Automatic speech analysis for early Alzheimer's disease diagnosis

Saeideh Mirzaei, Mounîm A. El‐Yacoubi, Sonia Garcia-Salicetti, Jérôme Boudy, Christian Kahindo Senge Muvingi, Victoria Cristancho‐Lacroix, Hélène Kerhervé, Anne‐Sophie RigaudHAL 2017

TL;DR

This paper explores using automatic speech analysis to detect early signs of Alzheimer's disease (AD) and Mild Cognitive Impairment (MCI). By analyzing acoustic features from read speech, researchers developed classification models that show promise in distinguishing between healthy individuals, MCI patients, and AD patients, with potential for low-cost, non-invasive early diagnosis.

What Problem Does This Paper Address?

Alzheimer's disease (AD) is a progressive neurodegenerative disorder that significantly impacts memory and cognitive functions. Early and accurate diagnosis is crucial for timely intervention and management, but current diagnostic methods like CT scans and MRIs can be invasive and costly. Mild Cognitive Impairment (MCI) is often a precursor to AD, and identifying individuals with MCI who are likely to progress to AD is a significant challenge. This paper addresses the need for accessible, non-invasive methods for early detection by investigating whether subtle changes in speech patterns can serve as reliable biomarkers for AD and MCI.

Methods

The researchers extracted a set of 15 acoustic features from read speech recordings. These features fall into three categories:

  • Fundamental Frequency (F0): Features related to the basic pitch of the voice, including minimum, maximum, mean, and standard deviation.
  • Speech Segments: Measures of speech and non-speech duration, and the number of speech interruptions.
  • Jitter and Shimmer: These quantify variations in voice pitch (jitter) and amplitude (shimmer) between consecutive speech cycles, indicating vocal instability.

Two classification algorithms were employed: K-Nearest Neighbors (kNN) and Support Vector Machines (SVM). To optimize the feature set, a wrapper feature selection method using backward elimination was applied, where features were removed if their elimination improved or maintained classification accuracy.

Datasets

The study utilized a dataset of voice recordings from 42 participants, equally divided into three groups:

  • Healthy Controls (HC)
  • Mild Cognitive Impairment (MCI)
  • Alzheimer's Disease (AD)

Each group consisted of 14 individuals over 60 years old, who were native French speakers with reading and writing abilities. Participants underwent a Mini-Mental State Examination (MMSE) with a score of 20 or higher, and were not taking any medications that could affect cognitive performance. The participants were asked to read a passage from "Le petit prince." The recordings were sampled at 44100 Hz.

Key Insights and Ideas

  • Speech as a Biomarker: The study reinforces the idea that subtle changes in vocal characteristics can reflect underlying neurological conditions like AD and MCI.
  • Feature Importance: While a comprehensive set of features was initially used, feature selection revealed that a smaller subset could achieve comparable or even better classification performance, suggesting specific acoustic markers are more discriminative.
  • kNN Outperforms SVM: In this study, the kNN classifier generally demonstrated superior performance compared to SVM, particularly in distinguishing between the groups.
  • Confusion Between Groups: A notable finding is the high confusion rate between the AD and HC groups, and also between MCI and HC groups, especially with the SVM classifier. This highlights the challenge in differentiating these conditions solely based on the analyzed speech features.

Key Results

  • Full Feature Set Performance: Using the full set of 15 features, kNN achieved a total accuracy of 55%, while SVM achieved 36%.
  • Selected Feature Set Performance: After feature selection, kNN used 10 features and achieved a total accuracy of 58%. SVM used 3 features and achieved a total accuracy of 55%.
  • Relative Improvement: Feature selection led to a relative accuracy improvement of 5% for kNN and a significant 53% improvement for SVM.
  • Classification Accuracy by Group: kNN showed the best performance in classifying the MCI group (68% accuracy with the full feature set). Both classifiers struggled more with distinguishing AD patients, often confusing them with MCI patients.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Heterogeneity of MCI: The study notes that the heterogeneity within the MCI group might contribute to classification difficulties.
  • Feature Set Limitations: The current set of acoustic features might lack sufficient discriminatory information, especially for differentiating AD from HC.
  • Future Work: The researchers plan to explore additional features, such as Mel-Frequency Cepstral Coefficients (MFCCs) and other temporal features. They also intend to use a larger dataset and investigate other classification algorithms to improve the discrimination between AD-MCI, MCI-HC, and AD-HC classes.

Summary updated

Abstract

International audience
BibTeX
@inproceedings{mirzaei2017automatic,
  author = {Saeideh Mirzaei and Mounîm A. El‐Yacoubi and Sonia Garcia-Salicetti and Jérôme Boudy and Christian Kahindo Senge Muvingi and Victoria Cristancho‐Lacroix and Hélène Kerhervé and Anne‐Sophie Rigaud},
  title = {{Automatic speech analysis for early Alzheimer's disease diagnosis}},
  booktitle = {HAL},
  year = {2017},
}
Figure