Machine Learning-Based Scoring System to Predict the Risk and Severity of Ataxic Speech Using Different Speech Tasks

Area of Science:

• Neurology
• Speech Science
• Biomedical Engineering

Background:

• Assessing speech in Cerebellar Ataxia (CA) is currently time-consuming and subjective.
• Objective, automated methods are needed for efficient and reliable diagnosis and severity rating of Ataxic Speech.

Purpose of the Study:

• To develop and validate a fully automated objective algorithm for the diagnosis and severity assessment of Ataxic Speech.
• To utilize machine learning models based on acoustic features for a 3-tier diagnostic categorization.

Main Methods:

• Employed acoustic features from time, spectral, cepstral, and non-linear dynamics in microphone data from Consonant-Vowel (C-V) syllable paradigms.
• Utilized mass univariate analysis and elastic net regularization for feature selection in diagnosis, and Spearman's rank-order correlation for severity.
• Developed and evaluated the algorithm on recordings from 126 participants (65 with CA, 61 controls).

Main Results:

• Achieved an Area Under the Curve (AUC) of 0.97 for Ataxic Speech diagnosis with high sensitivity (97.43%) and specificity (85.29%).
• Obtained a mean AUC of 0.74 for Ataxic Speech severity estimation.
• Prediction nomograms showed high efficacy with C-indexes of 0.96 for diagnosis and 0.81 for severity prediction.

Conclusions:

• The automated algorithm demonstrates strong classification ability for identifying and monitoring Ataxic Speech.
• The findings support the framework's usefulness in clinical settings for objective speech assessment in Cerebellar Ataxia.
• Decision curve analysis confirmed the value of acoustic features from repeated C-V syllable paradigms.