Spotted this teaser for new AI/ML app coming later this year:
Will be interesting to see how it differs from XSS
the claim is combining
to feed AI/ML algorithms and predict:
For someone in touch with their feelings, I’m skeptical. In agreement with Ivy and Amber’s comments on podcast 354.
Even if OTS is the same as Xert’s XSS, the claim is combining training stress with recovery data to provide guidelines on today’s training.
cool, i only read as far as the blurb about fatigue during the ride affecting tss generation with XSS also includes that TSS doesn’t. I just saw it includes carb intake as well? Will be interesting to see how that all gets taken into account. How precise will you have to be with the carb distribution during a before/after a ride I wonder?
Interesting. Last week’s FasCat podcast talked about their work analyzing data for one of the pro women’s teams. I suspect this is an offshoot.
or it started with Frank’s long-time interest in Whoop and power and work/recovery, and he picked up Dr Skiba to help, and they are developing in parallel with their data scientist role at Human Powered Health pro women team.
I’m sure the story will unfold over time, that blog post was just a teaser.
That seems super interesting. I am very curious to what degree this will augment his training program that he offers. Personally, I find his reliance on TrainingPeak’s CTL, etc. data a bit old school and in need for an update. So this is definitely an interesting development.
I’ve been checking my HRV with HRV4Training for the last two years.
It tracks pretty well with sore legs from prior training or I drank too many glasses of wine the night before.
Still, it doesn’t really tell me much that I don’t know intuitively.
One issue for me is that mental and physical readiness have many dimensions. On some days my legs and body feel very tired, but looking at my heart rate data, I know it is actually doing ok or even well. Other days I am struggling physically, but mentally I can (should?) push through. Depending on where the main weakness lies that day, it could make sense or make no sense at all to e. g. reduce the intensity. (For me personally, if I am having a tough time mentally, reducing the intensity usually breaks me.)
The piece of data that makes the most sense is sleep data: if the athlete hasn’t slept enough, I think it’d be good to remind athletes of catching enough sleep independently of whether that has any effect on the workouts.
I think the real problem with all of these approaches will be getting enough people who will be willing to alter their training based on what the AIML model says to validate if it improves performance or not. That is : you will need a large group who will ignore if they are feeling good, but the model says to back off, to actually back off (or vice versa) of their training, versus a control group. And by large, you really need hundreds or thousands.
This is actually one place I’m surprised that TR isn’t running its FTP detector in the background for every athlete every week or so. Doing this would generate a massive dataset of how well different people react to different training regimes. Now, if only TR sucked in sleep, RHR, weight, etc. you could start to develop closer to individualized training guidelines / what the work to recovery week ratio should be / etc.
I think they definitely are though. I think the nod towards using heart rate data also clues that they are looking at a bunch of features when using their ai approach and seems like probably an rnn but i will leave the speculation up to the experts.
Another issue is the placebo effect: if the app is telling you you are tired or not well rested based on objective metrics, you might be prone to believing it. Rather than validating something the athlete feels, this could be the cause for the athlete feeling worse.
Yes, and I reckon this might be something TR is already doing in the background. They could roll out an AT beta to a small subset of users and compare that to the rest, which is the control group. If the changes are very subtle, users might not really notice anything at all.
How you do know they don’t? I’ve always been assuming that they use their full dataset to validate their FTP predictors as well as other ML models.
Mixing power data wizardry with food is never a good idea me thinks. How am I supposed to feed good data to the model, excuse the pun. We are having problems with so called ±1% devices amd now I have to trust my mouth too?!
Also they need to take into account acute fatigue too. I am not factory resetting during the night.
Years ago I tried to program realTSS. It was supposed to take into accound how crappy I woke up that morning and how the ride sucked increasingly more. It did not work.
Ots is halfway there as long as I am concerned.
Not sure if you know or I’m missing it but does the athlete enter how much they ate post-ride? Or do they plan to model it somehow during a fitness assessment? Or “they haven’t gotten that far, slow your roll”
(just yell if I missed this up-thread…I skimmed but didn’t see). Pretty cool stuff.
I don’t know. Had a 1-1 call with my coach a couple weeks ago, asked in general about the initiative and he had to sign an NDA a while back. Told him “that’s cool, lets skip this topic and I’ll wait for Frank to release info.” We ended with him saying that coached athletes would get access. So it was an amazingly short discussion topic.
Looks like Frank will have a series discussing OTS in more detail.
Quick link to related podcast:
Edit: OTS discussion starts about 37 mins into the episode.
I can ride at ~95% FTP on pretty much all fats (lab test) - how would this be accounted for? Note: not asking you specifically @WindWarrior
Moreover, I not convinced we know enough about how sleep (and other included metrics) influence performance/stress/strain and the score/number returned is likely to be based on false assumptions.
Great line of investigation, however, but I reckon it needs a lot more work to be usable.
It wouldn’t and you would find yourself ensnarled in a “definition of threshold” discussion.
How does this lab measured FTP relate to 30min, 35 min, 40 min mean max power? (Any of them in that range)