Staircase classification with acoustic representations for ALS severity recognition

Tools

Syed, Zafi Sherhan and Perry, Ross and Conway, Frank and Lowit, Anja and Cohen, Wendy and Di Caterina, Gaetano (2026) Staircase classification with acoustic representations for ALS severity recognition. In: IEEE International Conference on ICT Solutions for eHealth, 2026-06-23 - 2026-06-26. (In Press)

[thumbnail of Syed-etal-2026-Staircase-classification-with-acoustic-representations-for-ALS-severity-recognition] Text. Filename: Syed-etal-2026-Staircase-classification-with-acoustic-representations-for-ALS-severity-recognition.pdf
Accepted Author Manuscript
Restricted to Repository staff only until 1 January 2099.
Download (156kB) | Request a copy

Abstract

Amyotrophic lateral sclerosis (ALS) is a neurological illness that affects motor functions including speech. The resulting dysarthria manifests as changes in articulation and voice production. There is a growing interest in developing AI models that use speech recordings to identify the severity of ALS with the aim of enabling remote telehealth monitoring, and the ICASSP 2026 Speech Analysis for Neurodegenerative Diseases (SAND) Challenge is one such endeavour. To that end, we propose a pipeline that uses speech signals alone to identify the severity of ALS. We compare domain-knowledge-based handcrafted acoustic features with deep acoustic embeddings (DAEs), and also investigate staircase classification as a means to exploit the ordinal structure of severity labels. Experimental results on the official validation partition of the SAND Challenge dataset show that DAEs outperform handcrafted features, while staircase classification further improves performance over the standard logistic regression classifier. The best models achieve macro F1-scores of 0.729 for Task 1 and 0.671 for Task 2 of the Challenge. Overall, the study highlights the importance of acoustic representation and ordinal modelling for speech-based ALS severity classification.

ORCID iDs

Syed, Zafi Sherhan, Perry, Ross ORCID logoORCID: https://orcid.org/0009-0008-5315-2987, Conway, Frank, Lowit, Anja ORCID logoORCID: https://orcid.org/0000-0003-0842-584X, Cohen, Wendy ORCID logoORCID: https://orcid.org/0000-0002-1271-9229 and Di Caterina, Gaetano ORCID logoORCID: https://orcid.org/0000-0002-7256-0897;
CORE (COnnecting REpositories)