Criteria for amnestic mild cognitive impairment (aMCI) have been devised in an attempt to capture the preclinical phase of Alzheimer's dementia. There is evidence to suggest, however, that individuals with aMCI according to the current criteria comprise a heterogeneous group, ^{Reference Blossom, Matthews, McKeith, Bond and Brayne1,Reference Chertkow, Nasreddine, Joanette, Drolet, Kirk and Massoud2} some of whom will progress to dementia with time, while others will not. ^{Reference Visser, Kester, Jolles and Verhey3} To maximise the diagnostic value of aMCI, criteria should identify a homogeneous group of people with preclinical dementia. To this end, cognitive criteria should be defined in a manner that reflects our current knowledge of their predictive value. As the prodromal phase of Alzheimer's disease is likely to extend beyond a 20-year period, ^{Reference Amieva, Jacqmin-Gadda, Orgogozo, Le Carret, Helmer and Letenneur4} studies with shorter follow-up periods are liable to underestimate the risk of conversion. In fact, only 2 of 14 clinic-based longitudinal aMCI studies report follow-up periods beyond 3 years (online Table DS1). ^{Reference Fox, Warrington, Seiffer, Agnew and Rossor5,Reference Tabert, Manly, Liu, Pelton, Rosenblum and Jacobs6} Moreover, low baseline levels of general cognitive functioning among participants with aMCI lead to a greater chance of neuropsychological tasks predicting dementia, because of the more advanced stage of disease in the cohort with aMCI. In 11 of 14 clinic-based longitudinal studies average baseline Mini-Mental State Examination (MMSE) ^{Reference Folstein, Folstein and McHugh7} scores for the group with mild cognitive impairment who converted to dementia (aMCI converters) fell below the higher screening cut-off for dementia (i.e. 27/30). In these studies, an underlying dementia might well have been suspected on the basis of such rudimentary cognitive screening instruments alone, begging the question of the ‘added value’ of a fuller cognitive evaluation. The comprehensiveness of the neuropsychological battery employed and the appropriateness (on both theoretical and empirical grounds) of test selection might also be expected to influence predictive power. For example, the combination of the Paired Associate Learning test (PAL), ^{Reference Blackwell, Sahakian, Vesey, Semple, Robbins and Hodges8} age and the Graded Naming Test (GNT) ^{Reference McKenna and Warrington9} give an overall classification accuracy of 100% over a 2.5 year follow-up interval. ^{Reference Blackwell, Sahakian, Vesey, Semple, Robbins and Hodges8} The same classification accuracy (i.e. 100%) has been reported ^{Reference Ahmed, Mitchell, Arnold, Nestor and Hodges10} for the use of the PAL in combination with the Addenbrooke's Cognitive Examination (ACE) ^{Reference Mathuranath, Nestor, Berrios, Rakowicz and Hodges11} and the Graded Faces Test (GFT) ^{Reference Thompson, Graham, Patterson, Sahakian and Hodges12} over shorter (1 year) intervals. However, these findings have not been replicated outside the test authors' group, in larger numbers of people with aMCI, across follow-up periods extending beyond 2.5 years, and where the mean general level of cognitive functioning at baseline (as indicated by performance on cognitive screening) falls above cut-off points for dementia. If replicable, such measures could be used in the neuropsychological assessment of aMCI within specialist memory clinic settings to provide information about differential diagnosis and prognosis.

Measurement of cognitive function represents just one, albeit an important, approach to detecting and diagnosing Alzheimer's disease at a very early and preclinical stage. Other work has looked at the ability of imaging (for the most part magnetic resonance image (MRI) scanning), biomarkers (i.e. total tau, beta amyloid 42 and phosphorylated tau) and changes of a behavioural nature to predict the future onset of clinically diagnosable Alzheimer's disease. A recent meta-analysis of imaging and biomarkers for Alzheimer's disease ^{Reference Schmand, Huizenga and van Gool13} indicated some promise for the cerebrospinal fluid (CSF) markers in so far as their overall predictive accuracy levels were similar to that of memory impairment 4 years prior to the point of diagnosis. Furthermore, the effect sizes for the CSF markers were largest when assessed longer before the point of diagnosis. However, atrophy of the hippocampus or other medial temporal lobe structures was found to be a less accurate predictor of future Alzheimer's disease than memory impairment, and the largest effect sizes, which are themselves likely to represent an underestimation owing to the removal of variability inherent in the inclusion of memory impairment as a selection criterion for a majority of studies, were seen in association with measures of delayed memory recall.

In this study we present a detailed neuropsychological and clinic-based cohort study, with an average of 4 years follow-up from baseline neuropsychological assessment until final review. It is the only study to date of people who can be classified as high-functioning aMCI converters (i.e. MMSE >27/30) extending beyond 3 years follow-up. Furthermore, it represents the first clinic-based study to investigate the robustness of the GNT, GFT and a combination of the ACE and PAL as predictors of conversion to dementia outside the original authors' research group, ^{Reference Mathuranath, Nestor, Berrios, Rakowicz and Hodges11} and to report the detailed fate of people who are aMCI non-converters in terms of their course of cognitive impairment.

Method

Materials, participants and outcome criteria

Participant characteristics (inclusion/exclusion criteria) and neuropsychological measures have been detailed previously. ^{Reference Lonie, Herrmann, Donaghey and Ebmeier14,Reference Lonie, Herrmann, Tierney, Donaghey, O'Carroll and Lee16,Reference Lonie, Tierney, Herrmann, Donaghey, O'Carroll and Lee17} In brief, the neuropsychological battery comprised measures of premorbid IQ (National Adult Reading Test; NART), ^{Reference Nelson and Willison18} episodic memory (PAL, ^{Reference Blackwell, Sahakian, Vesey, Semple, Robbins and Hodges8} Hopkins Verbal Learning Test – Revised (HVLT–R), ^{Reference Benedict, Schretlen, Groninger and Brandt19} Rey Complex Figure Test (RCFT), ^{Reference Rey20} semantic memory (GFT, GNT, Category fluency; animals), visuospatial function (RCFT copy), psychomotor processing speed (Trail Making Test Part A) ^{Reference Reitan21} and attention/executive function (Dual Task, ^{Reference Della Sala22} Controlled Oral Word Association Test (COWAT: F, A and S); ^{Reference Benton and Hamsher23} Trail Making Test Part B). ^{Reference Reitan21} Amnestic mild cognitive impairment was defined in accordance with the revised criteria set out by Petersen et al. ^{Reference Petersen, Thomas, Grundman, Bennett, Doody and Ferris15} Demographic characteristics of the aMCI converters group, the aMCI non-converters group and the normative sample group, together with their respective baseline mean performances on cognitive screening and selected neuropsychological measures, are summarised in Table 1.

Table 1 Demographic and baseline neuropsychological data for healthy elderly control group, group with amnestic mild cognitive impairment (aMCI) who converted to dementia and group with aMCI who did not convert

Variable (maximum score)	Healthy elderly control group, mean (s.e.)	aMCI converter group, mean (s.e.)	aMCI Non-converter group, mean (s.e.)	Converter group v. non-converter group, t-test	d.f.	P
Demographic information (confounders)
Age	70.8 (7.8)	76.0 (1.6)	73.2 (5.4)	–1.53	42	0.14
National Adult Reading Test IQ	118.5 (3.3)	116.4 (2.0)	117.4 (1.3)	0.44	41	0.66
Months of follow-up	28.0 (9.1)	51.4 (3.0)	52.4 (2.7)	0.23	42	0.82
Cognitive screening
Addenbrooke's Cognitive Examination, total (100)	94.5 (3.2)	86.6 (1.3)	91.3 (0.9)	2.98	42	0.006a
Episodic memory
PAL 6 box errorsb	7.9 (6.7)	23.2 (3.2)	13.5 (2.5)	–2.32	42	0.02
HVLT–R delayed recall (12)	8.1 (2.7)	3.5 (0.7)	5.8 (0.7)	2.38	41	0.02
HVLT–R discrimination index (12)	9.9 (1.8)	7.2 (0.7)	9.2 (0.4)	2.81	41	0.008a
Semantic memory
Graded Naming Test (30)	23.8 (3.1)	20.3 (1.2)	21.0 (0.8)	0.56	39	0.58
Graded Faces Test (30)	20.7 (3.6)	15.3 (1.3)	18.1 (0.9)	1.84	41	0.08
Attention/executive
Trail Making Test Part B b,c	88.7 (30.7)	152.7 (19.6)	100.9 (9.8)	–2.58	42	0.014

Statistical analysis

Independent t-tests were conducted to compare the baseline performances of the aMCI converters group and the aMCI non-converters group on the demographic indices of age, NART full-scale IQ (FSIQ), and years of follow-up and on the neuropsychological measures of ACE total score, PAL, HVLT–R delayed recall and discrimination index (a measure of accuracy of delayed recognition), GFT, Category fluency and Trail Making Test Part B. The alpha level was adjusted to control for multiple comparisons using Holm's sequential Bonferroni correction method. ^{Reference Holm24}

Baseline ages, time of follow-up and premorbid IQ were selected as potential confounders for their established influence on risk of developing late-onset dementia. ^{Reference Mebane-Sims25–Reference Petersen, Smith, Waring, Ivnik and Kokmen27} Seven neuropsychological measures (ACE, PAL, HVLT–R discrimination index and delayed recall, GNT, GFT and Trail Making Test Part B) were selected from a total of 18, on the basis of their known sensitivity to aMCI relative to other measures of cognitive functions established by our own cross-sectional analyses, or of their high levels of predictive validity as established by one or more clinic-based longitudinal study of neuropsychological predictors of dementia (online data supplement). Participants who were clinically diagnosed as having dementia at study end-point were identified. Participants who had not received a clinical diagnosis of dementia were further classified as ‘normal’ or having ‘persisting aMCI’ based on their neuropsychological profile at end-point, and ‘progressive’ or ‘non-progressive’ based on the longitudinal course of cognitive function during their years of study participation. ‘Abnormal’ neuropsychological performance was defined by a performance at the 7th centile or lower in two or more of the 18 neuropsychological tasks (this would occur by chance in approximately 1 of 22 participants with mild cognitive impairment without a diagnosis of dementia). Mild cognitive impairment decline was defined by cognitive deterioration of a magnitude seen in fewer than 2.5% of a sample of healthy elderly people over an average 28-month period on at least two measures of semantic memory or executive functioning. Cognitive domains other than episodic memory were selected for this criterion, because of the baseline floor-level performances on episodic memory tasks of many participants with aMCI.

Sensitivity, specificity, positive and negative predictive values, together with the overall percentage of classification accuracy in predicting conversion or non-conversion to dementia, were determined using a combination of a total score of <88/100 on the ACE or a performance of 2 standard deviations or more below controls on the PAL, as this combination of measures has previously been associated with 100% sensitivity and negative predictive values. ^{Reference Ahmed, Mitchell, Arnold, Nestor and Hodges10} These values were also determined for face (GFT) and object (GNT) naming measures.

Neuropsychological measures, for which the baseline performances of those in the aMCI converters group and the aMCI non-converters group were significantly different, were entered simultaneously alongside the putative confounders ‘age’, ‘NART FSIQ’ and ‘years of follow-up’ into a logistic regression analysis. A backward stepwise procedure using the likelihood ratio was applied to determine model content and levels of overall classification accuracy. Criteria for entry and removal were set at P = 0.05 and P = 0.01, respectively, using 20 iterations (SPSS 17 for Windows).

Fig. 1 Flow chart of amnestic mild cognitive impairment (aMCI) end-point classification in accordance with clinical diagnoses in medical notes.

Results

Forty-one percent (18/44, 95% CI 28–56%) of participants with aMCI received a clinical diagnosis of dementia (most often Alzheimer's disease) at some point prior to study end-point (i.e. on average 4.33 years after entry into the study), giving an average annual conversion rate of 11.4% (95% CI 4–23%). Fifty-nine percent (26/44) of participants had not received a clinical diagnosis of dementia. Medical notes were missing or not accessible for the remaining 2/44 individuals. For these participants, the most up-to-date information available at the study end-point was that obtained at final study attendance. Of the participants who had not received a clinical diagnosis of dementia, 8/26 (31%) were stable, 10/26 (38%) progressive and 8/26 (31%) reverted to normal, according to the criteria defined above (Fig. 1).

Following adjustment of the alpha level, ^{Reference Holm24} significant differences in the baseline performances of the aMCI converters group and the aMCI non-converters group were found on the ACE (t ₍₄₂₎ = 2.98, P<0.01) and HVLT–R discrimination index (t ₍₄₁₎ = 2.81, P<0.01) (Table 1).

Only one participant obtained a GNT score at baseline below the 2nd centile of our healthy elderly control group. None of the participants with aMCI performed below the 2.5th centile of our control group on the GFT at baseline. Univariate sensitivity, specificity, negative (NPV) and positive (PPV) predictive values for conversion from aMCI to dementia were therefore based on a cut-off performance at the 7th centile (s.d.>1.5 below the mean) of age norms, and can be summarised as follows: GNT sensitivity 38%, specificity 68%, PPV 43%, NPV 63%; GFT 44%, 68%, 50%, 63%. Using a combination cut-off of ACE <88/100 or PAL >14 errors, the overall rate of classification accuracy was 68%, with sensitivity 72%, specificity 65%, PPV 59% and NPV 77%.

Backward logistic regression with age, NART FSIQ, years of follow-up and the neuropsychological measures for which baseline performance differentiated converters and non-converters (HVLT–R discrimination index and ACE total score) resulted in a final model, completed after five iterations, including the variables ACE total score and HVLT–R discrimination index score only, yielding an overall classification accuracy (aMCI converters group v. aMCI non-converters group) of 74%, sensitivity 65%, specificity 80%, NPV 77% and PPV 69% (Table 2).

Table 2 Summary of final regression model^a

				95% CI for Exp(B)
	B	s.e.	Exp(B)	Lower limit	Upper limit
Constant	15.32*	6.74	4 501 208
Addenbrooke's Cognitive Examination, total score	–0.15*	0.74	0.86	0.75	1.00
Hopkins Verbal Learning Test – Revised, discrimination index	–0.32*	0.16	0.73	0.53	1.00

Discussion