According to national recommendations, CSF analyses are routinely used during testing of patients with cognitive disorders in French memory centers [4,17]. Patients were recruited between 1 September 2012 and 1 September 2013 from three French clinical and research memory centers specializing in the care of patients with cognitive disorders. These three centers are based in Paris, Lille and Montpellier (the PLM Study) [18-20]. All patients who were prospectively included in the study had cognitive disorders and received a lumbar puncture. Patients with unknown clinical diagnoses were excluded. Patients with mild cognitive impairment (MCI) represented a heterogeneous group, including patients with and without AD-related brain lesions, and thus were excluded from this study.
All patients had a thorough examination, including clinical and neuropsychological evaluations, biological measurements and brain imaging. Patients with AD were included according to the criteria for probable AD as defined by the National Institute on Aging-Alzheimer’s Association . On the basis of all available elements, which included the results for CSF biomarkers, patients were classified into two groups: AD and non-AD. Complex or unclear cases were discussed, and diagnoses were made by a multidisciplinary team of neurologists, geriatricians and neuropsychologists. Non-AD subjects included subjects with cognitive disorders other than AD (see Additional file 1), such as frontotemporal dementia, Lewy body disease, Parkinson’s disease, Creutzfeldt-Jakob disease and non-degenerative dementia (that is vascular dementia, alcohol-related dementia, normal pressure hydrocephalus, infectious disease and psychiatric disorders) .
This research study was approved by the local ethics committees of each hospital (see Additional file 2). All patients agreed to CSF collection, assessments and analyses by providing a written informed consent.
Cerebrospinal fluid analysis
Lumbar punctures were performed on fasting patients within 1 month following their clinical diagnosis, usually between 9:00 am and 12:00 pm. In an effort to reduce the intersite variability in CSF readings previously reported , all three centers used a common 10-ml polypropylene tube to collect the CSF (catalog number 62.610.201; Sarstedt, Nümbrecht, Germany) . Each CSF sample was transferred at 4°C to the corresponding local laboratory within 4 hours after collection and was then centrifuged at 1,000 g for 10 minutes at 4°C. A small amount of CSF was used to perform routine analyses, including total cell count, bacteriological examination and total protein and glucose levels. The CSF was aliquoted in 0.5-ml polypropylene tubes and stored at −80°C to await further analysis. CSF Aβ40, Aβ42, total tau and p-tau181 were measured in each local laboratory using a commercially available sandwich enzyme-linked immunosorbent assay (INNOTEST; Fujirebio Europe NV, Gent, Belgium) according to the manufacturer’s instructions. Concentrations of total tau greater than 1,200 pg/ml were well above the detection limit and thus were not recalculated after dilution because of the constraints in the procedure. The biological teams involved in CSF analyses were blinded to the clinical diagnoses.
Cerebrospinal fluid cutoff determination and interpretation
CSF cutoffs used in the analyses were determined on the basis of the population included in the present study. We computed both overall and local cutoff receiver operating characteristic (ROC) curves using STATA software (v10, StataCorp, College Station, TX, USA). ROC curves were built for each biomarker by plotting sensitivity and 1 − specificity to discriminate AD from non-AD patients. Optimum cutoff values were determined using two methods: the highest Youden index (that is, sensitivity + specificity − 1) and lowest distance between ROC plot and point (0.1). We checked that both approaches led to similar findings, and the presented results correspond to the highest Youden index. STATA code used for the determination of optimum cutoff is detailed in the Additional file 3.
CSF results were defined as negative (that is, CSF p-tau181 below cutoff and CSF Aβ42 above cutoff) or positive (that is, CSF p-tau181 above cutoff and CSF Aβ42 below cutoff). An indeterminate CSF profile was defined by the association of one positive biomarker and one negative biomarker.
We chose to evaluate tau pathology based on CSF p-tau181 levels alone because p-tau181 and total tau levels were highly correlated in our sample (Pearson’s correlation coefficient = 0.87). Furthermore, p-tau181 has previously been reported to be more discriminant than total tau . Analyses based on total tau led to findings and conclusions similar to those obtained with p-tau181.
The subjects’ characteristics are presented overall, by diagnosis (AD vs. non-AD) and by CSF collection centers. The various groups (diagnoses, centers) were compared using analysis of variance for continuous variables and χ2 test for categorical variables.
We computed ROC curves to evaluate the capacity of each CSF biomarker to discriminate AD from non-AD subjects. Analyses were performed in the overall study population and then stratified by centers. Optimum cutoffs for each biomarker were defined using the highest Youden index to discriminate in the best possible manner AD from non-AD subjects, thus maximizing sensitivity and specificity of the analyses based on ROC curve analyses. Then we used local optimum cutoffs for further analysis because intersite differences were present.
To evaluate the interest of ratio Aβ 42/40 versus CSF Aβ42 alone, we compared three methods to interpret CSF results:
Method 1: interpretation based on CSF p-tau181 and on CSF Aβ42
Method 2: interpretation based on CSF p-tau181 and on Aβ 42/40 ratio
Method 3: a decisional algorithm based, in a first step, on CSF p-tau181 and on CSF Aβ42. Then, in case of discrepancy between p-tau181 and Aβ42, we used, in a second step, the Aβ 42/40 ratio in place of Aβ1–42
For all three methods, we first determined the percentage of AD, non-AD and indeterminate biological CSF profiles. Proportions of indeterminate profiles according to the methods were compared using the McNemar test. Then we used a net reclassification improvement (NRI) method to compare method 2 to method 1, and method 3 to method 1, among AD patients and non-AD patients (based on clinical diagnosis). Briefly, NRI is a statistical tool used to assess improvement in model performance offered by a new method of classification compared with a reference method [23,24]. The NRI compares the frequency of appropriate reclassification to inappropriate reclassification with the use of a new model of classification. The NRI is based on reclassification tables constructed separately for participants with or without the interest event (that is, diagnosis of AD or non-AD) and quantifies the correct movement in categories, up for events and down for non-events. Upward movement (up) is a change into a higher category, and downward movement (down) is a change into a lower category based on the new algorithm. The NRI is calculated as follows:
NRI = P (up | event) ‐ P (down | event) + P (down | non ‐ event) ‐ P (up | non ‐ event) 
The null hypothesis for NRI = 0 was tested using Z-statistics following McNemar asymptotic test for correlated proportions . We provide an Excel file (Microsoft, Redmond, WA, USA) that allows the assessment of NRI with its standard error and its P-value .
All resulting P-values were two-tailed, and P ≤ 0.05 was considered statistically significant. Statistical analyses were performed using STATA version 10 and SAS version 9.2 (SAS Institute, Cary, NC, USA).