What is Enfusion™and why should I care? Part 2
In the world of endurance sports, Dr. Eric Banister’s Impulse-Response model has formed the foundation for, or had significant influence on nearly every model of performance currently used by training software and websites today. In fact, the vast majority of training websites and software packages are based on Dr. Andy Coggan’s adaptation of Impulse-Response into the Training Stress Balance approach. Enfusion is the first available model of human performance to use machine learning to step away from this methodology. The purpose of this series of blog posts is to explain what problems Enfusion helps to solve, and more importantly, how Enfusion can make you faster. In Part 1 of What is Enfusion and why should I care?, we discussed the benefits and limitations of using “stress based” metrics to track training load, as well as how Enfusion solves the problems associated with trying to track stress. Here in Part 2, we’re going to discuss the pros and cons of using “stress based” models to make estimations of overall fitness, or “form.” Then we’ll demonstrate how Enfusion’s machine learning technology addresses these limitations, and how this allows Enfusion to be the first system in the world able to track qualitative changes in performance on a daily basis.
Estimating “Form” by using a Training Stress Balance approach: Uses and Limitations
In Part 1 of What is Enfusion and why should I care?, we discussed the advent of Training Stress Score (TSS) and other similar stress metrics, and how they have given coaches and athletes the ability to more effectively track how hard someone is training. We also discussed the pros and cons to having a single value describing training “stress.” One of the major benefits to this approach is that the single-value stress metric can be used in a relatively simple model to estimate fitness, fatigue, and ultimately “form.” However, as you’ll see below, models like these can suffer from important shortfalls. Again, Dr. Andy Coggan lead the way in development of what he called the Training Stress Balance (TSB) approach, which he adapted from the pivotal Impulse-Response model put forward by Dr. Eric Banister in 1975. Knowing that Dr. Banister’s model was too complicated to have mainstream use outside of the laboratory, Dr. Coggan distilled the model down to its essence:
Form (or the ability to perform) = Fitness – Fatigue
According to Banister’s original work, fitness takes longer to build and is also longer lasting. In contrast, the fatigue impulse is thought to accumulate, as well as dissipate, more quickly than fitness. With this in mind, Coggan proposes the following as proxies for Form, Fitness, and Fatigue:
TSB = CTL – ATL
Here, Coggan estimates fitness through Chronic Training Load (CTL) and fatigue by Acute Training Load (ATL). CTL is a rolling average of TSS for the last 6-weeks and is indicative of how much fitness an athlete could have gained over the last month-and-a-half. ATL is the rolling average of TSS for the most recent week. Given the assumption that the most recent training would have the biggest impact on how tired an athlete is, ATL is used as an estimate of fatigue. In this case, a positive TSB value indicates that an athlete is possibly well-trained and also rested, while a negative TSB suggests that the athlete is likely to be tired and not perform at his/her best. As you can see, TSB could be a useful method for coaches and athletes to help inform whether they are rested or not prior to key events. However, this model is not without substantial limitations. Below, we’ll outline some of these limitations and explain how Enfusion addresses each one.
Stress Balance Models assume that fitness and fatigue affect everyone the same way
Because the 6-week CTL window and 1-week ATL windows are applied broadly to all users, regardless of talent level or training history, there is an underlying assumption that all athletes will respond in a similar manner to racing/training. However, we know that there is a great deal of heterogeneity within the population-at-large, and people respond differently to training based on a host of factors. How then can we know that a 6-week CTL and a 1-week ATL is the best option for you? The short answer is we can’t. The best that can be done is to take a trial-and-error approach to try and refine the process. However, this is extremely time-intensive and risky. More importantly, there is still no way to determine objectively if the right decision was made due to the difficulty of isolating which variable(s) were critical for any positive or negative change in performance.
Enfusion makes no assumptions about how anyone will respond to training
As we began to develop Enfusion, one requirement for us was that the model must not make assumptions about how any individual athlete would respond to anything. In other words, our model would have to be able to figure out what combination of factors lead to good versus bad performances for each respective user. To accomplish this, we chose to step away from the world of performance models that are derivatives of Banister’s Impulse-Response model, and move toward a machine learning approach. Machine learning requires the computer to learn by training a model, and Enfusion trains a model for each individual user. A second critical benefit of using machine learning is automatic and continual refinement. This means that as an athlete’s abilities change over time, the model will incorporate this new information and adjust the performance estimations accordingly. Therefore, there is no trial-and-error to figure out how to make the model work best for you, as is the case when using TSB.
Stress Balance Models cannot quantify top form or recovery from overtraining, illness or injury
Because the 6 week CTL window versus the 1 week ATL is largely fixed, it makes the model susceptible to problems when less than ideal training conditions present themselves. For example, let’s assume you’re super motivated to tackle a huge training block in the build up to your next priority objective for the year. You and your coach decide to push the limits compared to what you did at this point last year. By the end of the training block you’re very tired, and your TSB is negative. You begin a rest period as planned, and your TSB changes from negative to positive. However, you still feel like bad and fear you’ve overtrained. What do you do? You likely need more rest, so you take more time off. TSB becomes an even larger positive number, and yet you’re still underperforming. How will you know when you’ve recovered completely? In short, you can’t. What the model reports and what the athlete is experiencing has become disconnected. Thus, there is no way to accurately quantify recovery even from normal training (let alone overtraining) using TSB. Conversely, the same exact same limitations mean that you also cannot identify when you’re on top-form either. TSB can only quantify the relationship between CTL and ATL, and not the complex underlying biology.
Enfusion is the only model of human performance that can estimate the quality of your performance from day-to-day
In addition to flexibility and continual refinement, another essential feature of Enfusion’s machine learning approach is the ability to measure qualitative changes in performance each day. Currently, Enfusion is the only available model of human performance able to do this. The primary output of our model is called the Enfusion Score. We define Enfusion Score as the quality of a given performance in the context of the fitness you have at that particular point in time. In other words, Enfusion Score allows you to know whether your body is responding to training or becoming fatigued, regardless of how difficult the workout was.
We accomplish this by plotting Enfusion Scores across time in our Performance Monitoring Chart (Figure 1). The Performance Monitoring Chart includes an automated statistical analysis of the athlete’s Enfusion Scores to determine the likelihood that a given Enfusion Score is normal, better than normal, or worse than normal. As shown in Figure 1, you can see that the Performance Monitoring Chart shows 3 levels of increasing likelihood for an Enfusion Score being better (top, green bands), or worse (bottom, red bands) than normal. The bands of increasing respective darkness (defined in legend above chart) correspond to >85%, >90%, or >95% chance of an Enfusion Score being different from normal. Typically, the statistical threshold for something to be significantly different is ≥ 95%. However, we purposefully chose a less stringent threshold, so that coaches and athletes could be alerted to possible differences (good or bad) as early as possible. This allows for the anticipation of potential problems, in the hopes that they can be avoided. For example, we’re all required to have smoke detectors in our homes, in the hopes that we’ll be able to prevent a fire that could destroy where we live. As you can see, arrow #1 (Figure 1) is pointing to an Enfusion Score that >95% likely to be better than normal. Arrow #2 shows a workout that falls within the range of normal, while Arrow #3 is >90% likely to be a worse than normal performance. As you can also see in Figure 1, Enfusion Score can fluctuate from day-to-day. Though the fluctuation in scores is normal, it can make it difficult to visualize the overall direction of change. To address this, we also perform a trend analysis (Figure 1, yellow line) to show the direction of the overall performance trend.
Does a single day’s score that’s above or below normal suggest a problem? Absolutely not. We recommend coaches and athletes look for persistent change in Enfusion Score (2-4 days) to help identify meaningful performance differences. In the example shown in Figure 1, we can see that Arrow #3 is sandwiched between 2 days where the rider likely performed better than normal. In fact, if we look at the 2-week period between 3/12 and 3/26, the data suggest that the rider that is actually performing quite well. There are 7 rides with >85% chance of being better than normal performances over the 2-weeks, yet only 2 rides are likely to be worse than normal. In contrast, the 2-week period that follows indicates a period of declining performance. Here, over half of the Enfusion Scores are near to or below the 85% likelihood of poor performance. In addition, the overall trendline is also in decline, suggesting fatigue is building and that it might be time to rest. Taken together, this example of tracking Enfusion Score with the Performance Monitoring Chart shows how you can determine when your body is responding well and when it is time to take a break. Enfusion is currently the only available system able to do this.
Stress Balance models do not tolerate missing data well
Because CTL and ATL are rolling averages, missing data can detrimentally alter the final value of each number. Indeed, Dr. Coggan himself has acknowledged that if more than 10% or more of data are missing, it can call the output of the TSB formula into question. For the 6-week chronic window, this means as little as 4-5 missing workouts can cause one to second guess what the model is telling you.
Enfusion can buffer itself from data gaps
As with any model, more data is always better. However, the Enfusion Score is not prone to the same type of problems of missing data, like ATL and CTL. Enfusion Score is based on your personal training history, and once we have sufficient data to calculate your Enfusion Score, we can calculate your score with limited impact from missing data, as will be demonstrated below. Of course, it is clearly not in anyone’s best interest to just leave out data. Significant and repeated data gaps (recurring periods of weeks or months) can impede the ability for the algorithm to adapt to your training progress, resulting in suboptimal Enfusion Scores. However, we have had athletes who suffer an injury requiring time off the bike, who were able to resume training with usable data on day 1 of their return and will walk through this example below.
Putting it all together:
In Part 1 of our blog series, we explained how you can use Enfusion’s personalized metrics to track your training with unprecedented precision, as well as to set concrete training objectives. Here in Part 2, we’ve discussed how Enfusion can use machine learning to evaluate your training to determine whether your body is responding to the training or becoming fatigued. With that, we wanted to show an example that demonstrates how you can use Enfusion Score and the Performance Monitoring Chart to distinguish good versus bad performances and how you can monitor recovery from injury or illness. In Figure 2, we show data from an amateur road racer who experienced all the highs and lows one can expect to experience in a year’s training.
Good Form: On the left, the Performance Monitoring Chart (June, 2016 through mid-July, 2016), we see the athlete is in good form. The majority of the Enfusion scores are within the normal range (white area), yet several Enfusion Scores are in the upper bound of the chart corresponding to >90% and even >95% chance of being better than normal. You can also see that only 2 Enfusion Scores are in the red area of the chart ( >85% chance of being worse than normal), which are weeks apart. During this period of time, we can also see that the overall performance trend (yellow line) is high.
Declining Form: However in mid-July, we can begin to see the overall performance trend start to decrease as we move toward August, 2016. Though the athlete is still having good performances during this period, there are an increasing number of Enfusion Scores in the red, and less in the green. By the time we get to early August, the athlete’s Enfusion Scores teeter between normal and worse than normal, with no Enfusion Scores in the range of “better than normal”. This suggests that the rider is becoming fatigued, as the body has lost the ability to respond to racing/training. At this point, the athlete was prescribed a “rest and build” period prior to the next block of fall racing. As the athlete recovers, the Enfusion Scores return to within normal range. Subsequently, the overall performance trend begins to increase, and the athlete starts to produce some Enfusion Scores with high likelihood of being better than normal.
Injury: Unfortunately, in late October the rider sustained a serious injury that required weeks off the bike. Recall as we discussed earlier, once we have sufficient training history to calculate Enfusion Scores, we are buffered against occasional data gaps — even a gap that is close to 1-month long! Because of this, the rider is able to begin collecting valid data once training is resumed. As shown in Figure 2 (second yellow arrow) the first ride back went poorly. This was in large part due to accommodation of the injury. However, as training continues, performance returns to within normal range. By January, the athlete appears to be responding to the training, as the frequency of Enfusion Scores in the green increases. Interestingly, the overall performance trend through February is largely flat. This could be due to inconsistent performances as a consequence of injury recovery mixed with hard training, since the athlete seems to be having a similar number of “better” versus “worse” than normal Enfusion Scores. In total, we can see that after having established the ability to report Enfusion Scores, the athlete was able to suffer an injury requiring close to 1-month of the bike and immediately resume tracking the quality of performance on the first day of training.
Illness: Despite having worked hard to rebuild from the injury, the athlete was again forced off the bike due to the flu (rightmost yellow arrow). Following a few days of no riding, training resume in late March. After a week of taking it easy, we see that the athlete’s score are consistently within the normal range. Upon resumption of regular training, the range of Enfusion Scores is similar to before the illness, suggesting the athlete has allowed for proper recovery. Hopefully, from here the bad luck is over, and it’s nothing but smooth sailing from here out.
Together with Enfusion’s suite of personalized metrics, Enfusion Score and the Performance Monitoring Chart give you a more complete picture of your training than just looking at raw numbers. If you broke a personal record, that would be a great performance. But did you have your best day? Is there more in the tank? On the other hand, it can be equally difficult to decide when to rest. How can you know that you’ve trained to your limit without going too far into the red? Enfusion is the next-generation methodology that can help you answer all of these questions and make data-driven decisions.
As a final important note, users should know that each athlete’s range of Enfusion Scores is unique to him/her. Therefore, it is not possible to make direct comparison of Enfusion Scores between different athletes, as we’re all different. However, it is possible to make other important comparisons, such as the likelihood that a given performance is better or worse than normal among a group of people, or between the direction of the athletes’ respective fitness trends. We believe this last point is critical, as the Performance Monitoring Chart could serve as an important resource for professional and elite cycling teams. For example, comparison of fitness trend and performance quality across teammates can help teams make data-driven decisions about rider selection prior to major events.
We hope this was helpful. Please look out for upcoming blogs, as we break some of these larger concepts down in and explore them in greater detail. Thanks for reading.
TSB, CTL and ATL are all trademarked terms of TrainingPeaks and PeaksWare, LLC