Abstract
Background:
While cutoffs for abnormal levels of the cerebrospinal fluid (CSF) biomarkers amyloid-β 1-42 (Aβ142), total tau (t-tau), phosphorylated tau (p-tau), and the ratios of t-tau/Aβ142 and p-tau/Aβ142, have been established in Alzheimer’s disease (AD), biologically relevant cutoffs have not been studied extensively in Parkinson’s disease (PD).
Objective:
Assess the suitability and diagnostic accuracy of established AD-derived CSF biomarker cutoffs in the PD population.
Methods:
Baseline and longitudinal data on CSF biomarkers, cognitive diagnoses, and PET amyloid imaging in 423 newly diagnosed patients with PD from the Parkinson’s Progression Markers Initiative (PPMI) cohort were used to evaluate established AD biomarker cutoffs compared with optimal cutoffs derived from the PPMI cohort.
Results:
Using PET amyloid imaging as the gold standard for AD pathology, the optimal cutoff of Aβ142 was higher than the AD cutoff, the optimal cutoffs of t-tau/Aβ142 and p-tau/Aβ142 were lower than the AD cutoffs, and their confidence intervals (CIs) did not overlap with the AD cutoffs. Optimal cutoffs for t-tau and p-tau to predict cognitive impairment were significantly lower than the AD cutoffs, and their CIs did not overlap with the AD cutoffs.
Conclusion:
Optimal cutoffs for the PPMI cohort for Aβ142, t-tau/Aβ142, and p-tau/Aβ142 to predict amyloid-PET positivity and for t-tau and p-tau to predict cognitive impairment differ significantly from cutoffs derived from AD populations. The presence of additional pathologies such as alpha-synuclein in PD may lead to disease-specific CSF biomarker characteristics.
INTRODUCTION
At least 75% of patients with Parkinson’s disease (PD) develop dementia long-term, greatly impacting their quality of life [1, 2]. Mild cognitive impairment (MCI) often precedes dementia in patients with PD, and about 25–30% of patients with PD without dementia experience MCI [3]. As over 4 million people worldwide are affected by PD and there are presently no effective treatments [4], the establishment of biomarkers to identify patients with a higher risk of early and rapid cognitive decline is key both in the research and clinical setting. Stratification of the diverse population of patients with early PD through the use of biomarkers would aid in the development and implementation of clinical trials for treatments targeting specific PD symptoms [5], and understanding what patients are likely to experience cognitive decline better prepares physicians for the communication of the prognosis to patients and their families [3, 6].
The underlying pathology of cognitive impairment in PD is often a coexistence of cortical Lewy body inclusions with alpha-synuclein (SYN), the most strongly correlated measure of pathology, with amyloid-β (Aβ) plaques and tau neurofibrillary tangles from Alzheimer’s disease (AD) [7]. About one third of PD patients with dementia have coexisting AD pathologies, which is correlated with increased SYN pathology and results in more rapid cognitive decline [7]. These misfolded proteins, in addition to changes in neuronal structure and activity, are thought to contribute to cognitive impairment, as the combination of the neuropathological markers of SYN, Aβ, and tau are hypothesized to work synergistically with the loss of synapses, impacted neurotransmitters, mitochondrial dysfunction, and neuroinflammation to impact cognition in PD [3]. The overlap between pathology in PD, PD with dementia (PDD), AD, and dementia with Lewy bodies (DLB), and understanding how this relates to cognitive impairment in the different diseases, is a current area of great interest.
Biomarkers that predict progression to cognitive impairment have been extensively studied in AD, and some of the findings are supported in PD. Decreased levels of amyloid beta 1-42 (Aβ142), increased levels of total tau (t-tau), and increased levels of phosphorylated tau (p-tau) in the cerebrospinal fluid (CSF) predict progression to cognitive impairment in both diseases, with Aβ142 being the strongest predictor in PD [2, 8]. Some studies show that pathologic levels of AD CSF biomarkers have poor prognostic associations with cognition in PD and DLB [8, 9]. To compare, the Aβ42/40 ratio has been shown to provide more specific differentiation of cognitive impairment in AD [10]. Recent research in the Parkinson’s Progression Markers Initiative (PPMI) cohort shows that levels of t-tau and p-tau in PD patients had minimal increase over time compared to healthy controls (HC) whose t-tau and p-tau levels consistently increased with a higher rate, indicating a possible departure from hypothesized longitudinal profiles of these biomarkers in AD [11]. The model for AD biomarkers suggests a temporal and progressive change, in which the amyloid biomarker of CSF Aβ142 becomes abnormally low first and measures of neurodegeneration, such as increased levels of CSF t-tau and p-tau, follow [12].
In AD, cutoffs for classification of abnormal levels of the three CSF biomarkers, t-tau/Aβ142, and p-tau/Aβ142 have been established for the high-precision automated Elecsys immunoassay [13, 14]. The t-tau/Aβ142 and p-tau/Aβ142 ratios have been shown to have more accurate performance than single biomarkers in AD [14], but there is limited data in Lewy body disorders (LBD). The role of Aβ and tau pathology in DLB was recently investigated [15]. Cut points for the CSF biomarkers Aβ and t-tau in LBD to predict postmortem AD and Lewy body SYN pathology were found to be lower than AD cut points and showed preliminary evidence of high accuracy to detect AD co-pathology in a relatively small rare-autopsy cohort [16]. However, direct assessments of the applicability of AD CSF biomarker cutoffs to predict AD neuropathology in living patients with PD have not yet been made. For Aβ, positron emission tomography (PET) imaging for Aβ deposition allows for a cutoff to be established using a direct measure of pathology and PET tau imaging is also recently available, but both are understudied in PD. Overall, the usage of Aβ, t-tau, p-tau, and their ratios as biomarkers contributes to the AT(N) (amyloid, tau, neurodegeneration) framework in AD, which centers the diagnosis of AD around biomarkers that are grouped into the categories of Aβ deposition, pathologic tau, and neurodegeneration [17]. Developing cutoffs for the biomarkers in PD would allow for more accurate cross-disease comparisons and solidify the shared language used in the development and testing of hypothesis about the association between CSF biomarkers and cognitive symptoms [17].
In this study, we aimed to assess the suitability and diagnostic accuracy of established AD-derived CSF biomarker cutoffs predictive of Aβ and tau pathology (in AD cohorts) in the PD population. We used two metrics (baseline and longitudinal cognitive impairment, as well as PET amyloid imaging) to examine this question. We first investigated whether binary classifications of the biomarkers based on the established AD cutoffs behave similarly to the continuous biomarker when predicting cognitive impairment in PD. If the AD cutoffs effectively predict cognitive impairment, then we expected to find that they would have similar predictive ability as the continuous measures of the CSF biomarkers, as suggested by the ATN framework. To more directly evaluate the appropriateness of applying CSF biomarker cutoffs from AD to patients with early PD, we calculated optimal cutoffs for the biomarkers based on the PPMI cohort using both cognitive testing and, in the subset of patients with amyloid imaging, in vivo detection as a proxy for plaque pathology. We then compared the magnitude and predictive ability of these optimal cutoffs to the AD cutoffs, with the goal of determining whether PD specific cutoffs should be established.
MATERIALS AND METHODS
Participants
The PPMI study is an observational multicenter study to identify PD biomarkers. The cohort consists of patients with early PD (N = 423) and HC (N = 196). All participants in the PD cohort for our analyses have a clinical diagnosis of PD and a positive dopamine transporter (DAT) SPECT. The mean [standard deviation (SD)] duration of PD diagnosis at baseline is 6.7 [6.5] months. The study aims, diagnostic criteria, methods, and subject retention are previously described [18]. The data was downloaded from PPMI database on June 1, 2020 and included baseline CSF measures and clinical data at baseline, and annual follow up visits at years 1, 2, 3, 4, and 5. All procedures were performed with prior approval from ethical standards committees at each participating institution and with informed consent from all study participants. Data used in the preparation of this article were obtained from the PPMI database (https://www.ppmi-info.org/data). For up-to-date information on the study, visit https://www.ppmi-info.org.
Clinical data and measures
Clinical data (including demographics such as age at baseline, disease duration at baseline, and sex) were acquired from the PPMI database as described previously [19]. Two different measures of cognitive impairment were used in the analysis. First, Montreal Cognitive Assessment (MoCA) scores patients on a scale of 0 to 30, adjusted for education. Those with scores of ≥26 are classified as cognitively normal, and with scores < 26 as cognitively impaired [20]. We denote this binary classification as MoCA or moca26. Second, the site investigator’s clinical diagnosis of cognitive state (Investigator Diagnosis, or inv_dx) classifies patients as having either normal cognition or MCI or PDD. We grouped MCI and PDD into cognitive impairment category when analyzing Investigator Diagnosis. Investigator diagnosis was made by the Movement Disorder Society (MDS) Task Force Level I criteria for MCI or PDD, in which provided guidelines for assessing cognitive decline, functional impairment, and interpretation of cognitive tests informed classification of patients as having normal cognition, MCI or PDD at each annual visit, starting with year 1 or 2 for most PPMI participants [19, 21].
CSF analysis
CSF biospecimens were collected at baseline (i.e., at disease onset, and prior to initiation of PD treatment) according to standardized protocol (found at https://www.ppmi-info.org) and shipped from the PPMI Biorepository Core laboratories to the University of Pennsylvania Biomarker Research Laboratory. Following from the standardized PPMI protocol, 15–20 mL of CSF fluid was collected, centrifuged at 2000×g for 10 min at room temperature, and aliquoted into 1.5 mL aliquots to be frozen and used for analysis with the assays. Levels of CSF Aβ142, t-tau, and tau phosphorylated at threonine 181 position (p-tau) were measured using Elecsys® electrochemiluminescence immunoassays on the cobas e 601 analysis platform (Roche Diagnostics) [11].
Established cutoffs for t-tau and p-tau based on the Elecsys assay of > 266 pg/ml and > 24 pg/ml were used to classify patients as t-tau and p-tau positive, while cutoffs of ≥0.27 and ≥0.025 were used to classify patients as t-tau/Aβ142 and p-tau/Aβ142 positive [14]. Pre-analytical factors in specimen collection and analysis across studies can influence CSF Aβ142 levels, making it difficult to test specific diagnostic thresholds established in AD cohorts to PPMI patients [22]. To address this issue, we converted the Elecsys values to AlzBio3 equivalents using the equation [x = (CSF Aβ142 + 251.55)/3.74] from Shaw et al. [23] to apply the established cutoff of < 250 pg/ml of AlzBio3 equivalent values (i.e., 683 pg/ml in Elecsys units) to these AlzBio3 equivalent values to classify patients as amyloid positive [13], as we have done previously [11]. This process results in a more conservative threshold (i.e., 683 pg/ml in Elecsys units) compared to direct application of Elecsys derived thresholds in AD (e.g., 977 pg/ml in Elecsys units for the ADNI cohort [24]) to help account for pre-analytical differences in CSF Aβ142 in the PPMI versus AD cohorts [22]. For clarity in the manuscript, we report Aβ142 values in Elecsys units.
PET analysis
Thirty-four patients underwent PET imaging for 18F-Florbetaben (FBB) as a convenience sample. FBB measurements of quantitative standardized uptake value ratio (SUVr) values were compared to the published cutoff for the global cerebellar reference region of ≥1.48 to classify patients as pathologically amyloid beta positive [25]. Since only 22 patients had Aβ142 measurements at the same visit at which they underwent PET imaging, the Elecsys measurement of Aβ142 within one year before or after the PET imaging was used in the analysis. Two of the 34 patients lacked Aβ142 measurements within one year of PET imaging, so they were excluded from the analysis.
Statistical analysis
We analyzed data using RStudio Version 1.3.959.
Two sample t-test was used to compare the mean values of CSF biomarkers at baseline between PD and HC. To analyze the cross-sectional relationship between the three CSF biomarkers and the two cognitive impairment outcomes at baseline, we conducted logistic regression analysis using univariate and multivariate (including age and sex as covariates) models with the biomarkers as predictors either as continuous measure or binary measures based on the established AD cutoffs. The area under the curve (AUC) of the receiver operating characteristic (ROC) curve and the standard error (SE) for the AUC was calculated.
Next, to evaluate the association between baseline CSF biomarkers and longitudinal incident cognitive impairment, we used the Generalized Estimating Equation (GEE), which fits longitudinal data and can account for missing data and correlations among repeated measures. In the model the response variable was a binary classification of cognitive impairment for each patient at each time point in the study, while the predictor variable(s) was the biomarker at baseline (either in a continuous measure or binary based on the AD cutoffs) in the univariate model and included age at baseline and sex in the multivariate model. The “pROC” R package was used to compare two ROC curves and DeLong’s test was used to determine the significance between the AUC values of the ROC curves with either continuous biomarkers or binary biomarkers [26].
Optimal cutoffs for Aβ142, t-tau/Aβ142, and p-tau/Aβ142 for the two cognitive outcomes and PET imaging as a primary analysis and t-tau and p-tau for the two cognitive outcomes as a secondary analysis were determined based on Youden’s index using the “pROC” R package [27]. The confidence interval (CI) for the optimal cutoff was obtained based on stratified bootstrapping [28–31]. Optimal cutoffs were calculated using leave-one-out cross validation to obtain a robust and generalizable optimal cutoff, and also avoid overfitting.
Using the cross validated optimal cutoffs, we re-fit the GEE and logistic regression models with the binary classification of the three biomarkers and calculated AUC values for the different cutoffs. As earlier, Delong’s test via the “pROC” R package was used to compare AUC values of the AD cutoff and the optimal cutoff.
The GEE model was also used to calculate the difference between percentages of patients classified as cognitively impaired based on MoCA or Investigator Diagnosis.
All statistical tests were two-sided. Statistical significance was set at the 0.05 level.
RESULTS
Demographics and characteristics
Demographics for PD and HC patients and baseline CSF biomarker characteristics are shown in Table 1. Based on the AD cutoffs of 683 pg/mL (i.e., 250 pg/ml in transformed Luminex equivalents) for Aβ142, 266 pg/mL for t-tau, and 24 pg/mL for p-tau, 31.5% of patients with PD were amyloid positive at baseline, while only 6.5% and 6.9% were t-tau and p-tau positive. Based on the cutoffs of 0.27 and 0.025, 12.6% and 10.7% of patients with PD were classified as t-tau/Aβ142 and p-tau/Aβ142 positive, respectively. T-tau and p-tau levels were lower in PD than HC (t = –3.4, df = 277.9, p = 0.0007 for t-tau; t = –3.9, df = 241. 4, p = 0.0001 for p-tau). Aβ142 level was lower in PD than HC (t = –2.6, df = 307.4, p = 0.009). T-tau/Aβ142 and p-tau/Aβ142 levels were not significantly lower in PD than HC (t = –0.9, df = 259, p = 0.4 for t-tau/Aβ142; t = –1, df = 220, p = 0.2 for p-tau/Aβ142). Across the five years of the study, MoCA classified a higher percentage of patients as cognitively impaired (22.0–34.4%) than Investigator Diagnosis (8.5–22.0%, Supplementary Table 1, p < 0.0001). The number of patients with PD at baseline and each follow up visit was summarized in Supplementary Table 1. Out of the 423 patients with PD, 316 patients had cognitive data as far as year 5 ( Supplementary Table 1).
Baseline demographics and characteristics of PD patients and healthy controls
PD, Parkinson’s disease; HC, healthy control; Aβ142, amyloid-beta 1-42; T-Tau, total tau; P-Tau, phosphorylated tau; FBB PET, PET amyloid imaging for 18F-Florbetaben. Data listed = mean (SD) for continuous variables and frequency (%) for categorical variables. *FBB PET was taken as a convenience sample across the five years of the study, so the values reported in Table 1 are not at baseline.
Cross-sectional CSF biomarker analysis
The logistic regression at baseline showed no significant association between the biomarkers and clinical measures of cognitive impairment at baseline (p > 0.1), except in the univariate model with MoCA as the outcome and binary (based on the AD cutoff) p-tau as the predictor (p < 0.05, Supplementary Table 2, positive p-tau, or patients with p-tau above the AD cutoff, was associated with cognitive impairment). The continuous Aβ142 was not significantly associated with the PET measurements of amyloid positivity in both the univariate (coefficient for Aβ= –1.00*10–03, SE = 1.30*10–03, p = 0.440, logistic regression) and multivariate model (coefficient for Aβ= –7.73*10–04, SE = 1.29*10–03, p = 0.547, logistic regression). The corresponding AUC (SE) were 0.585 (0.136) in the univariate model and 0.711 (0.113) in the multivariate model including age at baseline and sex.
Comparison of predictive ability in longitudinal incident cognitive impairment between binary (based on AD cutoffs) and continuous CSF biomarkers measured at baseline
In the univariate models of the longitudinal GEE analysis of the clinical measures of cognitive impairment, the AUC for the model with binary Aβ142 was significantly lower than with continuous Aβ142, although this significance is marginal due to similar AUC values (p = 0.03 for Investigator Diagnosis, p < 0.001 for MoCA, Table 2). Binary t-tau/Aβ142 and p-tau/Aβ142 resulted in significantly lower AUC values using MoCA (p < 0.03), but not Investigator Diagnosis (p > 0.05, Table 2), as the measure of cognitive impairment. These significances are also marginal due to similar AUC values. There was no significant difference between the AUC for binary t-tau and continuous t-tau for both cognitive outcomes in the models (p > 0.1), and likewise for binary p-tau and continuous p-tau for MoCA and Investigator Diagnosis (p > 0.1, Table 2).
AUC values of the longitudinal GEE analysis when differentiating abnormal and normal cognition over time using continuous CSF biomarkers at baseline and binary biomarkers at baseline based on the AD cutoffs
moca26, Montreal Cognitive Assessment, below 26 cognitively impaired; inv_dx, site investigator diagnosis of cognitive state; univariate, univariate model; continuous, continuous measure of biomarker; bin., binary measure of biomarker based on the AD cutoffs provided in Table 1; SE, standard error.
Table 3 presents the regression coefficients with their standard errors and p values of the longitudinal GEE analyses. Lower or abnormal Aβ142 was significantly associated with longitudinal cognitive impairment measured by Investigator Diagnosis in the univariate and multivariate models in both the continuous and binary form (p < 0.01). Aβ142 was not significantly associated with MoCA over time (p > 0.09). Both t-tau/Aβ142 and p-tau/Aβ142 were significantly associated with longitudinal cognitive impairment as measured by Investigator Diagnosis in all models (p < 0.001), while they are significantly associated with MoCA in the univariate (p < 0.03) but not multivariate models (p > 0.1). T-tau was not significantly associated with either of the longitudinal cognitive outcomes in any of the models (p > 0.06). Higher or abnormal p-tau was significantly associated with longitudinal cognitive impairment measured by both longitudinal cognitive outcomes in the univariate model in both the continuous and binary form (p < 0.05) and abnormal p-tau was additionally associated with longitudinal cognitive impairment of MoCA in the multivariate model when in the binary but not continuous form (p < 0.05).
Regression coefficients and P values of biomarkers in the longitudinal GEE analysis when differentiating abnormal and normal cognition over time using continuous and binary (based on AD cutoffs) CSF biomarkers at baseline
moca26, Montreal Cognitive Assessment, below 26 cognitively impaired; inv_dx, site investigator diagnosis of cognitive state; univariate, univariate model; multivariate, multivariate model which included biomarker, age at baseline, and sex as independent variables; cont., continuous biomarker; bin., binary biomarker based on the AD cutoffs provided in Table 1; coefficient, regression coefficient of the biomarker; SE, standard error of the regression coefficient.
Optimal cutoffs
Primary analysis
Table 4 displays the optimal cutoffs for Aβ142 based on leave-one-out cross validation for the clinical cognitive outcomes and PET imaging. For the calculated optimal cutoffs, patients with levels less than or equal to the cutoff are classified as having abnormal levels. The optimal cutoffs calculated for Aβ142 based on predicting the longitudinal cognitive measures were higher than the AD cutoff of 683 [710 (CI: 635, 725) for Investigator Diagnosis and 1162 (CI: 523, 1338) for MoCA], but 683 was within the CIs for the optimal cutoffs. The optimal cutoff for Aβ142 based on predicting PET imaging of amyloid positivity was 945 (CI: 927, 1057), and the AD cutoff of 683 was not within this CI for the optimal cutoff, which indicates that the optimal cutoff is significantly higher than the AD cutoff. There was no significant difference between the AUCs using the AD cutoff versus the optimal cutoff (p > 0.1) when predicting PET amyloid imaging positivity or the longitudinal cognitive measures.
Optimal cutoffs and AUC values for Aβ142 when differentiating abnormal and normal cognition over time or PET-amyloid (in vivo measure of plaque pathology). Comparison of AD cutoffs to optimal cutoffs and AUC values
inv_dx, site investigator diagnosis of cognitive state; moca26, Montreal Cognitive Assessment, below 26 cognitively impaired; Optimal cutoff, leave-one-out, optimal cutoffs calculated using leave-one-out cross validation; SE, standard error; CI, confidence interval. Multivariate model included biomarker, age at baseline, and sex as independent variables.
As shown in Table 5, the optimal cutoffs for t-tau/Aβ142 and p-tau/Aβ142 were all lower than the AD cutoffs of 0.27 and 0.025. For the calculated optimal cutoffs, patients with levels greater than or equal to the cutoff are classified as having abnormal levels. While the optimal cutoffs for t-tau/Aβ142 calculated with Investigator Diagnosis [0.215 (CI: 0.191, 0.344)] and MoCA [0.219 (CI: 0.160, 0.318)] were lower than the AD cutoff of 0.27, their CIs overlapped with the AD cutoff. When PET imaging of amyloid positivity was used as the gold standard for AD pathology, the optimal cutoff of 0.178 (CI: 0.136, 0.259) was significantly lower than the AD cutoff of 0.27 and the AD cutoff was not within the CI for the optimal cutoff. For p-tau/Aβ142, the optimal cutoffs calculated with Investigator Diagnosis [0.0179 (CI: 0.0164, 0.0292)] and MoCA [0.0158 (CI: 0.0134, 0.0290)] again were lower than the AD cutoff of 0.025 but their CIs overlapped with the AD cutoff. However, PET amyloid imaging resulted in a significantly lower cutoff of 0.0187 (CI: 0.0122, 0.0233) than the AD cutoff of 0.025 and the AD cutoff was not within the CI for this optimal cutoff. There was no significant difference in AUC values using the AD cutoffs versus the optimal cutoffs for t-tau/Aβ142 (p > 0.05) and p-tau/Aβ142 (p > 0.05).
Optimal cutoffs and AUC values for t-tau/Aβ142 and p-tau/Aβ142 when differentiating abnormal and normal cognition over time or PET-amyloid. Comparison of AD cutoffs to optimal cutoffs and AUC values
inv_dx, site investigator diagnosis of cognitive state; moca26, Montreal Cognitive Assessment, below 26 cognitively impaired; Optimal cutoff, leave-one-out, optimal cutoffs calculated using leave-one-out cross validation; SE, standard error, CI, confidence interval. Multivariate model included biomarker, age at baseline, and sex as independent variables.
Secondary analysis
Table 6 shows the optimal cutoffs and AUC comparisons for t-tau and p-tau, calculated using the clinical cognitive outcomes. For the calculated optimal cutoffs, patients with levels greater than or equal to the cutoff are classified as having abnormal levels. The optimal cutoffs for t-tau and p-tau were all significantly lower than the AD cutoffs of 266 and 24, respectively. Specifically, the optimal t-tau cutoffs of 112 (CI: 84.7, 128) and 148 (CI: 135, 223) for Investigator Diagnosis and MoCA were lower than the AD cutoff of 266 and 266 was not within the CIs for the optimal cutoffs. For p-tau, the optimal cutoffs of 17.6 (CI: 13.0, 21.0) and 13.0 (CI: 11.1, 19.7) were lower than the AD cutoff of 24 and 24 was not within the CIs for the optimal cutoffs. There was no significant difference in the AUC values for the AD cutoff versus the optimal cutoff for t-tau (p > 0.2) and p-tau (p > 0.05).
Optimal cutoffs and AUC values for t-tau and p-tau when differentiating abnormal and normal cognition over time. Comparison of AD cutoffs to optimal cutoffs and AUC values
inv_dx, site investigator diagnosis of cognitive state; moca26, Montreal Cognitive Assessment, below 26 cognitively impaired; Optimal cutoff, leave-one-out, optimal cutoffs calculated using leave-one-out cross validation; SE, standard error; CI, confidence interval. Multivariate model included biomarker, age at baseline, and sex as independent variables.
DISCUSSION
In this study, we evaluated the appropriateness of applying CSF biomarker cutoffs from AD to predict cognitive decline and cerebral amyloidosis among patients with early PD. Our research question asked whether established AD-derived CSF biomarker cutoffs predictive of Aβ and tau pathology in AD cohorts were diagnostically accurate in the PD population. We recognize that cognitive impairment and dementia in PD is due to heterogenous underlying neuropathological substrates, including AD pathology [7]. Thus, we addressed this question by testing biomarker cutoffs established in AD to predict both non-disease specific clinical diagnosis of cognitive impairment and a biological measure of AD (i.e., gold-standard PET amyloid positivity) in patients with early PD. We acknowledge that several neuropathophysiological factors can lead to cognitive impairment in PD, but postmortem studies show that AD pathology is a strong and significant driver of cognitive impairment in PD [7] and thus important to include in our study. Importantly, our findings suggest that the cutoffs for CSF biomarkers to predict both cognitive impairment and PET amyloid in this PD cohort are significantly different from those established in AD patients, indicating that biological factors linked to PD and underlying alpha synucleinopathy likely influence AD CSF biomarker analytes. We contend that PD-specific cutoffs to predict cognitive decline and cerebral amyloidosis for use in PD cohorts need to be defined and independently validated for future use in research and eventually clinical care.
Using logistic regression models, we found that Aβ142, t-tau, and p-tau at baseline were not significantly associated with cognitive impairment at baseline in almost all situations. This lack of significance was unsurprising given that the biomarkers have been previously shown to predict longitudinal, not baseline, cognitive impairment in PD, and the PPMI cohort contains newly diagnosed patients with PD (78% participants were cognitively normal based on MoCA at baseline) [2, 8]. This led us to focus on the relationship between baseline CSF biomarkers and longitudinal incident cognitive impairment.
In the longitudinal GEE model, the AUC of binary Aβ142 was significantly lower than continuous Aβ142 (Table 2), indicating that some information is lost when Aβ142 is converted from a continuous measure to a binary measure based on the AD cutoff. There was, however, no significant difference between the AUC of binary and continuous t-tau and p-tau. We also found that Aβ142 was significantly associated with longitudinal Investigator Diagnosis in all situations (univariate, multivariate, continuous, binary), t-tau had no significant associations with either of the longitudinal cognitive outcomes, and p-tau was significantly associated with the longitudinal cognitive outcomes when it was in the univariate but not multivariate model (Table 3). This indicates that Aβ142 is a stronger predictor of progression to cognitive impairment than t-tau and p-tau, agreeing with previous studies in patients with PD [6, 8]. Since t-tau and p-tau levels were found to be lower in patients with PD than in HC (Table 1) while levels of t-tau and p-tau are higher in patients with AD than in HC, t-tau and p-tau may act fundamentally different in patients with PD than in patients with AD [32]. Cross-sectional studies at baseline show that drug naïve PD has lower p-tau and t-tau compared to HC on a group-level [33], and previous research in the PPMI cohort has found that t-tau and p-tau levels only mildly increased around years 2–3 of the study in PD and to a lesser extent than observed in HC [11]. Thus, PD pathology may partially suppress the increasing levels of t-tau and p-tau seen in AD and aging [11]. It is important to note these data are on a group level and there is individual patient heterogeneity. The high neuropathophysiological heterogeneity of cognitive impairment with PD, where only a proportion of patients have comorbid AD-like pathologies, means that individuals may not have higher p-tau and t-tau levels compared to PD without AD co-pathology, but there is limited PET amyloid data in the current study to test this in more detail. The mechanism for lower observed t-tau and p-tau levels in PD vs. HC is not entirely clear, so mechanistic work in model systems is needed to further elucidate these findings. Additionally, p-tau was more frequently significantly associated with the cognitive outcomes when in the binary rather than continuous measure, demonstrating that it may predict progression to cognitive impairment in a binary method.
After evaluating the predictive ability of the five CSF biomarker cutoffs from AD, we calculated optimal cutoffs based on the PPMI cohort. We focused on cutoffs for Aβ142, t-tau/Aβ142, and p-tau/Aβ142 as a primary analysis since we have gold-standard PET imaging for amyloid positivity, and we completed a secondary analysis of t-tau and p-tau with cognitive measures as the outcomes as the PPMI cohort lacks tau imaging. As the optimal cutoffs for Aβ142 were higher and for t-tau/Aβ142, p-tau/Aβ142, t-tau, and p-tau were lower than the AD cutoffs, the optimal cutoffs classify more patients as having abnormal biomarker levels (Tables 4–6). This looser criteria for abnormal levels of CSF biomarkers in patients with PD may account for co-pathology of Aβ plaques and neurofibrillary tangles with alpha-synuclein inclusions underlying cognitive impairment in patients with PD [3], but the mechanisms for this potential interaction and influence on AD biomarker levels in LBD are unclear.
We found that the CI of the optimal cutoffs for Aβ142, using either of the cognitive outcomes, overlapped with the AD cutoff of 683 pg/ml, and that there was no significant difference in AUC values using the different cutoffs (Table 4). Although the CI of these cutoffs overlapped with the AD cutoff, for MoCA specifically the CI is quite wide, indicating a high variability in the association between Aβ142 values and MoCA scores. The similar AUC values using the optimal (based on cognitive outcomes) and AD cutoffs correspond to the finding that the cutoffs themselves did not differ significantly. With PET imaging of the pathology of amyloid positivity, the optimal cutoff was again above the AD cutoff, but the CI did not overlap with the AD cutoff. This indicates that the optimal cutoff is significantly higher than the AD cutoff. While the AUC values using the optimal cutoff were not significantly different from those of the AD cutoff possibly due to low statistical power, the AUC improvement is large (for example, in the univariate model for Aβ142 there was a 32% increase in the AUC when using the optimal cutoff, from 0.515 with the AD cutoff to 0.678 using the PPMI-derived optimal cutoff, Table 4). This significantly higher cutoff derived from PET imaging may indicate a difference between Aβ deposition in AD and PD. Our findings are in the opposite direction to the lower threshold found of Aβ (compared to the AD cutoff) for predicting postmortem plaque pathology in LBD [16]. For the ratios of t-tau/Aβ142 and p-tau/Aβ142, we found that PET imaging, but not Investigator Diagnosis or MoCA, resulted in optimal cutoffs which were significantly different from the AD cutoffs of 0.27 and 0.025, respectively (Table 5). With all optimal cutoffs, the AUC values were not significantly different than those with the AD cutoffs. Similarly to Aβ142, this can be explained by the small sample size of patients with PET imaging and the similarity of cutoffs derived using MoCA and Investigator Diagnosis. PET imaging, which evaluates the pathology of Aβ deposition, is a more accurate measure of amyloid pathology over clinical observations of cognition since clinical features can be influenced by a wide range of factors, and it identifies more mature plaques than autopsy samples. However, in our study, the small sample size for PET imaging limits our ability to extrapolate these findings, so calculating an optimal cutoff for Aβ deposition in PD should be replicated in a larger cohort. Differences between postmortem pathology and the more mature plaques detected by PET amyloid may explain our higher optimal cutoff. And although PET imaging for FBB is linked to CERAD pathology (neuritic plaque density) level at autopsy and the cutoff should be the same regardless of diagnosis [25], imaging-to-autopsy data for FBB in PD patients is limited.
In our secondary analysis, we found that the CI of the optimal cutoffs for t-tau and p-tau did not include the AD cutoffs of 266 pg/ml and 24 pg/ml, respectively, indicating significant difference (Table 6). But, there was no significant difference in the AUC values using the AD versus optimal cutoffs. We may have found significantly different cutoffs but similar AUC values because the concentration zones around the cutoffs for t-tau and p-tau are relatively low, meaning that few patients have t-tau and p-tau levels close to the cutoffs and that there wouldn’t be a large difference in the number of patients classified as having abnormal levels when using the different cutoffs [14].
Our main limitations in our study include the small sample size of patients with PET imaging data (32 patients), as well as the lack of in vivo measurements of tau pathology. As PET imaging is a more accurate way of measuring amyloid positivity, our results should be replicated with a larger sample size of PET data to confirm that the optimal cutoffs of Aβ142 for patients with PD are significantly higher than in patients with AD and that the optimal cutoffs of t-tau/Aβ142 and p-tau/Aβ142 for patients with PD are significantly lower than in patients with AD. We also lack a in vivo measure of tau pathology. This measure would give a more precise evaluation of the ability for t-tau and p-tau at baseline to predict cognitive impairment/decline and tau pathology in PD. CSF t-tau is shown to be associated with PET tau as measured by 18F-flortaucipir retention, indicating that the method would be valuable in confirming cutoffs for t-tau’s predictive ability [34]. Our study is additionally limited by the small number of patients classified as cognitively impaired, which did not change much over the study after a jump in year 1 ( Supplementary Table 1). We note that even though the change in MoCA over time is small, it is significant [35], and our longitudinal statistical models incorporate all time points into the analysis. As the PPMI cohort enrolls newly diagnosed PD patients, has a relatively young age at disease onset and high education levels, and currently includes relatively short follow up, higher magnitudes of longitudinal change are expected to come as patients’ disease progresses [11]. There are missing data in cognitive measures partially due to dropouts in the study (from 423 participants at baseline to 316 participants at year 5), which also impacts this small percentage which is classified as cognitively impaired ( Supplementary Table 1). Finally, we cannot account for all possible pre-analytical factors between AD and PPMI clinical studies which could influence CSF Aβ levels and resultant discrepancies between optimal cutoffs observed here [22]. For example, CSF concentrations can be affected by factors including the type of aliquot tube it is stored in and plasma contamination, making the normalization of collection protocols key in future CSF research [36].
In summary, we evaluated the applicability of AD cutoffs for the CSF biomarkers of Aβ142, t-tau, p-tau, t-tau/Aβ142, and p-tau/Aβ142 to predict cognitive decline and cerebral amyloidosis in patients with early PD. Based on PET imaging of amyloid positivity the optimal cutoff of Aβ, t-tau/Aβ142, and p-tau/Aβ142 for the PPMI cohort may be significantly different than the AD cutoff, and the cognitive-outcome-based optimal t-tau and p-tau cutoffs may also be different for patients with PD. This study provides new insight into the cross disease comparison of biomarkers with AD, supporting evidence that the presence of SYN in PD and other LBD leads to a distinct behavior in CSF biomarkers from that in AD [16]. As this research has highlighted, updating and establishing cutoffs for these biomarkers across PD cohorts is a key step in aiding in the assignment of patients with early PD to clinical trials based on their disease course and working towards a deeper understanding of cognitive impairment in neurodegenerative diseases.
Footnotes
ACKNOWLEDGMENTS
This work was supported by funding from U.S. National Institute of Health grants R01-NS102324, AG072979, and AG062418. PPMI, a public-private partnership, is funded by the Michael J. Fox Foundation for Parkinson’s Research funding partners 4D Pharma, Abbvie, Acurex Therapeutics, Allergan, Amathus Therapeutics, ASAP, Avid Radiopharmaceuticals, Bial Biotech, Biogen, BioLegend, Bristol-Myers Squibb, Calico, Celgene, Dacapo Brain Science, Denali, The Edmond J. Safra Foundaiton, GE Healthcare, Genentech, GlaxoSmithKline, Golub Capital, Handl Therapeutics, Insitro, Janssen Neuroscience, Lilly, Lundbeck, Merck, Meso Scale Discovery, Neurocrine Biosciences, Pfizer, Piramal, Prevail, Roche, Sanofi Genzyme, Servier, Takeda, Teva, UCB, Verily, and Voyager Therapeutics.
CONFLICTS OF INTEREST
SW, DJI, and PZ report no disclosures.
DW serves on the steering committee for the PPMI study and receives grant funding from NIH grant AG062418.
LMS receives the following research support: NIH/NIA, U01 AG024904ADNI Biomarker Core-QC and analyses for biofluid AD biomarkers in CSF and plasma; P30AG010124 Penn ADRC Biomarker Core-CSF and plasma based AD biomarker studies; MJFox Foundation for Parkinson’s Research, grant #13637, CSF Biomarkers in PD; Roche IIS AD CSF biomarker studies. Also receives the following teaching/honorarium: Biogen teaching program on fluid biomarkers; Fujirebio teaching program on fluid biomarkers.
AS has been a consultant to the following companies in the past year: Biogen, Wave Life Sciences, Prevail, Bial Biotech and Takeda. He has served on DSMBs for the Huntington Study Group and The Healey ALS Consortium (Massachusetts General Hospital). He has received grant funding from the Michael J. Fox Foundation, NIA and NINDS.
SXX receives grant support from NIH grants R01-NS102324, AG10124, and AG062418.
