We compare polarized training, threshold training and AI optimized endurance training. AI optimized training yields the best results, followed by polarized training with threshold training in third. The results are inline with current exercise physiology research. If the training composition is not optimized to the individual athlete, substantially smaller gains are to be expected.

TL;DR

• AI optimized training promises on average 11.1% performance gains in FTP and 5k pace in a simulated study with 126 runners, cyclists and triathletes over a training plan of 8 weeks at the same overall training volume the athletes are accustomed to (3.3% error, improvement over the baseline significant at greater than 3-sigma, p<0.002).
• Providing an athlete with a personalized training plan that is not biased towards a training philosophy increases performance gains over the one-size-fits-all approach (about 2 times larger gains in the central value, significant at about 2-sigma, p<0.05).
• AI optimized training is studied in the context of polarized and threshold training principles. For 89% of athletes an AI optimized polarized training composition is most beneficial while for 11% an AI optimized threshold training composition is optimal. On average, polarized training is found to be superior over threshold training in agreement with exercise physiology research, see references [1-9].
• Compared to conventional randomized controlled trials, this study leverages the 'digital twin' approach instead: each athlete is represented by a machine learning model that executes a selection of training routines via inference. This yields significantly larger statistics compared to conventional studies.

A new way of studying the effectiveness of different training routines

In conventional exercise physiology research studies, randomized controlled trials are performed on a population of participants. For instance, the participants perform certain training routines and the results are analyzed as factual changes of a performance indicator on the population, see for example [1-9].

Here we employ a different approach: every participant has a 'digital twin' - a machine learning model/neural network that is trained on the individual's historical exercise data. The model represents how the athlete individually responds to different training inputs in terms of training volume and composition. It can be used to predict the athlete's performance changes when following a certain training routine, within a margin of error.

This approach allows orders of magnitude higher statistics compared to conventional studies, as the digital twin can execute thousands of different training routines. On the other hand, a human participant can only execute at most a few. Conventional studies are tedious, expensive, take a long time, and have generally low number of participants. We rely on neural networks to simulate the complex exercise response cycle of the human body. A limitation of the digital twin approach, however, is that a certain bias is introduced via the choice of data processing and machine learning model.

AI optimized training

AI optimized training (AIO) as used by AI Endurance is based on the individual athlete's historical data. In the first step, the machine learning algorithm 'learns' to be able to predict the athlete's performance by fitting the athlete's data and evaluating the prediction error. In the second step, AI Endurance finds the optimal training routine to reach peak performance at the athlete's goal event date by selecting the training plan with the largest predicted performance gains.

AI Endurance training plan is based on 5 intensity zones (Endurance, Tempo, Threshold, VO2Max, Anaerobic). For cycling, we define these as less than (75, 90, 106, 120, -) percent of FTP. For running, less than (75, 85, 95, 100, -) percent of 5k/VO2Max pace. Read more on AI Endurance's training zones.

A crucial problem when looking for the optimal training input is the extrapolation problem in machine learning [10]: one can only infer reliable predictions reasonably close to the data the neural network has been trained on. In this case, we restrict the possible input data up to one standard deviation away from the mean of the athlete's historical data. For instance, take an athlete with a mean of 1 hour and a standard deviation of 10 minutes for a Tempo intensity zone. The algorithm will explore the 50-70 minute range when looking for an optimal training plan. Conversely, if an athlete has never spent any time in a certain zone, the plan finding can't reliably explore this zone as there is no input data to learn from and hence we discard the zone. These limitations are important when discussing more static training inputs such as polarized and threshold training in the next sections.

AIO is the optimal training plan that can be inferred from an input space defined by the variety of the athlete's historical data. The algorithm is un-biased in that it is agnostic towards any training principle or philosophy. It simply tries to optimize the athlete's goal event performance by finding an optimal training plan in terms of times spent in certain intensity zones.

Polarized training

Polarized training (POL), see for example [1,2], is a training principle that favours a balanced approach between low intensity endurance efforts (typically around 80% of training time) and high intensity interval work (typically around 20% of training time) while almost completely cutting out the intermediate intensities. This approach is favoured by current exercise physiology research [5-9] and is used by many elite endurance athletes [6].

Polarized training is often described in a three zone system with Zones 1, 2, and 3 representing low, intermediate, and hard intensity, respectively. To compare with AIO, we identify Zone 1 with the Endurance zone, Zone 2 with the Tempo and Threshold zones combined, and Zone 3 with the VO2Max and Anaerobic zones combined. These identifications are sensible due to significant intensity overlap.

To explore the predicted performance gains from POL within this study, we restrict the AIO training input space by suppressing training time in the Tempo and Threshold zones. The Endurance, VO2Max and Anaerobic zones are unconstrained.

We study two versions of polarized training: AI optimized polarized (AI-POL) with the same amount of iterations as AIO (30,000) and un-optimized polarized (uPOL) which only uses 3,000 iterations to find the optimal training input. AI-POL represents an optimized polarized training plan that not only follows the polarized principle but also fine-tunes the work in the different sub zones to achieve maximal performance gains. uPOL on the other hand is much less optimized and is more representative of an off-the-shelf, one-size-fits-all polarized training plan.

Threshold and pyramidal training

Threshold training (THR), see for example [3,4], is a training principle that favours interval work at Threshold and Tempo intensities. Typically the balance between Endurance and the Threshold and Tempo intervals is in the 60/40 realm, with little to no intensity in the zones above threshold.

To explore the predicted performance gains from THR within this study, we restrict the AIO training input space by suppressing training time in the VO2Max and Anaerobic zones. The Endurance, Tempo and Threshold zones are unconstrained.

As with POL, we study two versions of threshold training: AI optimized threshold (AI-THR) with the same amount of iterations as AIO (30,000) and un-optimized threshold (uTHR) which only uses 3,000 iterations to find the optimal training input. AI-THR represents an optimized threshold training plan that not only follows the threshold principle but also fine-tunes the work in the different sub zones to achieve maximal performance gains. uTHR on the other hand is much less optimized and is more representative of an off-the-shelf, one-size-fits-all threshold training plan.

Pyramidal training [13] does not aim to avoid the high intensity zones but instead proposes a pyramidal structure with most time spent at low intensities, followed by less but still a substantial amount of time at intermediate intensities and the least amount of time at high intensities. As we are not enforcing a strict static training composition due to the extrapolation constraints discussed above, we group both threshold and pyramidal training under the THR keyword. This is also justified by the optimal THR training compositions for the subjects of this study which can be classified as somewhere in between threshold and pyramidal, see Technical Details section below.

Results: Predicted performance gains

We analyze the predicted improvements for a sample of 126 athlete digital twins which are either runners, cyclists or triathletes. Improvements are measured as changes in 5k pace (runners, triathletes) and/or FTP (cyclists, triathletes) over a training program duration of 8 weeks. The changes are relative to the last recorded performance before the start of the training program.

We limit each individual athlete's training volume by the volume they have trained over the past 8 weeks. On average, the training volume of the population was 4.7 hours/week. This is to see what improvements can be achieved with the same training volume an athlete is already accustomed to and might be restricted to. The average FTP prior to the training program is 248 Watts and the average 5k pace is 4:33 min/km (7:20 min/mi).

As a baseline, we let each athlete repeat the last 8 weeks of their training and infer the predicted change in performance. We use AI Endurance's optimization engine to create a training plan for each of the training routines discussed: AIO, AI-POL, AI-THR, uPOL, and uTHR. Predicted performance changes are inferred from these training plans with an individual prediction error for each athlete. We aggregate these results into mean predicted improvements (see Technical Details section below).

The baseline result is that if the athletes repeat their training no performance improvements are seen. There is a very significant (greater than 3-sigma, p<0.002) finding that AIO training leads to performance improvement relative to the baseline. For the training programs, there is a clear hierarchy as AIO and AI-POL training predict larger gains than the less optimized uPOL and uTHR routines.

Unsurprisingly, AIO achieves better results than AI-POL and AI-THR in the fully optimized category as it has a larger parameter space available than the two latter. AI-POL and AI-THR are pigeon-holed into the respective training philosophies while AI is completely unbiased towards these. However, for most athletes if you would choose a training philosophy bias, POL is much better than THR.

AI-POL realizes better gains than AI-THR for 89% (112) of athletes while AI-THR is better than AI-POL in only 11% (14) of cases. Also the gains are generally higher for AI-POL compared to AI-THR. These two findings combined, explain the strong overlap for predicted performance changes between AIO and AI-POL.

For more information on the errors and machine learning techniques see Technical Details below.

Comparison with exercise physiology research

More recent endurance exercise physiology research generally favours POL over THR, see for example [2, 5-9]. [5] is a study with 48 endurance athletes, finding significant improvement with POL in VO2Peak (+11.7%, p<0.01) and velocity/power at 4 mmol/L (+8.1%, p<0.01), and other indicators, while finding no significant improvements from THR. [6] finds that the elite endurance athletes generally follow the POL approach. [7] is a meta analysis which combines 4 randomized controlled trials with a combined 329 results, finding that POL is preferred over THR at moderate effect size. [2] finds THR leads to 3.6% improvements, while POL leads to 5.0% improvement in a study of 10k times with 30 recreational runners.

Performance indicators in these studies are not identical with this study (FTP, 5k pace). However, on a qualitative level, the results are in agreement. In particular, the lack of evidence supporting THR and the general trend that POL is preferred over THR is reflected in our results.

We would further like to comment on the sample sizes. The meta analysis of [7] includes a total of 329 results. Our study, while only using the historical exercise data of 126 athletes already includes 6 x 126 = 756 results (not counting the tens of thousands of optimization steps). This clearly outlines the potential of machine learning based or supplemented studies to overcome the cost and duration challenges of conventional randomized controlled trials. A future avenue could be to anchor machine learning based simulations with 'real life athlete' study results to get the best of both worlds. Similar possible ways to use neural networks in this context have been suggested in [11,12].

Results: How important is personalization?

As displayed in Table 1, uPOL and uTHR have central values of almost a factor 2 (1.7) less than their AI counterparts. These improvements are due to more nuanced adjustments of an individual's training plan as opposed to just roughly following the training philosophy. uTHR is not found to be a significant improvement over the baseline.

The improvement of AIO over uPOL is significant at 1.5-sigma while AIO over uTHR is significant at 1.9-sigma. As discussed in the previous section, POL generally yields better results for a higher number of athletes than THR. However, from the perspective of an individual athlete, they might be part of the 11% that benefit from a more THR-like training composition. Hence, an unbiased and personalized search for the ideal training composition is important.

How you can benefit

AIO as described in this study is enabled for every AI Endurance user by default. You can benefit from this truly data-driven, predictive and research-proven approach to training by signing up to AI Endurance.

Participate in future studies

The number of participants is a crucial factor for the quality and implications of a study such as this one. If you are using AI Endurance, you can allow us to use your data to publish research studies about the effectiveness of endurance training such as this one. These studies can help move exercise research forward and help your fellow athletes achieve their goals. We are extremely conscious of your privacy concerns. Your data would always remain anonymous, i.e. your name or any performance values will never be published. The data would only be published in aggregate form, such as: 'On average x training composition helped athletes improve by y percent.'

If you want to allow AI Endurance to use your data in studies such as this one, simply check this box in your Account page and press 'Save Changes'

By default this box is unchecked, i.e. your data will never enter any public research without your consent.

Technical details

This section provides additional technical background on this study:

Scope

• We conduct this study with the digital twins of 126 athletes (runners, cyclists and triathletes).
• The predicted performance changes are measured over a time frame of 8 weeks.
• The training volume for each athlete for the 8 weeks training block we study is guided by the volume averaged over the last 8 weeks of training, for each individual athlete.
• For runners, we measure performance changes in 5k pace; for cyclists, as changes in FTP; for triathletes as changes in 5k pace and FTP combined at equal weight. Changes are measured from the last measured performance to the predicted performance at the end of 8 weeks of training. The current level is conservatively estimated as the maximum of either the actual recorded athlete performance or the machine learning model prediction for the current performance level.
• The average level of cycling performance in this study prior to the training block is 248 Watts for FTP. For running performance the average 5k pace prior to the training block is 4:33 min/km (7:20 min/mi). The average optimal training time of each athlete in this study is 4.5 hours/week.
• For 6% (7) of athletes, no improvement, i.e. a negative central value for the predicted performance change, was found. This can be due to a convergence failure of the optimization algorithm to find an ideal training composition or genuine constraints from the athlete's physiology. For instance, if the athlete is 'riding off a peak' with more intense training prior to the last 8 weeks it might not be possible to improve or even hold the current form on the same training volume.

Training inputs & machine learning

• For machine learning inference, extrapolation is restricted to inputs that are reasonably close the athlete's historical data. The lower limit of the extrapolation range is the minimum of either the athlete's lowest recorded input or one standard deviation subtracted from the mean. Conversely, the upper limit of the extrapolation range is the maximum of either the athlete's highest recorded input or one standard deviation added to the mean.
• For AIO we leave the available input parameter space unrestricted.
• For POL we suppress time spent in the Tempo and Threshold training zones. Time in Endurance, VO2Max and Anaerobic zones are unrestricted. To avoid extrapolating into a regime where the model has little predictive power, these cannot simply be set to an arbitrarily low value and are instead set close to the athlete's lower extrapolation bound.
• For THR we suppress time spent in the VO2Max and Anaerobic training zones while leaving the Endurance, Tempo and Threshold zones unrestricted. The same extrapolation restrictions apply as for POL.
• The optimization process to find the ideal training input is run for 30,000 iterations for AIO, AI-POL and AI-THR. We also run a less optimized version with 3,000 iterations, representing uPOL and uTHR.
• The aggregated intensity zone distributions for each training composition are summarized in the Table 2. The AI optimized variants generally have a larger spread due to increased personalization.

Prediction errors

• The average prediction error is 5.47%. For each athlete, the prediction error is calculated as the cross-validation error of the neural network's predictions compared to actual performances of the athlete the model was blind to when being fitted. For more information, see this article about cross-validation.
• The error on the average prediction improvements is calculated taking into account potentially correlating effects. Even though predictions for different athletes are in principle not correlated as far as their underlying physiology is concerned, there exists a modest correlation due to the choice of the machine learning model and the implied data structures. We very conservatively estimate covariance via the sample correlation coefficient from a combined sample of the training predictions and optimization runs with different hyper parameters.

Disclaimer

This study was performed by AI Endurance. To protect the privacy of participants, we are not publishing any absolute performance values, only relative improvements and aggregate data on training compositions. To protect our intellectual property we are not planning on publishing any further details on the underlying AI algorithm.

References

1. Impact of training intensity distribution on performance in endurance athletes - Jonathan Esteve-Lanao, Carl Foster, Stephen Seiler, Alejandro Lucia - J Strength Cond Res 2007
2. Does polarized training improve performance in recreational runners? - Iker Muñoz, Stephen Seiler, Javier Bautista, Javier España, Eneko Larumbe, Jonathan Esteve-Lanao - Int J Sports Physiol Perform 2014
3. The significance of the aerobic-anaerobic transition for the determination of work load intensities during endurance training -W Kindermann, G Simon, J Keul - Eur J Appl Physiol Occup Physiol 1979
4. Metabolic and performance adaptations to interval training in endurance-trained cyclists - C Westgarth-Taylor, J A Hawley, S Rickard, K H Myburgh, T D Noakes, S C Dennis - Eur J Appl Physiol Occup Physiol 1997
5. Polarized training has greater impact on key endurance variables than threshold, high intensity, or high volume training - Thomas Stöggl, Billy Sperlich - Front Physiol. 2014
6. The training intensity distribution among well-trained and elite endurance athletes - Thomas Stöggl, Billy Sperlich - Front Physiol. 2015
7. Polarized vs. Threshold Training Intensity Distribution on Endurance Sport Performance: A Systematic Review and Meta-Analysis of Randomized Controlled Trials - Michael A Rosenblat, Andrew S Perrotta, Bill Vicenzino - J Strength Cond Res 2019
8. Polarized and Pyramidal Training Intensity Distribution: Relationship with a Half-Ironman Distance Triathlon Competition - Sergio Selles-Perez, José Fernández-Sáez and Roberto Cejuela - J Sports Sci Med. 2019
9. Current Scientific Evidence for a Polarized Cardiovascular Endurance Training Model - Jay R Hydren, Bruce S Cohen - J Strength Cond Res 2015
10. Extremal learning: extremizing the output of a neural network in regression problems - Zakaria Patel, Markus Rummel - arXiv:2102.03626
11. The analysis and utilization of cycling training data - Simon A Jobson, Louis Passfield, Greg Atkinson, Gabor Barton, Philip Scarf - Sports Med 2009
12. A Mine of Information: Can Sports Analytics Provide Wisdom From Your Data? - Louis Passfield and James G. Hopker - Int J Sports Physiol Perform 2017
13. Heart rate and performance parameters in elite cyclists: a longitudinal study - A Lucía, J Hoyos, M Pérez, J L Chicharro - Med Sci Sports Exerc 2000