At present, increasing amounts of data are collected in many disciplines including running, in particular through sensors (e.g., Global Positioning Sensor: GPS, accelerometer, heart rate monitor…), connected objects (e.g., smart meters such as watches, glasses, textiles, insoles…) (37–39) content published on databases (e.g., performance, split times during competitions, results…). If we are interested, for example, at these data commonly collected in running via connected devices and/or smartphone applications (e.g., Strava®, Garmin®, Runtastic®…) (40), the latter can make it possible to analyze, or even predict, performance by monitoring several variables (e.g., pace or speed of movement, variation in the altitude difference of the course, amplitude and frequency of strides…) in a non-invasive way in real conditions (i.e., with possible real-time feedback) and especially outside a laboratory (14, 34, 37, 41–43). The research of Emig and Peltonen (34), Smyth and Muniz-Pumares (30) has notably highlighted mathematical modelling based on the use of connected wearable technologies such as wrist devices (e.g., connected watches) or smartphones to correlate performance indices with the volume and intensity of training in order to quantify, for example, the optimal training load. These studies (30, 34) have chosen to integrate in their algorithms (already taking into consideration past references on the target distance), the runners’ training regime (e.g., distance, time, running pace and elevation gain) 6 weeks before the prepared competition. The use of these connected technologies and Big Data, suggests new ways of quantifying and predicting athletic performance in real conditions. Other technologies, still not widely used in the running world, such as connected insoles could also allow the collection and exploitation of new data in order to understand and optimize sports performance (e.g., integration of these data in prediction algorithms) (44, 45). However, beyond the quality of the recovered data (i.e., precision, accuracy with over or underestimation of raw values, like recorded energy expenditure depending running intensities, but also distances significantly underestimated and less accurate in the forest areas to the road area, that can highlight the limitations of connected objects) (46–48), the exploitation of these data could be complex in view of the quantity of available data (Big Data) obtained through connected technologies. What methodology/approach could then respond to this Big Data problem?

AI used in many fields of science (e.g., meteorology, medicine, sport sciences…) (32, 49, 50) suggests new ways of quantifying and predicting sports performance in real-life conditions, given the scientific publications obtained in recent years. Indeed, Hammerling et al. (20) performed predictions of race times in the 2013 Boston Marathon for all runners who reached the halfway point of the race but did not have the opportunity to cross the line due to an attack (i.e., interruption of the race following the explosion of two bombs placed near the finish line) and thus to recognize the achievements of these runners. To make these predictions, Hammerling et al. (20) used a database of all previous years’ performances at the same event (i.e., Boston Marathon of 2010 and 2011) and took into account the “real” times (i.e., time intervals of 5 km as well as the final 2,195 km) of the runners engaged in this event before they were interrupted by the organization. The use of different AI algorithms including K-Nearest Neighbors (KNN), made it possible to predict the performance of all the runners (i.e., a final time based on the prediction of intermediate times), i.e., the time they could have achieved over the distance, to establish a ranking of the runners involved in this event. To do this, authors created an independent validation dataset from a fraction of the runners (25.7%) who finished the race and then ran dropout simulations at various points of the race on these runners in the same proportion as the true runners who unfortunately did not finish. The predicted finish times of this sample of fake runners who drop out were then compared with the actual finish times of these runners to assess the effectiveness of the statistical approaches. These predictions shown to be relatively accurate (i.e., Mean Absolute Error of 1 min 30 s on average) with an increased accuracy for runners who had to abandon later. In addition, beyond the prediction of performance, via AI, based on real data recovered during competitions, other work (14) has highlighted the use of supervised learning algorithms based on real training conditions in amateur runners to predict marathon performance, for example.

Given the multifactorial aspect of performance (25–31, 51) and the fact that performance prediction is a subject of great interest to athletes and coaches, we could “legitimately” ask ourselves what future prediction models might look like? Taking into account previous work in this field of AI and the current evolution of connected technologies such as textiles or insoles, would the challenge then be to think of an approach, an equation that is able to “simply” optimize a large number of factors correlated with past data (e.g., data based on training program or even the start of a race), “actual” (i.e., data obtained in real time, such as heart rate) and/or future (e.g., estimates based on future conditions and forecasts during a race, such as changes in weather conditions) of running performance by connected objects? Could we not try to propose a multifactor equation:Perf=FactPhys×FactBiom×FactPsy×FactEnv×FactTechwhere each physiological (FactPhys), biomechanical (FactBiom), psychological (FactPsy), environmental (FactEnv) and/or technological (FactTech) factor would be expressed through the intermediary and preponderance of indices conducive to performance without taking the risk of straying into “prediction-fiction”. In this case, it would be a matter of extracting and using data from wearable devices to identify potential performance indices and then transform them into significant parameters with the ultimate objective of designing a “fair” and “accurate” modeling of running performance. This would be an “intelligent” model capable of taking into account a large amount of information based on physiological factors (e.g., values of critical speed, HR, acceleration, muscle oxygenation, body temperature, hydration rate…), biomechanics (e.g., values of amplitude and frequency of the stride, strength, muscle power, foot placement on the ground…), psychological (e.g., stress indices, motivation, psychological state or personality trait related to the challenge of the competition), environmental (e.g., weather indices, course profile, context of the race…) and technological (e.g., energy storage/return values of shoes, aerodynamic values of textiles) in order to be able to “coach” the athlete at the present time (“T” time), either to indicate to him/her, for example, to accelerate, stabilize or reduce his/her running speed according to the effort he is making and the effort he/she will still have to make with the “ultimate” objective of optimizing sports performance (Figure 1). To develop such a formula, we could use, for example, multiple regression to extract and use the relevant data in the model. However, we should be careful about the risk of multicollinearity if one of the explanatory variables in a model is a combination of one or more other explanatory variables in that model, thus distorting the coefficient estimates.