Fifty-nine elderly MCI individuals were recruited from a dementia and memory disorder clinic at Seoul National University Hospital. MCI was diagnosed accordingto current consensus criteria for amnestic MCI:³³ 1) memory complaint corroborated by an informant, 2) objective memory impairment for age, education and gender, 3) essentially preserved general cognitive function, 4) largely intact functional activities, 5) not demented. In terms of criterion 2), a performance score for at least one of the four episodic memory tests included in the Korean version ofthe Consortium to Establish a Registry for Alzheimer's disease (CERAD) neuropsychological battery [namely, Word List Memory (WLM), Word List Recall (WLR), Word List Recognition (WLRc) and Constructional Recall (CR) test] was 1.5 SD below the respective age-, education- and gender-specific normative mean.³⁴³⁵ All MCI individuals had an overall CDR index³⁰ of 0.5.

The following exclusion criteria were applied to all subjects: any evidence of present serious medical, psychiatric, or neurological disorders that might affectmental function; any evidence of focal brain lesions other than white matter changes on MRI; illiteracy; severe visual or hearing loss; no reliable informants. The Institutional Review Board of Seoul National University Hospital approved the study and written informed consent was obtained from each participant (IRB No. H-1506-131-683).

All subjects were examined by neuropsychiatrists with advanced training in dementia research and CDR assessment³⁰ according to the protocol of the CERAD, and received MRI and laboratory tests. To acquire accurate information, a reliable informant was necessarily interviewed as well as the participant. A panel consisting four neuropsychiatrists made clinical decisions including the assignment of CDR. The CERAD neuropsychological battery,³⁵³⁶ including Verbal fluency (VF), 15-item Boston naming test (BNT), MMSE, WLM, WLR, WLRc, Constructional Praxis (CP), and CR test, were also applied by experienced clinical psychologists who were blinded to the neuropsychiatrist's clinical assessment. In addition to the CERAD battery, the Neuropsychiatric Inventory (NPI)³⁷ and a structured, anchored version of the 17-item Hamilton Rating Scale for Depression (HRSD)³⁸ were administered to evaluate non-cognitive neuropsychiatric symptoms or depression.

Each subject underwent subsequent clinical assessments by a neuropsychiatrist according to the CERAD protocol at 24-month intervals. After each follow-up evaluation, the consensus panel reviewed all the available clinical data obtained current follow-up evaluation and made a clinical diagnosis and assigned a CDR rating. Separately from the clinical assessment process, at each visit, the subjects were administered the same neuropsychological tests as in the baseline evaluation by clinical psychologists blinded to the neuropsychiatrist's clinical evaluation.

The subject's condition was considered to have progressed to AD dementia if he or she met the diagnostic and statistical manual of mental disorders (DSM-IV TR) criteria for dementia³⁹ and the National Institute of Neurobiological and Communication Disorders and Stroke and the Alzheimer's Disease and Related Disorders Association (NINCDS-ADRDA) criteria for probable or possible AD⁴⁰ within 2 years.

The subjects were classified as 2 groups according to the clinical state at the time of 2 years follow-up evaluation: those who progressed to AD dementia (MCIp) and those who did not progress to AD dementia (MCInp). Between-group comparisons for baseline continuous data including demographic and clinical data were done by two-tailed t tests. Baseline categorical data were analyzed by the χ² test.

Logistic regression analyses were conducted to examine the ability of various clinical measures or their combinations to predict progression to AD dementia in MCI individuals. We used the differences of -2 log likelihood (-2LL) to statistically compare the predictive ability of various models with different numbers of independent variables.⁴¹ The -2LL is aquantity generated by the logistic regression procedure and is directly proportional to the contribution of variables to the separation of groups. A smaller -2LL means abetter predictive ability of the model. The probability distribution of -2LL difference between simple (model 1) and more complex model (model 2) can be approximated by a chi-square distribution with (df2-df1) degrees of freedom, where df1 and df2 are the degrees of freedom of models 1 and 2 respectively. Therefore, the -2LL difference allows the direct comparison of prediction models of different complexities.⁴² The level of statistical significance was set as two-tailed p<0.05.

All subjects (n=59) were in amnestic MCI at baseline assessments. After the 2 years follow-up period, 22 (37.3%) progressed to AD dementia and 37 (62.7%) did not (Table 1). Among MCInp group, 32 were in amnestic MCI, 5 were in non-amnestic MCI, and 2 were in non-specific cognitive impairment (CDR=0.5, all Z-score ≥-1.5 in the CERAD neuropsychological battery).

Baseline demographic and clinical characteristics of MCIpand MCInp individuals are shown in Table 1. There were no statistically significant differences between MCIp (n=22) and MCInp (n=37) in regard of age, gender, education, BDS-ADL, HRSD, NPI total score, and mHIS. MCIp had significantly greater mean CDR-SOB scores than MCInp. In terms of CDR subscale scores, MCIp individuals showed significantly higher CDR Orientation scores than MCInp ones, while there were no significant differences in other CDR subscale scores between the two. There were no significant differences in all CERAD neuropsychological scores between the two. MCIp showed marginally higher MMSE scores compared to MCInp.

We first selected CDR Orientation score, CDR-SOB score, and MMSE score referring to the results from the comparison of baseline clinical characteristics between MCIp and MCInp as candidate-independent variables for logistic regression analyses for model comparison. A series of logistic regression analyses were conducted in three steps (Table 2). In the first step, we tested the following three “one candidate models”: “Model O including CDR Orientation”, “Model S including CDR-SOB”, and “Model M including MMSE”. All three models were statistically significant, and “Model O” had the highest classification accuracy and smallest -2LL among three “one candidate models”. In the second step, we compared the-2LL between each of the “one-candidate models” with the corresponding “two candidate models”, which included CDR Orientation and either CDR-SOB or MMSE. The AD dementia prediction accuracy of either “Model OM including CDR Orientation and MMSE” or “Model OS including CDR Orientation and CDR-SOB” was not significantly different from that of “Model O”, which had the highest classification accuracy and smallest -2LL among the three “one candidate models”. In the third step, the “three-candidate model” (Model OSM), which included all three variables, was compared with “Model O”. The “Model OSM” was not significantly different from “Model O”. Table 3 shows the finally selected logistic regression model (Model O) for AD dementia prediction in amnestic MCI at two-year follow-up.