Deep Stops Increases DCS

[UWSojourner]

Yes, those numbers look about right. But I'm not going to provide you with the other details requested. I'm not interested starting a squabble over 1 pascal. You should use your own tools to verify it with. Just put it in a heat map


I'm guess others would like to verify... so here we go:

The minor line shows cell #10. This has almost the same half time as the 'slow' cell in the A2. 146 v 160 min. You can see the same details how the A2 slow cell line and the #10 tissue gradient line, align closely.


.

I don't believe it's about 1 pascal.

I was hoping you'd post what you had for the raw tissue compartment N2 pressure (not the amount over inspired pressure if I'm reading your chart right).

Another way to say it: For the controlling compartment represented in your supersaturation chart what is your N2 pressure (not gradient) at 0ft, 10ft, and 20ft and what compartment is controlling?

 

[Dr Simon Mitchell]

Igor,

One point that is probably so obvious to you that you did not mention it in acting as a conduit for Bruce Wienke's notions, is that like Ross, he makes money out of promoting deep stop approaches to decompression. I think that should be on the table before any of his claims are evaluated.

The proof is in the pudding -- thousands of divers
employ deep stops and have used them safely without anything but nominal DCS in the process for many years now. No probs.

There are multiple problems with this statement. First, even if it were true, it would not mean that the deep stop approach is the best. Yes, many dives are done uneventfully by users of deep stop algorithms. So what? It tells us nothing about the best approach to decompression. Second, the "without anything but nominal DCS" bit is indisputably wrong. Bruce is not a physician Igor. He does not treat sick divers. I do and have many colleagues who also do, and who have their metaphorical fingers on the pulse of DCS occurrence among technical divers. There a many cases of serious DCS (some fatal) among users of all decompression algorithms. To try to claim that deep stop approaches are somehow immune to this is just disingenuous or naïve or both.

No matter what model (USN, ZHL, VPM, RGBM) somebody will always get hit, but the concern is how many and how often. To date that number is very small by DCS incidences.

There are essentially no hard estimates of the incidence of DCS in technical diving, and the estimates that exist vary widely according to the environment, type of diving, measurement methodology and other factors. In any event, once again, even if it were true, it tells us nothing about what is the optimal approach to decompression.

Deep stops are faster and get us out of the water quicker
so we use them. Period.

Wow, even Ross stopped saying this quite some time ago. They probably get you out of the water quicker, but do they get you out of the water safer or "as safe"? Where are the data that justify the implication? Every bit of properly gathered comparative data published to date in relation to decompression diving suggests the opposite.

deep stop meters, tables and software have now been
around for 10s of years without reported problems with DCS; and judging from the licenses we issue will enjoy even more usage

At least you admit to profiting financially from promoting deep stop decompressions. But "without reported problems with DCS"?! Seriously??? I put it to you that it is unethical for someone effectively selling deep stops to post such misleading claims. What about the diver I presented at the deep stops workshop who completed a flawless RGBM-controlled decompression dive and died 15 minutes later from fulminant cardiopulmonary DCS? Balestra's data in the report you have attached suggests DCS in 1.75% of dives in which an RGBM computer was used - hardly "without reported problems with DCS". Your own data has cases of neurological DCS.

data is data no matter what generates it and for whatever reason and LANL computer downloaded data is headed to DAN as I have time to write translation software for 3000+ profiles

The first part of this statement sounds like a justification for dodgy data, but I don't really understand what you mean here to be honest.

both VPM and RGBM work well and have no reported
DCS spikes (unless misused) across meters, tables and software renderings;

Well, VPM had a DCS spike before it became VPM-B. But leaving that aside, even if this statement were true, it proves nothing about how these algorithms would compare to approaches with less emphasis on deep stops in properly conducted studies with comparisons of outcomes. And just to pre-empt the response, this does not include your LANL paper(s) which is not a proper prospective comparative study.

Balestra of DAN did a study of DCS rates in ZHL and RGBM computers and found DCS incidence rates almost exactly equal and small;

Actually, they were neither equal nor small. The DCS rate for RGBM computers was 1.75% which was 30% greater than the ZHL computers. Having said that, I doubt that much can be inferred from these numbers because the methodology and data were not described well enough to make sense of things.

reading thru ScubaBoard a bit I would only say, without
names, folks who denigrate LANL Data Bank (3000+ computer downloaded profiles in 10 sec intervals going to DAN) are "uninformed" to put it politely. These profiles were discussed a bunch at the Deep Stops Wkshp and published numerously in followup papers -- some appended.

I certainly did not denigrate the LANL database; I merely wanted to know how the profiles were gathered. It is good that you have confirmed that you had the "computer downloaded profiles in 10 second intervals" at the Deep Stops Workshop back in 2008.

Attached papers are some published work that was peer reviewed by medical, physiological, engineering, computer and physical scientists -- all real divers too.

I have to observe that only one of the documents appears to have been published in a peer reviewed journal.

Thanks for the response.

Simon M

 

[igor p]

Igor,

One point that is probably so obvious to you that you did not mention it in acting as a conduit for Bruce Wienke's notions, is that like Ross, he makes money out of promoting deep stop approaches to decompression. I think that should be on the table before any of his claims are evaluated.

Simon M

Waw, do you really thnink anyone here is unaware BRW is making money from his work same as you? That for sure is known to anyone, so doesent need to be on the table everytime some part of his work is posted.

For other your comments I am not the one to answer them, but if BRW will want to comment he wil do it himselfe or I will pos answer if he asks me.

 

[RainPilot]

Um, there is a big difference between being paid for work and having a financial interest in the outcome of a test/study. Reference the tobacco industry-sponsored "independent" studies that show that smoking prolongs your life (tongue in cheek)

Not all viewers of this forum may be aware that Dr Wienke earns money from the sale / licensing of his proprietary RGBM algorithm so it is fair to mention it.

Likewise, everyone here knows that Ross makes money from VPM planning software but that , I feel , is not the primary reason he defends VPM so strongly. His Multideco software will still be used by many divers for whatever algorithm they want to even if, for example, VPM was abandoned by divers en-masse for whatever reason.

Disclosure of financial interest in the results of a study are a critical first step to evaluating the overall integrity and "believability" of any study. Its also important not to throw the baby out with the bath water. Just because I make money from something doesnt mean Im not right.

 

[rossh]

I don't believe it's about 1 pascal.

I was hoping you'd post what you had for the raw tissue compartment N2 pressure (not the amount over inspired pressure if I'm reading your chart right).

Another way to say it: For the controlling compartment represented in your supersaturation chart what is your N2 pressure (not gradient) at 0ft, 10ft, and 20ft and what compartment is controlling?

Oh I'll get onto that right away.... my friend

.

 

[BRW]

All,
Don!t get out here much so just a
few quick comments while I have time:
1) -- RGBM licenses are necessary legal
comtracts to protect both parties and the
integrity of the software whether we give
It away or sell it;
2) -- would love to see and hear more about those
VPM DCS spikes alluded to -- details and
documentation. Haven't seen nor heard
of same;
3) -- will post peer reviewed versions of some documents
Igor blogged when I can, plus others if anybody
wants or needs them;
4)-- DCS prevalence in Balestra study is higher than
background rate in DAN studies for whatever biases
but are relatively small for both ZHL and RGBM meters;
and likely statistically the same (check with Balestra);
5) -- bottom line on deep stops and shallow stops is that
both work when used in models correlated with data;
6) -- at the SAME risk level (see papers for how risk is computed
for both), deep stops get you out of the water faster
than shallow stops -- why many mission oriented operations
may find them advantageous and useful. Plus correlated
software;;
7)-- "what works works" as Bill Hamilton remarked years ago
about real diving and outcomes, the multiplicity of which
is beyond the pale.
Thanks for allowing me this short post and very best to all divers
out thee. Sorry can't visit more often and join the fun
BW

 

[leadduck]

@leadduck, you posted a link on BRW's paper that both deep and shallow stops are ok but deep stop is more efficient. Any thoughts on their findings?

Do you mean the paper
BR Wienke: "Deep Stop Model Correlations". J Bioengineer & Biomedical Sci 5:155. doi:10.4172/2155- 9538.1000155

You can download as PDF at:
Deep Stop Model Correlations | Open Access | OMICS International

The authors conclude that VPM and RGBM correlate strongly with the LANL database, whereas USN and ZHL16 do so only weakly. There was a discussion about the submission on rebreatherworld.com but not with too many details about the applied methods and their validity.

If you are interested in the statistical methods of this paper, I recommend reading another paper first:
E.D. Thalmann et al: "Improved probabilistic decompression model risk predictions using linear-exponential kinetics". Undersea and Hyperbaric Medical Research Society, 1997.
http://www.diverbelow.it/attachment...diction using linear-exponential kinetics.pdf

Regarding the first paper of Bruce Wienke, I see two issues: (1) the definition of risk function over compartments is unclear, and (2) the result interpretation is flawed.

(1) Thalmann describes in his paper how the risk function depends on the supersaturation of each compartment. There's a threshold and weight for each compartment, so that the total risk function has 2*n parameters to be fitted for n compartments. In Thalmann's case it's only three compartments, hence 6 parameters for the risk function (plus 3 or 6 for the gas kinetic model).

In Wienke's paper there are only two parameters κ and ω. Why so few? How are the 16 compartments weighted? He writes "The asymptotic exposure limit is used in the likelihood
integrals for risk function, r, across all compartments, τ". But this compartment index τ doesn't show up anymore.
Did he just sum up compartments with no weights? Such a risk function would probably not fit well to the data, no matter what model you use, and you can see that in the results.

(2) In the results please see Wienke's Table 4 and Thalmann's Table 3. Both use a null model for comparison, i.e. a reference model that returns a constant DCS risk independent of the profile. Thalmann calls it "NULL", Wienke "1-step set".
The "1-step set" is a trivial model+risk function that returns P(DCS)=0.0077 for any profile. This is just the average DCS rate over the whole data. In Thalmann's paper it's called a NULL model with p=0.00003.
The "6-step set" is a trivial model+risk function that returns a constant PDCS depending on the dive depth. For example in the 0-199fsw subset ( Table 3 column 1), it's 5/(268+213+10+22+12) = 0.00952381
6-step model will be slightly better than 1-step, but obviously both of these "models" are pretty bad. They are used as a reference, because any model that predicts DCS risk from the dive profile should do better than trivially assuming a constant risk independent of the dive (1-step), or a risk that depends only on depth (6-step).

Thalmann get's that right and all of his models are significantly better than the NULL model.
Wienke get's it wrong, he writes "The canonical value, Ψ6 , is the LL for the 6-step data set. No fit value, Ψ, will better the canonical value, Ψ6". He seems to think that the 6-step model is the best possible one, maybe because all of his models are worse than 6-step. But that's nonsense, LL can be arbitrarily negative.

So, the conclusion from Wienke's Table 4 should have been: all models are useless (RGBM, USN, ZHL16, VPM). They are better than assuming a constant risk (1-step), but none of the tested models can predict the DCS risk better than guessing DCS risk from the depth of the dive alone (that's what 6-step does).

I guess the reason why no model beats 6-step in Table 4 is a useless choice of risk function.

 

[RainPilot]

I think the Doc is just miffed because others is making a living with tried and tested software and he ain't......

What amazes me is how much time and effort Dr Mitchell puts into dealing with endless, pedantic, arrogant arguing over these matters when, if anything, it costs him money. I suspect that his career as an eminent respected anaesthesiologist and research scientist receives no benefit at all from his efforts here.
For the record, I don't like the way Ross debates and argues, the combativeness etc etc. I also don't agree with him on his opposition to these trials and feel there is a lot that is applicable to us as divers and this bears further investigation. I have already made some prudent (I feel) changes to my deco strategy to maximise my chances of enjoying my diving.
I'm just tired of the ad-hominem attacks all the time and rehashing all the inflammatory nonsense.

 

[David Doolette]

Do you mean the paper
BR Wienke: "Deep Stop Model Correlations". J Bioengineer & Biomedical Sci 5:155. doi:10.4172/2155- 9538.1000155

You can download as PDF at:
Deep Stop Model Correlations | Open Access | OMICS International

The authors conclude that VPM and RGBM correlate strongly with the LANL database, whereas USN and ZHL16 do so only weakly. There was a discussion about the submission on rebreatherworld.com but not with too many details about the applied methods and their validity.

If you are interested in the statistical methods of this paper, I recommend reading another paper first:
E.D. Thalmann et al: "Improved probabilistic decompression model risk predictions using linear-exponential kinetics". Undersea and Hyperbaric Medical Research Society, 1997.
http://www.diverbelow.it/attachments/article/131/Thalmann et alii. Improved probabilistic decompression model risk prediction using linear-exponential kinetics.pdf

Regarding the first paper of Bruce Wienke, I see two issues: (1) the definition of risk function over compartments is unclear, and (2) the result interpretation is flawed.

(1) Thalmann describes in his paper how the risk function depends on the supersaturation of each compartment. There's a threshold and weight for each compartment, so that the total risk function has 2*n parameters to be fitted for n compartments. In Thalmann's case it's only three compartments, hence 6 parameters for the risk function (plus 3 or 6 for the gas kinetic model).

In Wienke's paper there are only two parameters κ and ω. Why so few? How are the 16 compartments weighted? He writes "The asymptotic exposure limit is used in the likelihood
integrals for risk function, r, across all compartments, τ". But this compartment index τ doesn't show up anymore.
Did he just sum up compartments with no weights? Such a risk function would probably not fit well to the data, no matter what model you use, and you can see that in the results.

(2) In the results please see Wienke's Table 4 and Thalmann's Table 3. Both use a null model for comparison, i.e. a reference model that returns a constant DCS risk independent of the profile. Thalmann calls it "NULL", Wienke "1-step set".
The "1-step set" is a trivial model+risk function that returns P(DCS)=0.0077 for any profile. This is just the average DCS rate over the whole data. In Thalmann's paper it's called a NULL model with p=0.00003.
The "6-step set" is a trivial model+risk function that returns a constant PDCS depending on the dive depth. For example in the 0-199fsw subset ( Table 3 column 1), it's 5/(268+213+10+22+12) = 0.00952381
6-step model will be slightly better than 1-step, but obviously both of these "models" are pretty bad. They are used as a reference, because any model that predicts DCS risk from the dive profile should do better than trivially assuming a constant risk independent of the dive (1-step), or a risk that depends only on depth (6-step).

Thalmann get's that right and all of his models are significantly better than the NULL model.
Wienke get's it wrong, he writes "The canonical value, Ψ6 , is the LL for the 6-step data set. No fit value, Ψ, will better the canonical value, Ψ6". He seems to think that the 6-step model is the best possible one, maybe because all of his models are worse than 6-step. But that's nonsense, LL can be arbitrarily negative.

So, the conclusion from Wienke's Table 4 should have been: all models are useless (RGBM, USN, ZHL16, VPM). They are better than assuming a constant risk (1-step), but none of the tested models can predict the DCS risk better than guessing DCS risk from the depth of the dive alone (that's what 6-step does).

I guess the reason why no model beats 6-step in Table 4 is a useless choice of risk function.

Leadduck


You are absolutely correct that any useful model must have a higher (less negative likelihood) than the null models such as the 1-step or 6-step sets. This same point was made earlier in this thread (back in post 424 on page 43) by Daniel Mewes. It is not clear what Wienke means by no fit value will be better than the 6-step set.

I made the same point in a lengthy commentary on this Deep stops model correlation manuscript (before it was published) and an earlier paper on another forum 2 years ago (still waiting to hear back from the author).

Correlation of popular diving models with computer profile data and outcomes

I will repost that commentary here, although it makes much the same point as you do, I have a bit of commentary on risk functions etc. that you might find interesting. I will have to split it up to stay under the 100000 character limit.

Any interpretation can only be based on what is written - and most of what is written is unclear - but from what is written it appears that the four models (USN, ZHL-6, VPM, RGBM) fit to the data set in the paper 1) are not actually the four algorithms of those names that are used to produce decompression schedules; 2) they differ from each other only by having different half-time compartments; and 3) none of them fit the data better than simply assigning identical risk to all dives in certain depth ranges, irrespective of how long the bottom time, how long the decompression, and what breathing gases are used (as you point out). This paper contributes nothing to the debate about whether deep stops or shallow stop schedules are more efficient.

David Doolette

 

[David Doolette]

Commentary on Correlation papers part 1

In the first paper (Comp Biol Med 2010), in an approach superficially similar to that pioneered by the U.S. Navy, the author has developed two “probabilistic” decompression models based on RGBM structure - one is a gas content (supersaturation) model and one is a bubble model. In both models, presumably using dive profile (a series of point-in-time nodes describing depth/breathing gas history) as input (it is never clearly stated), tissue gas tensions are calculated for a collection of compartments – presumably with the RGBM half-times. In the gas content version, the instantaneous hazard (h) is a function of compartment supersaturation and in the bubble version, the instantaneous hazard is a function of the RGBM bubble volume. Both hazard functions have two fitted parameters: a scaling parameter and a time-dependent threshold. The contribution of each compartment to the risk of DCS is a function of the time integral of the hazard, 1-exp(-⌠hdt). Exactly how each compartment contributes to the overall probability of DCS is not made clear, but the implication is that only one compartment contributes to the hazard at any time (which is not how these risk functions should be constructed).

The values for the two fitted parameters in the hazard functions are found by fit of the models to a database comprising 2879 dive profiles and outcomes (DCS=1, no DCS=0) in which there are 20 DCS . It is in this fit of the model to the data that the author’s approach differs from that of the U.S. Navy. The U.S. Navy approach is to fit the parameters of the biophysical model that underlie the hazard, such as compartment half-times, gas diffusivities, nuclei radius, etc., and in this process the biophysical model is altered to provide the best possible description of the data. The author’s approach is purposefully not to alter his biophysical models; his fitted parameters only scale the model outputs. As such, he appears to be seeking to demonstrate how well his unaltered biophysical models describe the data, and I believe this is what he means by correlation.

So how well do his models describe the data? Apparently not very well. In the first paper (Comp Biol Med 2010) the author compares his fitted models to a model based on an arbitrary grouping of the data into OC nitrox, Rebreather nitrox, OC trimix, Rebreather trimix, OC heliox, Rebreather heliox (“6 step set” that appears on the first line of table 3). In this six-parameter model, all OC nitrox dives have a probability of DCS equal the observed incidence in that group (8/344=0.0232), all Rebreather trimix dives have a probability equal the observed incidence in that group (2/754=0.0027), etc. and the specifics of dive profile has no influence on risk. The author compares this “6 step set” model to the fitted models by comparing the log-likelihoods. This can be done informally – a higher (less negative) log-likelihood indicates a better fit to the data (more explanation later). However, the author uses formal log likelihood ratio tests (the assumption that the test statistic is chi-square distributed in this particular case is questionable, but let us take the tests as, at worse, informal evaluations). The author’s principal finding is that his probabilistic version of RGBM is slightly worse, but statistically indistinguishable from the “6 step set” – this means RGBM is no better at describing the 2879 dive data set than assigning risks based on what scuba set and diluent gas was used. The probabilistic version of RGBM performs better than the author’s probabilistic gas content, but neither is demonstrated to link the information in the dive profiles to the observed incidence of DCS.


In the follow-on manuscript (now published in an open access journal as Deep stops model correlations. Bioengineering & Biomnedical Sciences 2015), the author appears to compare the probabilistic gas content model described above, to probabilistic gas content models based on the ZH-L16, VPM and Workman half-times. Each model has the same gas content hazard function described above. Therefore, the models appear to differ only in the compartment half-times. I assume the models are fit to actual dive profiles in the LANL data base, but this is not clear, and based on that assumption I cannot see that the permissible supersaturations (m-values) of any of the models come into the calculations. The data base appears to be a slightly expanded data set from the one used in the 2010 paper. The data set now is reported in the abstract to be 2994 dive profiles and 23 DCS (but table 4 shows 3004 dives). After fit to this data set the parameter estimates and log-likelihoods for all four models are quite similar. This is not surprisingly as the models appear to be so similar. In this case, the author compares the fitted models to a model based on grouping of the data into depth strata, this is his “6 step set” in the first line of table 4. In this six-parameter null model, dives to 199 fsw or less have a probability of DCS equal the observed incidence in that group (5/525=0.0095), all dives to 600 fsw or more have a probability equal the observed incidence in that group (1/2=0.5), etc. and the other specifics of the dive profile has no influence on risk. The author compares the fitted models to this 6-parameter model using log-likelihood ratio tests and each of the four models has a slightly worse, but statistically indistinguishable, log-likelihood than the 6 step set model. Again this indicates these models do not usefully link the information in the dive profiles to the observed incidence of DCS.