Periodic Reporting for period 1 - SPEARS (Skill Performance Estimation from cARdiac Signals)
Período documentado: 2024-01-01 hasta 2025-06-30
Worldwide, billions of people spend regular time training, to acquire new knowledge or improve their physical skills (e.g. running). Training apps are relatively new ways to do so, with for instance 150 millions sport trackers sold yearly. Such apps typically offer personalized training solutions that adapt the sequence of training exercises to each user to maximize training efficiency and effectiveness. One way to optimize training is to maximize difficulty while avoiding failure (to increase motivation and optimize progression). However, current solutions typically consider users’ past performance only to decide the next exercise to propose, ignoring that different learners will have different capacities at a given time. Learners can indeed be tired, stressed or very motivated during one exercise, or even sick one day, but not the next. Current training apps are thus bound to propose frustrating exercises not adapted to the users’ current capacities nor motivation, or even possibly dangerous ones, with a risk of injury for sport training apps. There is thus a need for personalized training apps taking into account the users’ current cognitive and/or physical states to enable optimally effective, efficient and safe training.
The scientific literature shows that users’ cognitive and physical states (ex : Stress, mental workload or fatigue) are reflected in their cardiac signals (CS). The development of consumer grade smartwatches (~220 millions users worldwide) embedding CS sensors allows such measures to be used globally. Most endurance sports athletes also wear CS sensors in smartwatches or chest belts. Thus, SPEARS proposes to use CS measured using smartwatches or chest belts, for cognitive and/or physical states estimation, to provide users with optimal training. Such apps could predict their users’ cognitive and/or physical performance from their CS to evaluate their capacity and adapt the subsequent training exercises accordingly. This should enable optimal cognitive and/or sport training, improving users’ education, cognitive and physical skills, and health. We notably explore such an industrial application with the Flit Sport company, commercializing an Artificial Intelligence (AI)-based app, FlitCoach, to prescribe individual training plans for runners.
The scientific literature shows that users’ cognitive and physical states (ex : Stress, mental workload or fatigue) are reflected in their cardiac signals (CS). The development of consumer grade smartwatches (~220 millions users worldwide) embedding CS sensors allows such measures to be used globally. Most endurance sports athletes also wear CS sensors in smartwatches or chest belts. Thus, SPEARS proposes to use CS measured using smartwatches or chest belts, for cognitive and/or physical states estimation, to provide users with optimal training. Such apps could predict their users’ cognitive and/or physical performance from their CS to evaluate their capacity and adapt the subsequent training exercises accordingly. This should enable optimal cognitive and/or sport training, improving users’ education, cognitive and physical skills, and health. We notably explore such an industrial application with the Flit Sport company, commercializing an Artificial Intelligence (AI)-based app, FlitCoach, to prescribe individual training plans for runners.
We proposed new approaches to estimate runners’ performances, and even predict their future performance, based on their past running speeds during exercises and their CS, using AI. This has been integrated into FlitCoach for internal evaluation by FlitSport.
We first implemented existing algorithms to estimate endurance athletes’ performances, i.e. CPW and W’ balance algorithms – based on runners’ speed, and assessed them on FlitSport’s runners database. We used the critical speed as metric of athletes’ performances, which is a threshold on the running speed below which a runner can theoretically run without stopping. We aimed at estimating and predicting this performance metric precisely.
The first algorithm (CPW) – considered the gold standard - can estimate a runner’s critical speed, but requires data from numerous (months of) exercises from that runner to do so and often underestimate the runners’ performance. Thus, it is not suitable to estimate performance from few data, nor to predict short term future performance, which is required for personalization, as proposed in FlitCoach. We thus proposed the CPW’ balance algorithm, combining CPW and the W’ balance algorithm. W’ balance tracks the expenditure of energy during an exercise, estimated from the runner’s speed. CPW’ balance can provide an estimate for the critical speed of a runner with substantially less data, typically from 3 weeks of data only. Its estimation performances, in terms of a mean squared error (MSE - difference with the CPW ground truth – the lower the better) is 0.28.
Then we proposed a 3rd algorithm based on Riemannian geometry AI, using the runners’ CS. It uses as input a representation of specific running exercises as covariance matrices, integrating CS and running data. Here, the heart rate (BPM, beat per minutes ) was estimated from runners’ smart watches or cardiac belts. Using this new approach, the critical speed can be estimated robustly with the data of a short period of exercises only. It proved very robust, reaching an MSE of around 0.13 only.
The last and main goal for physical performance prediction, was to predict the runner’s performance in the distant future (~6 months in the future), to prepare a competition. A model to do so is available in literature, the Banister model, but fitting its parameters works fine only for some runners but absolutely not for others. Moreover, this model is not very good at predicting performances in a far future, with an MSE of 0.34. Thus, we developed a new approach using AI taking as input the volume of training for the last month and predicting the critical speed for the next month. Thus, if the runner knows his/her planned training volume for each coming month, his/her performance can be predicted in a far future. Moreover, contrary to the Banister model, our new AI model does not require to be trained and fit on each runner’s data, it works across runners, and can thus predict performance for a new runner without data from that runner. This makes it more applicable in practice. It also proved very robust, with an MSE of around 0.05 only.
We also conducted pilot work on cognitive performance prediction. To do so, we had a user perform the so-called N-back tasks. With an N-back task, letters are displayed on screen one by one, and the user is required to identify if the current letter is the same as the letter N characters before. The user was wearing a consumer grade cardiac belt to record BPM. Then, an AI classifier was used to predict the cognitive performance of the user from BPM features, i.e. how well he successfully identified the letters. Our preliminary results suggest that CS can be used to predict – to some extent - success or failure at these cognitive tasks, with a classification accuracy of 62%.
We first implemented existing algorithms to estimate endurance athletes’ performances, i.e. CPW and W’ balance algorithms – based on runners’ speed, and assessed them on FlitSport’s runners database. We used the critical speed as metric of athletes’ performances, which is a threshold on the running speed below which a runner can theoretically run without stopping. We aimed at estimating and predicting this performance metric precisely.
The first algorithm (CPW) – considered the gold standard - can estimate a runner’s critical speed, but requires data from numerous (months of) exercises from that runner to do so and often underestimate the runners’ performance. Thus, it is not suitable to estimate performance from few data, nor to predict short term future performance, which is required for personalization, as proposed in FlitCoach. We thus proposed the CPW’ balance algorithm, combining CPW and the W’ balance algorithm. W’ balance tracks the expenditure of energy during an exercise, estimated from the runner’s speed. CPW’ balance can provide an estimate for the critical speed of a runner with substantially less data, typically from 3 weeks of data only. Its estimation performances, in terms of a mean squared error (MSE - difference with the CPW ground truth – the lower the better) is 0.28.
Then we proposed a 3rd algorithm based on Riemannian geometry AI, using the runners’ CS. It uses as input a representation of specific running exercises as covariance matrices, integrating CS and running data. Here, the heart rate (BPM, beat per minutes ) was estimated from runners’ smart watches or cardiac belts. Using this new approach, the critical speed can be estimated robustly with the data of a short period of exercises only. It proved very robust, reaching an MSE of around 0.13 only.
The last and main goal for physical performance prediction, was to predict the runner’s performance in the distant future (~6 months in the future), to prepare a competition. A model to do so is available in literature, the Banister model, but fitting its parameters works fine only for some runners but absolutely not for others. Moreover, this model is not very good at predicting performances in a far future, with an MSE of 0.34. Thus, we developed a new approach using AI taking as input the volume of training for the last month and predicting the critical speed for the next month. Thus, if the runner knows his/her planned training volume for each coming month, his/her performance can be predicted in a far future. Moreover, contrary to the Banister model, our new AI model does not require to be trained and fit on each runner’s data, it works across runners, and can thus predict performance for a new runner without data from that runner. This makes it more applicable in practice. It also proved very robust, with an MSE of around 0.05 only.
We also conducted pilot work on cognitive performance prediction. To do so, we had a user perform the so-called N-back tasks. With an N-back task, letters are displayed on screen one by one, and the user is required to identify if the current letter is the same as the letter N characters before. The user was wearing a consumer grade cardiac belt to record BPM. Then, an AI classifier was used to predict the cognitive performance of the user from BPM features, i.e. how well he successfully identified the letters. Our preliminary results suggest that CS can be used to predict – to some extent - success or failure at these cognitive tasks, with a classification accuracy of 62%.
Overall, we proposed 3 new algorithms: CPW’ balance, Riemannian performance estimation and long-term performance prediction based on AI.
Our Riemannian algorithm using CS requires a minimal amount of training data, can estimate performance from a short period of running exercises (when the stat-of-the-art requires months of exercises), and very robustly so (MSE around 0.13).
Our algorithm to predict performance in a far future outperforms the state-of-the-art, requiring few training data and being very accurate (MSE around 0.05 VS an MSE of 0.34 for the state-of-the-art Banister model). Moreover, it works across athletes, not requiring a new calibration for each new runner like the Banister model. This makes is much more practical, enabling long-term predictions for a new runner without data from that runner.
For better assessing and improving these models, more tests in real time, in FlitCoach, with various new runners’ profiles are required. They are currently between tested internally on FlitSport new runners’ data, on real use cases. If their robustness is confirmed with these tests, they could be transferred to FlitSport.
Regarding the cognitive performance estimation and prediction, the results are more at an exploratory stage, requiring further research and evaluation first.
Our Riemannian algorithm using CS requires a minimal amount of training data, can estimate performance from a short period of running exercises (when the stat-of-the-art requires months of exercises), and very robustly so (MSE around 0.13).
Our algorithm to predict performance in a far future outperforms the state-of-the-art, requiring few training data and being very accurate (MSE around 0.05 VS an MSE of 0.34 for the state-of-the-art Banister model). Moreover, it works across athletes, not requiring a new calibration for each new runner like the Banister model. This makes is much more practical, enabling long-term predictions for a new runner without data from that runner.
For better assessing and improving these models, more tests in real time, in FlitCoach, with various new runners’ profiles are required. They are currently between tested internally on FlitSport new runners’ data, on real use cases. If their robustness is confirmed with these tests, they could be transferred to FlitSport.
Regarding the cognitive performance estimation and prediction, the results are more at an exploratory stage, requiring further research and evaluation first.