Use of Neural Networks to Detect Impaired Cheating on the Computerized CERAD Word List Memory Test.

The computerized CERAD Word List Memory Test is one of several tools used by the Long Term Care insurance industry to screen applicants for mild cognitive impairment. A method to detect cheating, by cognitively impaired individuals, has been devised using a two stage approach: 1) distinguish cheaters from non-cheaters 2) distinguish impairedcheaters from unimpaired-cheaters. We employed the use of neural networks (NN) and compared our results with an unpublished statistical method. We found that our NN did a better job in stage 1 (2Stage1) by virtue of having a dominating average ROC curve and narrow confidence bands. We explored a one stage approach (1Stage) and produced a model with a greater average AUROC than in 2Stage2. We compared the two stage and one stage NN using specific cutoff values and found that the former approach marginally outperformed the latter approach by having a higher sensitivity with roughly equivalent positive predictive value and specificity.