As the population in developed countries ages, larger numbers of people are at risk of developing dementia. In the near future large-scale time- and cost-efficient screening methods will be needed. Speech can be recorded and analyzed in this manner, and as speech and language are affected early on in the course of dementia, automatic speech processing can provide valuable support for such screening methods.

We have developed acoustic and linguistic features for dementia screening and established that a combination of acoustic and linguistic features provides the best results. However, our full set of 429 fine-grained features from 15 feature types is too large to train a robust model on limited training data. We therefore need to select features to use for dementia screening. We employ forward feature selection nested in a cross-validation and identify the most commonly selected features. Both acoustic and linguistic features from seven different feature types are selected. Using sets of these features we obtain a 0.819 unweighted average recall which is a strong improvement over previous results.