The Preclinical Phase of Alzheimer Disease: A 22-Year Prospective Study of the Framingham Cohort

Objectives  To relate performance on tests of cognitive ability to the subsequent development of probable Alzheimer disease (pAD) and to identify the pattern of earliest changes in cognitive functioning associated with a diagnosis of pAD.

Design  From May 1975 to November 1979, a screening neuropsychological battery was administered to Framingham Study participants. They were followed up prospectively for 22 years and examined at least every 2 years for the development of pAD.

Setting  A community-based center for epidemiological research.

Participants  Subjects were 1076 participants of the Framingham Study aged 65 to 94 years who were free of dementia and stroke at baseline (initial) neuropsychological testing.

Main Outcome Measure  Presence or absence of pAD during a 22-year surveillance period was related to test performance at initial neuropsychological testing.

Results  Lower scores for measures of new learning, recall, retention, and abstract reasoning obtained during a dementia-free period were associated with the development of pAD. Lower scores for measures of abstract reasoning and retention predicted pAD after a dementia-free period of 10 years.

Conclusions  The "preclinical phase" of detectable lowering of cognitive functioning precedes the appearance of pAD by many years. Measures of retention of information and abstract reasoning are among the strongest predictors of pAD when the interval between initial assessment and the development of pAD is long.

THE "PRECLINICAL phase" of dementia refers to a period of cognitive decline that precedes the onset of Alzheimer disease (AD).¹ Early detection of AD will be increasingly important as research advances with respect to prognostic methods and therapeutic interventions.^2,3 Therefore, it is important to determine the time line and the earliest evidence of cognitive decline signaling the preclinical phase of AD and to identify specific neuropsychological tests that have clinical utility in the prediction of this disorder.

Lower performance levels with respect to the following cognitive abilities have been associated with the subsequent development of AD in one or more studies: verbal and visuospatial episodic memory^3-7; abstract reasoning^2,8; new learning⁹; verbal abilities, including category and letter fluency^7,10,11; and visuospatial and executive functioning.^2,7 Relations between measures of attention (eg, digit span forward and backward) have been seen in some studies¹² but not in others.^6,13

Many studies have been characterized by relatively small samples with few AD cases. Most have related baseline measures to probable AD (pAD) diagnosed 2 to 5 years after initial neuropsychological assessment.¹² To address these problems, Linn et al³ obtained baseline neuropsychological test measures for a dementia-free cohort of participants in the Framingham Study and then followed up subjects with respect to incidence of pAD for 13 years. Findings were consistent with those reported in the literature, ie, lower test scores for measures of new learning and immediate recall, visual reproductions from memory, verbal associations, and abstract reasoning were associated with a later diagnosis of pAD.

Given the scarcity of prospective studies with long AD surveillance periods and large samples, we believed it important to extend the surveillance period from 13 years to 22 years for the subjects initially followed up by Linn et al. The number of pAD cases increased from 55 in the study by Linn et al to 109 in the present study as the result of incident AD during the longer surveillance period.

Because of the relatively long surveillance period and the increase in sample size, we were able to examine the pattern of test results predicting incident pAD for 2 cohorts of individuals. These individuals were selected on the basis of the number of years after baseline neuropsychological assessment that they remained free of any form of dementia: (1) at least 5 years and (2) at least 10 years. These dementia-free intervals were defined as preclinical intervals.³ We predicted that a subset of the tests predicting pAD during the shorter preclinical period would predict pAD during the longer preclinical period. Presumably, abilities measured by this smaller set of tests will be among those providing an early and robust prediction of pAD.

We used the same methods for neuropsychological assessment at baseline and the same surveillance methods described by Linn et al,³ with 2 exceptions: (1) we used the actual year of diagnosis of pAD as the dependent variable, whereas Linn et al, working retrospectively from the year of actual diagnosis of pAD, used estimates of the year of onset of AD; and (2) we added occupation as a statistical control variable.

Beginning in 1950 and extending to the present time, comprehensive physical examinations have been conducted by a physician every 2 years (biennially), although dementia surveillance is an ongoing process that allows identification of the onset of dementia by the year of occurrence. From examination 14 through examination 15, 2898 individuals were examined.^14,15 Of these, 2123 individuals completed the tests used to predict incident dementia during the 22-year surveillance period.

We defined the sample in the same way as Linn et al. After persons younger than 65 years were excluded, there were 1281 individuals (age range, 65-94 years). Beginning with these 1281 subjects, 37 individuals who met the criteria for probable dementia (at any level of severity), based on neuropsychological test scores, and 2 individuals who were demented before administration of the neuropsychological battery (examinations 14 and 15) were excluded. Of the remaining 1242 individuals, 54 prevalent stroke cases were excluded, as was 1 individual for whom no follow-up examinations were conducted. Of the remaining 1187 individuals, 111 were excluded because they did not complete every test in the neuropsychological test battery. This exclusion replicated the method of Linn et al and allowed comparisons of the pattern of test results for a sample of persons who completed all tests. It also resulted in equal sample sizes for each test and eliminated a subset of individuals who did not complete tests as a result of physical disabilities.^14,15

The remaining sample of 1076 individuals was followed up prospectively for up to 22 years (the pAD surveillance period). During this time, 109 study participants (10.1%) met the criteria for a diagnosis of pAD. When covariates (described in the "Statistical Methods" subsection) were used in the regression models, sample size was reduced to 1043 subjects because of missing data on 1 or more covariates. Of this remaining sample of 1043 individuals, 106 individuals met the criteria for a diagnosis of pAD.

Table 1 lists the neuropsychological tests used and shows the unadjusted means and SDs for each test score for the full sample. This battery^14,15 consists of standardized tests taken from the Wechsler Memory Scale,¹⁶ the Wechsler Adult Intelligence Scale,¹⁷ and the Multilingual Aphasia Examination.¹⁸

Beginning in May 1975 (examination 14), Framingham Study participants were examined for the presence of dementia for 22 years. Since May 1981 (examination 17), the Mini-Mental State Examination¹⁹ has been administered. When longitudinal data on the Mini-Mental State Examination became available, a standard procedure (regardless of age and education level) was used; a score of 23 or below, a fall of 3 or more points between successive examinations, or a decline in 5 points with respect to any previous examination prompted retrospective and prospective in-depth tracking and examinations for dementia. Possible dementia cases also came to the attention of investigators as the result of reports of memory loss or cognitive deficit symptoms from the participant, the participant's family, primary health care providers, and Framingham Study physicians. If a subsequent examination by a neuropsychologist and a neurologist failed to indicate unequivocal dementia, the study participant was scheduled for another review within 1 year or at the next biennial examination, whichever came first. If dementia was suspected (based on Diagnostic and Statistical Manual of Mental Disorders, Revised Third Edition, and, later, Fourth Edition, criteria),^20,21 all pertinent case records available were reviewed by a panel of neuropsychologists and a neurologist. These data included (1) neurological findings, (2) neuropsychological data, (3) Framingham Study and hospital records, and (4) results of brain imaging (computed tomography or magnetic resonance imaging) studies when available. The following determinations were made: (1) presence of dementia, (2) the type of dementia, and (3) the year of diagnosis. Criteria established by the National Institute of Neurological and Communicative Disorders and Stroke and the Alzheimer's Disease and Related Disorders Association²² were used. Based on these data, the review panel assigned the year of dementia onset and the subtype of dementia.^23,24 In this study, we defined incident dementia in terms of the year in which the criteria for a diagnosis of pAD were first fully satisfied.

Cox regression (proportional hazards) analyses²⁵ were used to relate the baseline neuropsychological test scores to incidence of dementia during the surveillance period. For these analyses, time from an event (neuropsychological testing) to an event (pAD) was the dependent variable. Survival time (time free of AD) was calculated as years from the baseline neuropsychological testing to the diagnosis of pAD. For non-pAD cases, including those with other types of dementia and nondemented individuals, survival time was censored at the year when non-pAD status was last determined, at the last clinic visit, or at the year of diagnosis of non-pAD dementia.

Separate multiple regression analyses were done for each test score. There was a relationship between age and education level. Subjects with less than a high school education were slightly younger (mean±SD age, 72.0±5.6 years) than those with a high school or higher education (73.0±5.8 years). Crude analyses were performed. These were followed by analyses with adjustment for age (years) and finally by analyses with adjustment for age, education level, occupation levels, and sex effects. Each covariate was controlled statistically for all others. Previous analyses indicated no sex differences in incidence of dementia or AD in this cohort.²⁴ Thus, all analyses were conducted with men and women combined.

Table 2 provides sample demographic statistics for subjects who were diagnosed as having pAD and those who remained noncases during the surveillance period. Stroke cases were excluded at baseline neuropsychological assessment (examinations 14 and 15). During the surveillance period, 13.8% of the individuals with pAD and 16.0% of the non-pAD cases experienced incident stroke. There were no statistically significant differences between the pAD and the non-pAD groups with respect to the proportion of subjects in the education, sex, and occupation groups. Subjects who developed pAD were older (P<.001) than subjects who remained free of pAD (mean age, 74.6 vs 72.3 years). Table 3 shows the frequency of pAD cases by year of diagnosis.

Table 4 summarizes the age-, sex-, education-, and occupation-adjusted relationships between the individual test scores and the AD outcome in terms of Cox regression coefficients (β), odds ratios, and 95% confidence limits (N=1076). Odds ratios reflect the estimated increase in risk of a diagnosis of pAD afforded by each 1-SD decline (Table 1) in a neuropsychological test score at the baseline. An odds ratio of 1.00 signifies no increase in risk. Statistically significant associations between lower test scores and AD were observed for Logical Memory–Retained, Similarities, Paired Associate Learning, and the Learning and Immediate Recall composite score.

There were no statistically significant interactions between education level and any test score (P>.15). There was a single interaction involving age groups and Controlled Oral Word Association scores, and thus we performed analyses stratified by age. For these analyses, the younger cohort (65-74 years of age) included 55 patients with pAD of 708 persons surveyed, and the older cohort (75-94 years of age) included 54 patients with pAD of 314 persons surveyed. No statistically significant association between Controlled Oral Word Association scores and pAD outcome (n=55 pAD cases) was observed for the younger cohort (odds ratio, 0.94; 95% confidence limits, 0.72, 1.23; P=.63). In contrast, lower Controlled Oral Word Association scores were associated with higher risk of a later diagnosis of pAD (n=54 pAD cases) for the older cohort (odds ratio, 1.45; 95% confidence limits, 1.08, 1.94; P=.03).

When we repeated the analyses and included the individuals who did not complete all the neuropsychological tests (n=111), the odds ratios were similar and the same pattern of significant associations was obtained, with 1 exception. The Visual Reproductions test was significantly (P=.03) associated with pAD rather than marginally (P=.06) associated (Table 4). Comparisons of individuals who did and did not complete all tests in the battery during the initial screening indicated that the incidence of pAD was similar (10.1% vs 10.8%).

Table 5 and Table 6 summarize statistically significant findings for the comparisons between groups formed of subjects who were dementia free for at least 5 years and at least 10 years, respectively. Of 941 persons in the first category, 105 were diagnosed as having pAD during the subsequent 17-year surveillance period. Of 577 persons in the second category, 73 were diagnosed as having pAD during the subsequent 12-year surveillance period. Table 5 and Table 6 summarize findings for the neuropsychological tests for which there were significant associations between test scores and pAD. For the cohort that was dementia free for at least 5 years (Table 5), lowered Logical Memory–Retained, Similarities, Paired Associate Learning, and Learning and Immediate Recall composite scores were significantly associated with the later development of AD. For the cohort that was free of dementia for at least 10 years (Table 6), only Logical Memory–Retained and Similarities were significantly associated with a diagnosis of pAD during the surveillance period.

In a final set of analyses, adjusted for all covariates, age at onset of pAD was used as the time-to-event dependent variable rather than year (of diagnosis) from baseline. Essentially, results were the same for these analyses.

Lower scores on measures of episodic and semantic memory processes, acquisition, and abstract reasoning are risk factors for the subsequent development of AD.^1-13 The present study indicates that these associations are robust during a 22-year surveillance period. The pattern of tests predicting pAD (Table 4) was similar to that reported in the 13-year study by Linn et al,³ with 1 major exception. We did not find that lower Digit Span Test scores were protective with regard to the development of AD. The most parsimonious explanation of the difference in findings between our study and the study by Linn et al was that Linn et al obtained spurious findings with their smaller sample of study participants. In previous studies, lower Digit Span scores have been associated with higher risk of developing pAD⁴ or, consistent with the present study, there has been no significant association between Digit Span scores and pAD.^4,6,13

Controlled Oral Word Association performance has been a predictor of pAD.²⁶ This was true for the 75- to 94-year-old study participants in the present study, but not for individuals 65 to 74 years of age. This finding may reflect the fact that the Controlled Oral Word Association Test, a measure of semantic processing, is particularly difficult for persons older than 70 years.²⁷ It is also possible, although arguable, that AD with onset at a younger age is predicted by a different set of variables than AD with onset at an older age and that preclinical decrements in semantic memory predict late-onset AD.

As hypothesized, analyses comparing the 2 cohorts who were free of pAD for at least 5 and 10 years indicated that fewer test scores persisted in discriminating pAD from non-AD cases during a longer preclinical period. For the sample including only persons who were dementia free for at least 5 years after baseline, lower scores on tests of learning, immediate recall, retention, and abstract reasoning were associated with higher risk of pAD. Only Logical Memory–Retained and abstract reasoning were predictive of pAD for the sample restricted to persons who were dementia free for at least 10 years after baseline neuropsychological assessment.

One would not expect all neuropsychological tests to be equal in their ability to forecast AD. It is important to know that poor performance on measures of retention (Logical Memory–Retained) and abstract reasoning (Similarities) herald pAD at least 10 years before it is diagnosed. However, it is possible, although speculative, that pAD cases recognized only after the 10th year of surveillance exhibited a subtype, or subtypes, of dementia quite different from those that were diagnosed at a point in time after the fifth year of surveillance.

In the Framingham Study, including the present investigation, subjects are carefully screened at baseline to exclude cognitive deficit at any level, including moderate impairment. However, once the surveillance period is initiated, diagnoses of pAD are assigned only when definite dementia is diagnosed. These conservative criteria were adopted to avoid false-positive test results associated with the diagnosis of mild pAD, but they impose a limitation on the conclusions we can reach about the prediction of mild dementia. Conservatively, the pattern of tests related to the prediction of pAD in the present study apply only to moderate to severe AD.

The concept of a preclinical period of AD has heuristic value with respect to research on cognitive deficits in pAD. However, it is very difficult, if not impossible, to separate the preclinical phase of AD from early mild AD. It is especially difficult when the average interval between baseline neuropsychological testing and the clinical diagnosis of pAD is relatively short, ie, 2 to 3 years. In the present study, decrements in abstract reasoning and retention were observed in individuals who were dementia free for a minimum of 10 years, thus reducing the likelihood that they were mildly demented in the years immediately after baseline testing.

The preclinical period may in effect be a subclinical period in which underlying disease processes leading to AD have already begun. At present, there is no direct evidence linking pathophysiological changes in the brain to preclinical decline in cognitive ability, but physiological and anatomical changes in the brain during the preclinical phase of AD have been identified, ie, metabolic dysfunction in temporoparietal brain regions^28-30; regional decreases in cerebral perfusion in the hippocampal-amygdaloid complex, the posterior cingulate, the anterior thalamus, and the anterior cingulate³¹; and cholinergic fiber loss associated with diffuse plaques.³²

Regardless of the mechanisms underlying the mild cognitive deficits, it is clear that lowered cognitive functioning (although within normal limits of performance) is a "harbinger of AD"³ and that scores on measures of new learning, retention, abstract reasoning, episodic memory, and semantic memory forecast a diagnosis of pAD. Although this study is not unique with respect to these findings, it lends important support for these conclusions given that a large community sample was followed up prospectively during a 22-year period and that the number of pAD cases is large compared with previous studies. This study also provides evidence that lowered cognitive performance on measures of abstract reasoning and retention of verbal material predicts AD over an extended period (10 years) in which persons are free of a diagnosis of dementia.

Accepted for publication January 20, 2000.

This research was supported in part by contract NIH/NHLBI N01-HC-38038 and research grants 5R01 AG08122-09 and 5R37 AG0355 from the National Institute on Aging and research grant 5R01-NS17950-16 from the National Institute of Neurological Disorders and Stroke, National Institutes of Health, Bethesda, Md.

We wish to express appreciation to Michael Robbins, PhD, Jeremy Aho, Penelope K. Elias, PhD, and Sudha Seshadri, MD, for reading and commenting on initial drafts of the manuscript.

Reprints: Merrill F. Elias, PhD, MPH, Department of Mathematics and Statistics, Statistical Consulting Unit, 111 Cummington St, Boston University College of Arts and Sciences, Boston, MA 02215 (e-mail: mfelias@aol.com).