A prediction model to calculate probability of Alzheimer's disease using cerebrospinal fluid biomarkers

Abstract

Background: We aimed to develop a prediction model based on cerebrospinal fluid (CSF) biomarkers, that would yield a single estimate representing the probability that dementia in a memory clinic patient is due to Alzheimer's disease (AD).

Methods: All patients suspected of dementia in whom the CSF biomarkers had been analyzed were selected from a memory clinic database. Clinical diagnosis was AD (n = 272) or non-AD (n = 289). The prediction model was developed with logistic regression analysis and included CSF amyloid β42, CSF phosphorylated tau181, and sex. Validation was performed on an independent data set from another memory clinic, containing 334 AD and 157 non-AD patients.

Results: The prediction model estimated the probability that AD is present as follows: p(AD) = 1/(1 + e (- [-0.3315 + score])), where score is calculated from -1.9486 × ln(amyloid β42) + 2.7915 × ln(phosphorylated tau181) + 0.9178 × sex (male = 0, female = 1). When applied to the validation data set, the discriminative ability of the model was very good (area under the receiver operating characteristic curve: 0.85). The agreement between the probability of AD predicted by the model and the observed frequency of AD diagnoses was very good after taking into account the difference in AD prevalence between the two memory clinics.

Conclusions: We developed a prediction model that can accurately predict the probability of AD in a memory clinic population suspected of dementia based on CSF amyloid β42, CSF phosphorylated tau181, and sex.