In the current study we demonstrate that (1) there are proteins that are highly correlated in plasma and serum and are associated with AD status across blood fractions, (2) these findings are replicable across independent cohorts, and (3) using these proteins, we generated a prediction model in the TARC cohort that, when combined with demographic and clinical lab data, yielded clinically significant classification accuracy in the ADNI cohort. To date, this is the first blood-based screener for AD developed that has been cross-validated in an independent large-scale cohort that also works across blood fractions. This work not only further supports the notion that an accurate blood-based screening tool for AD can be generated, but also that such an algorithm can be applied across serum and plasma mediums. Our 11-protein serum-plasma risk score alone yielded an AUC of 0.70 accuracy that was enhanced by the addition of demographic (i.e. age, gender, education, APOE*E4
status) and clinical lab (i.e. glucose, triglycerides, total cholesterol, and homocysteine) data. In , the addition of clinical lab data did not improve the overall accuracy of the algorithm beyond demographic information, which is largely driven by the APOE*E4
rates in the ADNI cohort. However, in our prior work 
, the use of clinical lab data improved overall accuracy and will likely contribute to the robustness of our approach as it is applied to other cohorts. It is certainly possible that inclusion of additional markers, not available in the current analyses, would increase the accuracy of that risk score, which is an additional advantage of our approach as it can be expanded or reduced as necessary to support the accuracy and cost-effectiveness of the algorithm. A single biomarker algorithm that works across both serum and plasma will offer laboratories options that may be preferable for a variety of reasons.
There are several implications for the current findings. There are a number of previously conducted research projects with stored blood biospecimens; however, there is little consistency between what medium was stored. The current findings open up the possibility of utilizing samples from such studies to further validate and refine our algorithm. Additionally, it is likely that the components of diagnostic algorithms will be different from the components of algorithms for progression and different from those predicting long-term risk. Our findings offer a novel approach to each of these questions as well. These findings also support the need for standard protocols to be generated for blood-based AD biomarker research as is currently underway for the CSF markers.
These results also support the robustness of our methodological approach. In our initial serum-based algorithm, the biomarker risk score alone yielded an AUC of 0.91 whereas the serum-plasma algorithm in the current study yielded an AUC of 0.70. While impressive, this overall accuracy is not clinically adequate. However, as with our prior approach, the combination of clinical lab data and demographic variables into the algorithm increased the precision substantially (AUC
0.88). In our prior work, the training and test sample were both based on serum assays and were from the larger TARC cohort; however, the derivation of the algorithm in the TARC cohort and validation in the ADNI cohort supports the robustness of this method. As we have previously argued, using only age, gender, education and APOE*E4
status, one can accurately classify a large number of AD cases when compared to controls. Therefore, consideration of such factors should be considered when examining biomarkers of AD status. We are not the first to demonstrate that inclusion of these factors into an algorithm can improve overall accuracy as others have suggested that a multi-marker approach is superior to single-marker approaches 
. As an example, Vemuri and colleagues found that including demographic factors with structural MRI added to the overall accuracy of disease-prediction models even when cases and controls were matched by these variables 
. This is important given that the TARC cohort did not match cases and controls whereas ADNI samples were matched. The robustness of our methodology may also provide an explanation for the lack of cross-validation of prior work 
. The utility of our algorithm for separating MCI cases from normal controls (and/or AD) remains unknown at present.
The current markers overlap with our prior serum-only based algorithm 
though they do not overlap with those found by Ray and colleagues 
, which may be due to the significant differences in assay platforms utilized. However, there is an existing literature directly or indirectly linking each of the 11 proteins identified in this study to AD. As with our prior work, many of the markers in the algorithm are inflammatory in nature, which we propose as evidence of an inflammatory endophenotype of AD 
. We, and others, have documented a link between CRP and AD 
. Based on the available data, we proposed that the link between CRP and the risk of AD changes over the life course with midlife elevations in CRP increasing risk for AD, but that this risk declines as one ages with decreased CRP related to AD status though elevations in CRP are still related to increased disease severity among cases 
. Adiponectin, an adipocytokine, is related to obesity, insulin resistance, metabolic syndrome, type 2 diabetes, and cardiovascular disease 
and was recently found to be elevated in plasma among MCI and AD cases 
. Therefore, adiponectin levels may be related to the documented links between changes in body composition (e.g. weight loss) seen in prodromal and early stage AD. Pancreatic polypeptide is also linked with diabetes and obesity 
and may provide a clue into the biological link between these conditions and AD. Fatty acid binding proteins, cytosolic proteins found in all cells utilizing fatty acids, are rapidly released into circulation following cell damage 
. Serum levels fatty acid binding proteins have been shown to be elevated among AD and other dementia cases as compared to normal controls 
. A recent meta-analysis showed a significant up-regulation in blood concentrations of IL-18 (as well as IL-6, TNFα, IL1, transforming growth factor, IL-12) among AD cases 
. β2 microglobulin is an amyloid protein 
that has been found to be elevated in the CSF of AD cases 
. Tenascin-C, an extracellular matrix glycoprotein, is involved in a number of biological processes that have been linked to AD including inflammation and angiogenesis 
, which may provide a biological mechanism linking AD to a broad spectrum of cardiovascular diseases and risk factors. The human cytokine I-309, a small glycoprotein, was recently found to be elevated in a proteomic study of CSF among AD cases and was also related to scores on a test of global cognitive functioning (i.e. Mini Mental State Examination [MMSE]) 
. Factor VII is a protein in the coagulation cascade that is required for thrombin generation, which has also been linked to AD 
. VCAM-1 is a member of the immunoglobulin superfamily that has been found elevated in plasma of AD cases 
. It has been proposed that MCP-1 plays a dominant role in the chronic inflammation seen in AD 
and has been found to be elevated in serum of patients diagnosed with MCI and mild AD 
Given the sheer volume of elders worldwide who are at risk for AD, there is an urgent need for a multi-stage approach to screening and diagnosis. There are insufficient numbers of dementia experts to meet the needs of all individuals at risk for the disease and prior work has demonstrated that non-experts are not completely accurate in diagnosing the disease 
, particularly in the earlier stages 
. Our blood-based screener fits into the existing medical infrastructure where screen positives can be referred for confirmatory diagnosis using clinical, imaging, and/or CSF analysis. As with any screening measure, one must consider acceptable levels of false positive and false negative rates of the instrument as well as overall disease base rates of the setting when deciding on appropriate cut-scores on any instrument 
. Therefore it is important that additional work be conducted to determine how this algorithm (and other previously published biomarkers) performs in community-based settings (e.g. primary care offices) as both the TARC and ADNI are clinic-based cohort studies. While sensitivity and specificity are not base rate dependent, accuracy of diagnosis (prediction of disease status present/absent) is a function of base rates of the disease within a given population therefore, overall accuracy of AD presence (i.e. true positives) will increase with advancing age while accuracy of AD absence (i.e. true negatives) will be higher with younger ages. As with age, APOE*E4
genotype, gender, and/or years of education are also important considerations, which is why these variables are included in the algorithm itself.
The independent cohorts strongly support the validity of the findings. These observations also justify further analysis examining a broader range of markers across serum and plasma to determine if the biomarker risk score can be further refined. Our results also suggest that further work in the field should specifically examine the performance of blood-based protein panels across serum and plasma.