Speech-based Estimation of Bulbar Regression in Amyotrophic Lateral Sclerosis

Alan Wisler, Kristin Teplansky, Jordan Green, Yana Yunusova, Thomas Campbell, Daragh Heitzman, Jun WangNAACL 2019

TL;DR

This paper explores the use of speech and articulatory movement data to automatically estimate the bulbar subscore of the ALSFRS-R, a key clinical measure for tracking the progression of Amyotrophic Lateral Sclerosis (ALS). The authors demonstrate that machine learning models can predict this score from speech, offering a potential non-invasive method for continuous monitoring of disease progression.

What Problem Does This Paper Address?

Amyotrophic Lateral Sclerosis (ALS) is a devastating neurodegenerative disease that progressively impairs motor function, including speech. Accurately tracking disease progression is crucial for treatment planning and therapeutic intervention. The current standard, the ALSFRS-R, is a self-reported questionnaire that, while reliable, has limitations in capturing subtle changes in motor function. Specifically, the bulbar subscore, which assesses functions like speaking, salivating, and swallowing, is of particular interest due to the direct impact of ALS on speech. While previous research has used speech analytics to detect neurological diseases, this paper tackles the novel challenge of using speech and articulatory data to quantify the progression of bulbar dysfunction in ALS, aiming to provide a more objective and continuous monitoring tool.

Methods

The researchers employed two regression models: a simple ridge regression and a support vector regression (SVR). They extracted two main types of features:

  • Acoustic Features: These were based on Mel-frequency cepstral coefficients (MFCCs), which capture spectral information related to the vocal tract shape and can reflect articulatory changes. They extracted 14 MFCCs per frame, along with their first and second temporal derivatives, and then computed summary statistics (mean, standard deviation, pairwise variability) for a total of 126 acoustic features.
  • Articulatory Features: Using the NDI Wave System, they tracked the 3D movement of the tongue tip, tongue back, upper lip, and lower lip. For each sensor, they calculated the distance traveled between consecutive points in time. From these distance matrices, they extracted eight summary statistics (mean, standard deviation, skewness, kurtosis, etc.). They also applied three normalization procedures to control for overall movement distance and articulation space size. Combining these with four sensors yielded 128 articulatory features.

The models were trained and evaluated using a leave-one-speaker-out cross-validation approach to ensure generalization to new speakers. Predictions were made at the phrase level and then averaged for each patient-session pair. Performance was measured using Root Mean Squared Error (RMSE) and the correlation between predicted and true ALSFRS-R scores.

Datasets

The study utilized data from 66 participants diagnosed with ALS at an early onset. For the regression analysis, a subset of 28 participants (15 male, 13 female, average age 57.3 years) with up to four data collection sessions was used. The ALSFRS-R bulbar subscore was collected from these participants. The speech stimuli consisted of 20 pre-selected sentences commonly used in Augmentative and Alternative Communication (AAC) devices. No new benchmark datasets were introduced; the study used existing ALS patient data.

Key Insights and Ideas

  • Speech as a Window into Bulbar Function: The core idea is that subtle changes in speech production, both acoustic and physical (articulation), directly reflect the underlying bulbar motor neuron degeneration in ALS.
  • Articulatory Movement Adds Value: Combining articulatory data (tongue, lip movements) with acoustic features significantly improved the prediction of the ALSFRS-R bulbar subscore compared to using acoustic features alone.
  • Tongue is King (for Articulation): When considering articulatory features in isolation, tongue sensors provided more predictive power than lip sensors, aligning with the tongue's primary role in speech articulation.
  • SVR Outperforms Ridge Regression: The Support Vector Regression model generally achieved better performance, suggesting its effectiveness in modeling the complex, non-linear relationships between speech features and ALS progression.
  • Potential for Remote Monitoring: The success of this approach opens doors for developing mobile applications that could allow patients to record speech and lip motion (via webcam) for remote, continuous monitoring of their disease progression.

Key Results

The highest performance was achieved by the SVR model using acoustic data combined with all articulatory motion data, resulting in an RMSE of 1.78 and a correlation of r = 0.64 with the true ALSFRS-R bulbar scores.

  • The combined "Acoustic + Tongue + Lips" feature group yielded the best results for both RMSE and correlation.
  • Acoustic features alone provided a decent baseline, but adding articulatory features significantly improved performance.
  • The tongue sensor data alone was more predictive than lip sensor data alone.
  • Interestingly, adding tongue data to the "Acoustic + Lips" group did not significantly improve performance beyond the "Acoustic + Lips" group, suggesting further investigation is needed.

Limitations and Future Work

  • Limited Resolution of ALSFRS-R: The authors acknowledge that the ALSFRS-R itself has a limited resolution (5-point scale) and may not capture very subtle motor function changes, which could affect the ground truth for their predictions.
  • Generalization to Non-Bulbar Subscores: This study focused specifically on the bulbar subscore. Future work could explore predicting other ALSFRS-R components (e.g., motor function, respiratory function).
  • Investigating Tongue Data: Further research is needed to understand the specific contributions of different aspects of tongue movement data and how to best incorporate them.
  • Exploring Advanced Models: The authors suggest exploring more complex models like Convolutional Neural Networks (CNNs), which have shown promise in ALS early detection.
  • Non-Speech Modalities: Future research could investigate incorporating non-speech information to predict broader ALS motor function decline.
  • Mobile Health Applications: The findings support the development of mobile apps for home-based data collection, potentially using webcams for lip motion tracking, which could be a more accessible and scalable approach.

Summary updated

BibTeX
@inproceedings{wisler2019speech,
  author = {Alan Wisler and Kristin Teplansky and Jordan Green and Yana Yunusova and Thomas Campbell and Daragh Heitzman and Jun Wang},
  title = {{Speech-based Estimation of Bulbar Regression in Amyotrophic Lateral Sclerosis}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2019},
}
Figure