Predict Alzheimer’s disease with a “100 percent accurate” test? That’s what an New York Times article reported on August 9, 2010.
We’ll now look into the basis of the claim as I promised in my “Brain Research Using Online Data Repositories: Predicting Alzheimer’s Disease II” blog post.
What does the original research article actually say?
(The paper “Diagnosis-Independent Alzheimer Disease Biomarker Signature in Cognitively Normal Elderly People” by Geert De Meyer and colleagues was published in the August 2010 Archives of Neurology.)
There are two exciting methods behind this paper. First, is the use of an online research data repository that brings together a mass of Alzheimer’s disease related clinical data from a number of labs known as the Alzheimer’s Disease Neuroimaging Initiative (ADNI) data repository.
The second is the use of what the authors’ call a “mixture modeling approach.” The approach boils down to associating each kind of protein measured in the cerebrospinal fluid of patients with a dimension in a clustering algorithm.
They clustered Amyloid beta (2 components of 1 biomarker) or Amyloid beta and tau (2 biomarkers) using an unsupervised learning method. Only after this was done were the clinical diagnoses looked at to see if data clusters were associated with normal, mild cognitive impairment, or Alzheimer’s disease. In all cases, two clusters emerged. One associated with healthy cognitive function and the other with Alzheimer’s disease.
Using the mixed modeling approach, the cerebrospinal fluid level of the two biomarkers in 57 patients with Mild Cognitive Impairment (MCI) predicted with 100% accuracy the individuals that would progress to Alzheimer’s disease (clinically measured) over the next 5 years. This is the data cited by the New York Times.
All of the patients in the population showing 100% predictability already had noticeable problems with memory and other cognitive abilities when the measurements were taken. The same mixed modeling approach showed a 94% accuracy when applied to a group of autopsy confirmed Alzheimer’s disease patients.
Also in the current paper the authors showed a 93% accuracy at predicting Alzheimer’s disease using just Amyloid beta as the biomarker (autopsy confirmed Alzheimer’s disease) which is essentially the same as the number obtained when using two biomarkers (94% above). Finally, previous papers have shown comparable numbers using other methods (Oskar Hansson et al., 2006; Sebastiaan Engelborghs et al., 2008).
In sum, it seems clear that this paper did not set out to demonstrate a more accurate method for the early prediction of Alzheimer’s disease. What the paper does seem to demonstrate is an unbiased way to parse data into clusters that may be shown to correlate with disease related outcomes.
Other related blog posts:
Brain Research Using Online Data Repositories: Predicting Alzheimer’s Disease
Brain Research Using Online Data Repositories: Predicting Alzheimer’s Disease II
Brain Research Using Online Data Repositories: Predicting Alzheimer’s Disease III