Are you triathlon training with AI adaptive training?

Triathlon Forum
1

Who here has fully adopted an AI training platform for long course (70.3-140.6) triathlon training? I see info on Humango (bunch of paid promotion on YouTube with GTN), AI Endurance, Athletica, PKRS.AI, 2Peak, Alta Brio, TriQ… I am probably missing some.

So the simple version of AI seems to adjust a training plan to a missed or changed workout. A training plan makes certain assumptions about how high a training load an athlete can handle in any given day, given week, given mesocycle, and how fast the ramp up in training load can happen. The missing link that a coach can provide over a training plan is to assess how the athlete is absorbing the training, are they recovering fully, are the over training, or could they take on more training, and if so, could that additional load be best added by more intensity in a fixed time period (lifetime constraints for many AGers) or by more duration in a more evenhanded 80/20ish intensity distribution.

What I am getting at here is that a key part of an AI adaptive training solution should factor in recovery in the best, most objective way possible. Sure someone’s own perception of how they feel can play an important role, but more and more we are seeing some real data resources that are attempting to track and measure recovery. Sleep tracking because sleep is the greatest contributor to recovery and the absorption of training. Overnight HRV because that measures the health of the autonomic nervous system. Resting heart rate because an increase here also can indicate overtraining or illlness. it would be great if the AI training platform would surface these inputs on recovery and make recommendations for modifications to the training accordingly. Do any of the current AI platforms do this legitimately?

As well, it would be great if the AI platform could look at the data from a workout and tell me if I am getting faster or fitter than I was 30 and 60 days ago, what I could potentially expect in terms of future improvements before race day based on how I have handled the training so far, how I have improved, and most importantly where I am in the training plan and what lies ahead in training. Can the AI engine bring greater insights to my post workout analysis, not just showing that I went farther, faster, with more power, for longer than in the past, but also there in untapped data around how my body handled the effort, was my HR lower for the same effort, or did my HR recover after each interval faster than in the past (normalizing for the training conditions.)

The AI engine could also generate insights across training years that are specific to you. You say this year you want to podium at 70.3 Indian Wells or Ironman California, the AI engine can look at who finished in the top 3 for the last couple of years, assess your strengths, and tell you if you have a chance and what that plan and what those gains (increase TP and FTP by 20%?) are you’d have to make. How would this year’s plan be different than what you did last year?

I am excited to see what gets developed in AI training for triathletes. Have you had a positive personal experience with any of the current offerings? No sales pitches from the companies here please. Just real world experience from AG athletes that were willing to put their faith in an AI platform 100% to get them to the start in line, and then share if it worked out well or not, and how/why.

2

I have a fundamental issue with AI training for high performance athletes. The data set that AI by necessity needs to pull from is not that of top level athletes. The data set is just too small. Furthermore, the training data of top performing athletes is not published readily, so it will need to use either predetermined training plans or make assumptions about the limits of the athlete based on available data from the majority of athletes.

This AI training will probably work very well for someone trying to finish an Ironman, or someone wanting to break 12 or 11 or maybe 10 hours. The margin for error there is much greater, so bad data coming in will just shift the workload within that margin. Someone trying to win big races needs to train much closer to their limit. If AI makes incorrect judgments about recovery 10% of the time that’s 2 missed workouts per week. An athlete/coach can look at HRV and say “that’s not right, I feel great”.

I do see a potential for AI in data collection and processing of high level training groups (looking at you, Norway). It’s possible if not probable that increased computing power can make more sense of 24h/d heart rate, glucose, etc and workout power, HR, lactate data. This is probably in a backwards looking way tho, seeing what training protocols provided gains in specific metabolic areas and how they can be utilized in future training.

Ultimately the body is very complicated, and the AI coach model is only as good as the information it has available. The computerized model of the body is imperfect and the data coming into the model is imperfect. It’s getting better but I’m not sold yet.

3

an area I see potential but like all things needs good data in (or as they say GIGO) would be to look at a person’s data and try to find correlations with specific type of training. In other words those who are responders for a certain type of training vs those things they do not respond well to. It could also determine the amount of rest needed based upon responses as well. The key here is good longitudinal data that is specific to the training and the person. Big data is where AI shines or can shine as well as being able to tease out not easily seen correlational patterns within a specific person’s data. Large Language models could be great for inputting and conversing in a way to be able to get more subjective data into a model that is more machine learning than probabilistic word based.

It will be interesting to see where the AI applications for endurance training go in the next few years.

4

Really excellent point about the biased dataset that the AI model would be working off of. For the higher performance athletes, I think there are a lot more nuances, like you said, that an AI model just more than likely won’t be able to tap into (looking at you qualitative feedback) in a reliable fashion. Definitely agree though that for the masses of people who want to finish an Ironman or 70.3, or be MOP, this could be a great resource and remove the barrier that so many see with coaching. I have dabbled with a few of the trials and honestly feel they have a lot of work to still be done. I guess only time will tell.

5

an area I see potential but like all things needs good data in (or as they say GIGO) would be to look at a person’s data and try to find correlations with specific type of training. In other words those who are responders for a certain type of training vs those things they do not respond well to. It could also determine the amount of rest needed based upon responses as well. The key here is good longitudinal data that is specific to the training and the person. Big data is where AI shines or can shine as well as being able to tease out not easily seen correlational patterns within a specific person’s data. Large Language models could be great for inputting and conversing in a way to be able to get more subjective data into a model that is more machine learning than probabilistic word based.

It will be interesting to see where the AI applications for endurance training go in the next few years.

Yeah that’s what I was thinking would be the next step. Backwards looking validation of methods. Of course with the human body there’s always a placebo and mental effect. If I really like Vo2s and hate tempo work the model may see a greater effect with Vo2s, even though there’s potential gains left on the table. The hope is that big computing can tease out the ideal metabolic profile for an individual person for an individual event. But the model has a ton of catching up to do to get there.

The other issue with this backwards looking model is that the body is dynamic year to year and only has so many years of peak performance. It’s all well and good if AI can look at 10y of training and determine what’s ideal for next year, but at a certain point it just takes too long. I suppose the hope is that it can learn faster and more accurately than a coach.

It’s like what Lionel said in a recent video. It’s great to know the exact lactate and heart rate and effort levels you need to hold, but you need to be going fast enough first. This is where AI could also lack. Sometimes you just need to nut up and go really hard.

6

an area I see potential but like all things needs good data in (or as they say GIGO) would be to look at a person’s data and try to find correlations with specific type of training. In other words those who are responders for a certain type of training vs those things they do not respond well to. It could also determine the amount of rest needed based upon responses as well. The key here is good longitudinal data that is specific to the training and the person. Big data is where AI shines or can shine as well as being able to tease out not easily seen correlational patterns within a specific person’s data. Large Language models could be great for inputting and conversing in a way to be able to get more subjective data into a model that is more machine learning than probabilistic word based.

It will be interesting to see where the AI applications for endurance training go in the next few years.

Yeah that’s what I was thinking would be the next step. Backwards looking validation of methods. Of course with the human body there’s always a placebo and mental effect. If I really like Vo2s and hate tempo work the model may see a greater effect with Vo2s, even though there’s potential gains left on the table. The hope is that big computing can tease out the ideal metabolic profile for an individual person for an individual event. But the model has a ton of catching up to do to get there.

The other issue with this backwards looking model is that the body is dynamic year to year and only has so many years of peak performance. It’s all well and good if AI can look at 10y of training and determine what’s ideal for next year, but at a certain point it just takes too long. I suppose the hope is that it can learn faster and more accurately than a coach.

It’s like what Lionel said in a recent video. It’s great to know the exact lactate and heart rate and effort levels you need to hold, but you need to be going fast enough first. This is where AI could also lack. Sometimes you just need to nut up and go really hard.
totally agree, and also am in the time of life where the body and mind no longer are in sync… But if the input data both objective and subjective had the right (and I stress the right) depth it would so cool to see what AI models could do. It is sort of like what I heard said recently about weather predicting, AI is already in the mix and it is getting better and better all the time and really is showing promise.

7

ey There,

full disclosure before I give my two cents. I am Athletica.ai ambassador and I help Athletica.ai to develop their product and I’m also co-hosting their new podcast The Athlete’s Compass. Sure, you could now completely bypass my reply as you said “no sales pitches from companies here please”. If you could bear with me for a while and I will share my journey which I believe can resonate with many age-group athletes.

Long story short: I did my first full Ironman 2019 (3rd place, was hooked, wanted to qualify for Kona) and was hooked. I went on to crushing myself in an environment that added more stress into already stressed out body and mind. Yes, I had a wonderful coach who emphasized Health First. I measure HRV, cut sleep short to get more hours into the day, and ignored ALL the warning signs that my body and mind was sending me. All of this ended in exhausted, dysfunctional HPA axis with a long list of weird symptoms, illnesses and injuries, and finally overtraining diagnosis and detraining. I was napping twice a day, sobbing on the coach and wondering if my triathlon and sporting career was over.

I sat with that feeling for a while. Allowed some self-pity and decided to take one day at a time and see what happen. 6 months later, I was allowed to come back to training.

I found Professor Paul Laursen and Athletica. He offered to coach me using Athletica, watching me closely on the background.

3 short months later, I snatched 70.3 IM podium and ticket to the World Champs. I had also full IM in the schedule 4 months later. Due to pollution in Dubai read environmental stress that body cannot differentiate I struggled with health once again. Did more yoga, walking and meditation than IM training. But again, I snatched podium (3) and ticket to WC(had to decline).

Needless to say my results speak for themselves. But let me give you my thoughts on your ideas.

You can have excellent coach and you can have an excellent training program from a said coach or AI platform. AND you can still blow it up if you as an athlete don’t develop the skill and willingness to trust your body, trust your FEEL. Relying too much on your coach (who may have as many as 30 athletes in their team) to pick up warning signals for you. YOU must do it and communicate it effectively to your coach. It is a relationship where communication is a two-way street.

Over-reliance of gagdets (like garmin watch measuring your overnight HRV etc) is a factor you have to be aware of. For example, Garmin has been completely off with its’ recommendations of taking it easy when my FEEL has been energetic, powerful, ready to go, and vice versa. Unfortunately the optical heart rate sensors are not quite there yet that they measure HR reliably. And I am not willing to wear chest strap 24hr!
My take: The human FEEL is still the most reliable way to gauge recovery status. Motivation, mood, your sex drive (mojo), overall energy all can signal your readiness to train. Do you know a tool can reliably measure these factors?

Integration to AI platforms are dependent on many factors that I am not even aware of - I leave that part to the experts. But it is coming. Just remember these things are reliant on how well do they truly measure HR and HRV? Imagine the problems it could potentially cause if they are not, in fact, measuring what they’re supposed to. Imagine if not-reliable overnight HRV measurement indicates excellent recovery and AI training plan sets you up with a hard HIIT session?

Athletica has indeed invisible threshold detection, whenever you do 20min HARD effort it calculates your FTP so when you go harder than last time you tested (you start with Test week when you start with Athletica) and the same with 5K run. So yes, as you improve, your thresholds also change, you get notified. Athletica also has performance power/pace profiles that gets updated so you can keep an eye on your overall improvements and can use the power/pace curves in your race pace planning. You can also geek out on the scientific principles that go into creating these profiles with their education systems. Everything is there for you to learn.

Wouldn’t it be wonderful to be able to state what you have to do to get to podium? Thankfully human body and mind are intrigued systems and you can never be 100% how you perform on the day, yet alone how the top 3 will perform, or who they are. I’ve seen it from my own training that it is nearly impossible to say how the race day will unfold only based on training data. Sure you can have indications but race day is a party day and podiums are never to be taken for granted.

I’ve tested a few other platforms and quickly noticed that they are not based on sports science. I have MSc in Sports Science. Often times I noticed there is way too much intensity and not enough Zone 1&2. Tempo after tempo, Threshold after threshold. Be careful on putting too much stress in your overall bucket. You can only handle so much. When you lack zone 1-2, you’ll quickly start falling apart, with illness and injuries, and if you don’t have the intuitive feel and trust in yourself, you’ll end up in the “overtraining train”. The Norwegian Method is all the hype nowadays, but what many people ignore is the fact that these boys have developed a HUGE AEROBIC BASE over the years with low intensity training (in fact, this started already in their childhood with walking&biking to school, playing outside with friends, being active in the mountains with family which is THE NORDIC WAY of living) which most of the people these days Lack. So thinking everyone can do a lot of zone 3&4 work is fundamentally flawed. And that’s what is so wrong with many of the current platforms. Athletica has a lot of zone 2 training paired with some really fun HIIT sessions. As you start your build phases, then you start training more around your target race pace/power intensity. It has worked wonders with me, but I have learned the hard way to keep close watch on my own signals and I trust myself to alter the plan when I need. BTW this is easy to do with Athletica with their Workout Wizard function - which gives you alternative options based on time available, need for variety, in case of illness/injury…

One more thing about the learning if you are doing better than previously. Athletica has a function called Workout Reserve which takes into account your last 6 weeks of data and it compares your current activity- it then draws a graph in your workout analysis to show “how deep you went” compared to last 6 weeks, indicating how much “battery” you had in the session. You go on negative, and you’ve gone faster/harder than previously, ultimately indicating improved performance. It was developed by a leading data scientist Andrea Zignoli (look him up and his work). IT is really cool to see as this often correlates well when you start feeling pretty drained. (I hear little birds singing that this function may be available “in-real-time” in the near future.

Athletica also adjust the plan overnight, which is great. IF you miss a session or two, it will recalculate the load in the near future, so you arrive on race day still in your top shape (unless you take way too many days off- because there’s only so much any AI plan can do if you are not doing what you’re supposed to-train! lol). i’ve been on that wagon as well…

I know I’ve gone for way too long. Thanks for reading this far. This is my real world experience. I know I am biased but I’ve been using Athletica for 2 years and I can confidently say I know the platform in&out and can speak to most of the functions. Yes, the interphase needs some tender loving care, and an app could be useful, but I’ve got shortcut saved to my phone so it basically works like an app for me at least.

Happy to discuss more :smiley:
Marjaana Rakai
aka Tired Mom Runs