Detecting early signs of dementia in conversation

Dementia can affect a person's speech, language and conversational interaction capabilities. The early diagnosis of dementia is of great clinical importance. Recent studies using the qualitative methodology of Conversation Analysis (CA) demonstrated that communication problems may be picked up during conversations between patients and neurologists and that this can be used to differentiate between patients with Neuro-degenerative Disorders (ND) and those with non-progressive Functional Memory Disorder (FMD). However, conducting manual CA is expensive and difficult to scale up for routine clinical use. This study introduces an automatic approach for processing such conversations which can help in identifying the early signs of dementia and distinguishing them from the other clinical categories (FMD, Mild Cognitive Impairment (MCI), and Healthy Control (HC)). The dementia detection system starts with a speaker diarisation module to segment an input audio file (determining who talks when). Then the segmented files are passed to an automatic speech recogniser (ASR) to transcribe the utterances of each speaker. Next, the feature extraction unit extracts a number of features (CA-inspired, acoustic, lexical and word vector) from the transcripts and audio files. Finally, a classifier is trained by the features to determine the clinical category of the input conversation. Moreover, we investigate replacing the role of a neurologist in the conversation with an Intelligent Virtual Agent (IVA) (asking similar questions). We show that despite differences between the IVA-led and the neurologist-led conversations, the results achieved by the IVA are as good as those gained by the neurologists. Furthermore, the IVA can be used for administering more standard cognitive tests, like the verbal fluency tests and produce automatic scores, which then can boost the performance of the classifier. The final blind evaluation of the system shows that the classifier can identify early signs of dementia with an acceptable level of accuracy and robustness (considering both sensitivity and specificity).