Early detection of cognitive impairment is of great clinical importance. Current cognitive tests assess language and speech abilities. Recently, we have developed a fully automated system to detect cognitive impairment from the analysis of conversations between a person and an intelligent virtual agent (IVA). Promising results have been achieved, however more data than is typically available in the medical domain is required to train more complex classifiers. Data augmentation using generative models has been demonstrated to be an effective approach. In this paper, we use a variational autoencoder to augment data at the feature-level as opposed to the speech signal-level. We investigate whether this suits some feature types (e.g., acoustic, linguistic) better than others. We evaluate the approach on IVA recordings of people with four different cognitive impairment conditions. F-scores of a four-way logistic regression (LR) classifier are improved for certain feature types. For a deep neural network (DNN) classifier, the improvement is seen for almost all feature types. The F-score of the LR classifier on the combined features increases from 55% to 60%, and for the DNN classifier from 49% to 62%. Further improvements are gained by feature selection: 88% and 80% F-scores for LR and DNN classifiers respectively.
Cite as: Mirheidari, B., Blackburn, D., O’Malley, R., Venneri, A., Walker, T., Reuber, M., Christensen, H. (2020) Improving Cognitive Impairment Classification by Generative Neural Network-Based Feature Augmentation. Proc. Interspeech 2020, 2527-2531, doi: 10.21437/Interspeech.2020-2433
@inproceedings{mirheidari20_interspeech,
author={Bahman Mirheidari and Daniel Blackburn and Ronan O’Malley and Annalena Venneri and Traci Walker and Markus Reuber and Heidi Christensen},
title={{Improving Cognitive Impairment Classification by Generative Neural Network-Based Feature Augmentation}},
year=2020,
booktitle={Proc. Interspeech 2020},
pages={2527--2531},
doi={10.21437/Interspeech.2020-2433}
}