TL;DR
This study explores whether specific acoustic and perceptual "deviance scores" from the MonPaGe-2.0.s protocol can automatically differentiate between various subtypes of motor speech disorders (MSDs). By training decision trees on these scores, the researchers achieved promising classification accuracies, suggesting a potential for more objective and efficient MSD diagnosis.
What Problem Does This Paper Address?
Diagnosing motor speech disorders (MSDs) and their subtypes is a significant challenge in clinical practice. MSDs, which affect the planning, programming, and execution of speech movements, can manifest in various ways, leading to overlapping symptoms across different conditions. Traditionally, diagnosis relies on auditory-perceptual assessments, which suffer from low inter-judge agreement and can be subjective. While acoustic analysis offers a more objective approach, distinguishing between the numerous subtypes of MSDs (like apraxia of speech and various forms of dysarthria) remains difficult. This paper aims to bridge this gap by investigating if a set of "clinically informed" acoustic and perceptual features, captured as deviance scores from a standardized protocol (MonPaGe-2.0.s), can effectively automate the classification of different MSD subtypes.
Methods
The core of this study lies in using machine learning, specifically decision trees, to classify patients into different MSD subtypes. The researchers employed a two-class classification strategy, meaning they trained models to distinguish between pairs of MSD subtypes. The key innovation here is the use of seven "deviance scores" (DevS) derived from the MonPaGe-2.0.s protocol. These scores represent how much a patient's speech characteristics deviate from age- and sex-matched normative data across dimensions like intelligibility, articulation, voice, speech rate, maximum phonation time, prosody, and diadochokinetic rate. Instead of feeding raw acoustic features into the models, they used these pre-computed, clinically meaningful deviance scores. Decision trees were chosen for their interpretability and ability to handle both numerical and categorical data. To prevent overfitting on a relatively small dataset, they used a leave-one-speaker-out cross-validation strategy and limited the tree depth.
Datasets
The study utilized data from 60 patients diagnosed with mild to moderate MSDs, drawn from the "MoSpeeDi dataset." These patients were categorized into six neuropathological groups:
- Post-stroke apraxia of speech (AoS)
- Hypokinetic dysarthria in Parkinson Disease (PD)
- Mixed dysarthria in amyotrophic lateral sclerosis (ALS)
- Flaccid dysarthria in Kennedy disease
- Ataxic dysarthria in spinocerebellar ataxia (SCA)
- Mixed dysarthria in Wilson disease
The MonPaGe-2.0.s protocol itself was previously normalized on a larger dataset of 404 neurotypical French speakers.
Key Insights and Ideas
- Clinically Informed Features are Powerful: The study demonstrates that using "clinically informed" deviance scores, rather than raw acoustic features, can be highly effective for classifying MSD subtypes. This suggests that features already interpreted in a clinical context hold significant diagnostic power.
- Diadochokinetic Rate and Articulation are Key Discriminators: For most pairwise classifications, the diadochokinetic (DDK) rate and articulation deviance scores emerged as the most important features for distinguishing between MSD subtypes. This highlights the critical role of fine motor control and rapid articulatory movements in differentiating these disorders.
- Some Features are Less Informative: Intelligibility, maximum phonation time (MPT), and prosody scores showed minimal contribution to the classification accuracy. This might be due to the mild severity of the patient cohort or limitations in how these specific aspects of speech are captured by the current MonPaGe protocol.
- Subtype-Specific Patterns Emerge: The study identified distinct patterns of deviance for different groups. For instance, Parkinson's patients showed primary abnormalities in articulation and voice, while AoS, ALS, and SCA groups exhibited a broader range of impairments.
- Potential for Objective Diagnosis: The promising classification accuracies suggest that this approach could lead to more objective and standardized diagnostic tools, reducing reliance on subjective perceptual judgments.
Key Results
The study achieved promising classification accuracies for differentiating between MSD subtypes:
- Overall accuracy: Ranged from 79% to 87% for dysarthria subtypes.
- Apraxia of Speech (AoS) classification: Was slightly lower, around 79%.
- Highest accuracies: Were observed when classifying SCA (ataxic dysarthria), PD (hypokinetic dysarthria), and Kennedy disease (flaccid dysarthria), with average performances around 85-87%.
- Specific Pairwise Classifications: Achieved 100% accuracy in distinguishing Kennedy disease (flaccid dysarthria) from ALS (mixed dysarthria) and very high accuracies (95%) for SCA (ataxic dysarthria) vs. AoS and PD (hypokinetic dysarthria) vs. Wilson disease.
- Lowest accuracies: Were around 70% for certain classifications, such as AoS vs. Kennedy disease or ALS vs. SCA.
- Feature Importance: DDK rate and articulation were the most important features for most classifications, followed by voice quality. Speech rate played a significant role in some distinctions, particularly between AoS and Wilson disease.
Limitations and Future Work
- Small Cohort Size: The study acknowledges that the classification was tested on relatively small cohorts (10 patients per group), which limits the generalizability of the findings.
- Mild to Moderate Impairment: The focus on mild to moderate MSDs might explain why some deviance scores (like intelligibility) were less informative.
- Feature Granularity: The authors suggest that further refinement of the existing deviance scores (e.g., breaking down composite scores like DDK rate into individual components) and the inclusion of additional features could improve classification accuracy.
- AoS Specificity: The lower accuracy for AoS classification suggests that current DevS might not fully capture the specific planning and programming deficits characteristic of AoS. Future work should aim to better differentiate impairments related to planning versus motor execution.
- Prosody Measurement: The current prosody measure, focusing on a specific linguistic contrast, may not capture broader dysprosodic features. Developing more comprehensive prosody measures is suggested.
- Discarding MPT: The Maximum Phonation Time (MPT) score was found to be largely uninformative and may be considered for removal in future protocol developments.
Summary updated