Quantifying training

RE: Quantifying training

Kevin in MD — Tue, 23 Oct 2012 10:07:29 GMT

Periodic performance tests are definitely a problem. Well to dredge this guy up again, I am about to launch a new initiative with my squad that hopefully will help us by not needing periodic performance tests to use with a Banister style impulse-response model. I've done some digging and found that swim performance correlates pretty well with heart rate variability measures. RMSSD and LF power mainly. This is isn't exactly new, what is new is that now anyone with a garmin or other ant+ heart rate monitor strap and an ant+ dongle for their computer cna get heart rate variability data with the help of free hrv analyzer software and a program called kubios. The test takes 5 minutes every morning after waking up and peeing but before coffee. There is also $10 software for the iphone and android called ithlete that does the same thing along with their own little dongle as well. If we are willing to accept that hrv measures correlate closely with swimming performance (most r values are 0.60 or higher) then we can use those hrv measures as input instead of performances. Making it much more accessible. This way we can get Tn (drop dead day) estimates and Tg (best day to workout) estimates for all swimmers and time tapers more exactly for a wide variety of people. That's the idea anyway. Here are the swimming correlation studies ukpmc.ac.uk/.../reload=0;jsessionid=0vuJxRaMhMRUnvurybk1.0 www.thieme-connect.com/.../s-2006-924490 here's one for running performance www.springerlink.com/.../fulltext.pdf here's one for changes in vo2max www.depts.ttu.edu/.../hrv athletes_cnrl periperal perform.pdf

RE: Quantifying training

Former Member — Wed, 11 Aug 2010 16:18:59 GMT

The least hassle might be the session RPE score. Rate how hard the session is and multiply * time. It is well regarded in the literature and seems to be quite good. Carl Foster is the original proponent I think. nice person with whom I collaborate on a regular basis did test it extensively I believe. Integrated in Banister's impulse-reponse model that is. If I had to choose between using HR Based Trimp or RPE based Trimp for swimming, I would probably favor the later, since it gives you more flexibility whilst being easier to collect the data. Hit this link here and download some stuff. It performs all the calculations for you. There's even a Skiba SS calculation sheet developed using VBA driven functions to compute IFs etc.... It's in Spanish but even I could find my way through it.

RE: Quantifying training

Former Member — Wed, 11 Aug 2010 16:06:47 GMT

I agree that for achieving a better fit, time constants should be adjusted on a per individual basis. It's done through a long process of trial and error involving reconciliation between the peaks that the graphs generated with the actual performance that was delivered. So is the recommendation expressed by the authors of most modern impulse-response models. That said though, and with all due respect for the quality of the research you've been doing so far, it is also strongly recommended to use weighted averaged scoring data to these models for better accuracy, as opposed to rely solely on distance sort of inputs. Moreover, since there's a very strong sprinting over shorter distances component to most swimmers, I would certainly prioritize using weighted avg scoring data over adjusting time constant for better accuracy.

RE: Quantifying training

Former Member — Wed, 11 Aug 2010 15:35:17 GMT

I'm still looking for a method with the right balance of detail without too much daily hassle. Is the biggest hassle data entry or data collection?

RE: Quantifying training

Former Member — Wed, 11 Aug 2010 14:35:04 GMT

in fact yardage is an adequate input that in my own work has shown no difference from sharp stress scores as inputs On this question, I have changed my mind a couple of times. At one point, I gave up tracking trimps of any sort, for exactly the reason you mentioned: my training was fairly unvaried, and trimps were pretty much proportional to yards. So it didn't seem like tracking trimps was worth the extra effort. I suspect this is the case for many masters swimmers with a single coach writing workouts that are not dramatically periodized. But I switched pools a year ago, and my training mix is very different now. Plus I swim with different coaches on different days of the week, with very different styles / workouts. Even if the week-to-week balance is about the same at the masters workouts, there are other variables. For example... ...I do weekly lake swims for half the year, and they have dramatically different trimp/yard ratio from my pool swims. ...At swim meets, just counting yards gives a ridiculously poor estimate of the training load. ...Last week I was on vacation and dropped in two other teams' workouts. Since I have been playing around with "energy points" and "pain points" recently, I can tell you that one (excellent) workout with GSMS in North Myrtle Beach earned me 20% more "energy points" per yard than I'm used to, and 2.4x as many "pain points" per yard as I'm used to! Given all of these sources of variability, I have stopped relying on yardage alone, and am back to using trimps / points / scores to track training stress. I'm still looking for a method with the right balance of detail without too much daily hassle.

RE: Quantifying training

Former Member — Wed, 11 Aug 2010 14:15:05 GMT

In the other thread you pointed out Rick Sharp's training stress score, which gives points for training time spent in different zones (aerobic, anaerobic power, lactate threshold, sprint). I had a hard time tracking down the original reference, so here's a link for anyone else that's curious: www.swimmingcoach.org/.../JSRVol9_1993.pdf I also updated the intro post in this thread to include some info on that method of assigning points to swims.

RE: Quantifying training

Former Member — Wed, 11 Aug 2010 13:51:38 GMT

Kevin, welcome to the thread. FYI, the best available public paper on this is here, www.ncbi.nlm.nih.gov/.../ Very nice paper, thanks. It's great to see someone with real statistics chops trying to validate these models. The results point out what we already know: the models are a little too simple. If you take the model too literally (for example by trying a full-stop taper after your drop-dead day) then you'll discover that there is plenty of wisdom accumulated in common sense coaching techniques that is not captured in the models.

RE: Quantifying training

Kevin in MD — Wed, 11 Aug 2010 12:54:21 GMT

Hit this link here and download some stuff. No link when I see it, is it Alejandro Martinez' site?

RE: Quantifying training

Kevin in MD — Wed, 11 Aug 2010 12:52:27 GMT

It's done through a long process of trial and error involving reconciliation between the peaks that the graphs generated with the actual performance that was delivered. That said though, and with all due respect for the quality of the research you've been doing so far, it is also strongly recommended to use weighted averaged scoring data to these models for better accuracy, as opposed to rely solely on distance sort of inputs. Moreover, since there's a very strong sprinting over shorter distances component to most swimmers, I would certainly prioritize using weighted avg scoring data over adjusting time constant for better accuracy. As for the long process of fitting, it takes the excel solver roughly 2 seconds to find the time constants and K factors, it's really not that bad. As I mentioned, getting the performance measures seems to be the limiter for most people. Selecting the performances is also important regarding what you mentioned. In one athlete for whom I had data at 100, 200, and 500 yards, the constants were different for each distance tested. So we have to pick intelligently a test performance distance that is relevant to the athlete. That's how I have handled it before. As for yardage vs swimscore or sharp score, I am pragmatic about it. I use what is available, if only yaradage that is fine, we get good data fits with yardage as input. If I have yaradage and intensity that's fine too, I also get good fits using that as input. I'll be doing a comparative study with my masters squad this fall, if everything goes right.

RE: Quantifying training

Kevin in MD — Wed, 11 Aug 2010 12:04:17 GMT

I'm still looking for a method with the right balance of detail without too much daily hassle. The least hassle might be the session RPE score. Rate how hard the session is and multiply * time. It is well regarded in the literature and seems to be quite good. Carl Foster is the original proponent I think. I have not seen much about using it as input into the Banister model and whether the answers are different. For that matter, no study I have seen anywhere has shown any given quantification method gives a different answer using any given model. Not to say it isn't true, but it hasn't been shown. So while we stand around and might say that Phil Skiba's swimscore is better than yardage as an input, it hasn't been demonstrated that they give different time constants. The same thing holds for trimps, tss, or any other method we can dream up.

RE: Quantifying training

Kevin in MD — Wed, 11 Aug 2010 11:45:31 GMT

The normal assumption is that these time constants are fairly transferable. You can (some have) fit the time constants, but it requires a lot of data and work. It's fairly reasonable to think that they should be constant. They have a fairly biological interpretation, and are related to the rates at which your body can build new mitochondria, or hemoglobin, etc. It seems reasonable to think that these biological rates would be pretty similar across individuals. I have to disagree with these statements emphatically. As shown in the Avalos and Hellard paper, the time constants are not the same from one person to the next. Even viewed in the most harsh light given the spread of parameter estimates in the Avalos paper, three out of nine people had estimates that did not cross over. Estimates in published work on weight lifters, throwers, runners bikes and swimmers all have given different time constants. To boot, the ratios of K1 to K2 for these models is always reported to "ONE" significant digit. Leading an observer to believe that the mathematical solutions are highly constrained. If the solutions were not so unconstrained, it is my opinion that the results would be even more widely diverging. In addition there is published work on multisport athletes showing that individual athletes have different time constants and gain values for the different sports. Lastly, those of us who have been using these models for multisport athletes see wide variation in the time constant from one sport to the next. As for fitting the models, it's not as hard as it might seem, the weekly or bi weekly performance test is the biggest hurdle in my opinion. Once you have that and model inputs, a little work in excel can get you where you need to go.

RE: Quantifying training

Former Member — Wed, 11 Aug 2010 06:54:11 GMT

His site is the reason I use google translate, we have been discussing these and similar issues on and off for a couple of years now. As a fellow coach and engineer, we seem ot be on the same page on many things. Well it ain't hard to be on the same page as Ale. He is incredibly pragmatic and fully committed to evidence base research. I recommend to anyone here that knows how to use Google Translate, interested in science driven training, to visit his blog. www.amtriathlon.com

RE: Quantifying training

Former Member — Wed, 11 Aug 2010 06:50:31 GMT

I assume that when I test well in practice I will swim well in the meet. To me the point is to figure out when to maximize and minimize training, not predict meet times. Yes, that's a good point. But you can only use meet times or practice performance tests to calibrate the model -- not both. The model tends to be quite stable when the number of data points is above 8, throw out one test and you get pretty close to the same result, in my experience and in the experience of the Hellard et al authors. Okay, I guess I meant ill-conditioned, not unstable. The same as your point that the models are overspecified. I suppose I'm not surprised that the parameters are all correlated. Someone who recovers quickly (small t2) probably also improves quickly (small t1). But I still wonder if some of that apparent ill-conditioning isn't due to the fact that the model is being fit to a very narrow slice of the input space.

RE: Quantifying training

Former Member — Wed, 11 Aug 2010 06:11:55 GMT

Race Day already factors in training test / racing test features. Not sure how though. There's even a feature allowing for predicting performances in upcoming meets based on the evolution of these tests. All that in a relatively simple to use but rather expensive package.

RE: Quantifying training

Former Member — Wed, 11 Aug 2010 05:24:22 GMT

The normal assumption is that these time constants are fairly transferable. You can (some have) fit the time constants, but it requires a lot of data and work. It's fairly reasonable to think that they should be constant. They have a fairly biological interpretation, and are related to the rates at which your body can build new mitochondria, or hemoglobin, etc. It seems reasonable to think that these biological rates would be pretty similar across individuals. I have to disagree with these statements emphatically. After reading the Hellard et al paper, I'd agree that the constants obviously differ a lot from athlete to athlete. I'll have to abandon my belief that they have a biological basis. Given the (ridiculously) big spread in values for the 9 elite swimmers in the Hellard et al paper, I wonder if it says more about the fact that the model is underfit, though. When I said the "normal" assumption is to keep them constant, I was thinking about their use in software like WKO+ for cycling. I have fit them for myself, and know first-hand that they can differ from the default values. As for fitting the models, it's not as hard as it might seem, the weekly or bi weekly performance test is the biggest hurdle in my opinion. I agree, on both counts. Again, my statement about "a lot" of work was relative to what most athletes are willing to put in. Those reading this thread are atypical. :) Periodic performance tests are definitely a problem. Especially for swimming: I can get close to my current running PR in a self-timed tempo run. I can never come anywhere close to a current swimming PR in practice, even when I dive from the blocks and have someone timing me. You could say that consistency in the performance test is all that's important, but then you can't include meet times in the model (when that was the entire point!). Plus, to make sure you've got enough data to fit the model well, you'd ideally want some performance tests after a wide variation of different training. But most of us aren't willing to suffer through goofy blocks of training just to determine the model parameters. My suspicion is that this is part of the reason the models seem to be unstable in many cases.

RE: Quantifying training

Former Member — Wed, 11 Aug 2010 05:06:26 GMT

No link when I see it, is it Alejandro Martinez' site?sites.google.com/.../ (sorry) And yes, it is indeed the Great Alejandro Martinez who is kindly putting this blog (one of the best at the moment) together. Ale is a humble person, doesn't make lots of noise, but he is an extremely qualified resource. Euh well. You seem to know him already so...

RE: Quantifying training

Former Member — Wed, 11 Aug 2010 05:02:53 GMT

Is the biggest hassle data entry or data collection? Both. Data collection is tough for anything involving HR. Data entry begins to be tough if I need to recall paces for individual components of lots of reps / sets. Some days I'm happy to geek out and log tons of details. Other days I don't have time.

RE: Quantifying training

Kevin in MD — Wed, 11 Aug 2010 02:25:23 GMT

sites.google.com/.../ (sorry) And yes, it is indeed the Great Alejandro Martinez who is kindly putting this blog (one of the best at the moment) together. Ale is a humble person, doesn't make lots of noise, but he is an extremely qualified resource. Euh well. You seem to know him already so... His site is the reason I use google translate, we have been discussing these and similar issues on and off for a couple of years now. As a fellow coach and engineer, we seem ot be on the same page on many things.

RE: Quantifying training

Kevin in MD — Wed, 11 Aug 2010 02:22:40 GMT

After reading the Hellard et al paper, I'd agree that the constants obviously differ a lot from athlete to athlete. I'll have to abandon my belief that they have a biological basis. Given the (ridiculously) big spread in values for the 9 elite swimmers in the Hellard et al paper, I wonder if it says more about the fact that the model is underfit, though. It takes me a bit to wade through all the statistics in that paper, but IIRC their contention is that the model is overspecified. In real practice overspecification is an issue as you have 5 constants in the equation, 90 days to get relevant tests and most people only do one test per week. So you have 12 data points being fit by 5 constants. As for why the constants vary so widely between people. I take it as the mathematical manifestation of what we already knew, some people recover more quickly than others, easy as that. If you want to be more clever you might say that recovery is a function of many systems, neurological, muscular, glycogen recovery etc, all with different time constants in different people and furthermore that in different people, different systems are rate limiting; all leading to wide variations. Periodic performance tests are definitely a problem. Especially for swimming: I can get close to my current running PR in a self-timed tempo run. I can never come anywhere close to a current swimming PR in practice, even when I dive from the blocks and have someone timing me. You could say that consistency in the performance test is all that's important, but then you can't include meet times in the model (when that was the entire point!). I assume that when I test well in practice I will swim well in the meet. To me the point is to figure out when to maximize and minimize training, not predict meet times. Plus, to make sure you've got enough data to fit the model well, you'd ideally want some performance tests after a wide variation of different training. But most of us aren't willing to suffer through goofy blocks of training just to determine the model parameters. My suspicion is that this is part of the reason the models seem to be unstable in many cases. The model tends to be quite stable when the number of data points is above 8, throw out one test and you get pretty close to the same result, in my experience and in the experience of the Hellard et al authors. The issue of variation is one that can catch you. I have varied the day of the week on which tests are given to try and get around it. Particularly if every monday is distance day, and you test every tuesday you might not get much variation. There is also the pace learning aspect of it, we get in our head that all out = this pace. And so week after week we might hit the same pace. I have definitely had that problem.

RE: Quantifying training

Chris Stevenson — Tue, 10 Aug 2010 07:31:38 GMT

For example in your own training, does your % of below threshold and above threshold yards vary significantly from one week to the next? Week to week? Probably not, other than random fluctuation. But over the course of a season it very well might change (eg, early in the training cycle might have more of an aerobic component than later in the season). I assume you are referring to a statement I made (or frequently make) that overall yardage is not a sufficient metric of training load. I stand by that statement, though I agree that for a given club there may be a pretty significant correlation between high-intensity work (eg race-pace training) and overall yardage. But if you want to compare across clubs then the correlation will be significantly less. I have swum with two different workout groups in the last 5 years. With my current group, my yardage has increased a little (maybe 15%) but the amount of high-intensity training we do has increased tremendously (2-3X as much). The positive impact on my performance has much more to do with the latter factor than the former, in my opinion. In other words, a statement from a person that they do (say) 15000 yards/week, by itself, is not sufficient to given an idea of their training load. At least, in my opinion. But I do agree with the general idea that, for an individual person, if there is a strong correlation between the two factors, you might as well use total yardage, since that is pretty easy to determine.

RE: Quantifying training

Kevin in MD — Tue, 10 Aug 2010 05:19:11 GMT

How did I miss this thread? I have been using these models for a few years now and love this stuff! FYI, the best available public paper on this is here, www.ncbi.nlm.nih.gov/.../ The paper is put forth as showing all the problems with the model, notably wide confidence intervals around the parameter estimates of Tn and Tg, what I call the drop dead day (the day after which exercising will only make you slower at the big race, theoretically) and the time of greatest benefit. However, even in this case of a relatively homogeneous group of swimmers, there are several instances of swimmers parameter estimates being outside each others' confidence intervals. In practical terms that means that if one followed the taper of the other, then they would have a sub optimal response. You can actually do these models for yourself and see how you respond to a taper. So you can find your own proper taper length.

RE: Quantifying training

Kevin in MD — Tue, 10 Aug 2010 05:03:31 GMT

Chris, I've used the Banister model for modeling swim performances for years and in fact yardage is an adequate input that in my own work has shown no difference from sharp stress scores as inputs. I think what is happening is that in the people I have used this on, the quality of training does not vary enough to make the differences significant. My own thought is that this would probably extend to other masters swimmers as well. For example in your own training, does your % of below threshold and above threshold yards vary significantly from one week to the next? In my own case it varies of course, there are tapers, there are times when I work on yards more than intensity, but for the most part those percentages are within 10% either way. So then when used an input into these models it just comes out in the wash.

RE: Quantifying training

Former Member — Sat, 26 Jun 2010 14:32:22 GMT

Thanks for the links. I think the thing we're disagreeing on is the meaning of "fatigue". I don't think so. It just means a few things and so we just gotta be careful to apply the correct meaning to the correct context. Not sure if you had read this short post when I first introduced it. Here just in case.... It's one of the best *short and understandable* explanation of what fatigue is. It was proposed by Coggan to describe everything he had to take into account to create a sound model: Not only that, but the algorithm isn't really based on production and clearance rates of lactate at all (or at least not directly). It's like I told Kirk Willett the other day: the algorithm is intended to "track" numerous physiological responses, but none of them in particular. Changes in steady-state (or quasi-steady-state) blood lactate concentrations were simply used as a proxy to estimate the degree of curvature of numerous physiological and metabolic responses that respond in a non-linear fashion as a function of exercise intensity. This approach/logic (which is also the basis for TRIMP) is made possible by the fact that they all seem to follow a comparable pattern - that is, there is a high correlation between blood lactate levels and the rate of lactate release from exercising muscle, between the rate of lactate release and the rate of lactate production/accumulation, between the rate of lactate production/accumulation and the rate of glycogenolysis, between the rate of glycogenolysis and changes in muscle "energy charge" (e.g., (=)/), between muscle "energy charge" and the rate of glucose oxidation, between the rate of glucose oxidation and the rate of glucose uptake, between the rate of glucose uptake and the rate of glucose production, between changes in sympathetic nervous system activity (as indicated by changes in plasma norepinephrine and epinephrine) and changes in carbohydrate utilization, etc., etc., etc., etc., etc., etc. To focus excessively on blood lactate (e.g., on the pattern of accumulation during the "30/30" intervals that Billat has studied) is, quite simply, missing the point. (With apologies to RapDaddyo and frenchyge, because in all fairness there's really no reason to expect them to have realized this, unless perhaps they happen to be trained in exercise physiology.)

RE: Quantifying training

Former Member — Sat, 26 Jun 2010 14:20:45 GMT

I agree that my lactate concentration, glycogen, soreness are all recovered within a few hours or days of a tough workout (fatigue in your sense). But that workout will prevent me from swimming peak times for a somewhat longer time (fatigue in the model sense). +1. In total agreement here. Good enough for booking more quality work doesn't mean good enough for peak performance. I intend to push the study of this last article much further. What I find particularly interesting is that (and it could be seen as a downside as well) the author seems to extend the definition of muscle damage well over the damage done at the muscle fiber level. His definition of muscle damage seems to include the sort of damage one would incur as part of a normal race pace set. Altered calcium balance is the sort of *damage* that anyone would get after a hard race pace set. is most likely due to several factors including the disruption of muscle calcium balance and energy production, the poor recovery of muscle energy during this period, and the decrease in muscle protein content In case you had missed this paragraph, he later describes how compromised muscle structures tend to be slower in reloading its glycogen level, which by the same token explains Maglischo's Graph (referred to in a previous post). Based on this graph, the swimmers' glycogen level was not completely recovered even after 48 hours. **edit** Analysis in progress. Too cool sjstuart. First reference: Muscle damage comparison between 2 groups. Dumbell curl light weight high endurance and dumbell curl heavy weight. Muscle damage was found in both groups although recovery time was significantly faster in the endurance group. That first reference kind of confirms what I first though: Subjects were previously untrained. That explains the exaggeratedly long recovery times put forward by the author. That is a huge weakness but still. Very nice material for anyone wanting to improve on the applied exercise physiology side.

RE: Quantifying training

Former Member — Sat, 26 Jun 2010 10:53:32 GMT

Thanks for the links. I think the thing we're disagreeing on is the meaning of "fatigue". The 7- and 15-day constants in the models weren't made up. They were fitted to real results from real athletes. (Mostly trained, but perhaps not always.) If there were no measurable effect on performance at 3 days out, then the time constant would be much shorter. You say that "in exercise physiology, fatigue simply refers to the phenomenon that forces you to slow down as Lactate Level increase." But in the context of these models, the quantity defined as the "fatigue score" (or ATL in Coggan's model) is something very different. It's the residual effect of training (biological source unspecified) that has a negative effect on performance. I agree that my lactate concentration, glycogen, soreness are all recovered within a few hours or days of a tough workout (fatigue in your sense). But that workout will prevent me from swimming peak times for a somewhat longer time (fatigue in the model sense).