The Bayesian approach for real-world implementation of plasma p-tau217 in tertiary care memory clinics in Thailand

Abstract

Introduction: Plasma phosphorylated tau (p-tau)217 is a promising biomarker for Alzheimer's disease (AD) diagnosis, but its clinical implementation remains challenging. We propose a strategy based on Bayes' theorem and test it in real-life memory clinics.

Methods: Memory clinic patients were evaluated by neurocognitive specialists for prespecified diagnosis and subsequently underwent blood collection for p-tau217, cerebrospinal fluid, or amyloid positron emission tomography. Using cross-validation, the Bayesian approach (pretest probability × individualized likelihood ratio) was compared to other models for AD diagnosis.

Results: The Bayesian strategy demonstrated an area under the receiver operating characteristic curve (AUC) of 0.98 (95% confidence interval [CI]: 0.96-1.0), significantly outperforming multivariable logistic regression (p-tau217, age, apolipoprotein E; AUC 0.95, p = 0.024) and p-tau217 alone (AUC = 0.94, p = 0.007). When applying the two-threshold approach, the Bayesian strategy yielded an accuracy of 0.94 (95% CI: 0.88-1.0) without requiring confirmatory tests in 62.9% of the iterations.

Discussion: The Bayesian strategy offers an effective and flexible approach to address the limitations of plasma p-tau217 in clinical practice.

Highlights: Incorporating pretest probability into the interpretation of plasma phosphorylated tau (p-tau)217 improves the diagnostic performance significantly. The strategy could obviate the need for confirmatory testing in most of the patients. Plasma p-tau217 proves useful as a biomarker for Alzheimer's disease in low- and middle-income country such as Thailand.

Keywords: Alzheimer's disease; Bayes’ theorem; diagnosis; low–middle income countries; plasma biomarkers; plasma phosphorylated tau.

FIGURE 1

The Bayesian approach. Using the probability density function, two Gaussian distributions were generated from log‐transformed plasma p‐tau values obtained from the training set, one for AD and another for non‐AD participants. These Gaussian distributions enable the likelihood estimation for new observations, such as those in the testing set, allowing for the calculation of individualized likelihood ratios. The product of the pretest odds of amyloid positivity/AD pathology and individualized likelihood ratio represents the posttest odds of AD after considering plasma p‐tau results. The former term is obtained from the prevalence of amyloid abnormality/AD pathology for a given age group across the AD clinical spectrum or clinical syndrome reported in the literature. Subsequently, the posttest odds were converted back to individualized predicted probabilities. AD, Alzheimer's disease; LR, likelihood ratio; MCI, mild cognitive impairment; p‐tau, phosphorylated tau.

FIGURE 2

The flow diagram of participant enrollment. AD, Alzheimer's disease; CSF, cerebrospinal fluid; PET, positron emission tomography.

FIGURE 3

The pooled receiver operating characteristic curves of each implementation strategy: the Bayesian approach, median AUC: 0.98, 95% CI: 0.96–1.0; multivariable logistic regression (plasma p‐tau217, age, and APOE status), median AUC: 0.95, 95% CI: 0.91–0.98; and p‐tau217 alone, median AUC: 0.94, 95% CI: 0.90–0.98 (A). The predicted probabilities from one of the iterations: data points from the testing set are shown relative to the thresholds from the corresponding training set (B). AD, Alzheimer's disease; APOE, apolipoprotein E; AUC, area under the receiver operating characteristic curve; CI, confidence interval; p‐tau, phosphorylated tau.

FIGURE 4

An online calculation tool available at https://trceid.org/diagAD. For example, a 68‐year‐old with mild cognitive impairment and a plasma p‐tau217 level of 13.3 pg/mL would have a 90% probability of biomarker‐defined AD, compared to the 50% prevalence in this age group and clinical spectrum. AD, Alzheimer's disease; p‐tau, phosphorylated tau.

FIGURE 5

Proposed workflow for implementing the Bayesian strategy for the future use of plasma p‐tau217 in memory clinics. After routine clinical evaluation and basic testing, a neurocognitive disorder specialist establishes a clinical diagnosis. Plasma p‐tau217 testing is recommended for patients with syndromes that have a high pretest probability of AD. After the test, the Bayesian posttest probability can be calculated using the online tool available at https://trceid.org/diagAD. The need for subsequent confirmatory testing will depend on both the Bayesian output values and the specific clinical decisions to be made. Note that the threshold values provided are conceptual and not intended as rigid recommendations. For syndromes with a lower prevalence of AD, p‐tau217 can be used to assess AD co‐pathology when the physician deems it helpful. Aβ, amyloid beta; AD, Alzheimer's disease; bvFTD, behavioral variant of frontotemporal dementia; CSF, cerebrospinal fluid; DLB, dementia with Lewy bodies; DMT, disease‐modifying therapy; FDG, fluorodeoxyglucose F18; MCI, mild cognitive impairment; MRI, magnetic resonance imaging; NPH, normal pressure hydrocephalus; PD, Parkinson's disease; PET, positron emission tomography; PPA, primary progressive aphasia; p‐tau217, tau phosphorylated at threonine 217; RT‐QuIC, real‐time quaking‐induced conversion; SCD, subjective cognitive decline. ^aSuch as complete blood count, kidney, liver, thyroid function, and syphilis serologies (not exhaustive). ^bVascular dementia in this context refers to cases with MRI abnormalities sufficient to cause the observed impairment, though identifying AD co‐pathology in these patients can still be useful. ^cDepending on the level of expertise, non‐fluent aphasia and logopenic aphasia can sometimes be clinically indistinguishable. ^dThis also applies to other future DMTs with cost or safety profiles similar to those of anti‐Aβ therapy. ^eFuture DMTs with lower cost and fewer serious adverse events. ^fClinical relevance and precise application are to be established. ^gThis may also apply to mixed causes of dementia, although their pretest prevalence of AD is not examined in this study.