There is a significant need for a fast and cost-effective means of screening the rapidly growing elderly segment of the population. The ideal AD biomarker would come from blood [29
], and we recently published a serum-based algorithm that yielded excellent diagnostic accuracy [8
]. However, that algorithm utilized over 100 proteins in the original model. In order to become more cost efficient, such an algorithm would ideally require a more focused set of markers. We utilized variable importance estimates from Random Forest in our initial publication and utilized the 30 most important markers to create a refined algorithm. In the initial analyses, our protein biomarker risk score yielded an observed SN = 0.80, SP = 0.91, and AUC = 0.91 with the current 30-protein risk score being very comparable (SN = 0.88, SP = 0.82, AUC = 0.91). As can be seen from table , the 30 proteins in our biomarker portion of the algorithm cover a range of biological processes. It is our hypothesis that such a broad scope in the biomarker risk score will be necessary for the generalizability of the algorithm and approach to other populations. In fact, a lack of such breadth may be one reason for the failure of prior attempts to cross-validate.
There are several advantages to our approach. One of the recommended criteria proposed by the Consensus Report of the Working Group on Molecular and Biochemical Markers of Alzheimer's disease [30
] was that biomarkers for AD have a SN and SP of >0.80. Even though the SP of our 30-protein risk score decreased to 0.82, the SN increased to 0.88, thereby providing a better balance across both estimates than the original biomarker risk score, which also meet the Consensus Working Group's criteria. The balance between SN and SP is another excellent feature of the current results as this also will provide a balance between positive and negative predictive power [31
]. An additional advantage of our work is the direct comparison of the biomarker risk score to the diagnostic accuracy of common demographic and clinical data. Given the significant difference in age, gender, education, and APOE4 frequency between AD cases and controls in those at risk for this disease, one can classify a large number of individuals without the use of biomarker data (blood, imaging, CSF, genetic or otherwise). In fact, using only age, gender, and education, we find an AUC of 0.80. This may be somewhat inflated by the group differences in demographics in this study; however, the addition of demographic factors adds ecological validity to our methodology. In fact, Vemuri et al. [32
] have shown that adding demographic factors to structural MRI diagnostics added to the overall accuracy of the models, even when cases and controls were matched by these variables. Others have also found that a multimodal approach to the search for biomarkers for AD is superior to any single method [33
]; our method adds the modalities of demographics and clinical labs to the algorithm, which are more cost and time efficient than adding additional biomarker modalities. While it is not necessary that the biomarker surpass the accuracy provided by demographic and/or clinical labs, it is necessary that the biomarker add unique information to the overall accuracy thereby improving the utility of the approach. As such, presentation of the biomarker results in the absence of such comparisons should be considered inadequate. When examining our biomarker risk score, both of our serum-based protein risk scores (1) yield better overall diagnostic accuracy than demographic or clinical variables alone, (2) contributed significantly and independently to case status from demographic factors, and (3) the combination of all modalities yielded far superior results. The combined multimodal nature of our algorithm also increases the likelihood of utility across settings and populations, which we are currently testing.
In the current analyses, the biomarker algorithm was also significantly related to neuropsychological status and disease severity. In fact, the biomarker risk score was most strongly associated with the cognitive domains of memory and language, which are among the first impacted by AD pathology. While this may be confounded by the fact that our cohort consisted of only AD cases and controls, such strong associations of the biomarker risk score with neuropsychological status and disease severity suggest utility of the algorithm to predict decline prospectively and we are working on those statistical models currently.
As can be seen from table , a large number of the proteins included in our biomarker algorithm are inflammatory in nature, which is consistent with our initial findings [8
]. Therefore, it is possible that our biomarker algorithm is detecting a globally dysregulated inflammatory system (and other biological pathways) and future work will include non-AD disease groups (e.g. Parkinson's disease, Lewy body dementia, vascular dementia) for comparison purposes. There is a large body of literature documenting a significant link between inflammation and AD. In fact, in our prior work, we have proposed the existence of an inflammatory endophenotype of AD [8
]. Such an endophenotype may explain inconsistent findings in the biomarker literature as well as the discrepancy between epidemiological studies demonstrating a protective effect of anti-inflammatory medications against AD development [36
] and the failure of therapeutic trials using these compounds [39
There are limitations to the current study. First, while the multiplex platform we utilized is superior to individual assay (e.g. ELISA) methodologies, we have not cross-validated the blood test on a separate platform. Additionally, we have not yet incorporated non-AD dementia cases into the analyses in an effort to determine the differential diagnostic utility of the algorithm. However, our neuropsychological findings that the biomarker risk score is most strongly related to the domains of language and memory may provide initial support for the notion of discriminative ability, though such analyses must be conducted. An additional limitation is the use of a clinic-based sample and our findings need to be tested in a population-based cohort. Lastly, our study is cross-sectional in nature and does not address the utility of the algorithm in predicting AD risk. Future work will include mild cognitive impairment cases as well as longitudinal data (controls, mild cognitive impairment, and AD) in order to determine the utility of the algorithm, or possibly the need for a separate algorithm, in predicting incident risk of AD.
Overall, the current results suggest that (1) the addition of standard clinical labs to the diagnostic algorithm yields increased overall accuracy, (2) the addition of clinical labs to the 30-protein algorithm (along with demographic data) results in excellent diagnostic accuracy, and finally (3) the biomarker risk score is a significantly associated with neuropsychological test scores, particularly memory and language function. While we must still apply our algorithm to an independent cohort, these analyses provide further support for our serum-based diagnostic algorithm for detecting AD.