Does Elevated Glucose Cause Dementia?

Does elevated glucose cause dementia?  Even in non-diabetics?

This would seem to be the initial conclusion based on a recent study published in the NEJM:

Glucose levels and risk of dementia.  




In case the typeface in my graphic is not clear, the Background, stating the purpose of the study:

Diabetes is a risk factor for dementia. It is unknown whether higher glucose levels increase the risk of dementia in people without diabetes.

So they used data from the Adult Changes in Thought study involving almost 2100 people without dementia at baseline for a median followup of around 7 years.

Dementia:  524 of the 2067 participants (25.4%)

• Non-diabetics:  450 of the 1724 participants (26.1%)
• Diabetics:  74 of the 343 participants (21.6%)

These are the actual incidence rates. Not simulated by some model. What actually happened. While they didn't specify whether the difference was statistically significant, I repeat in bold, large red font:

The diabetics had a lower recorded incidence of dementia compared with non-diabetics.

And I do this because, despite using "35,264 clinical measurements of glucose levels and 10,208 measurements of glycated hemoglobin levels from 2067 participants", and for all the discussion of controlling for this and that (even more in the supplementary material), separate analyses excluding some patients, differing time frames for prior-to-baseline average glucose ... splines this and Bayesian that ... the authors do not address this glaring inconsistency in this most concrete of results.   I'm not sure if they even tested the statistical significance of this difference in the recorded incidence rate of dementia.  This result alone could have warranted a title of "Dementia risk decreased in diabetes".  Indeed depending on the basis, you either had:

• 21% increased incidence of dementia in non-diabetics vs. diabetics, or 
• 17% decreased incidence of dementia in diabetics vs. non-diabetics.  

But leaving this aside, they evaluated a 5 year average glucose (eAG) which works out as follows:

• Non-diabetics:  Median = 101 mg/dL  (Q1 = 96, Q3 = 108) 
• Diabetics:  Median = 175 mg/dL  (Q1 = 153, Q3 = 198)  

Converting back to HbA1c this would equate to a median 5.1% (5.0 , 5.4) for non-diabetics and 7.7% (7.0 , 8.5) in the diabetics.

Here is the major outcome from the study in tabular/graphic form. I've chopped the tables and graphs together to group non-diabetics and then diabetics.





Notice anything?  Why does the diabetic plot have that "U" shape where the lower average glucose increases the hazard ratio?  At glucose levels, I might add, significantly higher than those of the non-diabetics?   If it is hyperglycemia per se that is causing dementia, then one would expect:

• Higher dementia in diabetics vs. non-diabetics.
• A hazard ratio plot for diabetics similar in shape to that of non-diabetics although perhaps lacking the apparent leveling off at higher levels 

The "U" in the diabetic model is addressed in the text:

Fig. S5

Some people with diabetes did have data that had a marked influence on model parameter estimates, and we reviewed their medical records. We repeated our primary analyses after excluding data from one participant with acromegaly (Fig. S4 in the Supplementary Appendix) and after excluding data from that participant and two other participants, each of whom had an atypical natural history of type 2 diabetes (Fig. S5 in the Supplementary Appendix). The exclusion of these data resulted in the near elimination of the suggestion of elevated risk at the lowest glucose levels.

Reading further in the supplement materials I believe a fatal flaw in the study analysis is unveiled:

Diabetes was not designated according to the extensive glucose level data and ADA or other diagnostic guidelines.

No, instead:

We classified participants as having treated diabetes on the basis of diabetes-related medication data from Group Health pharmacy records (Table S2 in the Supplementary Appendix). At least two filled prescriptions per year were required for the classification, with the onset date for treated diabetes defined as the date when the second prescription was filled. Once a participant was classified as having treated diabetes, the classification was retained for the remainder of the study.

This is important, because they discuss in detail those three subjects that were "responsible" for the apparent increased risk amongst those diabetics with lower glycemia.  You see their diabetes wasn't "typical" because they had been misdiagnosed/treated or had transient hyperglycemia and were not being treated with diabetes medication during the study.  So ... if you remove these three from the diabetic group, the inverse behavior at the low end of the curve almost disappears.  This would imply that these three all developed dementia ... but they were really not diabetic.  Huh??!!

Further, as was related in the supplementary materials, fully one third (111 of 343) of those designated as having diabetes, "developed" the condition during the course of the study.  I put the quotations around developed because the big problem with this study was that a designation of a subject as diabetic or non-diabetic was apparently unrelated to the mountain of glucose measurements -- from fasting to random to HbA1c or other glycation -- they had for these subjects.  A mountain they appeared to be boasting about to open their Methods section, I might add.  No:

Our operational definition of diabetes was treatment with an antidiabetic medication. 

Which brings me to Supplementary Table 4.  Now, the title of this table differs from the rest, and creates some confusion.  To me this would indicate we are looking at the glycemic distribution of those 524 subjects that developed dementia, but there's no indication of that in the text and all other "5 years" in this study seem to be prior to baseline.  I'm confused.  This isn't a very clearly written study I'm afraid.  (Please, save me the critique on the clarity of my writing.  This is a blog, not an article in a well respected peer review journal)  

If this is the glycemic distribution for those developing dementia it would seem to indicate that the incidences are clustered in the lower glucose levels and not the upper ones.  This doesn't make sense and I will try and contact the author and try to find out.  In the meantime ....

... let's work with what we have here.  While the Q1/Med/Q3 values match up to those in the text for the non-diabetics (96/101/108) they do not for the diabetics (149/165/182 vs. 153/175/198 in the text).  Something's not adding up, but if there's anything to that other than a typo, it would appear that those that developed dementia had slightly better (lower) glycemia than those that did not!

I've revamped the table above and included the approximate number of the subject (ascending order) corresponding to the percentile cut-offs provided, and shaded the quartile cut-offs.  Quartiles would include approximately 113 and 19 subjects for non-diabetics and diabetics respectively for the 524 diagnosed with dementia, or 294 and 86 subjects respectively for all 2067 participants.  I included the subject numbers for all subjects as my hunch is this is what this data reveals, otherwise dementia diagnoses would seem oddly clustered on the lower glycemic end.

Presuming this is the distribution for the entire study group, the top 5% of non-diabetics, 59 people, are likely diabetic and you have at least the three non-diabetics that don't belong in there with the diabetics (according to the study authors) and likely a few more mischaracterized.  

The authors also plot this data in the Supplement  to show the distribution  and this,  Figure S1, makes no sense to me.  Your typical box-plot uses the 5-number summary of min-Q1-med-Q3-max.  I plotted these values and matched up as best I could to the graphic (colored dots are mine, black box plots and dots are theirs).   The min-Q1-med-Q3's match up but where they get the "whisker" end for the fourth quartile is a mystery.  For the non-diabetics, it matches with the 95th percentile but that could be coincidence as that does not work for the diabetics (orange) ... it appears they plotted an arbitrary line symmetrical to the first quartile range?  If anyone reading this can make any better sense of their plot I'm all ears!

So ... with apologies for the graphics heavy post here, I repeat the initial graphic for a reason.  If difficult to read, here are the Conclusions:

Our results suggest that higher glucose levels may be a risk factor for dementia even among persons without diabetes.

While the wording leaves room for investigation of diabetics, the study aim seemed to be to look at those without diabetes.  And the title was simply "Glucose Levels ...", not "Glucose Levels and Diabetes Diagnosis Status ..." or something like that.

Which begs the question, why didn't they just pool all of the results and look at them together?  If they really wanted to ascertain the role of non-diabetic glucose levels on developing dementia, then only those for whom glucose tests met diagnostic criteria without the help of any hypoglycemic agents should have been considered.  Seems simple enough.  This is how diabetics and non-diabetics are enrolled in RCT's and selected for studies of this nature all the time.  If it's just glucose levels, then what does it matter whether or not the person is diabetic and/or their medication schedule?   If types of meds were an issue that could be a subject for another study, but as it was, this wasn't analyzed for anyway.

Given the problems that only three subjects threw into modeling for the diabetic group, I wonder why none of the fourteen authors even considered doing an analysis of all of the data lumped together.  Or at least used diagnostic criteria to divide the group.  Why not?  This would seem to make more sense than trying to explain away why the curve for diabetics went up on the low end.  It would also remove any ambiguity of the diagnosis and duration for the 111 subjects that moved from non-diabetic to diabetic status during the course of the study.

After all was said and done, the relative risk for dementia for non-diabetics was estimated by the chosen model to be 1.18 for those with an eAG of 115 vs. 100 mg/dL.  For diabetics the risk hazard ratio was 1.40 for eAG of 190 vs. 160 mg/dL.  All of which seems odd comparing within groups at all when the real observed dementia rate in all non-diabetics was lower than that in diabetics.   with a median eAG of 175 (text, or 165 Table S4) compared to only 101 in non-diabetics.



One more graphic ...   This is a table from the main article text, I chopped it up and rearranged to juxtapose the time frames for the two diabetic classifications.  They also reported eAG for the time frame between 5 to 8 years prior and repeat the data for the 5 prior years.  You would think that if elevated glucose were the issue, having them develop sooner, and thus persist for longer, should amplify the effect.

Average glucose levels were highly correlated for the two time periods (r = 0.85). Including glucose levels for both periods in regression models resulted in  somewhat attenuated estimates of associations between more recent elevations in glucose levels and risk of dementia. 

So just 5 years immediately prior, we see an effect of glucose levels increasing hazard ratio.  Combine the 5-8 years prior with the 5 years (my understanding meaning 8 years prior) and the risks are attenuated.  Look at just the eAG for a "distant" preceding time from 5 to 8 years before, and you see essentially no effect despite the fact that these levels correlated with those for the 5 year timeframe.  What do you make of this?  It seems to be yet another counterintuitive "general" finding of this study.

Bottom line ... In this current environment of glucose is going to kill you, drive you crazy and make you lose your mind, the abstract of this study makes it look like great ammunition for those looking to blame blood sugar levels.  A closer look at the full text and supplementary materials raise more questions than answers are provided.