Optimizing athletic performance and recovery using machine learning at WHOOP

WHOOP, a start-up founded out of the Harvard iLab, uses machine learning to improve athletic performance and possibly the lives of diabetes patients.

While the technology in wearable health trackers such as Fitbit, Apple Watch, and Garmin devices are becoming a commodity, a Boston based-startup named WHOOP, founded out of the Harvard iLab, is changing the game. WHOOP offers a device targeted at athletes to monitor strain, recovery, and sleep using proprietary algorithms based on Heart Rate Variability (HRV), the variation in time between each heartbeat, as well as four other variables tracked 100 times per second [1].


Improving wrist-based heart rate measurement

While chest straps have been able to provide more accurate HR data for decades, getting this information from a device that people will wear 24/7 is crucial to answer meaningful questions for optimizing performance. Wrist-based devices have exhibited a less accurate heart rate (HR) during physical activity [2]. However, WHOOP claims to use machine learning with 250 features, including accelerometer and other physiological data, improve upon its measurement of your heart rate [3].

Figure 1: Correlation matrix of 250 features to estimate heart rate [3].


HRV and training strategy

Monitoring HRV is a well-known strategy to measure recovery and prescribe high/low intensity training or rest [4]. WHOOP has implemented supervised learning to create metrics for strain (training load) and recovery (readiness for stress) based on HRV and determined an optimal level of stress to optimize performance given an athlete’s level of recovery [5].

Figure 2: Expected change in HRV based on recovery and strain of workout [5].
Delivering insights through machine learning is the main value proposition for WHOOP. A device collects over 100MB of data per day to better understand individuals or teams of athletes. In the short term, WHOOP is focused on these performance optimization questions such as “is it better for me to work out in the morning or before bed?” or “how should I schedule my workouts around athletic events?” [6].


A foray into healthcare?

To begin to tackle these questions, WHOOP partnered with the Korey Stringer Institute (KSI) to track 40 UCONN athletes over eight months during the 2016-2017 season. In addition to the WHOOP data, KSI was able to collect athlete demographic data, blood-biomarkers, training and competition loads, and fitness and hydration status [7]. WHOOP hopes to demonstrate correlation between its device metrics and athlete performance data to offer better insights to its customers.

Over the medium term, WHOOP will try to use its data on recovery to improve health outcomes outside of athlete performance. WHOOP is participating in a study with Evidation and Tidepool to explore the linkage between nocturnal hypoglycemia, next-day behavior, sleep patterns, and heart rates [7,8]. Additionally, one of WHOOP’s users is Ryan Reed, a NASCAR Driver with Diabetes, whose diabetes management has improved through the use of WHOOP to encourage proper recovery [9]. If WHOOP continues to collect more demographic and performance data, they can keep feeding this back into their product development process.


Publish some papers already!

WHOOP shows promise as a standard device that health researchers can utilize to measure stress and recovery in longitudinal studies. However, WHOOP first needs to validate the accuracy and precision of its device and algorithms before it can be trusted by comparing the estimated HR with that measured by an ECG (such as described in [2]).

Secondly, the problems that WHOOP is trying to solve are complex and have so many variables that they should continue to grow their user base and collect additional data. They have demonstrated that the product is sticky, so I would add surveys and other integrations (such as MyFitnessPal for nutrition data) to collect relevant auxiliary data to improve the models. WHOOP could also open up their platform to allow physiologists to run opt-in studies on users to crowdsource innovation. This may be necessary especially because WHOOP has failed to publish results from their study a couple years ago.

Lastly, WHOOP should invest more in unsupervised learning. If WHOOP is truly measuring valuable signals, this could allow for some less intuitive and more interesting insights. For example, WHOOP may be able to predict when a user is beginning to get sick and what exercise or sleep regimen he or she could implement to heal quickest. Currently these findings are coming through intentional studies such as when the CEO discovered that if he did a really short workout before a flight that involves jetlag, he would have a better recovery the next day.


The future of WHOOP

Given the recent shift from a standalone unit cost to a subscription model, WHOOP’s core competency seems to be less about their patented device than the analytics platform. Despite this, should WHOOP open up its platform and data and if so, to who? Also, what do you think about WHOOP’s choice to use supervised vs. unsupervised learning.


(787 words)



[1] Goode, L., Goode, L., Ceres, P., Barrett, B., Ceres, P., Perlmutter, K., Ceres, P. and Staff, W. (2018). Here’s One Way to Keep Wearables on Wrists: Subscriptions. [online] WIRED. Available at: https://www.wired.com/story/whoop-wearable-subscription/ [Accessed 13 Nov. 2018].

[2] Wang, R., Blackburn, G., Desai, M., Phelan, D., Gillinov, L., Houghtaling, P. and Gillinov, M. (2017). Accuracy of Wrist-Worn Heart Rate Monitors. JAMA Cardiology, 2(1), p.104.

[3] WHOOP. (2018). Improving Heart Rate Accuracy: Your WHOOP is Getting Smarter! – WHOOP. [online] Available at: https://www.whoop.com/the-locker/improving-heart-rate-accuracy-whoop-getting-smarter/ [Accessed 13 Nov. 2018].

[4] KIVINIEMI, A., HAUTALA, A., KINNUNEN, H., NISSILÄ, J., VIRTANEN, P., KARJALAINEN, J. and TULPPO, M. (2010). Daily Exercise Prescription on the Basis of HR Variability among Men and Women. Medicine & Science in Sports & Exercise, 42(7), pp.1355-1363.

[5] Capodilupo, E. and Lee, T. (2018). TRAINING WITH WHOOP: USING RECOVERY AND STRAIN TO UNLOCK YOUR POTENTIAL. [online] Whoop.com. Available at: https://www.whoop.com/wp-content/uploads/2018/08/180806_whoop_training_with_whoop.pdf [Accessed 13 Nov. 2018].

[6] CAPODILUPO, J. (2018). The CTO and the Analysis of the Human Body – WHOOP. [online] WHOOP. Available at: https://www.whoop.com/the-locker/the-cto/ [Accessed 13 Nov. 2018].

[7] Businesswire.com. (2018). Business Wire. [online] Available at: https://www.businesswire.com/news/home/20170731005519/en/WHOOP-Korey-Stringer-Institute-Conduct-Largest-Athlete [Accessed 13 Nov. 2018].

[8] Draper, S. (2018). Evidation Health Teams Up with Tidepool to Use Connected Devices for Studying Sleep, Type 1 diabetes. [online] Wearable Technologies. Available at: https://www.wearable-technologies.com/2018/08/evidation-health-teams-up-with-tidepool-to-use-connected-devices-for-studying-sleep-type-1-diabetes/ [Accessed 13 Nov. 2018].

[9] Evidation. (2018). Evidation Health, Tidepool Partner to Use the Data Generated Every Day by People with Diabetes to Improve Clinical Research – Evidation. [online] Available at: https://evidation.com/news/evidation-health-tidepool-partner/ [Accessed 13 Nov. 2018].

[10] Van Deusen, M. (2018). WHOOP helps NACSAR driver Ryan Reed train, and manage his diabetes. [online] WHOOP. Available at: https://www.whoop.com/the-locker/9-questions-ryan-reed-nascar-driver-diabetes/ [Accessed 13 Nov. 2018].

[11] PATRICK PULLEN, J. (2018). Why Professional Athletes Love This Fitness Band. [online] Available at: http://time.com/4744459/whoop-strap-fitness-tracker-band/ [Accessed 13 Nov. 2018].


“Just let the Machines do the Work”: Marriot’s Opportunity to Build on Starwood’s Digital Expertise


Open Research: Advancing Innovation in Public Health

Student comments on Optimizing athletic performance and recovery using machine learning at WHOOP

  1. Very interesting product, but I’m extremely skeptical. This seems like a higher end fitbit, and the transition from a standalone device to subscription model is a move that every technology company in the world is trying to do. Maybe I need to do some more research into the product and tracking capabilities, but sleep, heart rate, and steps are some of only thousands of factors that would influence my performance. For example, diet is likely a huge factor which would be almost impossible to track unless manually inputted. Blood work would also be another important factor. Mental stress and fatigue would also be factors which wouldn’t necessarily be picked up in the criteria the Whoop tracks.

    To answer your question, I completely agree that Whoop should open up its data to allow doctors and physicians to test and use the data to see if there is any statistically significant information in terms of predicting sickness or providing strain on the body. As to this, there should be a combination of both supervised and unsupervised learning. Supervised learning would come in the form of doctors and physicians doing tests as mentioned previously. Unsupervised learning should come in the form of machine learning and using the data to make predictive estimates of fatigue, sickness, and strain on the human body and tracking that against actual results. Lot of exciting potential here, just skeptical given the wave of wearable devices that have come before which have largely fallen short of expectations.

    1. I am a big believer in Whoop’s value proposition for individuals based on the information you laid out above. However, I do have concerns as it relates to the sensitivity of the data they collect. Even before they decide whether they should “open up” their data, how does the company deal with the issue of sharing an athlete’s personal data with teams? Would an athlete want his or her data to be shared with a team ahead of a big contract negotiation? Further, who would have access to an individual’s data outside of a team context?

  2. It seems like there’s a fine line between fitness insights and medical advice that WHOOP will need to pay attention to. In the example you mention about possibly offering guidance on the best rest regiment for recovery, how will WHOOP handle potential backlash from people who feel the advice they received wasn’t useful?

    And on the topic of potentially sending out surveys to customers to augment their current data set, I wonder if there are risks if people check the wrong box or overestimate a metric. As this data will go into the machine learning algorithms, poorly written questions may yield data that would be hard to unwind from the ML algorithms after the fact.

  3. This is a very insightful article into the use of machine learning in wearable healthcare technologies. In reading this article, the questions that are presented around the future of WHOOP is very thought provoking. I agree that their unique and innovative platform is their true strength and product differentiator. However, I do not this that this should be open sourced and a platform that is shared publicly. This type of machine learning algorithm is what set’s WHOOP apart and, as mentioned in the article, until this is validated and deemed precise this should be kept in-house. Likewise, I believe their choice to use supervised learning is very smart given the limited variation and consistency within the measured attributes although there are numerous variables. The data will tend to be structured and in turn supervised learning will be able to adapt and leverage this data in a more dynamic and streamlined fashion relative to unsupervised.

  4. I agree that WHOOP should begin to release its data publicly, perhaps in a controlled fashion to credible researchers. In my opinion, the company’s primary competitive advantage is in its proprietary algorithms that translate a wearer’s health data into meaningful and informative metrics like “Strain” and “Recovery” — not in the quantity or quality of the data collection itself. Thus, WHOOP could benefit from releasing these data to a broader audience in order to gather valuable external feedback how accurate their predictive claims on health measures like “Strain” and “Recovery” are, without compromising their competitive advantage in the marketplace. And in turn, the company could work to refine its algorithms to increase their predictive power on health outcomes.

  5. WHOOP has some very interesting potential. I’m skeptical of the healthcare benefits but extremely intrigued by the commercialization potential of the data and the associated analytics. Imagine a world where athletes could commercialize their performance data. Concerned about your Fantasy Lineup for the weekend? What does WHOOP say about the recovery rate of each of your players. Want to bet on a team to make a deep run into the playoffs? What does the WHOOP data say about their endurance level and recovery programs. With this in mind, I think WHOOP should be very careful about outsourcing its innovation and data. Instead, it should look to partner with major athletics leagues, star athletes, and other professional sports to keep both its data collection and analytics proprietary – those two facets provide tremendous value to the company.

  6. In my perspective, one of the main challenges that WHOOP needs to overcome in order to be successful is to define a broad target market. Although this product seems highly technical and suitable for high-performance athletes, it would be beneficial for it to also become appalling to regular people that practice sport. In this way, WHOPP would be able to substantially improve its amount of data gathering, which will in turn feed the overall system. As many other projects that involve machine learning, the amount of data is crucial to offer a useful and reliable product.

  7. It seems like WHOOP is only focus on one aspect of the performance measurement – optimize window of training. A function can become an add on function to other wearable product in the market. With so many fitness wearable product in the market, WHOOP should think about its market position and how to debut itself. On anther perspective, WHOOP is potential diabetes wearable device seems much more prominent and interesting. I think WHOOP should more focus on the healthcare aspect of its innovation and become first mover in this market.
    Thanks for sharing.

  8. I agree that wearables have become somewhat of a commodity, and although they have garnered relative success, they still don’t possess the functionality to drive widespread adoption. WHOOP, however, seems like it might be a game-changer, given it is based in machine learning and thus has higher potential for functionality. This is especially interesting given the broader movement toward a healthier, more active lifestyle. As it seems like this product would warrant a high sticker price, I think the initial target would be high-performance athletes or those with health issues – e.g. diabetes. However, once initial adoption takes place, I could see word spreading to a more general population that recognizes the higher functionality / reliability of the product.

  9. Curious to what degree the WHOOP platform would benefit from a standardized baselining workout in order to provide more reliable information. It seems to me that having standardized workouts that can be tracked and compared would provide a better baseline for athletic performance. From there individuals can be categorized based on performance and then assigned more meaningful steps to improve.

  10. Thanks for sharing. One question I had when reading this is the success of the subscription model if (A) They have no academic papers to prove its efficacy and (B) as pointed out in other posts, whether they are really differentiated from the Fitbits of the world. One ideal I had is for Whoop to integrate it’s proprietary hardware with more outcomes focused software programs. For example, could it partner with virtual reality training programs? One company called STRIVR offers immersive training for athletes, signing big users such as the NFL. Professionals may be more willing to shell out a subscription for a premium product, especially if it integrates with what they are already using.

  11. Very exciting concept. Though it sounds like they are utilizing unique metrics to improve athlete performance, I am curious how the company plans to build a mote around these metrics to protect themselves from the much larger competitors (Nike, Fitbit, etc.) from incorporating similar metrics into their wearable devices. Additionally, Nike and Fitbit have a meaningful competitive advantage attributable to the data they collect from millions of customers. Whoop is in a situation where they need to quickly scale their proprietary databases, which is not a simple task.

    I am hopeful that the Company can continue to use machine learning and AI to refine their metrics and out-compete their larger rivals.

  12. Personally, I would love to have a more nuanced view of my health using a data-driven approach that is customizable. I think tying the physiological data to other types of inputs like diet, stress levels, and other user inputs could produce interesting insights. Given that the metrics WHOOP is tracking don’t seem particularly novel (there are already a myriad of wearables that track similar health metrics), I do think that the analytics side of WHOOP will be an important differentiator, if they can provide interesting ways to combine and interpret user data. It seems the company is trying to position itself as a serious medical tool rather than appealing to the more casual Fitbit type of user. To that end, I wonder if their platform is rigorous enough to merit engagement from the medical community – it will be interesting to see if doctors will actually support this tool as a way to monitor conditions like diabetes.

  13. I love this topic! The study of health, activity, and utilizing data to optimize performance brings to mind so many ways that society can benefit and the general population can improve their health (and disease prevention) through activity tracking. The anecdote about the CEO determining how to fight jet lag through personal experience exemplifies the need for this technology to be applied to the broader masses for unsupervised learning beyond professional athletes. Clearly athletes are a perfect market to start with: they have required (sporting) activity and training, and closely track performance. However it seems clear to me that if the general population starts tracking daily life with the WHOOP technology, questions such as: What are the physical signs one may notice leading up to a heart attack? When is a person getting “enough” physical activity to prevent coronary disease? How many steps a day are needed for a unique individual to stay fit? can be answered. I see this technology revolutionizing the way that people think about health. Right now the general belief and stated guidelines are to exercise plenty, but the inability for doctors or practitioners to truly qualify how much and what type is “correct” remains a huge open question for society that WHOOP technology seems poised to solve. I see this adding years to many individuals lives but correcting their activity patterns to optimize health!

  14. Very cool! I find this tech fascinating, especially since they seem to be a rare bird in the fitness tech space by seeking to leverage other hardware’s data rather than build their own. It simultaneously lowers their R&D costs while also reducing concerns that users will feel that their fitness tech experience is too fragmented across platforms.

    It also seems like an issue to me though; if they’re dependent on data from other platforms, a) they’re vulnerable to being shut out, b) especially in the case where Fitbit or another hardware company start to compete in the analytics space as well, or start to pursue horizontal integrations (such as the one you suggested, with MFP). An interesting line to toe 🙂

Leave a comment