Methods
Study Population
Data came from the PRE-AL Study (PREdiction of AD). Briefly, in 2001 and 2002, 251 patients with MCI were recruited and followed up semiannually over a period of 3 years. Subjects were recruited from memory clinics of 14 expert centres in the field of dementia across France (see Acknowledgments). Patients were enrolled according to the following criteria: (1) a subjective memory complaint; (2) an objective memory impairment documented by at least one word missing in the three-word recall of the Mini-Mental State Examination (MMSE) or a score under 29 on the Isaacs-set test, or both; (3) a preservation of general cognitive functioning (MMSE between 25 and 29/30); (4) a normal score or only one item impaired at the first level in the four Instrumental Activities of Daily Living (IADL); and (5) the absence of the Diagnostic and Statistical Manual of Mental Disorders, 3rd edition, revised (DSM-III-R) criteria for dementia. Patients with focal lesions in cerebral imaging or documented depressive symptoms or with other medical conditions, which could interfere with memory performance or follow-up, were excluded. Each subject signed an informed consent form once the nature of the procedures had been fully explained. The study was approved by the Ethics Committee of the 'La Salpêtrière' Hospital.
Out of the 251 participants included in the PREAL study, patients who had had no follow-up (n=17), those with only one visit at 6 months (n=13), and patients with no measurements for any of the neuropsychological tests during their follow-up (n=3), were excluded from our analyses. As AD was the primary outcome of the study, patients who converted to a non-AD dementia (n=6) were excluded from further statistical analyses. Finally, the present analyses were carried out on a total of 212 participants.
Data Collection
Dementia diagnosis and definition of prodromal-AD Patients were seen every 6 months for 3 years. At each visit, clinical assessment included the recording of all medical events, current treatment and complete neurologic examination. Activities of daily life were rated with the IADL scale during an interview with the patient and a knowledgeable collateral source (a spouse or a child). Memory complaint was assessed by a specific questionnaire. The Clinical Dementia Rating (CDR) scale was completed at each visit during follow-up. During the follow-up, when patient meets the clinical DSM-III-R criteria for dementia and/or the score of 1 at the CDR scale, he/she was considered as 'converted' to dementia. In order to minimise intercentre variability, all clinical records were reviewed by an Expert Committee composed of neurologists (n=3), neuropsychologists (n=3), geriatricians (n=3) and psychiatrists (n=3). They re-examined whether clinical criteria for dementia were satisfied using DSM-III-R criteria and classified the patient as AD using the NINCDS-ADRDA criteria. Patients who converted to AD during the 3-year follow-up were retrospectively classified as patients with prodromal AD; other subjects were designated as MCI non-AD.
Assessment of neuropsychological performance All subjects were tested at inclusion and annually using a standardised neuropsychological battery. In the event of a suspected conversion, the patient underwent an additional neuropsychological evaluation 6 months later. Six cognitive domains commonly affected by ageing and AD were assessed using 13 scores described in Table 1. The mean duration of the administration of the complete predictive battery was 91 min (SD=15).
Statistics We used a nonlinear mixed model for multivariate longitudinal data involving a latent process (illustrated in figure 1) to analyse the cognitive trajectory over time of the MCI subjects during the follow-up. The statistical model assumes that the correlation between the neuropsychological tests is induced by a latent cognitive process (LCP) representing the common cognitive part measured by the neuropsychological tests. The model is divided into two parts: (a) a linear mixed model describes the change over time in the LCP and evaluates the common effects of covariates on this latent cognitive trajectory, and (b) test-specific nonlinear measurement models relate each administration of the neuropsychological tests with the LCP by describing and accounting for the metrological properties of the tests, and evaluating test-specific associations with covariates (here specifically prodromal AD status). Therefore the impact of a given covariate can be explored on the LCP and on each test.
(Enlarge Image)
Figure 1.
Conceptualisation of the nonlinear mixed model involving a latent process to model cognition from several neuropsychological tests.
Effect of the covariates on latent cognitive change with time The LCP trajectory was modelled using a linear mixed model, which evaluates changes of a repeated outcome over time (ie, the LCP) and accounts for correlations between the repeated measures of each subject. The linear mixed model included a subject-specific random intercept and a slope for time after inclusion (in years of follow-up), as well as the following covariates: age, sex, educational level, prodromal AD status and their interaction with the time elapsed since inclusion. To define the LCP dimension, we assumed it followed a Gaussian distribution N(0.1) at baseline for the reference state of the linear mixed regression model (ie, women, with a low baseline level, MCI non-AD of 55 years old).
Metrological properties and test-specific effects The test-specific measurement models consisted in flexible transformations linking the neuropsychological tests with the LCP. These transformations, which capture the metrological properties of the tests, are covariate-and-time-invariant parametric functions depending on parameters that are estimated simultaneously with the other parameters of the model. β cumulative distribution functions were chosen as flexible and parsimonious transformations. Indeed, these functions offered a large variety of shapes (concave, convex, sigmoid or simply linear) and thus modelled the curvilinearity of the tests. The complete methodology was developed elsewhere. This part of the model also evaluated the test-specific effects of covariates on each test after adjustment for their effect on the LCP.
Mean annual change for each neuropsychological test according to the occurrence of AD during the follow-up (in LCP units) This was calculated by adding together (a) the common effect of time on the LCP, (b) the test-specific effect of time on each test and the test-specific effect of the interaction between time and prodromal AD status on each test.
Statistical tests were performed at the conventional two-tailed α level of 0.05. Data were analysed using SAS Enterprise Guide V.4.3 (SAS Institute, Cary, North Carolina, USA) and a FORTRAN90 executable for the nonlinear mixed model with the latent process (program HETMIXSURV available at http://biostat.isped.u-bordeaux2.fr/).