-wear. Such inversion events can affect the accuracy of accelerometer-based predictions and physiological measurements. Therefore, the system can utilize the rolling average of one or more accelerometer axial components for detection and achieve automatic correction of signal polarity to maintain data integrity throughout the long monitoring period.

[0025] Furthermore, the system can utilize electrocardiogram (“ECG”) data acquired during the wearing period to inform the analysis and / or generate correlations between ECG-derived analysis results and accelerometer-based analysis results. In some examples, the system leverages machine learning systems to achieve this, including the use of individual trained machine learning models / algorithms. Machine learning systems can improve accuracy by combining accelerometer-based predictions with ECG-based predictions, for example, through heuristic combination schemes. By applying machine learning algorithms trained on low-sampling-rate training data and enhanced with ECG-derived predictions, the techniques described herein enable the generation of analysis results under constraints that cannot be achieved using conventional methods that only downsample high-frequency algorithms.

[0026] The system can also be configured to present the generated analysis results, such as in a comprehensive report output during and / or at the end of the wearing period. For example, the report may include various information, such as a summary section depicting aggregated user status data and / or a daily breakdown section correlating sleep and activity information with physiological data such as heart rate patterns. These reports enable users and healthcare providers to visualize patterns and trends in patient behavior and physiological responses over longer monitoring periods, thereby facilitating informed clinical decision-making.

[0027] In this way, the techniques, methods, and systems described herein offer significant advantages over conventional systems by substantially reducing power consumption while improving monitoring accuracy. Therefore, these techniques enable continuous monitoring for extended periods of wear (e.g., 14 days) without the need for charging, overcoming the limitations of traditional devices that rely on daily charging. Furthermore, the techniques described herein support the generation of analytical results and / or adjustment of signal processing based on one or more interrelated factors such as sleep / activity status, body posture, device orientation, and ECG measurements to provide clinically relevant and actionable information.

[0028] In some aspects, the technology described herein relates to a method comprising: receiving low-sampling-rate accelerometer data acquired by a wearable device attached to the skin surface of a user's chest region during a wear period, the low-sampling-rate accelerometer data having a sampling rate below a sampling rate threshold; processing the low-sampling-rate accelerometer data to extract one or more motion-derived parameters, the one or more motion-derived parameters characterizing the user's temporal motion patterns during the wear period; and generating, based on the one or more motion-derived parameters, analytical results related to the user's state during the wear period for presentation.

[0029] In some respects, the techniques described herein relate to a method in which low-sampling-rate accelerometer data is acquired at a sampling rate of approximately 1.56 Hz and worn for a period of 1 to 14 days.

[0030] In some respects, the technology described herein relates to a method in which the analysis results include predictions of a user's sleep, activity, and inactivity states during one or more time intervals of the wearing period.

[0031] In some respects, the techniques described herein relate to a method in which one or more motion-derived parameters include: an acceleration amplitude over a specific time interval of the wearing period, the acceleration amplitude being calculated as the square root of the sum of squares of one or more acceleration components of low-sampling-rate accelerometer data; and an activity parameter being calculated as the standard deviation of the acceleration amplitude over the specific time interval, wherein a relatively low standard deviation corresponds to the user being stationary and a relatively high standard deviation corresponds to the user being in motion.

[0032] In some respects, the technology described herein relates to a method in which the analysis results include the user's body angles or body postures during one or more time intervals of the wearing period.

[0033] In some respects, the techniques described herein relate to a method in which one or more motion-derived parameters include a reference vector corresponding to the user’s upright posture and generated based on a portion of low-sampling-rate accelerometer data indicating relatively high activity, and body angles are calculated as polar angles between the reference vector and the position vector at a specific moment during the wearing period in a spherical coordinate system.

[0034] In some respects, the techniques described herein relate to a method in which the analysis results include the detection of inversion events of a wearable device during a period of wear, and the analysis results are generated based on motion-derived parameters, which include the rolling average of the accelerometer axis components of low-sampling-rate accelerometer data over a specific time interval of the period of wear.

[0035] In some respects, the techniques described herein relate to a method that also includes configuring the analysis results to be presented in a user interface as part of a report, which includes: a summary section depicting aggregated user status data over the wearing period; and a daily detail section depicting the user status in temporal relation to physiological data collected during the wearing period.

[0036] In some respects, the technology described herein relates to a method that also includes receiving electrocardiogram (“ECG”) data acquired by a wearable device, and wherein generating analysis results is also based on the ECG data.

[0037] In some aspects, the technology described herein relates to a processing device comprising: one or more processors; and a memory storing computer-readable instructions executable by the one or more processors to perform operations including: receiving accelerometer data with a sampling rate below a sampling rate threshold, the accelerometer data being acquired by an accelerometer of a wearable device attached to the surface of a user's skin during a wear period; processing the accelerometer data to extract one or more motion-derived parameters characterizing a user's temporal motion pattern during the wear period; and generating, based on the one or more motion-derived parameters, analysis results related to the conditions during the wear period for presentation.

[0038] In some respects, the technology described herein relates to a processing device in which accelerometer data is acquired at a sampling rate of approximately 1.56 Hz and worn for a period of 1 to 14 days.

[0039] In some respects, the technology described herein relates to a processing device in which the analysis results include the user's sleep state, activity state, and inactivity state throughout the entire wearing period.

[0040] In some respects, the technology described herein relates to a processing device in which the analysis results include the user's body angles or body postures throughout the wearing period.

[0041] In some respects, the technology described herein relates to a processing device in which the analysis results include the detection of whether an inversion event occurs during the wearing period of a wearable device.

[0042] In some respects, the technology described herein relates to a processing device in which operation further includes receiving electrocardiogram (“ECG”) data acquired by an ECG sensor of a wearable device and generating analysis results in part based on the ECG data.

[0043] In some respects, the technology described herein relates to a processing device in which operation further includes configuring the analysis results to be presented in a user interface as part of a report, the report including: a summary section depicting aggregated user status data over the wearing period; and a daily details section including heart rate data overlaid on sleep and activity data.

[0044] In some aspects, the technology described herein relates to a system comprising: an accelerometer sensor of a wearable device configured to acquire low-sampling-rate accelerometer data via contact with a user's skin surface during a period of wear; and one or more processors configured to: receive the low-sampling-rate accelerometer data, the sampling rate of which is below a sampling rate threshold; process the low-sampling-rate accelerometer data to extract one or more motion-derived parameters characterizing a user's temporal motion patterns during the period of wear; and present an analysis result generated based on the one or more motion-derived parameters, the analysis result indicating the user's state during the period of wear.

[0045] In some respects, the technology described herein relates to a system in which the analysis results include one or more of the following: the user's sleep or activity state during the wearing period, the user's body posture, or device inversion events of the wearable device.

[0046] In some respects, the technology described herein relates to a system that also includes one or more electrocardiogram (ECG) sensors in a wearable device, wherein one or more processors are configured to receive ECG data acquired by the ECG sensors and determine analysis results based on the ECG data and low-sampling-rate accelerometer data.

[0047] In some respects, the technology described herein relates to a system in which one or more processors are configured to process low-sampling-rate accelerometer data by applying a trained machine learning algorithm to extract one or more motion-derived parameters and generate analysis results. The trained machine learning algorithm has been trained based on historical accelerometer data and corresponding user state labels.

[0048] Figure 1 This is a block diagram of a non-limiting example 100 of an environment in which the system described herein can be employed. Example 100 includes a person 102, shown wearing monitoring device 104 (i.e., a wearable device). The environment also includes an analytics platform 106. Analytics platform 106 may be directly connected to monitoring device 104 via one or more wireless connections, or via one or more wired and / or wireless connections and one or more intermediate devices, such as computing devices, network routing devices and equipment, server equipment and / or the Internet, etc., associated with person 102.

[0049] Monitoring device 104 can be used to monitor one or more aspects of person 102, for example, to generate measurement results 108. For example, in some cases, monitoring device 104 can be configured to record the electrocardiographic activity of person 102 over an observation period (e.g., lasting several seconds or minutes, lasting several days, etc.). As an example, person 102 may have the amplitude of his or her electrocardiographic potential monitored over time to generate one or more electrocardiograms (ECGs), which can be used to predict any of a plurality of events. In at least one example, monitoring device 104 is configured to record accelerometer measurements and / or electrocardiogram (“ECG”) measurements over an observation period. Alternatively or additionally, monitoring device 104 can be used to output measurement results 108 (e.g., a time series of measurement results such as a time series of potential measurements), which can indicate the observations or be used to generate predictions for one or more events.

[0050] Instructions may be provided to person 102 in conjunction with the monitoring device 104, instructing person 102 on how to operate the monitoring device 104 and / or how to behave (e.g., sleep, engage in activities) while wearing the monitoring device 104. In one or more embodiments, these instructions may be provided as part of a kit (e.g., written instructions). Alternatively or additionally, the analysis platform 106 may enable the instructions to be transmitted and output (e.g., for display and / or audio output) via a computing device associated with person 102. In one or more embodiments, the analysis platform 106 may provide these instructions for output only after a predetermined duration (e.g., two days) of observation period in which the monitoring device 104 is worn has elapsed and / or based on patterns in various aspects of person 102 measured.

[0051] Monitoring device 104 can be configured in various ways to monitor one or more aspects of personnel 102. Furthermore, Figure 1 and Figure 2 The external dimensions shown are merely exemplary, and the external dimensions of monitoring device 104 may vary in variations. It should be understood that monitoring device 104 may be configured with one or more sensors, examples of which include one or more of the following: multiple electrodes (e.g., which may be placed on a person's skin), an accelerometer and a pulse oximeter (e.g., for measuring and recording blood oxygen saturation (SpO2) and / or generating a photoplethysmogram of person 102), etc. Of course, monitoring device 104 may be configured with any of a variety of sensor types without departing from the described technique.

[0052] While the monitoring device 104 can be configured in a manner similar to that of monitoring devices used for clinical monitoring of patients, in one or more embodiments, the monitoring device 104 may be configured differently from devices used for clinical monitoring and / or diagnosis of patients. By way of example and not limitation, the monitoring device 104 may be configured as a form factor such as a ring, watch, patch, and / or strap. Alternatively or additionally, the monitoring device 104 may have a form factor similar to that of a clinical setup but with different functions, such as the function of preventing the wearer from viewing the measurement results 108.

[0053] In one or more embodiments, the monitoring device 104 may be configured to unload the measurement results 108 during the observation period. As an example, the monitoring device 104 may unload these measurement results by transmitting them to an external computing device via a wired or wireless connection, for example, during a predetermined time interval and / or in response to establishing or re-establishing a connection with the computing device. In one or more embodiments, the measurement results 108 and / or other data from the monitoring device 104 may be compressed by the monitoring device 104 for wireless transmission (e.g., using one or more of a variety of data compression techniques). Compressing sensor data in this way can reduce the battery consumption of the monitoring device 104 during the observation period and facilitate wear during physiological condition assessments.

[0054] In cases where monitoring device 104 can be configured to store measurement results 108 for the entire observation period, in one or more embodiments, monitoring device 104 can be configured without wireless transmission means; for example, without any antenna for wirelessly transmitting measurement results 108, and without hardware or firmware for generating such wireless transmission packets. Alternatively, monitoring device 104 can be configured with hardware for transmitting measurement results 108 via a physical wired connection. In these cases, monitoring device 104 can be "plugged in" to retrieve measurement results 108 from its storage.

[0055] Therefore, the monitoring device 104 may be configured with one or more ports to enable wired transmission of measurement results 108 to an external computing device. Examples of such physical coupling may include a micro universal serial bus (USB) connection, a mini USB connection, and a USB-C connection, etc. Although the monitoring device 104 may be configured to retrieve measurement results 108 via a wired connection as described above, in different situations, the monitoring device 104 may alternatively or additionally be configured to offload measurement results 108 via one or more wireless connections.

[0056] Once the monitoring device 104 generates a measurement result 108, it provides the measurement result 108 to the analysis platform 106. As described above, the measurement result 108 can be transmitted to the analysis platform 106 via one or more wired and / or wireless connections.

[0057] For example, where the analysis platform 106 is partially or wholly implemented on the monitoring device 104, the measurement results 108 can be transferred from the device's local storage to the device's processing system via a bus. Where the monitoring device 104 is configured to generate one or more prediction results 110 by processing the measurement results 108, the monitoring device 104 can also be configured to provide the generated one or more prediction results 110 as output (e.g., by transmitting one or more prediction results 110 to an external computing device). In other cases, the measurement results 108 can be processed by an external computing device configured to generate one or more prediction results 110. For example, the measurement results 108 can be processed by a smartphone associated with a user, a smartphone or other dedicated device associated with the monitoring device 104, and / or one or more server computers in a data center or other location, which can be used by an entity associated with the monitoring device 104, etc. In other words, these other devices can implement at least a portion of the analysis platform 106 and / or the prediction system 114.

[0058] In one or more embodiments, the monitoring device 104 is configured to transmit measurement results 108 to an external device via a wired connection (e.g., via USB-C or some other physical communication coupling). Here, a connector can be inserted into the monitoring device 104, or the monitoring device 104 can be inserted into a device having a socket that engages with a corresponding contact of the device. The measurement results 108 can then be obtained from the storage device of the monitoring device 104 via this wired connection, for example, by transmitting the measurement results to the external device via the wired connection. This connection can be used in situations where, after an observation period, a person 102 mails the monitoring device 104 to, for example, a healthcare provider, a telemedicine service provider, the provider of the monitoring device 104, or a medical testing laboratory.

[0059] Alternatively or additionally, monitoring device 104 may provide measurement results 108 to analysis platform 106 by transmitting measurement results 108 via one or more wireless connections. For example, monitoring device 104 may wirelessly transmit measurement results 108 to external computing devices, such as mobile phones, tablets, laptops, smartwatches, other wearable health trackers, etc. Therefore, monitoring device 104 may be configured to communicate with external devices using one or more wireless communication protocols or technologies. As an example, monitoring device 104 may communicate with external devices using one or more of Bluetooth (e.g., Bluetooth Low Energy link), Near Field Communication (NFC), Long Term Evolution (LTE) standards such as 5G. In the case of transmitting measurement results 108 to external devices for processing, monitoring device 104 may be configured with corresponding antennas and other wireless transmission devices. In these cases, measurement results 108 may be transmitted to analysis platform 106 in various ways, such as at predetermined time intervals (e.g., daily, hourly, or every five minutes), in response to the occurrence of some event (e.g., the storage buffer of monitoring device 104 is filled), or in response to the end of an observation period, etc.

[0060] Therefore, regardless of where the analysis platform 106 is implemented (e.g., at monitoring device 104, at a smartphone associated with person 102, or at a server device), the analysis platform 106 obtains the measurement results 108 generated by monitoring device 104. In one or more embodiments, the analysis platform 106 also obtains additional measurement results generated by monitoring device 104 and / or any other device used during the observation period (e.g., smartwatch, chest strap, etc.).

[0061] In one or more embodiments, the analysis platform 106 may be implemented wholly or partially at the monitoring device 104. Alternatively or additionally, the analysis platform 106 may be implemented wholly or partially using one or more computing devices external to the monitoring device 104, such as one or more computing devices associated with person 102 (e.g., mobile phone, tablet, laptop, desktop computer, or smartwatch), or one or more computing devices associated with a service provider (e.g., healthcare provider, telemedicine service provider, service provider corresponding to the provider of monitoring device 104, medical testing laboratory service provider, etc.). In the latter case, the analysis platform 106 may be implemented at least partially in one or more server devices.

[0062] In Example 100, the analytics platform includes a storage device 112. According to the described technology, the storage device 112 is configured to maintain measurement results 108 and / or other measurement results or information processed by the prediction system 114 to generate one or more prediction results 110. The storage device 112 may represent one or more databases and other types of storage devices capable of storing measurement results 108 and / or other types of measurement results. The storage device 112 may also store various other data, such as personal information, demographic information describing person 102, information about healthcare providers, information about insurance providers, payment information, prescription information, determined health indicators, account information (e.g., username and password), etc. The storage device 112 may also maintain data on other users within the user group and / or data used to support the operation of one or more machine learning systems.

[0063] In the illustrated example 100, the analysis platform 106 also includes a prediction system 114. The prediction system 114 represents the functionality for processing measurement results 108 to generate one or more prediction results 110. Alternatively or additionally, the prediction system 114 may output one or more time series indicating observations or predictions of one or more events over time. It should also be understood that, in various variations, the prediction system 114 may output different combinations of multiple prediction results.

[0064] In at least one embodiment, the prediction system 114 uses machine learning to generate one or more prediction results 110. By way of example and not limitation, the prediction system 114 may include one or more neural networks trained based on historical measurement results and historical result data of a user group. The prediction system 114 may include one or more machine learning models (e.g., a collection of models and / or algorithms). Alternatively or additionally, the prediction system 114 may include logic for preprocessing the acquired measurement results (machine learning models and / or other types of logic), such as extracting various cardiovascular features and / or other features from the sequence of measurement results. Example 100 also includes prediction results 110 corresponding to the output of the prediction system 114.

[0065] In various examples, the prediction result 110 includes and / or represents the analysis result 116. The analysis result 116 may represent a determination, classification, and / or analysis derived from the measurement result 108, which provides information about the status during the wearing period, such as the status and / or characteristics of the person 102 and / or the monitoring device 104 during the observation period. By way of example and not limitation, the analysis result 116 may include one or more of sleep / activity status 118, body posture 120, device orientation 122, and / or multimodal relationship 124.

[0066] As described in further detail below, sleep / activity status 118 can represent a classification of the behavioral or physiological state of person 102 during the observation period, such as whether person 102 is asleep, actively engaging in physical movement, or awake but inactive. Body posture 120 can represent the determination of the body orientation or posture of person 102 during the observation period, such as whether person 102 is upright, tilted, or lying down. Device orientation 122 can represent the determination of the physical location or alignment of monitoring device 104 during the observation period, such as whether the device has been inverted or rotated from its expected orientation. For example, multimodal relationships 124 can represent the correlation or integration between multiple data sources such as accelerometer data and ECG data, thereby enabling improved predictive accuracy through combined analysis of different physiological and kinematic parameters.

[0067] In various examples, one or more operations of the analysis platform 106 and / or the prediction system 114 can be performed substantially in real time by the monitoring device 104 and / or by one or more devices not physically connected to the monitoring device 104 and / or as post-processing operations.

[0068] Figure 2 A non-limiting example 200 of a monitoring device is shown. Example 200 shows a monitoring device 104.

[0069] According to the described technology, the monitoring device 104 includes one or more sensors 202, examples of which include, but are not limited to, a pair or more pairs of electrodes, an accelerometer, a pulse oximeter, and a sweat sensor, etc. The monitoring device 104 may also include a transmitter 204. In this example 200, the monitoring device 104 also includes one or more adhesive portions 206. In operation, the monitoring device 104 is configured to adhere to the skin via the one or more adhesive portions 206, such that, for example, one or more sensors 202 are positioned to detect and record the electrical activity of the heart of the person 102, for example, to generate an electrocardiogram (ECG and / or EKG). In at least one embodiment, the monitoring device 104 can be removed by peeling the one or more adhesive portions 206 from the skin.

[0070] It should be understood that the monitoring device 104 and its various components are only one form factor, and the monitoring device 104 and its components may have different form factors without departing from the spirit or scope of the described technology.

[0071] In one or more embodiments, monitoring device 104 may include a processor and / or memory (not shown). Monitoring device 104 may utilize the processor to generate measurement results 108 based on communication with one or more sensors 202, which indicate aspects of the person 102, such as the physical activity of the person 102's heart, sleep state, electrical activity, etc. In one or more embodiments, the processor also generates one or more communicable data packets including one or more measurement results 108 and / or other measurement results, such as low-sampling-rate accelerometer data and ECG measurements. Alternatively or additionally, the processor generates and / or stores other data that can be used to predict the classification of physiological conditions, such as sleep apnea.

[0072] In embodiments where monitoring device 104 is configured for wireless transmission, transmitter 204 can wirelessly transmit measurement results 108 as a data stream to a computing device. For example, in one or more embodiments, monitoring device 104 is configured to transmit (e.g., send and / or receive) information (e.g., potential measurement results) via a Bluetooth Low Energy (BLE) connection. Alternatively or additionally, monitoring device 104 can buffer measurement results 108 (e.g., in memory) and cause transmitter 204 to subsequently transmit the buffered measurement results at various intervals (e.g., time intervals (per second, every thirty seconds, per minute, every five minutes, per hour, etc.), storage intervals (when the buffered measurement results reach a threshold data amount), etc.).

[0073] Physiological monitoring using low sampling rate accelerometer data

[0074] The following discussion describes techniques that can be implemented using the aforementioned systems and devices. Each aspect of a process can be implemented using hardware, firmware, software, or a combination thereof. These processes are shown as a set of boxes specifying operations that can be performed by one or more devices, and these processes are not necessarily limited to the order in which the operations are performed by the respective boxes. For example, one or more boxes of a process specify operations that can be programmed into instructions by hardware (e.g., a processor, microprocessor, controller, firmware) to create a dedicated machine for executing the algorithm shown in the flowchart. Therefore, these instructions can be stored on a computer-readable storage medium that enables the hardware to execute the algorithm. Reference will be made to the following sections. Figures 1 to 14 .

[0075] Figure 3 A non-limiting system for physiological monitoring using low-sampling-rate accelerometer data is shown in example embodiment 300, wherein a more detailed description is provided. Figure 1The operation of the prediction system 114. In various examples, the prediction system 114 represents, supports, and can be implemented by and / or (partially or wholly) included by a wearable device such as monitoring device 104.

[0076] For example, prediction system 114 represents a system architecture for processing low-sampling-rate accelerometer data and generating various analytical results 116 related to the conditions during the wearing period, such as one or more user conditions when monitoring device 104 is attached to the user's skin surface. Prediction system 114 is shown to include multiple interconnected modules, such as sensor module 302, analysis module 304, and presentation module 306, which can be configured to acquire, process, and present physiological data collected from the user during extended wearing periods (e.g., 1 to 14 days) and derive analytical results from that physiological data. In various examples, prediction system 114 can record data for extended wearing periods on a single battery without requiring recharging; therefore, the techniques described herein achieve continuous monitoring while maintaining effectiveness through low-sampling-rate data acquisition and processing techniques.

[0077] For example, sensor module 302 may house one or more data acquisition components for obtaining physiological measurements from a user. As shown, sensor module 302 includes an accelerometer 308 configured to acquire accelerometer data 310 during the period of wear. Accelerometer 308 may be implemented as one or more of a variety of motion sensing devices, such as a microelectromechanical system (“MEMS”) accelerometer, a piezoelectric accelerometer, or a capacitive accelerometer, and may be configured to detect acceleration forces along one or more axes. In some examples, accelerometer 308 may be a triaxial accelerometer including x, y, and z components, but single-axis, dual-axis, or various multi-axis configurations are also contemplated.

[0078] For example, accelerometer data 310 is low-sampling-rate accelerometer data below a sampling rate threshold. As an example and not a limitation, the sampling rate threshold can be approximately 5 Hz, 3 Hz, 2 Hz, or 1.56 Hz, but other sampling rates are conceivable without departing from the described technical scope. In various embodiments, the prediction system 114 maintains the sampling rate threshold within a specific range, for example, between 1.0 Hz and 2.0 Hz. In at least one example, accelerometer data 310 is sampled at a variable rate throughout the wearing period, for example, due to the detection of one or more conditions. The relatively low sampling rate maintains sufficient data resolution to support the generation of analysis results 116 according to the data processing techniques described herein while achieving long monitoring periods.

[0079] Sensor module 302 is also shown to include ECG sensor 312. For example, ECG sensor 312 includes one or more electrodes and / or processing elements configured to acquire electrocardiogram (“ECG”) measurements and generate ECG data 314. ECG sensor 312 can be implemented using various electrode configurations (e.g., two-electrode systems, three-electrode systems, other multi-electrode arrangements, etc.) that can be positioned on a user’s chest region to detect cardiac electrical activity. In some embodiments, ECG sensor 312 may include signal conditioning circuitry, an amplifier, and an analog-to-digital converter to process raw electrical signals from the heart into digital ECG data 314.

[0080] ECG data 314 can capture various cardiac parameters, including heart rate, rhythm patterns, and waveform characteristics, which can be analyzed in conjunction with accelerometer data 310 to provide multimodal physiological analysis results. ECG data 314 can be sampled at various sampling rates and processed using various filtering and / or noise reduction techniques, for example, to improve signal quality during prolonged wear periods. As described in more detail below, the prediction system 114 can generate various multimodal analysis results based on accelerometer data 310 and ECG data 314.

[0081] For example, analysis module 304 can receive one or more accelerometer data 310 and / or ECG data 314 and perform various data processing operations to generate analysis results 116. For example, analysis results 116 may represent determinations, classifications, and / or analyses derived from measurement results 108, providing information about the status during the wearing period, such as the status and / or characteristics of the person 102 and / or monitoring device 104 during the observation period. As an example and not a limitation, analysis results 116 may include one or more of sleep / activity states 118, body posture 120, device orientation 122, and / or multimodal relationships 124.

[0082] To this end, analysis module 304 processes low-sampling-rate accelerometer data 310 to extract one or more motion-derived parameters 316, which characterize the user's temporal motion patterns during the wearing period. For example, motion-derived parameters 316 represent calculated values ​​extracted from raw accelerometer data 310 that quantify various aspects of the user's movement and posture over time. Analysis module 304 may utilize one or more of state module 318, posture module 320, orientation module 322, multi-parameter module 324, and / or machine learning system 326 to generate analysis results 116 based on various motion-derived parameters 316. Therefore, the specific motion-derived parameters 316 generated can vary based on the specific analysis results 116 generated, allowing different sets of parameters to be calculated depending on whether prediction system 114 is generating sleep / activity state 118, body posture 120, device orientation 122, multimodal relationship 124, or a combination thereof.

[0083] For example, the state module 318 can analyze motion-derived parameters 316 to determine the user's sleep / activity state 118 throughout the wearing period, such as sleep, activity, and / or inactivity. In the example, the state module 318 generates and / or processes motion-derived parameters 316, such as acceleration amplitude and activity parameters, to classify the user's state. For example, the state module 318 can calculate the total acceleration amplitude as the square root of the sum of the squares of the acceleration components along the three orthogonal axes of the accelerometer data 310. The state module 318 can also generate activity parameters as the standard deviation of the acceleration amplitude over a predetermined time window. In some examples, a relatively low standard deviation can indicate that the user is at rest, corresponding to an inactive or sleep state, while a relatively high standard deviation can indicate motion, corresponding to an active state.

[0084] The state module 318 can also distinguish between awake and inactive states by applying a threshold to the motion-derived parameter 316, thereby enabling the classification of sleep / activity states 118 at minute-level resolution throughout the entire extended wear period. In some examples, the state module 318 can incorporate additional analysis results 116 from other modules within the analysis module 304 (such as body angle information from the posture module 320) to improve the accuracy of determining the sleep / activity state 118 by providing contextual information about the user's posture during different activity periods. The following section combines... Figure 4 The functions of the status module 318 will be discussed in further detail.

[0085] The posture module 320 can analyze motion-derived parameters 316 to determine the user's body angles and / or body posture 120 throughout the wearing period. In the example, the posture module 320 generates and / or processes motion-derived parameters 316, including a reference vector and / or a position vector, to calculate body posture information. For example, the posture module 320 can calculate a reference vector based on periods of high activity where the user is likely to be upright, and calculate a position vector corresponding to the user's posture at various times throughout the wearing period. The posture module 320 can also calculate body angles as the polar angle between the reference vector and a specific position vector in a spherical coordinate system. This polar angle can range from 0 degrees (e.g., the position vector is parallel to the reference vector) to 90 degrees (e.g., the position vector is substantially perpendicular to the reference vector).

[0086] The posture module 320 can also determine the user's body posture 120 by classifying the calculated body angles into posture categories using predefined angle thresholds. For example, body angles between 0 and 15 degrees can correspond to an upright or standing posture, intermediate angles greater than 15 degrees and less than 85 degrees can correspond to a tilted posture, and angles between 85 and 90 degrees can represent a lying or supine posture. The above are merely examples and not limitations, and various angle threshold ranges are considered.

[0087] In some implementations, the posture module 320 uses the determined body angles and body postures 120 to inform sleep and activity predictions generated by other modules within the analysis module 304. For example, the posture module 320 can determine that a supine posture during a period of relatively low movement indicates a sleep state, an upright posture with relatively low movement indicates a wakeful but inactive state, and an upright posture with increased movement can correspond to an active state. The posture module 320 is capable of processing accelerometer data 310 to calculate body angles and detect the user's body postures 120 in various time intervals during the wearing period. The following is in conjunction with... Figure 5 Exemplary functions of gesture module 320 are discussed.

[0088] The orientation module 322 can analyze motion-derived parameters 316 to determine device orientation 122 and detect inversion events of the wearable device during wear periods. For example, an inversion event refers to a situation where the wearable device is attached to the skin surface in an "upside-down" position, such as during initial device application and / or wear periods after device removal. In this example, the orientation module 322 generates and / or processes motion-derived parameters 316, including rolling averages of one or more accelerometer axis components, to monitor device positioning and identify orientation changes. For example, the orientation module 322 can calculate rolling averages of the x-accelerometer and y-accelerometer components over a predetermined time interval (e.g., a 24-hour period) to establish a baseline device orientation pattern and detect deviations indicating inversion events.

[0089] Orientation module 322 can also determine device inversion events by comparing the range of the rolling average with a predefined threshold. For example, if the range of the rolling average exceeds the threshold, orientation module 322 can identify that a device flip or inversion event has occurred. In some embodiments, orientation module 322 uses the detected inversion event to inform signal processing corrections (such as automatic adjustment of ECG signal polarity and / or correction based on accelerometer characteristics) for other modules within analysis module 304 to maintain data integrity over long monitoring periods. Thus, orientation module 322 can process accelerometer data 310 to detect changes in device orientation and provide inversion event information that can be used to improve the accuracy of other physiological measurements and predictions during the wearing period.

[0090] The multi-parameter module 324 can combine additional sensor data to analyze motion-derived parameters 316, generating analysis results 116 based on a multimodal data source. In this example, in addition to generating and / or processing ECG data 314, the multi-parameter module 324 also generates and / or processes motion-derived parameters 316 to create a comprehensive analysis result 116 utilizing information from accelerometer-based monitoring and cardiac monitoring. For example, the multi-parameter module 324 can combine acceleration amplitude and activity parameters with heart rate variability indices to distinguish different types of inactivity states, such as differentiating states like "sound sleep" and "sedentary wakefulness" based on movement patterns and heart rhythm characteristics.

[0091] The multi-parameter module 324 can also implement machine learning algorithms trained on multimodal datasets (e.g., machine learning algorithms from machine learning system 326) to optimize the integration of accelerometer-derived features and ECG-derived features. For example, the multi-parameter module 324 can apply a weighted combination scheme that dynamically adjusts the importance of accelerometer-based predictions relative to ECG-based predictions based on signal quality, user-specific patterns, computational resource consumption, and / or temporal context during the wearing period. In some implementations, the multi-parameter module 324 utilizes analysis results generated by one or more other modules within the analysis module 304 (such as by combining body posture information from the posture module 320) to provide contextual information about the user's posture during various heart rhythm patterns, thereby improving the accuracy of sleep phase detection. In this way, the multi-parameter module 324 can process both accelerometer data 310 and ECG data 314 to generate multimodal relationship 124 analysis results, thereby providing robust and clinically relevant predictions.

[0092] The analysis module 304 can also utilize the machine learning system 326 to generate one or more analysis results 116. For example, the machine learning system 326 can apply trained algorithms to generate and / or process motion-derived parameters 316 and other sensor data to generate accurate user state classification and physiological analysis results. The machine learning system 326 may include one or more trained models optimized for processing low-sampling-rate data, such as neural networks, decision trees, support vector machines, or ensemble methods combining multiple algorithmic approaches.

[0093] In some implementations, the machine learning system 326 can employ separate models for different prediction tasks, such as specialized models for sleep detection, activity classification, body posture determination, and / or multimodal prediction. These models can be trained on domain-specific datasets to improve the accuracy of their respective functions. The machine learning system 326 can also incorporate adaptive learning capabilities that adjust model parameters based on user-specific patterns observed during wear time to achieve personalized predictions that consider individual differences in movement patterns, sleep behavior, and physiological responses. Furthermore, the machine learning system 326 can implement feature selection and dimensionality reduction techniques to optimize processing efficiency while maintaining prediction accuracy in situations where wearable devices have limited computational resources. Examples of this are illustrated below. Figure 6a and Figure 6b Further details are provided on additional features and examples of machine learning system 326.

[0094] Therefore, the analysis result 116 can provide a variety of meaningful information about various aspects of the user's physiological state and device performance during the monitoring period. The analysis result 116 can be generated using various computational methods (e.g., algorithmic processing, statistical analysis, machine learning techniques, or combinations thereof), and can include both real-time determination and post-processing analysis. In various implementations, the analysis result 116 can be presented in different formats such as numerical values, categorical classifications, graphical representations, or text summaries, and can be customized for specific audiences such as healthcare providers, researchers, and end users.

[0095] The presentation module 306 can format the analysis results 116 and configure them for output to users, healthcare providers, or other systems. In various implementations, the presentation module 306 can generate visual displays, text summaries, graphical representations, or interactive interfaces to present processed physiological data in a clinically meaningful format. The presentation module 306 can also implement customizable reporting capabilities, allowing for the emphasis or filtering of different types of information based on specific use cases or user preferences.

[0096] For example, presentation module 306 can generate a report 328 that includes one or more analysis results 116. Report 328 can be configured in various formats, such as a PDF document, a web-based interactive interface, a mobile application display, or other digital or print formats suitable for healthcare providers and users. In some implementations, report 328 may include a summary section providing aggregated data for the entire wearing period, a daily detail section displaying time patterns, and graphical visualizations that correlate multiple data types, such as heart rate patterns overlaid with sleep and activity information.

[0097] In some examples, the prediction system 114 can also dynamically adjust device attributes or operating parameters based on the generated analysis results 116. For example, the prediction system 114 may modify the sampling rate of one or more sensors during a period of wear in response to a detected user state, such as increasing the accelerometer sampling frequency to capture detailed motion patterns during active periods, or decreasing the sampling rate to conserve battery power during sleep. The prediction system 114 may also trigger the activation or deactivation of one or more sensors based on the analysis results 116. In some implementations, the prediction system 114 may adjust signal processing parameters in real time and / or implement power management strategies that optimize battery usage by selectively enabling or disabling various device functions based on one or more analysis results 116.

[0098] In various implementations, the processing operations performed by the analysis module 304 can be performed during and / or after the wearing period, and can be performed synchronously and / or asynchronously. For example, the extraction of motion-derived parameters 316 and the generation of analysis results 116 can be performed in real time on the monitoring device 104 during the wearing period to provide immediate feedback or trigger additional sensor operations based on the detected user status. Additionally or alternatively, accelerometer data 310 and / or ECG data 314 can be transmitted wirelessly or via a wired connection to one or more external computing devices (e.g., smartphones, tablets, or server systems) for analysis. In at least one example, accelerometer data 310 and ECG data 314 are stored locally on the monitoring device 104 during the wearing period and downloaded at the end of the wearing period.

[0099] Figure 4 A non-limiting example 400 of physiological monitoring using low-sampling-rate accelerometer data is shown, wherein analysis results 116 including sleep / activity status 118 are generated.

[0100] Example 400 includes a first scenario 402 and a second scenario 404, which illustrate how the analysis module 304 processes accelerometer data 310 collected by the wearable device to generate a sleep / activity state 118 based on the user's temporal movement patterns during the wearing period. In the first scenario 402, the user is in a supine position for approximately fifteen minutes, representing a period of relatively minimal physical activity (e.g., a resting state). The accelerometer data 310 collected by the monitoring device 104 during this period includes three orthogonal acceleration components (e.g., x, y, and z components) corresponding to the respective axes of the monitoring device 104. During the resting state, the accelerometer data 310 remains relatively constant on the three axes, exhibiting only relatively small fluctuations due to natural physiological movements such as breathing or slight postural adjustments.

[0101] Analysis module 304 processes accelerometer data 310 from first case 402 to calculate one or more motion-derived parameters 316, such as acceleration amplitude 406 and activity parameters 408. Acceleration amplitude 406 can be calculated as the square root of the sum of squares of the individual acceleration components of the low-sampling-rate accelerometer data 310. For example, for an acceleration vector at a given time t... The acceleration amplitude of 406 can be calculated as follows: .

[0102] For a specific time interval 410 of the wearing period, the activity parameter 408 can be calculated as the standard deviation of the acceleration amplitude 406 within the time interval 410. In the first scenario 402, the relatively constant nature of the accelerometer data 310 results in a low standard deviation for the active parameter 408 (e.g., <threshold), the lower standard deviation corresponds to the user being at rest. Based on these motion-derived parameters 316, the analysis module 304 generates analysis results 116 indicating the sleep / activity state 118 corresponding to the inactivity state.

[0103] In the second scenario 404, the user engages in a running activity lasting fifteen minutes. During this activity period, the acceleration values ​​of the accelerometer data 310 change significantly along each axis. The analysis module 304 processes the accelerometer data 310 as described above to calculate the acceleration amplitude 406 and the activity parameters 408.

[0104] The changes in accelerometer data 310 during the running activity resulted in a relatively high standard deviation for activity parameter 408 over the time interval 410 (e.g., >Threshold). A relatively high standard deviation corresponds to user motion and indicates a large amount of physical activity. Therefore, the analysis module 304 processes these motion-derived parameters 316 to generate an analysis result 116 that classifies the sleep / activity state 118 as an activity state. In this way, Example 400 demonstrates how the prediction system 114 can accurately distinguish various user states by analyzing the motion-derived parameters 316 extracted from low-sampling-rate accelerometer data 310, thereby supporting reliable classification of the sleep / activity state 118 under power-constrained conditions.

[0105] Figure 5 A non-limiting example 500 of physiological monitoring using low-sampling-rate accelerometer data is shown, wherein body angles and body postures are generated based on the accelerometer data.

[0106] Example 500 illustrates three different postures for a person 102 wearing monitoring device 104, with body angles determined based on low-sampling-rate accelerometer data. Example 500 includes: an upright posture diagram 502, showing a person 102 wearing monitoring device 104 in a vertical orientation; a supine posture diagram 504, showing a person 102 wearing monitoring device 104 in a horizontal orientation; and a tilted posture diagram 506, showing a person 102 wearing monitoring device 104 at an angle of approximately 45 degrees. These diagrams illustrate various vector components, such as a reference vector 508, a first position vector 510, a second position vector 512, and a third position vector 514.

[0107] For example, reference vector 508 represents a baseline orientation measurement used as an anchor point for subsequent body angle calculations. In various embodiments, reference vector 508 can be calculated based on accelerometer data 310 obtained during periods of high activity (e.g., walking or other movement), thus corresponding to the "upright" posture of person 102 (e.g., 90 degrees). The above are merely examples and not limitations, and in various examples, reference vector 508 is calculated using statistical analysis of accelerometer data 310 during relatively low activity periods (e.g., sleep periods), thus corresponding to the "tilted" posture of person 102 (e.g., zero degrees).

[0108] Various statistical analyses can be used to determine the reference vector 508, such as machine learning algorithms, pre-calibrated procedures, or other calculation methods used to identify when a user is most likely in a particular orientation / body angle. As an example, and not a limitation, the reference vector 508 can be calculated as the average of acceleration measurements taken during periods above an activity threshold level and / or indicating vertical orientation within the monitoring period. In various examples, one or more components of the accelerometer data 310 can be weighted during the calculation of the reference vector 508, for example, to emphasize one or more axial or directional components that are relatively more indicative of upright posture during periods of high activity, thereby improving the accuracy of body angle calculations throughout the wearing period.

[0109] To determine the body angle of person 102 at a specific moment, prediction system 114 calculates a position vector representing the orientation of monitoring device 104 based on the three-dimensional accelerometer data components at that specific moment. For example, the position vector can be derived from the raw acceleration measurement by processing the x, y, and z acceleration components to create a direction vector indicating the spatial orientation of the device. Prediction system 114 can then determine the body angle based on the polar angle between reference vector 508 in the spherical coordinate system and the specific position vector.

[0110] For example, in the example where reference vector 508 represents an upright posture, the polar angle can range from 0 degrees in the basic upright posture to 90 degrees in the basic supine posture. In the example shown, the polar angle in upright posture diagram 502 ( The polar angle is approximately 0 degrees, the polar angle in the supine position diagram 504 is 90 degrees, and the polar angle in the inclined position diagram 506 is 45 degrees.

[0111] Furthermore, the prediction system 114 is capable of generating a classification of the user's body posture 120 based on the determined body angles. As an example, and not a limitation, the prediction system 114 can classify the body posture 120 into discrete states such as upright, sitting, tilted, or supine orientations based on predetermined angle thresholds. For example, in the upright posture diagram 502, the body angle is approximately 0 degrees, and the prediction system 114 determines that the person 102 is in a standing or upright posture. In the supine posture diagram 504, the body angle is approximately 90 degrees, so the prediction system 114 determines that the person 102 is in a lying or horizontal posture. In the tilted posture diagram 506, the body angle is approximately 45 degrees, so the prediction system 114 determines that the person 102 is in an intermediate posture between an upright and supine orientation.

[0112] Figure 6a , Figure 6b and Figure 6cNon-limiting examples 600a, 600b and 600c of physiological monitoring using low sampling rate accelerometer data are shown, wherein a machine learning system 326 is trained, configured and implemented to generate analysis results 116 based on various parameters.

[0113] In Example 600a, the machine learning system 326 is configured to train one or more models to process low-sampling-rate accelerometer data and / or ECG data, thereby generating analysis results 116. As described in further detail below, the machine learning system 326 may include and / or represent one or more types of machine learning models, including but not limited to neural networks, decision trees, support vector machines, random forests, ensemble methods, convolutional neural networks, recurrent neural networks, long short-term memory networks, gradient boosting algorithms, logistic regression models, k-nearest neighbor classifiers, etc.

[0114] In this example, firstly, the machine learning system 326 includes a training module 602 that receives various types of training data and applies machine learning techniques to develop / train models for processing various data modalities to generate analysis results 116. For example, the training module 602 uses accelerometer sleep period model (“ASPM”) training data 604 to train an accelerometer sleep period (“ASP”) model 606, for example, to generate a trained ASP model 608. The training module 602 is also capable of using ECG sleep period model (“ESPM”) training data 610 to train an ESP model 612 to generate a trained ESP model 614. Depending on computing resources and system requirements, this training can be performed synchronously, asynchronously, in parallel, sequentially, and / or using various other timing schemes.

[0115] ASPM training data 604 may include training accelerometer data 616, which represents low-sampling-rate accelerometer measurements acquired from the wearable device during previous monitoring periods. ASPM training data 604 also includes a ground truth label 618, which corresponds to verified conditions during the period in which the training accelerometer data 616 was acquired. The ground truth label 618 may include various conditions during the wearing period, such as sleep state, active state, inactive state, body posture, device orientation, etc.

[0116] Training module 602 processes training accelerometer data 616 and ground truth labels 618 to train ASP model 606, thereby generating analysis results 116 based on low-sampling-rate accelerometer measurements. The training process can employ various machine learning techniques to adjust one or more weights and / or learn one or more parameters of model 606 as part of the training. In some implementations, training module 602 may utilize cross-validation techniques, feature selection algorithms, or data augmentation methods, for example, to improve model performance and generalization ability across different user groups and monitoring conditions.

[0117] Training module 602 also utilizes ESPM training data 610 to train ESP model 612, for example, to generate trained ESP model 614. For example, ESPM training data 610 includes ECG measurements with corresponding labels (e.g., sleep and activity labels), which are used to train ESP model 612 to generate predictions. In the illustrated example, ESPM training data 610 includes PSG-annotated ECG training data 620 and ASPM-annotated ECG training data 622.

[0118] For example, PSG-annotated ECG training data 620 represents ECG data collected during a polysomnography (“PSG”) study, with corresponding sleep stage labels validated, for example, according to clinical sleep monitoring standards. ASPM-annotated ECG training data 622 represents ECG measurements annotated using predictions generated by a trained ASP model 608. For example, machine learning system 326 uses the trained ASP model 608 to process additional accelerometer data 624 collected along with the corresponding ECG data to generate labels (e.g., sleep labels and activity labels), which are then applied to the corresponding ECG data. In this way, combining ASPM-annotated ECG training data 622 with ESPM training data 610 ensures that wake periods are adequately represented during the training of the ESP model 612.

[0119] Training module 602 processes PSG-annotated ECG training data 620 and ASPM-annotated ECG training data 622 to train ESP model 612, thereby generating analysis results 116 based on the collected ECG data 314. The training process may involve iteratively adjusting model parameters, such as neural network weights, decision tree thresholds, or support vector machine hyperparameters, to minimize prediction error and optimize performance on the training dataset. Various optimization techniques can be employed during training, including gradient descent algorithms, backpropagation methods, regularization methods, or ensemble learning strategies that combine multiple models to improve prediction accuracy and robustness across different user groups and monitoring scenarios.

[0120] Reference Figure 6b Example 600b illustrates implementations of a trained ASP model 608 and a trained ESP model 614 to process input data. For example, the trained ASP model 608 receives and processes accelerometer data 310 to generate accelerometer-based predictions 626, such as identifying patterns associated with the user's sleep, activity, and inactivity states during the wearing period by analyzing motion-derived parameters extracted from the accelerometer data 310. The trained ESP model 614 receives and processes ECG data 314 to generate ECG predictions 628. For example, the trained ESP model 614 can detect sleep states and / or sleep stages by analyzing heart rate variability, heart rhythm patterns, and other physiological indicators present in the ECG data 314.

[0121] Machine learning system 326 can combine accelerometer-based predictions 626 and ECG-based predictions 628 according to combination scheme 630 to generate a comprehensive prediction result 632. For example, combination scheme 630 represents a framework for integrating these predictions to improve the accuracy and reliability of analysis result 116. For example, combination scheme 630 can select between predictions generated by trained ASP model 608 and trained ESP model 614, and / or can combine two or more predictions based on various considerations. Figure 6c An exemplary combination scheme 630 is shown, and the combination scheme is discussed in further detail below.

[0122] Therefore, the overall prediction result 632 may include values ​​from either or both of the accelerometer-based prediction result 626 and / or the ECG-based prediction result 628. In one or more examples, the overall prediction result 632 includes analysis results 116 over the duration of the wearing period, such as minute-level classifications of user states (e.g., sleep periods, active periods, and inactive periods). In some examples, the overall prediction result 632 may combine body posture information and / or device orientation prediction results.

[0123] Reference Figure 6cExample 600c illustrates a combination scheme 630 for a machine learning system 326 to integrate accelerometer-based predictions 626 and ECG-based predictions 628, such as generating a comprehensive prediction 632 using heuristic decision logic. For example, the combination scheme 630 determines, for a specific time interval (e.g., minute intervals of a wearing period), whether to use a specific accelerometer-based prediction 626, a specific ECG-based prediction 628, or a combination of both. For example, the combination scheme 630 can be used to determine minute-level predictions over the duration of a wearing period. The combination scheme 630 can determine which predictions to include in the comprehensive prediction 632 based on considerations such as signal quality, prediction confidence level, and consistency with expected physiological patterns.

[0124] As an example, the machine learning system 326 can identify anomalies that may be inaccurate in the accelerometer-based prediction 626 and / or the ECG-based prediction 628. Such anomalies may include, but are not limited to, situations where the ECG-based prediction 628 (and / or ECG data 314) indicates arrhythmias above a threshold level, which may interfere with the ECG-based sleep detection algorithm. Additionally or alternatively, anomalies may occur when the deviation between the accelerometer-derived sleep / wake pattern and the expected pattern exceeds a significance threshold. Furthermore, such anomalies may also occur when the deviation between the ECG-based sleep / wake pattern and the expected pattern exceeds a significance threshold, which may indicate the presence of cardiac abnormalities or signal quality degradation affecting accuracy.

[0125] Machine learning system 326 utilizes a combination scheme 630 to reconcile this anomalous situation. For example, as shown in the first node 634, if machine learning system 326 detects that the percentage of arrhythmias in ECG data 314 during a specific time period is higher than a threshold, then combination scheme 630 selects the accelerometer-based prediction result 626 to include in the comprehensive prediction result 632. As further shown in the first node 634, if the deviations of both the accelerometer-based prediction result 626 and the ECG-based prediction result 628 from the expected values ​​are higher than a significance threshold, then combination scheme 630 selects the accelerometer-based prediction result 626. This could be because accelerometer data 310 is less susceptible to arrhythmia-related interference compared to ECG data 314.

[0126] As shown in the second node 636, if the deviation between the accelerometer-based prediction 626 and the expected nighttime pattern exceeds a significance threshold, while the ECG-based prediction 628 remains within an acceptable range, then the combination scheme 630 selects the ECG-based prediction 628 to include in the comprehensive prediction result 632. This may occur when device displacement, abnormal user motion patterns, or accelerometer sensor problems impair motion-based sleep detection, while the heart rate pattern remains stable and reliable.

[0127] As shown in the third node 638, if the conditions of the first node 634 or the second node 636 are not met, the combination scheme 630 can use alternative criteria to select the prediction result. In some examples, this may include combining the accelerometer-based prediction result 626 and the ECG-based prediction result 628, such as by averaging two or more values ​​over a specific interval using a weighted scheme. Additionally or alternatively, the combination scheme 630 may select the prediction result based on the rolling average generated by either modality (e.g., the 24-hour rolling average of the “closest” target value (e.g., 1 / 3, which represents approximately 8 hours of sleep per day). For example, if the corresponding 24-hour rolling average of the accelerometer-based prediction result 626 indicates 7.8 hours of sleep, while the ECG-based prediction result 628 indicates 5.2 hours of sleep, the combination scheme 630 may select the accelerometer-based prediction result 626 because it is relatively closer to the target value of 8 hours per day.

[0128] Therefore, the machine learning system 326 provides a robust framework for generating accurate analysis results 116 by utilizing specialized training methods to process multimodal data. Furthermore, the implementation of a combination scheme 630 that dynamically selects and / or integrates prediction results based on signal quality improves the reliability of sleep and activity classification under various monitoring conditions. In this way, the system can maintain high accuracy under power-constrained conditions, thus supporting long wearing periods without compromising prediction quality.

[0129] Figure 7 A non-limiting example 700 of physiological monitoring using low sampling rate accelerometer data is shown, in which device inversion may lead to inaccurate measurements.

[0130] Example 700 illustrates an acceleration component graph 702, a body angle graph 704, and a user state graph 706. Acceleration component graph 702 represents the y-component of the accelerometer data 310 as a rolling average over time. Acceleration component graph 702 also illustrates an inversion event 708, which is visually represented when the value of the y-component drops below zero.

[0131] Due to the inversion event 708, body angle diagram 704 illustrates that body angle calculations may become inaccurate when monitoring device 104 is inverted. This is, for example, because the reference frame used to determine the user's position relative to gravity has changed. User state diagram 706, indicating sleep / activity status throughout the monitoring period, also illustrates how predictions of sleep and activity status may be affected when device inversion occurs. For example, in various embodiments, sleep / activity status 118 is partially based on body posture 120, so device inversion event 708 may lead to incorrect analysis results associated with sleep / activity status 118. Therefore, as described in more detail in the following examples, prediction system 114 may detect and take into account inversion event 708 during the generation of analysis results 116.

[0132] Figure 8 A non-limiting example 800 of physiological monitoring using low-sampling-rate accelerometer data is shown, which includes the detection of device inversion events.

[0133] In this example 800, the prediction system 114 calculates and analyzes rolling averages of one or more acceleration components of accelerometer data 310 to identify when the wearable device is flipped or inverted during the monitoring period. For example, the prediction system 114 analyzes the y-component rolling average 802 and x-component rolling average 804, plotted over time, to detect changes in device orientation. The prediction system 114 uses rolling averages (e.g., over a 24-hour period) to smooth short-term variations while preserving long-term trends corresponding to device orientation.

[0134] If the values ​​of the y-component rolling average 802 and / or the x-component rolling average 804 exceed one or more thresholds, the prediction system 114 can determine that an inversion event has occurred. In at least one example, if the range of the y-component rolling average 802 and / or the x-component rolling average 804 exceeds 0.8, the prediction system 114 can determine that an inversion event has occurred. The prediction system 114 can also identify x-flip points 806 and / or y-flip points 808 at locations where the corresponding acceleration components cross predetermined thresholds, as shown in the illustrated example.

[0135] Therefore, the prediction system 114 is able to determine the number and timing of inversion events occurring throughout the entire wearing period. This can be output by the prediction system 114, such as included in report 328, to provide healthcare providers with context regarding device positioning during the monitoring period. This information can also be used to automatically adjust signal processing parameters for other physiological measurements, such as ECG polarity correction and body angle recalibration, to maintain data integrity throughout the wearing period.

[0136] Figures 9a to 9eA non-limiting example of physiological monitoring using low sampling rate accelerometer data is shown, wherein a report including various analytical results is output in health reporting interfaces 900a to 900e.

[0137] Reference Figure 9a The health report interface 900a displays comprehensive sleep and activity monitoring data collected during the wearing period. For example, the health report interface 900a shows a summary section depicting the aggregated user status data during the wearing period. For example, the health report interface 900a includes a sleep data graph 902 presenting sleep duration measurements for each day of the wearing period. An activity data graph 904 shows the corresponding activity duration measurements within the same time period. Both the sleep data graph 902 and the activity data graph 904 also include statistical data such as daily averages, minimums, and maximums.

[0138] The health report interface 900a includes a legend panel 906 that provides contextual information about the data aggregation method, including definitions of activity status, inactivity periods, and sleep categories. The legend panel 906 also includes color-coded information, such as visual representations of activity, inactivity, and sleep that may appear in the health report interfaces 900b to 900e.

[0139] For example, health report interfaces 900b to 900e represent a daily detail section that provides refined analysis results for each day of the wear period (e.g., day 1 to day 14). In one or more examples, health report interfaces 900b to 900e may be included together with health report interface 900a in report 328, such as in a single viewable document, PDF, etc. Therefore, health report interfaces 900b to 900e can be viewed in the user interface by means of methods such as "scrolling down". Additionally or alternatively, one or more of health report interfaces 900b to 900e represent separate pages accessible via one or more optional markers (e.g., page links, drop-down arrows, etc.).

[0140] Figures 9b to 9e The health report interfaces 900b to 900e shown each include a daily report section 908, a daily summary section 910, and a legend 912, which includes time-segment representations to explain the daily report section 908. The daily report section 908 displays cardiac measurements (e.g., heart rate expressed as beats per minute) plotted over time, with corresponding activity indicators below the heart rate data. The daily report section 908 in each interface demonstrates how heart rate data can be overlaid on sleep and activity data to provide healthcare providers with detailed analysis of a user's physiological responses during different behavioral states.

[0141] The Daily Summary section 910 includes summary sleep and activity metrics for each corresponding monitoring day. The Daily Summary section 910 may include a summary of daily sleep duration, activity time, and inactivity time. In the example shown, the Daily Summary section 910 includes the total sleep duration and total activity duration for the corresponding day. Health reporting interfaces 900a to 900e collectively provide comprehensive visualization tools for outputting various analytical results 116, enabling healthcare providers and users to analyze patterns and trends in patient behavior and physiological responses over longer monitoring periods, thereby supporting clinical decision-making and personalized health assessments.

[0142] Figure 10 A flowchart is shown illustrating an algorithm as step-by-step process 1000 in an example implementation, which can be executed by a processing device to generate analysis results related to the wear period based on low-sampling-rate accelerometer data from the wearable device.

[0143] In this example, firstly, low-sampling-rate accelerometer data is received from a wearable device attached to the user during the wearing period. This low-sampling-rate accelerometer data has a sampling rate below a sampling rate threshold (box 1002). For example, the accelerometer data is generated by monitoring device 104 during the wearing period. In various examples, monitoring device 104 uses accelerometer 308 to detect changes in the movement and orientation of person 102 and generates accelerometer data 310 based on the detected movement patterns. In at least one example, the sampling rate is approximately 1.56 Hz, and the wearing period can last from 1 to 14 days.

[0144] The low-sampling-rate accelerometer data is then processed to extract one or more motion-derived parameters that characterize the user's temporal motion patterns during the wearing period (box 1004). For example, motion-derived parameter 316 is calculated by prediction system 114 to quantify various aspects of the user's motion and positioning. For example, analysis module 304 calculates acceleration amplitude, activity parameters, reference vectors and / or position vectors for body angle calculations, rolling averages of accelerometer axis components for device orientation detection, etc.

[0145] Then, based on one or more motion-derived parameters, analysis results related to the conditions during the wearing period are generated (box 1006). For example, analysis result 116 is generated by prediction system 114 by analyzing the extracted motion-derived parameters. In various examples, analysis result 116 includes predictions of the user's sleep, activity, and inactivity states within the time interval of the wearing period. Analysis result 116 may also include body angle measurements and / or body posture classifications indicating whether the user is upright, tilted, or lying down at a particular time. Additionally or alternatively, analysis result 116 may include detection of inversion events of the wearable device during the wearing period. In some embodiments, prediction system 114 implements machine learning system 326 to process motion-derived parameters using a trained machine learning algorithm / model configured to analyze low-sampling-rate accelerometer data.

[0146] The analysis results are further configured for presentation (box 1008). For example, the prediction system 114 can format the analysis results 116 into a user interface display that can be viewed by a healthcare provider or user. In various examples, the prediction system 114 creates a comprehensive report that includes a summary section depicting aggregated user status data during the wearing period, and a daily detail section depicting the user's status in temporal relation to the physiological data collected during the wearing period. Report 328 can overlay heart rate data on sleep and activity information to provide a holistic view of the user's health patterns. The presentation may also include graphical representations of changes in body angle over time, changes in activity levels, and detected device orientation events that occurred during monitoring.

[0147] The prediction system 114 can also perform various functions based on the analysis results 116, such as adjusting device operating parameters during the wearing period, triggering additional sensor measurements based on the detected user status, generating alarms or notifications when specific conditions are identified, correlating the analysis results 116 with other physiological data streams, automatically calibrating signal processing algorithms based on the analysis results 116, providing real-time feedback to the user, storing the analysis results 116 for long-term analysis, sending the analysis results 116 to healthcare providers for remote monitoring, modifying data collection strategies based on detected user behavior, and integrating the analysis results with electronic health records for comprehensive patient assessment.

[0148] Figure 11A flowchart is shown illustrating an algorithm as step-by-step process 1100 in an example implementation, which can be executed by a processing device to generate analysis results related to user state based on low-sampling-rate accelerometer data. For example, process 1100 can be implemented as one or more sub-steps of one or more flowcharts described above or below. In various examples, process 1100 represents decision tree logic implemented by an algorithm that sequentially evaluates motion-derived parameters to determine whether the user is asleep, active, or inactive during the wearing period.

[0149] In this example, firstly, low-sampling-rate accelerometer data acquired by a wearable device attached to the user is processed (box 1102). For example, the low-sampling-rate accelerometer data may be acquired by a monitoring device at approximately 1.56 Hz, and this low-sampling-rate accelerometer data is processed to extract motion-derived parameters 316 characterizing the user's temporal motion patterns during the wearing period. Motion-derived parameters 316 may include: an acceleration amplitude 406, which is calculated as the square root of the sum of squares of one or more acceleration components of the low-sampling-rate accelerometer data; and / or an activity parameter 408, which is calculated as the standard deviation of the acceleration amplitude. For example, a relatively low standard deviation may correspond to the user being stationary, and a relatively high standard deviation may correspond to the user being in motion.

[0150] Then, based on the processed accelerometer data, it is determined whether sleep has been detected (box 1104). Sleep detection assessment may involve analyzing stationary patterns and body angle information derived from motion-derived parameters 316. In some cases, analysis module 304 compares activity parameters 408 to thresholds, where a low standard deviation of acceleration amplitude 406 within time interval 410 can indicate user stillness consistent with a sleep state. Sleep detection may also incorporate body posture information 120, where a tilted body angle close to 90 degrees relative to an upright reference vector can support sleep state classification.

[0151] If sleep is detected (e.g., "Yes" at box 1104), the user state is identified as a sleep state (box 1106). Sleep / activity states 118 can be classified as sleep states for specific time intervals of recording and incorporated into the analysis results 116 generated by the prediction system 114. In various examples, sleep state identification can trigger additional processing operations, such as one or more additional sensors and / or adjustment of data acquisition parameters during the detected sleep period.

[0152] If sleep is not detected (e.g., "No" at box 1104), the prediction system 114 continues to evaluate whether activity is detected (box 1108). Activity detection evaluation may involve comparing motion-derived parameters 316 to an activity threshold, such as to identify walking speeds of approximately 2 mph or higher. In some examples, the analysis module 304 examines the standard deviation of acceleration amplitude 406 over one or more time intervals 410, where a relatively high standard deviation indicates user movement consistent with an active state. Additionally or alternatively, activity detection may be based in part on body posture 120, such that an upright posture may indicate activity, while a supine posture may indicate inactivity.

[0153] If activity is detected (e.g., "Yes" at box 1108), the user state is identified as an active state (box 1110). This activity state classification can be incorporated into the sleep / activity state determination 118 and recorded as part of the analysis results 116 for the corresponding time interval. Activity state identification can also influence subsequent processing operations, such as calculating body posture 120 or assessing device orientation 122, which may be influenced by the user's movement patterns.

[0154] If no activity is detected (e.g., "No" at box 1108), the user state is identified as inactive (box 1112). For example, an inactive state represents a period of time when the user is awake but not engaging in movement that meets the activity threshold criteria. In various examples, the inactive state classification may correspond to sedentary behaviors such as sitting still or tilting, without a static pattern characteristic of a sleep state. Therefore, the technique described herein supports comprehensive user state classification based on a sequential evaluation of motion-derived parameters 316 extracted from low-sampling-rate accelerometer data throughout the entire wearing period.

[0155] Figure 12 A flowchart is shown illustrating an algorithm as step-by-step process 1200 in an example implementation, which can be executed by a processing device to determine body angles and body posture based on low-sampling-rate accelerometer data acquired during a prolonged period of wear. For example, process 1200 can be implemented as one or more sub-steps of one or more flowcharts described above or below.

[0156] In this example, firstly, the low-sampling-rate accelerometer data is processed to identify high-activity regions (box 1202) during the user's wearing period. For example, high-activity regions correspond to periods when the person 102 is performing motion patterns indicating an upright posture or active state. Identification of high-activity regions may involve analyzing the accelerometer data 310 for variations in acceleration amplitude, frequency domain characteristics, and / or statistical measurements within a defined time window.

[0157] Based on low-sampling-rate accelerometer data from high-activity areas, a reference vector representing the user's upright posture is generated (box 1204). For example, reference vector 508 serves as a baseline orientation marker, relative to which subsequent body angle calculations are performed. In various examples, posture module 320 calculates reference vector 508 by averaging or processing one or more acceleration components from accelerometer data 310 from identified high-activity periods to establish a consistent upright reference frame. Reference vector 508 can be calculated using one or more statistical methods, such as mean vector calculation, principal component analysis, or weighted averaging techniques.

[0158] Based on low-sampling-rate accelerometer data, a user's position vector is generated for a specific moment (box 1206). For example, position vector 510 represents the orientation of monitoring device 104, and thus the body posture of person 102 at a specific moment during the wearing period. In various examples, analysis module 304 processes accelerometer data 310 to extract three-dimensional acceleration components defining the position vector 510 for each time sampling point (e.g., per minute). In some examples, multiple position vectors 510 are calculated across consecutive time points to enable continuous monitoring of changes in body posture throughout the wearing period.

[0159] The user's body angle at that specific moment can then be determined as the polar angle between the reference vector and the position vector in a spherical coordinate system (box 1208). For example, body angle calculations quantify the angular deviation relative to an established upright reference to provide a numerical measurement of body tilt. The range of body angle determination can be from 0 degrees in a fully upright posture to 90 degrees in a basic supine posture, for example, to provide a continuous measurement of postural orientation.

[0160] Then, the user's body posture is classified based on body angles and one or more body angle thresholds (box 1210). For example, body posture classification converts body angle measurements into discrete posture categories. In various examples, posture module 320 applies predetermined thresholds to classify body angles into categories such as upright, sitting, tilted, and lying postures. The classification process can utilize multiple threshold boundaries to create different ranges for different posture states, thereby enabling the differentiation of subtle changes in body posture.

[0161] Once body angles and / or body postures are generated, they are output (box 1212). In various examples, presentation module 306 formats the body angle and body posture data to include them in a report 328 along with other analytical results 116, such as sleep / activity status 118 information. The output may include time series of body posture classifications, statistical summaries of posture patterns, or correlations between body postures and other monitored parameters.

[0162] Figure 13 A flowchart is shown illustrating an algorithm as step-by-step process 1300 in an example implementation. This algorithm can be executed by a processing device to use machine learning to integrate ECG and accelerometer data, thereby enhancing sleep and activity prediction. For example, process 1300 can be implemented as one or more sub-steps of one or more flowcharts described above or below. Process 1300 demonstrates how a single trained machine learning algorithm can process multimodal data to generate a comprehensive classification.

[0163] In this example, firstly, low-sampling-rate accelerometer data and ECG data are received from a wearable device attached to the user's skin surface during the wearing period (box 1302). In various examples, monitoring device 104 uses accelerometer 308 to detect motion patterns and uses ECG sensor 312 to detect electrical activity of the heart to generate accelerometer data 310 and ECG data 314.

[0164] The low-sampling-rate accelerometer data is processed using a first trained machine learning algorithm to generate accelerometer-based predictions (box 1304). For example, the first trained machine learning algorithm is a trained ASP model 608, which is configured to analyze motion-derived parameters 316 extracted from the accelerometer data 310 to generate predictions, such as classifying user states.

[0165] The ECG data is processed using a second trained machine learning algorithm to generate ECG-based predictions (box 1306). For example, the second trained machine learning algorithm is a trained ESP model 614 configured to analyze cardiac parameters to generate predictions, such as classifying user states. The second trained machine learning algorithm can be trained using a combination of ECG training data 620 annotated with polysomnography and ECG training data 622 annotated with ASPM, for example, to ensure that both sleep and wake periods are represented in the training set.

[0166] A combination scheme is implemented to integrate accelerometer-based and ECG-based predictions (box 1308). For example, combination scheme 630 applies heuristics to determine which prediction source to use for each time interval of the wearing period. In some cases, combination scheme 630 evaluates factors such as the incidence of arrhythmias, signal quality, and consistency between the two prediction sources to make the integration determination. Alternatively or additionally, combination scheme 630 may combine the predictions generated by the first and second machine learning algorithms, for example, using weighted averaging or other integration techniques on a minute-by-minute basis.

[0167] A comprehensive prediction result for the user during the wearing period is generated based on the integrated prediction results from the combined scheme (box 1310). For example, comprehensive prediction result 632 represents a unified assessment that leverages the advantages of both accelerometer and ECG data sources, which can provide higher accuracy compared to a single-modal approach. In various examples, comprehensive prediction result 632 includes indications of sleep periods, activity periods, and inactivity periods throughout the wearing period.

[0168] Then, a comprehensive sleep and activity classification is output (box 1312). For example, analysis module 304 displays the comprehensive prediction result 632 as part of report 328 via presentation module 306. In some implementations, the comprehensive prediction result 632 is incorporated into report 328 to overlay heart rate data with sleep and activity information. In various examples, the output includes summary statistics for the entire wearing period and / or daily details describing the temporal correlation between physiological data and behavioral states.

[0169] Figure 14 A flowchart is shown illustrating an algorithm as step-by-step process 1400 in an example implementation, which can be executed by a processing device to detect device inversion events during a long monitoring period. For example, process 1400 can be implemented as one or more sub-steps of one or more flowcharts in the foregoing flowchart.

[0170] In this example, the accelerometer data is first processed to generate a rolling average of the axis components (box 1402). In various examples, process 1400 calculates a first rolling average of the x-component of the accelerometer data 310 over a predetermined time interval (e.g., a 24-hour period) and a second rolling average of the y-component of the accelerometer data 310 over the same predetermined time interval. For example, the rolling average smooths out variations associated with typical user activity patterns while preserving long-term trends indicating changes in device orientation.

[0171] Then, the range of the calculated rolling averages is determined (box 1404). For example, prediction system 114 calculates a first range of the first rolling average and a second range of the second rolling average. In some cases, range calculation involves determining the difference between the maximum and minimum values ​​of each rolling average component during the wearing period. These range values ​​provide a quantitative measurement of the variation in the accelerometer axial components, which can indicate whether the device has experienced a significant orientation change during monitoring.

[0172] Then, process 1400 evaluates whether the calculated range exceeds a predetermined threshold (box 1406). For example, a threshold may be set to distinguish between normal positional changes and inversion events. In various implementations, the threshold comparison may involve evaluating whether either a first range or a second range exceeds the threshold. In at least one example, the threshold is approximately 0.8.

[0173] When the range does not exceed the threshold (e.g., "No" at box 1406), process 1400 determines that no device inversion event was detected (box 1408). In these cases, standard analysis algorithms can be used to process the accelerometer data 310 without orientation correction.

[0174] Conversely, when the range exceeds a threshold (e.g., "Yes" at box 1406), process 1400 determines that a device inversion event has occurred (box 1410). Detection of an inversion event can trigger additional analysis to characterize the timing and nature of the orientation change. In some embodiments, the number of inversion events can be calculated, and the specific moment of orientation change can be identified by analyzing the crossover of the rolling average relative to a calculated inversion threshold.

[0175] Then, an indication of the device inversion event is output (box 1412). For example, the prediction system 114 generates an indication substantially in real time, such as to alert personnel 102 that device 104 has been inverted. Additionally or alternatively, the prediction system 114 configures the device inversion event to be output in a user interface, such as in report 328.

[0176] Furthermore, signal processing correction is performed based on the detected device inversion event (box 1414). For example, correction may include automatically adjusting how accelerometer data 310 and / or ECG data 314 are processed. As an example, the processing of ECG signal polarity is modified in response to the detection of an inversion event, as changes in device orientation can affect the characteristics of the electrical signals captured by ECG sensor 312. Despite the possibility of orientation changes, signal processing correction helps maintain the accuracy of the derived analysis results 116 throughout the entire wearing period.

[0177] Machine learning and AI in physiological monitoring using low-sampling-rate accelerometer data

[0178] The foregoing examples, such as for machine learning system 326, trained ASP model 608, and / or trained ESP model 614, describe various instances of artificial intelligence (“AI”) models and / or machine learning models. In one or more examples, an AI model (e.g., a machine learning model) refers to a computer representation that is capable of being tuned (e.g., through training and retraining) based on inputs without being actively programmed by a user to automatically and without user intervention to approximate an unknown function. For example, the term “machine learning model” includes models that utilize algorithms to learn and relearn by analyzing training data to learn from and make predictions based on known data, thereby generating outputs that reflect the patterns and properties of the training data.

[0179] In the context of physiological monitoring using low-sampling-rate accelerometer data, machine learning models can be implemented (e.g., by one or more processing devices of the prediction system 114) to analyze motion patterns and physiological data, thereby generating analytical results 116 related to the user's state and device status during prolonged wear periods. For example, the state module 318, posture module 320, orientation module 322, and / or multi-parameter module 324 can each utilize one or more machine learning models to process the low-sampling-rate accelerometer data 310 and ECG data 314 acquired by the monitoring device 104. Examples of machine learning models suitable for low-sampling-rate accelerometer analysis include neural networks, convolutional neural networks (CNNs), long short-term memory (LSTM) neural networks, generative adversarial networks (GANs), decision trees (e.g., for sleep / activity state classification), support vector machines, linear regression, logistic regression, Bayesian networks, random forest learning, dimensionality reduction algorithms, boosting algorithms, deep learning neural networks, etc.

[0180] For example, a machine learning model can be configured using multiple layers, each with multiple nodes. These multiple layers can be configured to include an input layer, an output layer, and one or more hidden layers. In the context of low-sampling-rate accelerometer monitoring, the input layer can receive various motion-derived parameters 316 generated based on accelerometer data 310, such as acceleration amplitude 406, activity parameters 408, reference vector 508, position vector 510, rolling averages of accelerometer axis components, etc. For example, the hidden layer processes these inputs through weighted connections to identify complex patterns indicating user states such as sleep / activity state 118, body posture 120, and device orientation 122—patterns that cannot be detected using conventional high-frequency accelerometer modal analysis. The output layer can generate classifications indicating sleep, activity, and inactivity states during processing by the state module 318, and / or provide body angle calculations and body posture classifications. Computations are performed by nodes within the layers via hidden states and through a weighted connection system “learned” during the training of the machine learning model to achieve various physiological assessment tasks under power-constrained conditions.

[0181] To train a machine learning model for low-sampling-rate accelerometer monitoring, training data is received that provides examples of "what the machine learning model should learn," i.e., the training data serves as the basis for learning patterns from data. For low-sampling-rate accelerometer applications, the training data may include labeled datasets such as ASPM training data 604 and ESPM training data 610. ASPM training data 604 includes training accelerometer data 616 with corresponding ground truth labels 618 from users whose sleep and activity states are known. ESPM training data 610 includes PSG-labeled ECG training data 620 and ASPM-labeled ECG training data 622. For example, a machine learning system 326 including a machine learning model acquires and preprocesses the training data, which includes input features (e.g., motion-derived parameters 316 extracted from 1.56 Hz accelerometer measurements, body angle calculations, device orientation indicators) and corresponding target labels (e.g., "sleep state," "active state," "inactive state," or a specific body posture classification, such as upright, tilted, or supine orientation).

[0182] The machine learning system 326 is also capable of initializing various parameters of the machine learning model, which can be used by the machine learning model as internal variables to represent and process information during training. These parameters can also be used to represent inferences obtained through training on low-sampling-rate data patterns that are significantly different from conventional high-frequency accelerometer analysis. In one or more embodiments, training data is divided into multiple batches to improve the efficiency of processing and optimizing the parameters of the machine learning model during training, which is particularly beneficial for model accuracy when processing long-term series data collected over a wear period of 1 to 14 days.

[0183] Then, the machine learning model receives training data as input and uses this training data as the basis for generating prediction results based on the current state of the model's layers and corresponding node parameters. The results are output as output data (e.g., accelerometer-based prediction result 626, ECG-based prediction result 628, comprehensive prediction result 632, etc.). For example, the prediction system 114 includes machine learning models such as a trained ASP model 608 and a trained ESP model 614, which are trained to identify patterns in low-sampling-rate accelerometer data 310 and ECG data 314 associated with a specific user state, thereby enabling the prediction system 114 to generate accurate analysis results 116.

[0184] Training a machine learning model can include calculating a loss function to quantify the loss associated with the operations performed by the nodes of the machine learning model. Loss functions can be configured in various ways to control the operation and / or functionality of the machine learning model. For example, a loss function can be designed to prioritize ensuring the accuracy of sleep state detection while minimizing false positives that might lead to misclassification of activity during periods of minimal movement. For example, calculating the loss function involves comparing the difference between a specified prediction in the output data (e.g., a predicted sleep / activity state or body posture classification) and a target label specified by the training data (e.g., a clinically validated sleep state or validated body posture). Loss functions can be configured in various ways, examples of which include regret loss, quadratic loss functions as part of a least-squares technique, cross-entropy loss, custom loss functions incorporating temporal consistency requirements for analysis over long periods of wear, and so on.

[0185] The configuration of training data can be used to support a variety of use cases in low-sampling-rate accelerometer monitoring. For example, machine learning models can be trained to detect specific patterns in motion-derived parameters 316 that indicate various sleep stages, identify motion patterns associated with different activity levels equivalent to walking speeds of 2 mph or higher, identify device orientation changes that indicate inversion events during the wearing period, or detect subtle changes in body posture that may be associated with sleep quality or health status. Models such as the trained ASP model 608 can be configured to operate within the computational limitations of the wearable device while providing accurate state classification throughout the extended wearing period. These models can also be integrated with ECG-based models (e.g., the trained ESP model 614) through a combination scheme 630 to provide multimodal analysis results that leverage the advantages of both accelerometer and cardiac monitoring. This adaptive approach enables efficient use of computational resources dedicated to the machine learning process while ensuring comprehensive physiological analysis using low-sampling-rate data acquisition techniques that support extended monitoring without requiring device recharging.

[0186] It should be understood that many variations are possible based on the disclosed content of this document. Although the features and elements have been described above in specific combinations, each feature or element may be used alone without other features and elements, or in various combinations with or without other features and elements.

[0187] Clause 1. A method comprising: receiving low-sampling-rate accelerometer data acquired by a wearable device attached to the skin surface of a user's chest region during a wear period, the low-sampling-rate accelerometer data having a sampling rate below a sampling rate threshold; processing the low-sampling-rate accelerometer data to extract one or more motion-derived parameters, the one or more motion-derived parameters characterizing a user's temporal motion pattern during the wear period; and generating, based on the one or more motion-derived parameters, analytical results relating to the user's state during the wear period for presentation.

[0188] Clause 2. The method according to Clause 1, wherein the low sampling rate accelerometer data is acquired at a sampling rate of approximately 1.56 Hz and the wearing period is 1 to 14 days.

[0189] Clause 3. The method described in Clause 1 or Clause 2, wherein the analysis results include predictions of the user's sleep, activity, and inactivity states during one or more time intervals of the wearing period.

[0190] Clause 4. The method according to any of the preceding clauses, wherein one or more motion-derived parameters include: an acceleration amplitude for a specific time interval of the wearing period, the acceleration amplitude being calculated as the square root of the sum of squares of one or more acceleration components of low-sampling-rate accelerometer data; and an activity parameter being calculated as the standard deviation of the acceleration amplitude for the specific time interval, wherein a relatively low standard deviation corresponds to the user being stationary and a relatively high standard deviation corresponds to the user being in motion.

[0191] Clause 5. The method described in any of the preceding clauses, wherein the analysis results include the user's body angles or body postures during one or more time intervals of the wearing period.

[0192] Clause 6. The method according to any of the preceding clauses, wherein one or more motion-derived parameters include: a reference vector corresponding to the user's upright posture and generated based on a portion of low-sampling-rate accelerometer data indicating relatively high activity; and a position vector of the user for a specific moment during the wearing period.

[0193] Clause 7. The method according to any of the preceding clauses, wherein the analysis results include the detection of inversion events of the wearable device during the wearing period, and the analysis results are generated based on motion-derived parameters, which include the rolling average of the accelerometer axis components of low-sampling-rate accelerometer data over a specific time interval of the wearing period.

[0194] Clause 8. The method according to any of the preceding clauses further includes configuring the analysis results to be presented in a user interface as part of a report, the report including: a summary section depicting aggregated user status data during the wearing period; and a daily detail section depicting the user status in temporal relation to the physiological data collected during the wearing period.

[0195] Clause 9. The method described under any of the preceding clauses further includes receiving electrocardiogram (“ECG”) data acquired by a wearable device, and wherein the generation of analysis results is also based on the ECG data.

[0196] Clause 10. A processing device comprising: one or more processors; and a memory storing computer-readable instructions executable by the one or more processors to perform operations including: receiving accelerometer data with a sampling rate below a sampling rate threshold, the accelerometer data being acquired by an accelerometer of a wearable device attached to the surface of a user's skin during a wear period; processing the accelerometer data to extract one or more motion-derived parameters characterizing a temporal motion pattern of the user during the wear period; and generating, based on the one or more motion-derived parameters, analysis results related to the conditions during the wear period for presentation.

[0197] Clause 11. The processing device as described in Clause 10, wherein the accelerometer data is acquired at a sampling rate of approximately 1.56 Hz and the wearing period is from 1 to 14 days.

[0198] Clause 12. The processing device as described in Clause 10 or Clause 11, wherein the analysis results include the user's sleep state, activity state, and inactivity state throughout the wearing period.

[0199] Clause 13. The processing device according to any one of Clauses 10 to 12, wherein the analysis results include the user's body angles or body postures throughout the wearing period.

[0200] Clause 14. The processing device according to any one of Clauses 10 to 13, wherein the analysis results include the detection of whether an inversion event occurs during the wearing period of the wearable device.

[0201] Clause 15. The processing apparatus according to any one of Clauses 10 to 14, wherein operation further includes receiving ECG data acquired by an electrocardiogram (“ECG”) sensor of a wearable device and generating analysis results in part based on the ECG data.

Claims

1. A method comprising: Receive low-sampling-rate accelerometer data collected by a wearable device attached to the skin surface of the user's chest area during the wearing period, wherein the sampling rate of the low-sampling-rate accelerometer data is lower than a sampling rate threshold; The low sampling rate accelerometer data is processed to extract one or more motion-derived parameters, which characterize the user's temporal motion patterns during the wearing period. as well as Based on the one or more motion-derived parameters, analysis results related to the user's state during the wearing period are generated for presentation.

2. The method according to claim 1, wherein, The low-sampling-rate accelerometer data was acquired at a sampling rate of approximately 1.56 Hz, and the wearing period was between 1 and 14 days.

3. The method according to claim 1 or 2, wherein, The analysis results include predictions of the user's sleep, activity, and inactivity states during one or more time intervals of the wearing period.

4. The method according to any one of the preceding claims, wherein, The one or more motion-derived parameters include: an acceleration amplitude for a specific time interval of the wearing period, the acceleration amplitude being calculated as the square root of the sum of squares of one or more acceleration components of the low sampling rate accelerometer data; and an activity parameter, the activity parameter being calculated as the standard deviation of the acceleration amplitude for the specific time interval, wherein a relatively low standard deviation corresponds to the user being stationary, and a relatively high standard deviation corresponds to the user being in motion.

5. The method according to any one of the preceding claims, wherein, The analysis results include the user's body angles or body postures during one or more time intervals of the wearing period.

6. The method according to any one of the preceding claims, wherein, The one or more motion-derived parameters include: a reference vector corresponding to the user's upright posture and generated based on a portion of the low-sampling-rate accelerometer data indicating relatively high activity; and the user's position vector for a specific moment during the wearing period.

7. The method according to any one of the preceding claims, wherein, The analysis results include the detection of inversion events of the wearable device during the wearing period, and the analysis results are generated based on motion-derived parameters, which include the rolling average of the accelerometer axis components of the low sampling rate accelerometer data within a specific time interval of the wearing period.

8. The method according to any one of the preceding claims, further comprising configuring the analysis results to be presented as part of a report in a user interface, the report comprising: The summary section describes the aggregated user status data for the wearing period; as well as The daily details section depicts the user's state in a temporal relation to physiological data collected during the wearing period.

9. The method according to any one of the preceding claims further includes receiving electrocardiogram (ECG) data acquired by the wearable device, wherein, The analysis results are also generated based on the ECG data.

10. A processing apparatus, comprising: One or more processors; as well as A memory having stored computer-readable instructions that can be executed by the one or more processors, the operations including: Receive accelerometer data with a sampling rate lower than a sampling rate threshold, the accelerometer data being collected by the accelerometer of a wearable device attached to the user's skin surface during the wearing period; The accelerometer data is processed to extract one or more motion-derived parameters, which characterize the user's temporal motion patterns during the wearing period; and Based on the one or more motion-derived parameters, analysis results related to the condition during the wearing period are generated for presentation.

11. The processing apparatus according to claim 10, wherein, The accelerometer data was collected at a sampling rate of approximately 1.56 Hz, and the wearing period was between 1 and 14 days.

12. The processing apparatus according to claim 10 or 11, wherein, The analysis results include the user's sleep state, activity state, and inactivity state throughout the entire wearing period.

13. The processing apparatus according to any one of claims 10 to 12, wherein, The analysis results include the user's body angles or body postures throughout the entire wearing period.

14. The processing apparatus according to any one of claims 10 to 13, wherein, The analysis results include the detection of whether an inversion event of the wearable device occurred during the wearing period.

15. The processing apparatus according to any one of claims 10 to 14, wherein, The operation also includes receiving ECG data acquired by the wearable device's ECG sensor and generating the analysis results in part based on the ECG data.

Images