F023 Predicting the clinical status of patients with Huntington’s disease from raw speech samples, a lightweight machine learning approach

<h3></h3> Assessing HD patients usually requires a battery of tests that are both time-consuming and challenging to implement. To facilitate this assessment, a number of works propose to use speech, an emerging biomarker of HD, to predict the clinical scores traditionally used for the disease characterisation [Nunes et al., 2024, PMID: 38322583; Riad et al., 2022, PMID: 35567614]. While achieving honourable accuracy in clinical score regression, these approaches require the potentially incomplete selection and costly extraction of hand-crafted features, mostly derived from human annotations, making these difficult to deploy in a care setting. We propose a machine learning-based method for inferring the clinical status of patients with Huntington’s Disease (HD) from raw voice recordings. Data were collected as part of the BioHD (NTC01412125) and RepairHD (NCT03119246) studies. They involved 124 speakers, 33 healthy and 91 with HD (age = 51.9 ± 10.5, TFC = 10.1 ± 2.1). In contrast to previous approaches, we make no a priori assumptions about the speech HD correlates. We trained a neural network in a 5-fold cross validation paradigm to predict HD scores from mel spectrograms. Preliminary results show that our approach reaches an accuracy at least comparable to the best models published to date [Riad et al., 2022, PMID: 35567614]. By leveraging the ability of neural networks to capture information within the speech signal, our approach overcomes the need for human-annotated features, bringing the clinical assessment of HD patients one step closer to automation.