Thus, four cosine-similarity

scores were computed for each PET scan using a residual vector of each type (i.e., two from the AD/NC projections and two from the MCI-n/MCI-c projections). Statistical analysis of measurements Cosine similarity scores were entered individually into logistic regression models with category membership (AD vs. NC or MCI-c vs. MCI-n) as the dependent variable. Age and sex were considered as potential covariates but were removed if they failed to improve the overall fit of the model. Scores on the MMSE and Functional Activities Questionnaire (FAQ) and interactions of these scores with cosine similarity scores were considered as covariates only Inhibitors,research,lifescience,medical for the MCI-c versus MCI-n logistic regression models. MMSE and FAQ scores were not included in the AD versus NC logistic regression model due to concern of circularity, because these diagnostic classifications were assigned when subjects originally entered the Inhibitors,research,lifescience,medical study, and these scores might have influenced the classification itself. Thus, the maximal possible logistic equations were learn more represented by Equation (1), where cosim represents the appropriate cosine similarity scores and the terms in parentheses were considered only for the MCI-c/MCI-n contrast. (1) Scores on the FAQ were obtained Inhibitors,research,lifescience,medical for each subject at baseline and at each follow-up visit. A linear mixed model was fitted using FAQ follow-up scores as the dependent variable, beginning with

a null Inhibitors,research,lifescience,medical model and refining it by the addition of subjects as a random effect. Fixed effects were then added and those that improved the model’s fit were left in. Candidate fixed effects included diagnostic group (NC, AD, MCI), baseline FAQ score, cosine similarity scores and their interactions, baseline MMSE score, and the interactions of each of these variables with time to follow-up (measured in months). Training classifiers The quality of the logistic regression models as classifiers was then evaluated by the following method. A logistic regression model (with the same variables that were Inhibitors,research,lifescience,medical chosen from the statistical analysis) was computed using all but one subject. Scores from the left-out subject

were then entered into the logistic model to compute an output between zero and one. This output was Mannose-binding protein-associated serine protease thresholded at 11 different levels on the interval between zero and one (with increments of 0.1) to derive predictions of the subject’s diagnostic or conversion status. The process was repeated for each subject, and prediction data were accumulated across all subjects. Sensitivity, specificity, and predictive value scores were calculated from the accumulated prediction data at each threshold level. Receiver operating characteristic (ROC) curves were constructed using the 11 different thresholds. The quality of the classifier at each threshold was determined by comparing it to a random classifier using McNemar’s chi-square and the best classifier was selected.