Biohacking the (Aging) Distance Runner’s Body with Artificial Intelligence

In this part of our CampIO blog series, Carmen Fontana shares how she biohacked the runner’s body using artificial intelligence and IoT wearables.

Part of our CampIO 2020 series.

In my head, I am still the runner I was twenty years ago: collegiate cross-country athlete, piling on the mileage and PRs effortlessly, nary an ache or pain.

My current body tells a different story: middle-age mom, slowing at a humbling rate, rickety and run-down.

When the legs lose their pep, what’s an aging distance runner to do? Step 1: Wallow in self-pity. Step 2: Embark in a multi-year biohacking journey, using IoT wearables and custom AI models.

Disclaimer: Your Mileage May Vary

Before I get too far ahead of myself, let’s start with a few disclaimers. Please, please, please consider the following before taking any of my biohacks seriously.

1. N=1. I customized the methods outlined below and their results to the data of a forty-something woman whose name is Carmen Fontana. If that is not you, then your mileage may vary.
2. I am a technology consultant. I am not a doctor, physical therapist, nutritionist, or running coach. Heed accordingly.
3. That said, those people are really smart, and at the end of the day, I certainly weigh their informed opinions ahead of a computer-generated model built in my spare time. I used the machine learning models to guide my biohacking, but always validated approaches against universally accepted training tactics. If I thought the medical people would find an approach was stupid, I wouldn’t do it.

Ok, disclaimers complete. Let’s get to talking about the fun stuff – biohacking running performance!

How To Run Faster

When I first started running, I thought the key to success was just to keep running faster and longer. That worked at first, but as I evolved from beginner to intermediate runner, my progress stalled out. Crossing finish lines quickly is a bit more complicated, as it turns out.

Several factors contribute to running faster:

• Training Load – This is your mileage’s quantity and intensity, both in individual workouts and cumulative over a training block.
• Nutrition – This means all things food and drink: quality and quantity of calories consumed, hydration, macronutrients and micronutrients.
• Recovery – Sleep is the primary driver in recovering from training loads, but things like stretching and hot and cold therapy can also contribute.
• Mental – This requires tricking your brain into doing incredibly painful things over and over. (People have written shelves upon shelves of awesome books on training the running mind, so I won’t even attempt to cover in this blog post. But this one and this one are two of my favorites).

All of these components intertwine, so to run faster, you need to optimize each of them. However, the human body is notoriously fickle, so no one perfect approach works for every single runner. Hence, the lucrative running technology industry, parting runners from their money for nearly thirty years.

The Evolution of Running Technology

• Early 1990s – As a high school cross-country runner living in a two stoplight Ohio town, my access to running technology was limited at best. I did have a cheap Timex and two fingers, allowing me to take my wrist pulse manually and roughly gauge recovery between those nasty 400m intervals my scholastic coach loved. GPS had not transitioned from the military into civilian life, so the only way to measure my running route distances was to have my dad drive me around in our minivan, noting the odometer’s mileage.
• Late 1990s – My collegiate coach somehow convinced the notoriously stingy athletic department to purchase Polar chest strap heart monitors. These were the early days of the technology, and they required much finesse to get them to work properly (including, I kid you not, licking the sensors with your tongue). Polar apparently only sized them for middle-aged men, and not for skinny 19-year women, so my teammates and I were constantly repositioning them in attempts to keep them from sliding off our rib cages. Despite their awkwardness, they were a serious upgrade from manually checking our pulse in workouts to determine training intensity.
• 2000s – GPS finally makes it off the battlefields and onto the wrists of runners everywhere. Never mind that the first GPS watches were clunky facsimiles of an Inspector Gadget watch, we were just giddy to know the exact distance we had run so far, and more importantly, how far we still needed to run.
• 2010s – GPS watches slimmed down, and finally becoming presentable to wear in public. More importantly, wrist-based heart rate sensors became mainstream. Tucked discreetly underneath the watch face, they tracked heart rate 24×7 without the need for a cumbersome chest strap.

As mentioned earlier, proper recovery is one of the tried and true methods of running faster over time. Since the days when I manually captured my heart rate with two fingers, runners would check their pulse first thing in the morning upon wakening. Resting pulse is a decent indicator of whether your body has recovered from previous runs. A few beats above your norm? You probably should take a rest day. At or below your baseline? Time to go hard!

The one flaw with this system is that it is reactive. If you wake up and see a heart rate indicating poor recovery, all you can do is take an easier run that day and lament the fact you should have gotten more sleep the night before. Knowing a bit about artificial intelligence, my wheels started turning. What if I could use the new 24/7 heart rate tracking technology to build a proactive system — one that predicts tomorrow’s recovery today? That way, if it predicts I’ll be in bad shape tomorrow, I can take steps to improve recovery today, like sleeping more.

I’m Going to Be Crazy Rich, aka My First Running AI Model

So, I did what every self-respecting techie runner would do — I built a machine learning model to process historical data from my running watch. Each day was an input record, including data such as mileage run, time spent in various heart rate zones, and hours of sleep.

The model was able to learn from the historical data better than I expected – I was able to predict tomorrow’s resting heart rate based on today’s activities with an accuracy of plus or minus one heartbeat.

This was huge. Every aging distance runner I knew was trying to figure out how to recover better to run faster. And, I had developed an algorithm that could tell you how much sleep you needed to do just that.

I was going to be rich.

I Did Not Become Crazy Rich, But I Did Get Smarter

As I was researching how to bring my amazing technology to market to make my millions, I stumbled upon a new wearable device called Whoop.

Whoop uses wrist-based heart rate monitoring, much like Fitbit and Garmin. However, Whoop took it one step farther and tracked heart rate variability or HRV. I am not going dive into the medical details here (once again, I am not a doctor), but the short story is HRV is more sensitive than resting heart rate in determining body strain. Whoop combined HRV and RHR measurements, as well as your daily activity, to predict how much sleep you need for peak performance tomorrow.

So basically, Whoop had already invented my invention and did it with better data.

I quickly got over my disappointment and started reveling in the information it provided me. Thanks to a daily diary feature, I could track how different habits impacted my ability to recover from training.

Here are some of the things I learned:

• My brain matters more than my body – having a sense of purpose, feeling in control, and efficacy (skills mastery) positively impacted my recovery more than anything else.
• For every 30 minutes I read in bed, my recovery increased seven percent.
• Less alcohol is better. No alcohol is best (Whoop can be a killjoy).
• My body responds best to a plant-based diet, and worse when I eat meat.
• Proper hydration and stretching also have positive impacts.

None of these insights were groundbreaking, yet seeing them quantified reinforced positive behaviors. In particular, I was interested in the data on my diet. I have been gradually going plant-based in the last few years, primarily for environmental reasons. It was encouraging to see I was also helping my body, in addition to the planet, with my food choices.

Food has always been a weakness when it came to my running regime, and I was motivated by the positive changes occurring thanks to this data-driven approach.

My Second Running AI Model

Thanks to my wearable devices, Garmin and Whoop, I had dialed in my training load, recovery, and mental health pretty well. And I had indicators about which overall diet composition worked for me. But could I go even deeper, using machine learning to determine ideal macronutrients and micronutrients in my diet?

(Once again, I am not a doctor, so here is a great explainer on what those are and why they are important to runners.)

I had been capturing my daily eats via the MyFitnessPal app for months. That, paired with my Whoop recovery information, became my training data for my newly created machine learning model.

In other words, using my historical data, the model would learn how today’s diet decisions would affect tomorrow’s recovery.

Technical details:

• The MyFitnessPal inputs (or “features” as we call them in machine learning speak) included daily macronutrients (percent of consumed calories from fat, protein, and carbohydrates) and micronutrients (vitamin C, cholesterol, monounsaturated fat, and so on).
• I used The next day’s Whoop recovery data as the optimized variable.
• I used Azure Machine Learning Studio to quickly stand up a model that tested five algorithms in parallel. I am a big fan of this low-code tool, as it allows you to drag and drop key machine learning components allowing you to create and change models with ease.
• I was able to test five algorithms in parallel, as a result, using Root Mean Square Error as the judge of algorithm accuracy.
• The algorithms I tested were: Linear Regression, Bayesian Linear Regression, Boosted Decision TreeRregression, Decision Forest Regression, and Neural Network Regression.
• I updated the training data set every week or so to incorporate my latest nutrition and recovery data. Over a dozen or so model executions, Decision Forest Regression tended to be the best performer.
• From there, I used the “Permutation Feature Import” functionality to rank the features based on their model influence.
• I identified the top three features, and then I calculated the target metric based on the top quartile of my recovery data.
• MyFitnessPal allows you to track three micronutrients on the home screen, so it was easy to track my daily progress against the goal.

Among the things I learned:

• Protein is overrated, at least for me. I consistently had better recoveries when my diet was about 15 percent calories from protein – far lower than the recommended 30 plus percent by some diets such as Keto.
• My most important micronutrients were more variable, perhaps because I modified my diet based on the model runs on a near-weekly basis over the experiment period.
• That said, fiber proved to be the most important micronutrient for me. Thanks to a plant-based diet, I already consumed a larger than the normal amount. But when I did not consume enough fiber, it showed in my recovery results.
• I also needed other important micronutrients, including calcium (I need a bit more than the recommended USDA daily amount) and sodium (I need much less — 1600mg versus the recommended 2500mg).
• Sugar did not have nearly the influence I thought it would. Hallelujah! This chocolate addict needs her daily ration.
• Calorie differential is important. My body was happier when the delta between calories consumed and calories expended was small. You can diet, or you can run fast, but trying to do both concurrently put too much strain on the body.

Biohacking: Results for a Rickety Runner

Since my high school days, I have been informally biohacking when I used my two fingers to gauge recovery between intervals at high school track practice. However, my efforts became more sophisticated in the last few years as I introduced custom-developed machine learning and IoT wearables into my training.

Unfortunately, due to COVID-19, I have not been able to participate in any races to test for running performance improvements. However, I have increased my mileage and broke the Strava top 10 for my hometown trail.

While Strava glory is certainly unquantifiable, I do have the following metrics that reinforce how well my biohacking efforts are paying off:

• 22 percent increase in training intensity. More intensity yields better running performance over time.
• 10 percent increase in heart rate variability. My body is adapting to physical and mental stress better. It is a positive sign that I am adapting to physical stress (i.e., the increase in Training Intensity). But, I am even more pleased that I could exhibit this much of an increase given the mental toil of this tumultuous year.
• Six percent decline in resting heart rate. This demonstrates that my body is recovering better, despite the increase in training volume and intensity.

No one can halt the passing of time. Sadly, I will continue to get older and slower. However, the mindful physical and mental habits acquired from AI and IoT-driven biohacking allow me to age more gracefully.