Fully Automated Speech-based Frontline Screening for Dementia

Johannes Tröger; Nicklas Linz; Alexandra König; Jessica Peter; Jan Alexandersson; Philippe Robert
In: Proceedings of 26th European Congress of Psychiatry. European Congress of Psychiatry (EPA-18), Elsevier, 2018.


Introduction: New telemedicine tools for effective frontline screening for dementia are needed. Automatic speech analysis can be a powerful solution to address this need as speech can be analysed via telephone. Objectives: To benchmark a fully automated dementia frontline screening based on automatic speech recognition and machine learning classification. Methods: 166 Elderly people diagnosed with either dementia (D; 79), Mild Cognitive Impairment (MCI; 47) or only subjective memory complaints (SMC; 40), were assessed at the memory clinic at the Institut Claude Pompidou in Nice, France. Within the scope of the Dem@care and ELEMENT projects participants performed a battery of speech-based cognitive tests. Speech was recorded and processed using automatic speech recognition (Figure 1). The experiment for this study is solely based on the recordings of the 60s semantic verbal fluency (SVF). Qualitative SVF features were extracted according to previous work (Linz et. al., 2017; Using Neural Word Embeddings in the Analysis of the Clinical Semantic Verbal Fluency Task). For classification, Support Vector Machines with 10-fold cross validation were used. Results: Despite imperfect ASR—mainly misses in SVF correct responses—for the screening scenario (SMC vs. MCI & D) the approach achieves a sensitivity of .99 and a specificity of .74. Intergroup classification results were as follows (Sensitivity/Specificity): SMC vs. D (.91/.66), SMC vs. MCI (.85/.62) and MCI vs. D (.86/.49). Conclusion: The results show high ecological validity of the proposed automatic analysis for frontline screening of dementia. The high sensitivity in the screening scenario underlines its feasibility in telemedicine applications.

Weitere Links