Most non-converters retain high functioning over the duration. Hence, therefore cognitive flexibility can be useful for discriminating future conversion outcomes, but does not appear as informative as episodic memory level 2. Finally, in Figures ?Figures8a8a and ?and8b,8b, for perceptual motor speed, note that there appears to be a subset of converters for whom perceptual motor speed becomes more impaired. While there are non-converters who also have low probability values, this number is outweighed by the converters over the duration. Moreover, as Table ?Table55 indicates, these converters are likely to also be relatively more impaired with episodic memory level 2 than the non-converters. This allows us to identify this particular combination of lower level functioning as being specifically associated with high risk for conversion.

Multivariate prediction using logistic regression models A multivariate model was fit that recognized the above findings, and included other variables such as gender, age, and educational level. The presence of an APOE4 allele was viewed as a binary variable, as well as whether or not a subject had attended college. Also, probability values for performing at a relatively high level for episodic memory level 2, cognitive flexibility, and perceptual motor speed were viewed as continuous explanatory variables. After an initial fit of a full model, gender, age, and educational level were clearly not significant predictors in the model (respective P-values were 0.96, 0.65, and 0.81; Wald’s test).

Using goodness-of-fit tests based on the test statistic of -2 times the difference in log-likelihood values to compare nested models, it appears the best fit is when episodic Drug_discovery memory level 2, perceptual motor speed, and APOE4 status are included in the model. Results are given in Table ?Table6.6. Note that episodic memory level 2, perceptual motor speed, and APOE4 status all are significant predictors (P-value <0.05; Wald's test). When cognitive flexibility is included with these variables, it is not significant (P-value = 0.26). Table 6 Multivariate logistic regression model with outcome as conversion to Alzheimer's disease (AD) from mild cognitive impairment (MCI) within 24 months from baseline Using the model-based estimated probabilities of converting to AD as a predictor, and for instance, using a cutoff value of 0.

55 or higher, classification accuracy is 66.8%, with positive predictive value of 61.5% (32 out of 52). Figure ?Figure99 displays receiver operating characteristic (ROC) curves for these logistic regression probabilities, as well as for the probabilities for episodic memory selleck kinase inhibitor level 2, perceptual motor speed, and cognitive flexibility. Values for the area under the curve (AUC) are 0.710, 0.678, 0.655, and 0.644, respectively. Using a multivariate approach can thus apparently improve prediction.