In today’s aging society, the number of neurodegenerative dis-eases such as Alzheimer’s disease (AD) increases. Reliable tools forautomatic early screening as well as monitoring of AD patients arenecessary. For that, semantic deficits have been shown to be usefulindicators. We present a way to significantly improve the methodintroduced by Wankerl et al. [1]. The purely statistical approach ofn-gram language models (LMs) is enhanced by using the rwthlmtoolkit to create neural network language models (NNLMs) withLong Short Term-Memory (LSTM) cells. The prediction is solelybased on evaluating the perplexity of transliterations of descriptionsof the Cookie Theft picture from DementiaBank’s Pitt Corpus. Eachtransliteration is evaluated on LMs of both control and Alzheimerspeakers in a leave-one-speaker-out cross-validation scheme. Theresulting perplexity values reveal enough discrepancy to classify pa-tients on just those two values with an accuracy of 85.6%at equal-error-rate.