Patent Application
20230148973
The present disclosure relates to improvements in systems and methods for physically identifying biological conditions where overtraining and injury will cause an alteration in fitness training activities, and the architecture and design of a system configured to reduce memory storage requirements for certain artificial intelligence-based applications on a mobile device.
Some athletes choose a goal competition (e.g., New York City Marathon), and attempt to implement a training program with little consideration as to a likelihood of injury. The athlete becomes extremely focused, and often misses cues, or simply is unaware or unable to determine signs of overtraining and the approach of an injury condition. It would be helpful for the athlete to be able to receive an overtraining indication from a device they carry with them on a training activity, such as a fitness tracker or similar device.
Artificial intelligence (“AI”) and machine learning algorithms are becoming more cemented in today's society to solve a number of problems. Conventional AI techniques utilise large data sets in order to identify patterns and draw conclusions. Conventional AI techniques are simply ill-suited for use in mobile devices of the type likely to be carried by an athlete for a variety of reasons. First, the typical memory requirement needed to accommodate large data sets ordinarily used by conventional AI would be quite large. Secondly, the processing capability needed to generate conclusions obtainable from conventional AI techniques is likewise large. Thirdly, when an athlete trains, typically they venture into areas where there is not a Wi-Fi network, and hence no internet. Where the athlete's device has provision for cellular coverage, there are still areas where cellular coverage is non-existent, or compromised. Even where cellular coverage is available, the cost of large data transmissions would make use of such cellular coverage for typical AI methodology cost-prohibitive and not pragmatic. What is needed is a system and method that encapsulates the ability to provide an analysis and warning or indication of an elevated overtraining condition so that modifications to a training program can be made before it is too late, and an overtraining or injury situation arises.
The present disclosure in one aspect provides for a less complicated system and method of providing an athlete/user with identification and advanced warning of an overtraining condition. In instances utilizing artificial intelligence (“AI”), the system and methodology provides for a configuration which reduces memory requirements, communications, and processing requirements to permit “micro-AI” to become a reality. In one aspect, the system and method utilize bifurcated data organized into “latent data” and “current data”. Latent data is further preferably organized into general historical athlete data (which may be categorical across a range of non-particular individuals), and historical data particular or specific to the individual athlete. The scope of data useable by micro-AI is reduced and organized for efficiency compared to traditional AI systems, which permits mobile applications in a realistic manner.
The present disclosure in one preferred aspect provides for a fitness tracker wearable around a portion of a user. The fitness tracking includes a heart rate sensor and a memory. The memory includes a latent memory component configured to retain latent data including non-individual data not particularly pertaining to the user, and includes personal data particular to the user. The latent memory is updatable only when the tracker is within an internet coverage area. The memory also includes a current memory component configured to retain current data that is updatable as the user is exercising. The tracker also includes a microprocessor including a classifier. The microprocessor is configured to determine an existence of an overtraining condition based on a prediction model utilizing data only in the memory irrespective of presence in an internet coverage area. The microprocessor is configured to provide an alert to the user after determining that the overtraining condition exists according to an output of the classifier.
The present disclosure in another aspect provides for a system for biologically monitoring the fitness of an athlete and providing a warning to the athlete when biological conditions indicate the athlete is overtraining. The system includes a remote processor accessible over the internet; a remote database accessible over the internet; and a wearable fitness monitoring device having a heart rate sensor, an onboard processor and a memory. The memory includes a latent memory updatable only when the device is in internet communication with the remote processor, and a current memory updatable with current training data measured by the monitoring device as the athlete is training. The latent memory is configured to store non-individual data not specific to the wearer of the fitness device. The non-individual data includes general historical athlete data. The latent memory is also configured to store personal data particular or specific to the wearer of the fitness device. The onboard processor includes at least one classifier. The onboard processor is configured to analyze the current data and compare the current data with the latent data. The onboard processor ascertains a likelihood of the athlete overtraining and becoming injured based on an output of the classifier. The onboard processor provides an alert to the athlete upon determining that the athlete is overtraining according to an output of the classifier. The system may further include a temperature sensor. The system may further include a moisture sensor configured to measure a sweat index.
The present disclosure in another aspect sets forth a method of generating a set of latent features relating to historical training patterns and outcomes; periodically refreshing a wearable fitness device worn by the athlete with the latent features; recording current training data of the athlete as the athlete is training to generate a set of current features, the current features including a heart rate; performing a statistical application with the latent features and the current features to generate a feature vector indicative of an overtraining condition; feeding the feature vector into a neural network or any other classifier selected residing on the wearable fitness device; and obtaining a result from the neural network as to whether an overtraining condition is present.
The following terms are defined to provide clarity when reading the present disclosure. References herein to “training” include exercise in general, and that a person partaking in training activities need not be a professional athlete, but anyone with an interest in exercising.
“Latent data” is data that is not constantly updated, but data that is updated occasionally (if at all), at convenient opportunities, for example, when in a Wi-Fi zone. Latent data may include general historical data and specific historical data, both described below.
“General historical data” is historical data that is not focused on only a single individual, but pertains to a grouping or segment of the human population.
“Specific historical data” is historical data that is personal to a particular person or individual, and pertains to the person's personal characteristics and/or experience.
“Current data” is data that is constantly or dynamically being updated when the data-gathering device is in an active state.
“Training intensity” is a measure of the exercise effort, or perceived exercise effort or exertion of an individual during exercise.
As used herein, “configured” includes creating, changing, or modifying a program on a microprocessor, computer or network of computers so that the processor, computer or network of computers behave according to a set of instructions. The programming to accomplish the various embodiments described herein will be apparent to a person of ordinary skill in the art after reviewing the present specification, and for simplicity, is not detailed herein. The programming may be stored on a computer readable medium, such as, but not limited to, a non-transitory computer readable storage medium. The system may be implemented on a field-programable gate array and graphics processing unit.
It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention, as claimed, unless otherwise stated. In the present specification and claims, the word “comprising” and its derivatives including “comprises” and “comprise” include each of the stated integers, but does not exclude the inclusion of one or more further integers. The claims as filed with this application are hereby incorporated by reference in the description. The entire disclosures of U.S. Pat. Nos. 10,013,638; 10,124,234; band 10,322,314 are hereby incorporated by reference herein.
The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate several embodiments of the invention and together with the description, serve to explain the principles of one or more forms of the invention.
Reference will now be made in detail to exemplary embodiments of the invention, some of which are illustrated in the accompanying drawings.
As described already, latent data 104 may include general historical athlete data, and specific historical data particular or personal to the individual athlete. General historical data is that which pertains, for example, to demographical data in a general human population, and may be categorized and sub-categorized. For example, if an athlete is a 40 year-old male runner, general historical data would focus on a subset of the human population directed to male runners in the 40-44 year-old age group. Other factors could take into account geographical location, exercise type (e.g., running, and whether marathoner or sprinter), demographical data (age, gender), experience in a particular activity (e.g., new to the activity, or years of experience in the activity), terrain (e.g., hilly, flat), and climate (e.g., hot, cold, humid).
Specific historical data is that which pertains to the athlete's own personal characteristics and experience. For example, such data items may include an injury prone factor (a person's susceptibility to being injured, or how durable they are from heavy exercise), family history/genetics, specific experience in the activity, and a training goal (fitness level or upcoming competition). For training organization, the athlete's goal, and time (e.g., time between exercise sessions, sleep time, and time remaining to a goal competition, such as how many days or weeks remaining to the competition date) may be accommodated.
Other indicators and factors pointing to an overtraining condition include measuring and recording a fatigue factor (a subjective factor as to how tired someone feels), muscle soreness, a sweat index (how much the athlete is sweating, quantity of excreted water through the skin, preferably taking into account humidity level), and VO2 max data (maximal oxygen uptake, which relates to the maximum amount of oxygen that a person can utilize during intense exercise). The VO2 max data is usually obtained by testing with an athlete on a treadmill in combination with a breath mask. As may be seen, features can be all objective, subjective, or a combination of objectively and subjectively derived features.
Current data 106 is data which is updated and stored on the athlete's fitness device, such as fitness device 124 (
Referring to
The timeframe for automatic deletion of current data may be based on the volume of current data. For example, 14 days of training data may be stored, regardless of the actual date of when the training occurred. That way, if it has been a week since the device was last used, there will still be two weeks' worth of current data on the device.
The system may be configured so that the current data is merged or integrated with the latent data residing on the remote server when internet is available to update the latent data. For example, when the latent data is refreshed in the memory of the device once device 124 is in an internet coverage area, portions of the current data may be added to the latent data, for example, the athlete's specific historical data is updated with recent activities from the current data. The refreshing or updating of the latent data (e.g., of the athlete's specific historical data with elements of the current data) may be based on the expired current data (data older than the predetermined time frame), and/or based on all of the athlete's training data (specific historical and current as of the most recent training activity). If desired, only the expired current data may be used to update the latent data so that a merger of expired current data and latent data is completed at the remote server, using archived current data where such current data was archived. In the event of utilization of archived data, a decompressor may be used prior to the merger/integration of data.
In a preferred embodiment, latent data is not constantly updated or refreshed with the current data, but maintained separately as a different component of the device memory. Therefore, for example, during training sessions, the amount of latent data remains static and substantially unaltered.
Heart rate data may be of two general types, active heart rate (the heart rate during training), and at-rest heart rate (e.g., a heart rate measured when an athlete first awakes at the start of the day). Active heart rate can be weighted or analysed against historical heart rate data (general category and/or specific individual).
An at rest heart rate, or a “waking heart rate” (WHR) is a very good indicator of fitness and overtraining. Usually taken when someone wakes up, a pulse of upper 30s beats per minute usually means that person is fit. Even a pulse in the 40s or 50s is a sign of good fitness. A waking heart rate may be measured by a fitness device having a heart rate sensor if the fitness device is worn to bed by the athlete, or may be manually derived from the athlete taking their pulse in a prone position.
A real time “active heart rate” monitoring (typically with a fitness tracking device), during exercise can also indicate fitness and overtraining, but is usually tied to physical exertion and effort. Usually, a heart rate of 60 to 90 beats per minute means someone is at a low activity (e.g., walking or sitting). A range of between 100-140 is usually an indicator of moderate activity. A range of around 140 to 160 is an indicator of harder activity. A heart rate of between 170 to 210 typically indicates peak exertion. A maximum heart rate depends on age. Younger people (in their 20s and 30s) should be able to have a max heart rate of around 210. The max heart rate has been shown to decrease with age. For someone in their 40s, the typical max heart rate might be 190.
Preferably, a set of 17 features relating to overtraining conditions, including 9 latent features (e.g., age, heart rate, age difference from the average age within a particular sport, exercise duration, exercise amount, past injuries, past intensity measurements, variables derived from statistical analysis based on historical data, parameters from statistical tests or regressions) and 8 current features (e.g., heart rate (active and at rest), exercise duration, exercise type, time between exercise sessions, exercise amount, intensity level, sleep time) are utilized for classification. Intensity level may be measured by a combination of heart rate, or heart rate and pace (speed) when the exercise involves travel over a distance, or weight amount for weight training.
It will be appreciated that varying combinations of features may be used without departing from the scope of this disclosure. For example, more or less than 9 latent features may be utilized. More or less than 8 current features may be utilized. The feature set may include any feature that contributes to identification of an overtraining condition. A few more important features in relation to detecting overtraining conditions include heart rate, training duration, and intensity (depending upon the form of exercise), in combination with historical and demographical data.
Referring to
A neural network (NN) typically involves an artificial neuron, where a set of inputs is applied, each representing an output of another neuron. Each input is multiplied by a corresponding weight which is analogous to synaptic strengths in biological neurons. The weighted inputs are summed to determine the net input of the neuron. This net input is processed further by using a squashing function (activation function), to produce the neuron's output signal. This function may be linear, nonlinear (e.g., step function) or sigmoid (S-shaped).
Examples of weighted inputs include heart rate being a given a higher weighting, while exercise duration, short term sleep (e.g., in the last 24 hours), and waking heart rate being given a lower weighting. For example, active HR may be assigned a weighting of 1.5, while WHR may be assigned a weighting of 1.0 or 1.1. Other features could be assigned a lower weighting, like the user's age being assigned a weight of 0.5.
Referring again to
A feature vector is preferably created for every condition associated with overtraining, and with the current features 106 plus the outputs from one or more classifier models 108 (e.g., the LR, DA and MARS models). In this step a file with vectors including, for example, 20 features for all the cases is created.
The features are preferably normalized before the feature file can be efficiently used by NN 112. The normalized combined feature-vectors are fed into the neural network 112 for further classification.
It will be appreciated that ways other than statistical applications may be used to provide further features, or second order features. For example, data mining techniques or other classifiers such as neural networks, decision trees, associated rule mining, and case-based reasoning, or a combination of any of the foregoing, may be used to provide second order features which may be eventually fed into a neural network.
Other applications may be used in place of, or in addition to the NN as a final classifier. For example, the NN may be used in combination with SVM, or substituted by SVM for final classification if desired.
If desired, a micro-AI analysis can be conducted in a real-time exercise situation, i.e., as the athlete is training. A micro-AI analysis during training consumes more device power and processing capabilities, but can be conducted if conditions are warranted (e.g., if a condition exists to trigger local during-training AI analysis). Several levels of triggers may be utilized. For example, a primary trigger may be active heart rate level. Heart rate is one of the most important metrics in ascertaining the existence of overtraining conditions. Second degree triggers may include one or more of training duration, short term sleep (e.g., sleep over the last 24 hours), training intensity, and wake up hear rate (WHR). Real-time micro-AI analysis may be initiated solely by a primary trigger (e.g., a heart rate going beyond a predetermined threshold), or a combination of secondary triggers (e.g., a combination of current training duration being greater than a threshold level, sleep level being less than a predetermined minimum, and a WHR above a predetermined threshold), or a combination of primary and secondary triggers (e.g., active HR in combination with a training intensity). When using active heart rate, the micro-AI analysis compares the current data of the active heart rate and training intensity level (in a running situation, running pace) with latent data pertaining to specific historical data of the individual (active heart rate and intensity level) to determine if the active heart rate is beyond a predetermined threshold, indicating an overtraining condition is present that could lead to an injury condition.
Once real-time micro-AI is triggered, the micro-AI analysis is conducted while the athlete is training. If the result is that an overtraining condition exists while the athlete is training, a warning is provided to the athlete while they are training. Such warning can be provided through fitness device 124 where it is configured to provide one or more of an audible signal, a vibratory indication, and/or visual warning (e.g., text message to a screen). The warning may be provided through the use of a paired smartphone should the smartphone be within range.
Referring again to
The latent memory is preferably updatable only when the tracker is within an internet coverage area. One result of this is that the amount and type of data in the latent memory is finite compared to cloud servers, reducing the quantity of data that is otherwise analyzed as part of any AI analysis when confining an AI analysis to only data in the device onboard memory.
The memory of device 124 further includes a current memory component configured to retain current data that is updatable as the user is training. Device 124 also includes a processor, such as a microprocessor, that includes a classifier. The processor is configured to determine an existence of an overtraining condition, and other determinations as needed, based on a prediction model utilizing data only in the device onboard memory irrespective of presence in an internet coverage area. The processor is configured to provide an alert to the user after determining that the overtraining condition exists according to an output of the classifier. In a preferred embodiment, the processor is configured to compare the current data with the latent data and determine an existence of an overtraining condition.
Having described the preferred components of system 100, a preferred method 200 for classifying an overtraining condition of an athlete is shown in
It will be appreciated that the steps described above may be performed in a different order, varied, or some steps added or omitted entirely without departing from the scope of the disclosure. For example, the method may include monitoring an athlete's heart rate during training and comparing the active heart rate with a historical intensity level, depending on the form of exercise. If the heart rate is above a threshold value, preferably taking into consideration one or more of intensity, speed, and duration, then an AI analysis is triggered to determine whether the elevated heart rate may represent an overtraining condition and indicate the approach of an injury condition while the athlete is training. If the analysis determines an overtraining condition is present, then the method may include sending an alert to the athlete, preferably using the wearable fitness device. The alert may be delivered as an auditory message, a vibratory indication, and/or a visual indication. The alert may specify that the athlete decrease training intensity, slow down, walk, or stop to reduce injury risk. If the analysis determines that an overtraining condition is not present, then the athlete's heart rate is continued to be monitored.
Remote platform 304 includes a raw population data database 306 configured to store general population data, which may include the general historical athlete data mentioned previously. Remote platform 304 also includes an individual historical database 308 configured to store or record an athlete's own personal characteristics and experiences, such as, for example only, an injury prone factor (a person's susceptibility to being injured, or how durable they are from heavy exercise), family history/genetics, specific experience in the activity, and a training goal (fitness level or upcoming competition).
Raw population data from database 306, and individual historical data from database 308 are used to generate a latent dataset 330, a heart rate (HR) model 310, a latent model 316 and a classification model 320 that may be exported/transferred to local storage within the micro-AI module to be carried by the athlete, and be used by a local microprocessor (described further below). The heart rate related data included in raw population database 306 and in individual historical database 308 is used to build a HR model.
Data from raw population database 306 is used in analysis 312 to extract one or more features. A descriptive feature set 314 is then created based on the features extracted from the population data, with application of descriptive statistical analysis. For example, a range of wake up heart rate and heart rate range during exercising. The descriptive feature set 314 is used to form part of the local latent data 330 (described further below). The features extracted at 312 may be used in latent model 316 to generate a set of outputs 326 (described further below).
The outputs 326 and the features extracted from 312 are utilised by classification algorithms (such as NN, SVM, LR, BN and hybrid algorithms) 318. Individual historical data from database 308 and population historical data (optional) may have a time series analysis applied at 322, the output of which may be used in combination with the algorithms at 318 to build the final classification model 320. The time series analysis may, for example, be a heart rate change pattern by time and/or sweat intensity pattern change by time.
Remote HR model 310, latent model 316, final classification model 320 and up to two or more data outputs from 324 and 328 may be generated by remote platform 304 are exported to local storage within the micro-AI module 302. These models and data may be updated when only in an internet-accessible area, such as a Wi-Fi area. When not in an internet area, data on latent database 330 does not change while the user is exercising or performing an exercise workout. Current data database 332 is configured to store or record current data when the device is in an active state, such as the user is performing a workout, training, or otherwise exercising. When not exercising, current data database 332 may actively store current data reflective of the user's at rest condition, such as a wake up heart rate (WHR), or sleep duration (as indicated by the WHR over a measured interval of time). Current data database 332 dynamically stores data as it is being generated, whether the user is at rest or exercising and retains that data for a predetermined time interval as earlier described, such as two weeks.
As shown in
Referring to
To train one or more of the classification models, including the ones in the latent model and the ones for building the final classification model, a supervised training model is preferred. The models (DA, LR, NN . . . ) included in the latent model generate the probability outputs (326 in
Let's say Joe Smith is a 40-year-old runner who has been running for 3 years. He lives in the Gold Coast hinterland of Australia (hilly area in a warmish climate). His dad was a professional runner. His mother was an average rower. Joe's goal is to run the Gold Coast Marathon in 3 months.
Latent Historical Data
(General) Historical patterns for a male 40-44 years old show that 40-50 Kilometers a week lead to a decent health fitness (e.g., average wake up heart rate (WHR) of 45 beats per minute).
(Specific) Joe's own historical data is based on his last 3 years. He averaged 40K per week, maintaining decent health (average wake up HR 50), but became injured one time when he ran 80K per week.
Current Data
Joe's training data is stored automatically on his fitness tracking device for up to 2 weeks every time he exercises. This includes his heart rate (active and wake-up). On a given day, Joe logged in a long 15K run to prepare for the marathon. This is one day after he ran 20K. He only slept 4 hours. His wake-up HR was only 62.
Conclusions: Joe's wake up HR is elevated compared to his personal “fit” HR, and deviates from the general athlete population of 40-year-old men “fit” HR too. The other factor is the two days in a row of longer than normal training. The HR combined with the duration of exercise suggests that Joe is over-training. Joe complained of sore muscles (sharp pain) after his second day 15K run. Joe has historically been injured when he has done too much without sufficient recovery between long or hard sessions. Based on the latent historical data (general and specific), and the current data (HR), the resulting conclusion from the micro-AI on Joe's device is that Joe is overtraining and needs to alter his training regimen. This conclusion is based on the technical conclusion further described below. The sharp pain is a flag that he is at high risk of injury, and needs to significantly alter his training by taking a day or two off.
Technical conclusion: an overtraining conclusion is derived from latent historical parameters using statistical applications stored on the fitness tracker, combined with current data on the fitness tracker. Current data HR, intensity, and training duration (time and/or distance), combined with analysis that compares historical individual trends and historical general population category trends is utilized.
Output: an analysis which shows that training is on track towards a goal (can be visually presented per a curve), or that training needs to be altered due to overtraining and/or elevated injury risk. This may be delivered as a warning on the fitness tracker (verbal, vibratory, and/or visual), or via another connected device.
Following on from the above example, Joe decides to go on a 10K run the day after his 15K run. His WHR is 65, and he had 7 hours of sleep. Feeling tired, he decides to go for an easy run. Ten minutes into the run, his active heart rate reaches 180 beats per minute as measured by the heart rate sensor on the fitness tracker worn by Joe. The micro-AI component of Joe's fitness tracker analyses Joe's current active heart rate and intensity level compared to latent specific historical data and determines that the heart rate is overly elevated given Joe's relaxed pace, activating the primary trigger. The processor on Joe's fitness tracker sends an audio, vibratory and visual warning to Joe alerting him that he ought to cease further training or risk becoming injured.
The system may be configured for use with activities other than athletic activities such as running, cycling, swimming or multisport. For example only, the system may be configured for use with forms of exercise such as kayaking, crew, hiking, yoga, weight training, core sessions, and other sports or athletic activities.
Other sensors may be used in combination with a heart rate sensor, or in instead of a heart rate sensor to help ascertain an overtraining condition and to minimize injury risk. For example, one or more movement sensors may be used to determine proper running gait, cycle form or swim stroke analysis to increase training efficiency and lower injury risk. Information or data normally stored at a remote physical location may be stored in the Cloud, considerably reducing the onsite hardware needed for memory requirements often associated with large volumes of data.
Where the system is configured to biologically monitor the fitness of an athlete and provide a warning to the athlete when biological conditions indicate the athlete is overtraining, the system may be configured to base the determination of overtraining on the active heart rate of the athlete while the athlete is exercising, and/or base the determination on an at rest heart rate of the athlete when the athlete awakes from sleeping, and/or a combination of the heart rate and training duration. The alert provided by the system may be configured to caution the athlete to alter training frequency, and/or the alert may be configured to warn the athlete to alter training duration.
If desired, elements of the micro-AI module may be implanted (e.g., sub-dermally) to increase the durability of the device to outdoor conditions.
The features described with respect to one embodiment may be applied to other embodiments, or combined with or interchanged with the features of other embodiments, as appropriate, without departing from the scope of the disclosure.
The present disclosure in one or more preferred forms provides the advantages of being more responsive, more convenient, involving reduced data transmission (saves on communication resources), and being capable of being used even when out of communications range, which can happen on training segments. Lessor memory is required since lesser bulk data is being stored for constant analysis.
Other embodiments of the disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the embodiments disclosed herein. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the invention being indicated by the following claims.
This application claims the benefit of U.S. Provisional Patent Application No. 62/982,660, and U.S. patent application Ser. No. 16/863,285, the entire contents of which are each hereby incorporated by reference herein.
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/US2021/019729 | 2/25/2021 | WO |
| Number | Date | Country | |
|---|---|---|---|
| 62982660 | Feb 2020 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 16863285 | Apr 2020 | US |
| Child | 17802955 | US |
