We present a novel approach for diagnostics of different forms of cognitive impairment (Alzheimer disease and Semantic Dementia) from the description of a picture. Traditional linguistic tests require a picture-proprietary list of Information Content Units (ICUs), either manually crafted or automatically generated from the reference text. We show that the requirement above is no longer necessary. A text classifier trained on some arbitrary scenes can detect participant's impairment by descriptions of other scenes, unseen in training. The same algorithm trained only on the control group can predict the age group of participant, his educational level, and gender. We compare two types of learning – a feature-based learning and deep convolutional networks trained on word vector representations of a raw text.