Analysis of Parkinson's speech patterns for digital biomarker discovery and intervention assessment

Parkinson's disease (PD) affects motor and cognitive abilities and leads to communication impairments that significantly impact on quality of life for patients and their caregivers. Current speech and language therapies rely on voice amplification training to improve voice clarity, however there are other conversation strategies that could benefit patients with different types of communication impairment. There is an unmet clinical need to fully characterise the spectrum of PD-related communication deficits that interfere with normal participation in everyday life and impact on quality of life. This would facilitate patient selection into treatment trials based on precise speech and communication characteristics (which could also be used as outcome measures for those trials). These clinical biomarkers could also be used in the diagnosis and/or monitoring of disease states.

To date, little research has been done on natural dialogical speech in PD, despite the fact that it is widely acknowledged that the study of communicative behaviour in interactive contexts is crucial to our understanding of cognitive-communication disorders1. Unfortunately, manual analysis of spontaneous speech communication in natural environments poses significant challenges in terms of resources and researcher time. The development of AI tools would represent a significant advance in speech therapy research, with the potential for translation into routine clinical practice.

Aims

This project aims to investigate novel methods for automatic annotation and analysis of PD patients' spontaneous speech, with a focus on dialogue, for the detection of conversation difficulties and speech features that may be useful as digital biomarkers. Specifically, the project aims to:

1. analyse connected speech samples from participants at different stages of PD as well as control speakers, in controlled (reading) tasks and spontaneous dialogues to extract vocalisation, speech turns, linguistic content and paralinguistic features for machine learning modelling, building on methods developed in our lab for AD2;

2. model PD speech for diagnosis, progression monitoring and classification, using existing datasets (such as DementiaBank, Italian Parkinson's Voice and Speech Database, and the MDVR-KCL dataset)

3. model PD speech in dialogues using data collected as part of a treatment study conducted by Dr. Roberts, consisting of 10 dyads, each comprising a person with moderate stage PD and their spouse. Conversation data were recorded (audio and video) in participants' homes as part of the baseline treatment data. Conversations were recorded in a naturalistic environment during typical family mealtimes using high fidelity recording equipment without the presence of researchers. Each conversation file contains ~60 minutes of conversation data. New data will be collected from cohorts in Scotland and Canada, and models will be tested on these data to assess whether temporal and prosodic differences affect the accuracy of segmentation and inference. Models will be assessed against the gold standard of manual annotation.

Training outcomes

The student will receive training in analysis of multimodal data sets including speech data for the creation of clinically relevant models. This will be done via advanced signal processing and machine learning approaches, including speech segmentation and clustering methods, deep neural networks, and vocalisation modelling3. The student will receive general training in machine learning methodology through attendance in DTP courses and through Dr Luz's lab, and specific training on communication disorders.