A Spreadsheet for Calculating Your Levine Phenotypic Age

JGC 

I believe It is a reasonable hypothesis and it wold be worth some research for which I lack time. AFAIK it is quite well established both MCV and RDW increase with age and are associated to mortality and morbidity as cardio vascular diseases (see also the other thread). As much of this is mediated by “inflammaging” it would be interesting to test whether the Senescence-Associated Secretory Phenotype of the cells (SASP) which characterizes senescence in particular in response to inflammatory signalling, likely in all tissues, incl. blood, is also associated to morphological changes which would mechanistically explain the high weights in the Levine’s regression in step 1. Note also that in the regression weights it is rather the variance of MCV (i.e. RDW) which is impactful rather than the MCV itself which makes me wonder, once again, about the higher heterogeneity of the aging population sample as I was mentioning in the previous thread. Worth to follow!

JGC 

The following paper adds to what mentioned previously and confirms the important role of RDW.  WRT mortality and morbidity the paper is no surprise as the meta-analysis is based on similar cohorts as the Levine et al paper. However an interesting comment on the mechanisms seems confirming what discussed: the role of inflammation (and oxidative stress), the higher heterogeneity point I made in my previous posts and the independent prediction capability:

“The mechanisms through which RDW increases with age and is associated with mortality have not been defined; however, it is possible that oxidative stress and inflammation play a role given that both can reduce RBC survival (16,17), leading to a more mixed population of RBC volumes in the circulation. In patients with conditions characterized by increased levels of oxidative stress, such as Down syndrome (18), poor pulmonary function (19), and dialysis (20), RDW values are elevated. Analyses of the NHANES III data showed that decreased serum antioxidant levels, including carotenoids, selenium, and vitamin E, were also associated with increased RDW, although adjusting for these antioxidants did not meaningfully attenuate the RDW-to-mortality association (5). In addition to reduced RBC survival, inflammation might further influence RDW levels by disrupting erythropoiesis. Indeed, it is believed that the increased prevalence of anemia with advancing age is due, in part, to the effects of proinflammatory cytokines on inhibiting the proliferation of erythroid progenitor cells and downregulating erythropoietin receptor expression (17,21,22). Perturbations in erythropoiesis can lead to more variation in cell sizes exiting the bone marrow and might increase RDW. Importantly, however, RDW was more strongly associated with mortality in nonanemic than in anemic older adults (Figure 4). In the NHANES III, adjustment for C-reactive protein, fibrinogen level, and white blood cell count did not substantially alter the effect of RDW on mortality, even though each of these factors was strongly related to RDW and mortality (5).”

Patel KV, Semba RD, Ferrucci L, et al. Red cell distribution width and mortality in older adults: a meta-analysis. J Gerontol A Biol Sci Med Sci. 2010;65(3):258-65.

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2822283/

JGC 

Yes, thank you. I did not realize that. MCV is included though. ML/AI are black boxes though which lack the explanatory power of system biology approaches but can generate hypotheses to test. On the other side Horvath’s reply to your note seems also to indicate we do not have a biological understanding underlying the step 1 algorithm neither. I also understand the role of DNA methylation in step 2 is also not completely clear (they try to understand in step 3) in particular its associative or causative role in aging.

JGC 

Along the lines of the answer you got from Dr. Horvath (thank you), Dr. Levine answered my note on the relative roles played by the clinical biomarkers used for the phenotypic age calculation. While recognizing the RDW's big impact on the estimate, she warns the weights listed in the table were not standardized to allow for a direct comparison.

JGC Thanks for the Aging AI page. I'm 52, the spreadsheet puts me at 40 and Aging.ai 3.0 puts me at 28, FWIW.

DanMcL 

DanMcL said:
JGC Thanks for the Aging AI page. I'm 52, the spreadsheet puts me at 40 and Aging.ai 3.0 puts me at 28, FWIW.

Interesting. Have you tried also the V 1.0, just curious? It includes 41 parameters (vs. 19 of V 3.0) and I think it should be more accurate. In my case V 3.0 has been giving systematically way too low values.

JGC 

JGC said:
albedo

Dr. Steve Horvath answered, saying that they did not have any explanation of the relative weight strengths because it was a numerical optimization that didn't involve human judgement. My own theory is that the two blood geometry factors are so important in determining the Phenotypic Ager because senescent blood cells have a different geometry from normal blood cells and their increased presence trends to increase both MCV and RDW.

Along the lines of the answer you got from Dr. Horvath (thank you), Dr. Levine answered my note on the relative roles played by the clinical biomarkers used for the phenotypic age calculation. While recognizing the RDW's big impact on the estimate, she warns the weights listed in the table were not standardized to allow for a direct comparison.

“...In statistics, standardized [regression] coefficients, also called beta coefficients or beta weights, are the estimates resulting from a regression analysis that have been standardized so that the variances of dependent and independent variables are 1.[1] Therefore, standardized coefficients refer to how many standard deviations a dependent variable will change, per standard deviation increase in the predictor variable…. Standardization of the coefficient is usually done to answer the question of which of the independent variables have a greater effect on the dependent variable in a multiple regression analysis, when the variables are measured in different units of measurement (for example, income measured in dollars and family size measured in number of individuals).”

https://en.wikipedia.org/wiki/Standardized_coefficient

This is something I would expect to be considered on future follow on works. Would you agree?

albedo I'll try the earlier version, but assuming they used machine learning (a deep NN) for this, while larger (more input parameters) is generally better even more it's the population data. For example, if I was doing this (I'm a machine learning engineer) I would just take the blood results and age from a large population and make a conv net from that. Easy, but the population has a hidden skew which is it is made up of people who eat the SAD (Standard American Diet) and probably few if nobody like us who practice extreme health habits. So we're probably outliers, my numbers (cholesterol, WBC, etc) are so low they flag in the reports. 

So your typical 28 year old they measured is still eating SAD, so perhaps we don't fit in the sample set. 

DanMcL 

DanMcL said:
albedo I'll try the earlier version, but assuming they used machine learning (a deep NN) for this, while larger (more input parameters) is generally better even more it's the population data. For example, if I was doing this (I'm a machine learning engineer) I would just take the blood results and age from a large population and make a conv net from that. Easy, but the population has a hidden skew which is it is made up of people who eat the SAD (Standard American Diet) and probably few if nobody like us who practice extreme health habits. So we're probably outliers, my numbers (cholesterol, WBC, etc) are so low they flag in the reports.

So your typical 28 year old they measured is still eating SAD, so perhaps we don't fit in the sample set.

Hi Dan. It is great to have you insight on these ML estimators as you work in the field and you must understand exactly what they are doing. I suggest you read both papers they refer to in their web site resp. for V 1.0 and V 3.0. But maybe in the meantime you might refer to the same group overview, see section "4.1.3. Multi-modal biomarkers" in the review paper:

Zhavoronkov A, Mamoshina P, Vanhaelen Q, Scheibye-knudsen M, Moskalev A, Aliper A. Artificial intelligence for aging and longevity research: Recent advances and perspectives. Ageing Res Rev. 2018;49:49-66.

https://www.ncbi.nlm.nih.gov/pubmed/30472217

albedo Thanks for the pointers, I'll try to find some time. I am focused on another project you folks might find interesting though, which is a DNN for doing regression analysis of biomarkers. For example, in another thread (under self testing) I talk about a bunch of biomarkers I measure every morning, based originally on Steve Perry's set and augmented. I've since added ECG, galvanic skin response (autonomous emotional state), EEG and a few others. Anyhow once I've got a collection I'll see what can be done with it. Regression analysis (e.g. how the variables change relative to one another) is one approach. The problem is of course is having a population of 1, and even if I put a web page up there are few people measuring this much data. I might ping Steve and get access to his data. Anyhow, WIP ...

DanMcL 

DanMcL said:
Thanks for the pointers, I'll try to find some time. I am focused on another project you folks might find interesting though, which is a DNN for doing regression analysis of biomarkers.....

Hi Dan, in case you will find some time to look at the two papers Insilico Medicine gives to reference resp. the v 1.0 and v 3.0 of their aging.ai predictor, maybe you can have a clue on why the two versions differ so largely. In my case v 3.0 is systematically lower than v 1.0, in average -33%. I only did a superficial reading of the two different references so maybe I am missing a crucial point likely in the ML process they use. Thank you !

albedo papers summarize and show results so unless they call out your question specifically you usually have to contact the authors directly. In this case the approach they took is standard and they have a large population (120kish) from three countries, should be pretty good. Unless you/we are 1/1M (1 in a million) instead of 1/111k or better the network should be a good estimator. 

Anyhow I’m definitely an outlier so the network is probably a poor predictor, and the discrepancy isn’t clear. Other than when networks mispredict they often do so badly.

28 for me seems too low. Or is it? My blood work is awfully good. They claim good generalization so are confident.

You might send an email, they might answer.

JGC 

I understand and I am a bit disappointed this is not addressed in the paper, neither it is in the Phenotypic Age paper yet to be published though.

For "standardized" I understand it as previously mentioned but of course I might be wrong and never went really through the details, I might ask the authors again.

“...In statistics, standardized [regression] coefficients, also called beta coefficients or beta weights, are the estimates resulting from a regression analysis that have been standardized so that the variances of dependent and independent variables are 1.[1] Therefore, standardized coefficients refer to how many standard deviations a dependent variable will change, per standard deviation increase in the predictor variable…. Standardization of the coefficient is usually done to answer the question of which of the independent variables have a greater effect on the dependent variable in a multiple regression analysis, when the variables are measured in different units of measurement (for example, income measured in dollars and family size measured in number of individuals).”

https://en.wikipedia.org/wiki/Standardized_coefficient

  albedo  JGC

Yes that's the correct quote in this context I believe. In simple terms they found the coefficients from a model fit (the model has many inputs and from a large sample set), so in this sense it's only relevant/accurate by plugging all your values into the resultant equation. To 'standardize' the variables would mean that you would be able to look at them in isolation. 

I think it gets back to my hack earlier; the most accurate way of learning how to 'optimize' your blood work would be to play with the numbers in the spread sheet and see how they affect the output. But going further and saying one variable is 'more important' than another is going out on a limb. 

JGC DanMcL

Thank you for your replies. Food for thought ...