Extracting Phonetic Posterior-Based Features for Detecting Multiple Sclerosis From Speech.

Area of Science:

• Neurology
• Speech Science
• Machine Learning

Background:

• Multiple sclerosis (MS) is a chronic central nervous system inflammatory disease impacting motor, cognitive, and speech functions.
• Speech changes in MS patients may indicate executive function limitations, affecting production, comprehension, and narrative structure.
• MS can cause dysarthria, a motor speech disorder potentially altering phonetic characteristics.

Purpose of the Study:

• To develop and evaluate a speech-based machine learning technique for distinguishing relapsing-remitting MS patients from healthy controls.
• To investigate the utility of phonetic posterior estimates derived from Deep Neural Network acoustic models for MS detection.
• To explore the potential of this method for automatic MS screening and disease progression monitoring.

Main Methods:

• Utilized a machine learning approach analyzing speech features derived from phonetic posterior estimates.
• Employed a Deep Neural Network acoustic model to generate posteriorgrams.
• Compared the performance of the proposed posteriorgram-based features against x-vector and openSMILE 'ComParE functionals' attributes.

Main Results:

• The posteriorgram-based feature extraction method demonstrated effectiveness in detecting MS.
• Achieved Equal Error Rate (EER) values as low as 13.3% and Area Under the Curve (AUC) scores up to 0.891.
• Exhibited competitive and more consistent classification performance compared to existing methods like x-vector and openSMILE.

Conclusions:

• The developed speech-based machine learning technique shows significant potential for accurate MS detection.
• The interpretable nature of phonetic posterior features supports its use in automatic MS screening and progression monitoring.
• Phonetic feature analysis may offer valuable insights for speech therapy interventions in MS patients.