Document resource
Background There is an urgent need for scalable, non-invasive and quantifiable biomarkers in neurodegenerative disorders (NDDs) including dementia. Speech is an attractive candidate, with potential for remote and cost-effective assessments. Development of robust models is currently limited by a lack of high quality clinically annotated speech data. We present a large prospective longitudinal speech cohort including people with dementia, MND, Parkinson’s disease (PD), and progressive multiple sclerosis (MS) assessed using novel machine learning (ML) approaches.Methods People living with NDDs and healthy individuals recorded longitudinal standardised recording tasks on an App co-produced with patients, aligned to contemporaneous deep clinical phenotyping (clinical rating scales, cognitive tests and blood-based biomarkers). Conventional and deep learning features were extracted for inputs in classification and regression ML models. Data for classification was matched and balanced by age and sex.Results 771 participants provided 4984 recordings over 1004 assessments. Disease vs healthy classifiers demonstrated promising accuracies: MND 0.81, Alzheimer’s disease 0.74, PD 0.74, secondary progressive MS 0.63. Speech predicted disease severity in MND (ALSFRS-R speech: R2 0.56).Conclusion Application of ML approaches to speech suggests potential diagnostic and monitoring utility in NDDs. Ongoing work includes scaling data acquisition and analysis across larger more diverse populations.jtam@ed.ac.uk