Voice Analysis for Neurological Disorder Recognition-A Systematic Review and Perspective on Emerging Trends

Pascal Hecker, Nico Steckhan, Florian Eyben, Björn W. Schuller, Bert ArnrichFrontiers Digit. Health 2022 · DOI

TL;DR

This paper presents a systematic review of how voice analysis is used to detect neurological disorders, synthesizing current practices and highlighting emerging trends. It provides a valuable overview for researchers planning new data collection efforts, emphasizing the potential of everyday-life data and multimodal approaches for more natural and accurate disorder recognition.

What Problem Does This Paper Address?

Diagnosing and monitoring neurological disorders often relies on clinical examinations that can be time-consuming and subjective. The paper addresses the challenge of developing more objective, unobtrusive, and scalable methods for disorder detection. Voice analysis offers a promising avenue due to its susceptibility to subtle physiological changes and the widespread availability of recording devices like smartphones. However, the field is fragmented, with varying approaches to data collection, feature extraction, and analysis across different disorders. This review aims to consolidate this knowledge, identify common practices, and point towards future directions.

Methods

The authors conducted a systematic literature review following PRISMA guidelines, searching PubMed, Web of Science, and IEEE Xplore for studies that collected original (newly recorded) datasets for voice analysis in neurological disorders. They extracted information on:

  • Study Setup: Patient assessment and cohort characteristics.
  • Voice Recording Setup: Speech tasks used, recording conditions (laboratory vs. everyday life), and the use of additional modalities.
  • Analysis Methods: Feature extraction toolkits and predictive modeling approaches.

They also extended a previous review's synthesis of acoustic features correlated with disorders, incorporating neurodegenerative conditions.

Datasets

The review identified 43 studies that reported on original, newly recorded datasets. The disorders covered included psychiatric (depression, anxiety, stress, etc.) and neurodegenerative conditions (Alzheimer's, Parkinson's, ALS, etc.), as well as speech impairments. Parkinson's disease was the most represented disorder with 19 datasets. The paper does not introduce new benchmark datasets but rather synthesizes findings from existing ones.

Key Insights and Ideas

  • Free Speech Dominates: Free speech tasks are the most frequently used elicitation method across disorders, likely due to their ability to capture natural speech patterns and subtle indicators like hesitations and pauses.
  • Toolkits vs. Custom: While established toolkits like OpenSMILE and PRAAT are common, a significant portion of studies employ custom feature extraction methods, particularly for Parkinson's disease, often tailored to specific treatment effects (e.g., LSVT).
  • Statistical Analysis and SVMs Prevail: Statistical analysis for individual feature significance and predictive modeling using Support Vector Machines (SVMs) are the most prevalent analytical approaches. However, deep learning methods, especially Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks, are gaining traction.
  • Everyday Life is the Future: A major emerging trend is collecting voice data in "everyday life" settings using smartphones and smartwatches. This allows for longitudinal data collection and captures more natural behavior, moving beyond controlled laboratory environments.
  • Multimodality Boosts Performance: Combining voice data with other modalities (e.g., video, physical activity, biosignals) is another key trend, showing promise for improving the accuracy of disorder recognition, especially for conditions affecting motor control or emotional expression.

Key Results

  • Parkinson's Disease Leads: Parkinson's disease has the highest number of studies and datasets (20 studies), followed by stress (6 studies) and depression (5 studies).
  • Feature Correlations: The review synthesizes correlations between acoustic features and disorders, extending previous work to include neurodegenerative conditions. For example, jitter and shimmer are consistently linked to ALS progression, while increased pause duration is characteristic of early AD.
  • Emerging Trends Highlighted: The paper identifies "everyday-life data collection" and "multiple modalities" as significant emerging trends, suggesting these will be crucial for future advancements.

Limitations and Future Work

  • Confounders: The paper acknowledges that factors like age, sex, native language, comorbidities, medication, and even dialect can confound voice analysis results and need careful control.
  • Data Imbalance: Imbalanced class distributions in datasets are a common challenge, and the paper suggests using metrics like unweighted average recall (UAR) for evaluation.
  • Custom Methods: The prevalence of custom feature extraction methods makes direct comparison across studies difficult.
  • Future Work: The authors strongly advocate for collecting data in everyday life to facilitate longitudinal studies and capture natural behavior. They also emphasize the need for further research into multimodal data processing to effectively fuse information from different sources and balance the benefits against increased complexity. The development of user-friendly recording protocols and robust validation procedures is also crucial for bridging the gap between research and real-world applications.

Summary updated

BibTeX
@article{hecker2022voice,
  author = {Pascal Hecker and Nico Steckhan and Florian Eyben and Björn W. Schuller and Bert Arnrich},
  title = {{Voice Analysis for Neurological Disorder Recognition-A Systematic Review and Perspective on Emerging Trends}},
  journal = {Frontiers Digit. Health},
  year = {2022},
  doi = {10.3389/fdgth.2022.842301},
}
Figure