Prognostic Power? Do the Plasma Biomarkers,Neurofilament Light and Phospho-Tau 181,Improve Prediction of Progression to Alzheimer’s Disease Using a Machine Learning Approach in the ADNI Cohort?

The need for effective, minimally-invasive and low-cost biomarkers which can accurately diagnose Alzheimer’s disease and track disease progression, has never been more acute. As access to potentially disease modifying therapies [1, 2] is growing, the necessity for sensitive and specific biomarkers for both diagnosis and prognosis, are needed to guide widespread use of emerging therapeutics. Plasma biomarkers have promise to meet the need for a precise understanding of the coordinated changes in disease associated proteins while avoiding invasive and resource consuming procedures such as cerebrospinal fluid (CSF) collection and brain imaging. However, a better grasp of the predictive power of biomarker levels as well the biological pathways underlying the changes in biomarker levels remains necessary.

A recent study published in the Journal of Alzheimer’s Disease [3] used data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) to train machine learning Random Forests algorithm to classify cognitively stable and participants with dementia diagnosis at a 2-year follow-up. These models were used to predict progression to dementia in mild cognitive impairment (MCI) individuals after 2 years. The reference model included demographics and APOE ɛ4 status (DA) and advanced models included MRI based hippocampal volume, CSF-derived biomarkers (A ₄₂, tau, and p-tau), and plasma-derived biomarkers (p-tau 181) and neurofilament light (NFL). The biomarkers were included in the reference model and compared to the predictive value of including neuropsychological tests in the model.

The reference model, which included only demographics and APOE ɛ4 status, was improved by inclusion of plasma p-tau 181 and NFL from AUC 0.69±0.02 to AUC 0.75±0.03. However, when CSF biomarkers Aβ₄₂, tau, and p-tau were included in the model, the AUC was improved from the basic model (AUC 0.81±0.02) but only nominally improved by the addition of plasma p-tau 181 and NFL (AUC 0.83±0.03). Finally, after inclusion of adjusted hippocampal volume and CSF biomarkers in the model, addition of plasma biomarkers nominally reduced the AUC from 0.90±0.01 to 0.89±0.02 indicating that inclusion of plasma p-tau 181 and NFL added no benefit to models which already included CSF biomarkers (Aβ₄₂, tau, and p-tau) and adjusted hippocampal volume.

Plasma biomarkers, NFL and p-tau181, showed as much benefit as the CSF biomarkers Aβ₄₂, tau, and p-tau, for prediction of conversion to dementia in the MCI population with an improvement over the prediction with the reference model which included only demographics and APOE ɛ4 status (AUC of 0.54±0.02 improved to 0.64±0.03 for both models). However, when CSF biomarkers and/or hippocampal volume were included in the model, the improvement by further including plasma NFL and p-tau181 resulted in a nominal improvement in the case of the model with CSF biomarkers alone and no improvement in the case of hippocampal volume and CSF biomarkers Aβ₄₂, tau, and p-tau (Table 3). These analyses were done in a population which included both Aβ–and Aβ+ as assessed by PET scan. Interestingly, analysis of the Aβ-positive subpopulation (MCI-Aβ+) showed only nominally improved prediction metrics for each of the models. In fact, inclusion of plasma biomarkers p-tau 181 and NFL nominally reduced the predictive value of the model which included CSF biomarkers Aβ₄₂, tau, and p-tau.

These models were used to predict progression to dementia independently of neuropsychological scores which have commonly been included in previous diagnostic and prognostic models. Using these combined biomarkers independently of neuropsychological scores, their model is within the range, but still lower than the use of neuropsychological assessments currently used for diagnostics which have a predictive AUC value of 0.95 in the training model. Despite ease of procuring, and availability of plasma biomarkers, CSF biomarkers and imaging biomarkers remain clinically invasive or laborious procedures which may be less desirable than neuropsychological assessments. Given the necessity for CSF biomarkers and hippocampal volume in these models, the improvement by inclusion of plasma p-tau 181 and NFL is modest in this study.

A more complete analysis would require a model with additional plasma biomarkers including plasma Aβ₄₂ and Aβ₄₀ [4–6] as well as glial fibrillary acidic protein (GFAP) [7], p-tau 217 [8], and p-tau 231 [9], which were not included in these analyses. This may be due to limited availability of these data in the ADNI dataset. Therefore, achieving harmonization of plasma, CSF, and imaging biomarkers across ADNI sites continues to be critical to address in the ADNI cohort as well as other populations with more diversity.

The analyses presented in this paper may suggest disease progression independence from amyloid status. The outcome in this study design, progression to dementia, does not depend on amyloid pathology detected by PET since both the cognitively stable and MCI groups harbor individuals with both Aβ–and Aβ+PET status and prediction metrics were not significantly improved in a sub-population of only those with Aβ+PET scan status. The threshold used for establishing Aβ positivity was determined at amyloid PET standardized uptake values ratio [10] >1.11. However, similar results were obtained with a higher cut-off value for Aβ positivity as shown in Supplementary Table 5. These analyses highlight that amyloid independent disease mechanisms may underly the critical biology leading to disease progression. However, it is not clear from these studies if inclusion of additional plasma biomarkers including p-tau 217, p-tau231, GFAP, Aβ₄₂, and Aβ₄₀ would improve the model. As data accumulate for plasma biomarkers, further analyses are needed. Ultimately, accessibility of non-amyloid plasma biomarkers such GFAP, p-tau 217, and p-tau 231, as well as emerging biomarkers including lipidomic and metabolomic features, may provide new insight into amyloid independent mechanisms of disease as well as prognostic value.

In the context of multiple studies which have analyzed the ADNI cohort, this study highlights the performance of plasma p-tau 181 and NFL in conjunction with CSF and imaging biomarkers. It remains to be determined if this model can be improved by including additional plasma biomarkers such as Aβ₄₂, Aβ₄₀, GFAP, p-tau 217, and p-tau 231 as these harmonized data become available. Further analyses with a broadened biomarker profile should also be explored including lipidomic and metabolomic data for insight into prognostic capabilities of a shift in profiles of multiple biomarkers.