Comparison of User Models Based on GMM-UBM and I-Vectors for Speech, Handwriting, and Gait Assessment of Parkinson's Disease Patients

Juan Camilo Vásquez-Correa, Tobias Bocklet, Juan Rafael Orozco-Arroyave, Elmar Nöth

Introduction

Evaluating the progression of Parkinson’s Disease (PD) traditionally relies on the MDS-UPDRS-III scale, a clinical assessment performed by neurologists. While effective, this process is often subjective, expensive, and requires patients to travel to specialized clinics. This paper explores an objective, multimodal approach to neurological state assessment by adapting "user modeling" techniques—originally designed for biometric speaker recognition—to analyze speech, handwriting, and gait signals.

Methodology: Adapting Biometrics for Health

The researchers propose a system that maps specific motor impairments to a patient's overall neurological state using two primary modeling techniques:

  1. GMM-UBM (Gaussian Mixture Model - Universal Background Model): A "Universal" model representing healthy patterns is trained on Healthy Control (HC) subjects. This model is then adapted to create a specific model for each PD patient. The Bhattacharyya distance between the healthy UBM and the patient’s adapted model serves as a proxy for disease severity.
  2. i-vectors: This technique compresses high-dimensional feature data into a low-dimensional "total variability space." Patient i-vectors are compared to reference vectors from age- and gender-matched healthy peers using cosine distance.
Proposed methodology for multimodal neurological state assessment.

Features were extracted across three distinct modalities:

  • Speech: Phonation (vocal fold vibration), articulation, prosody (rhythm/intonation), and phonological posteriors.
  • Handwriting: Kinematic features such as pen pressure, velocity, and acceleration captured via a digital tablet.
  • Gait: Harmonic (spectral) and non-linear features (complexity/stability) captured via wearable ankle sensors.

Data and Experimental Setup

The study utilized an extended version of the PC-GITA corpus, involving 106 PD patients and 87 Healthy Controls, all native Colombian Spanish speakers. The dataset included 16kHz speech recordings, online handwriting data from a Wacom tablet, and 3D-accelerometer/gyroscope gait data. Professional neurologists provided ground truth labels for every patient using the MDS-UPDRS-III scale.

Key Results and Discussion

The experimental results highlighted several critical insights into digital PD diagnostics:

  • GMM-UBM Superiority: For this specific task, GMM-UBM models were more accurate than i-vectors. The authors suggest that i-vectors, which reduce a population's distribution to a single vector, may lose statistical nuance when training data is relatively limited.
  • Gait as a Primary Indicator: Gait features (specifically harmonic features) showed the highest individual correlation with the clinical scale ($\rho=0.619$). This is consistent with the MDS-UPDRS-III scale's heavy focus on lower-limb movement and balance.
  • The Power of Multimodal Fusion: No single signal captures the full complexity of PD. By combining the distances from all three modalities using linear regression, the researchers achieved a robust Spearman correlation of $\rho = 0.634$ with the clinical scores.
Correlation between predicted and real MDS-UPDRS-III scores using multimodal fusion.

While speech and handwriting showed weaker individual correlations (e.g., Kinematic handwriting at $\rho = 0.261$), they provide complementary information. Speech impairments represent only a small fraction of the total MDS-UPDRS-III scale, explaining why speech alone is an insufficient proxy for the entire neurological state.

Conclusion and Future Work

This study demonstrates that adapting biometric user models for multimodal analysis provides a viable path toward objective PD assessment. While gait remains the strongest individual predictor, the fusion of multiple motor signals is essential for a comprehensive diagnostic tool. Future work will focus on larger datasets to improve i-vector performance and the exploration of deep learning architectures, such as autoencoders, to automatically discover more nuanced neuromotor features.

Summary updated

BibTeX
@inproceedings{vasquezcorrea2020comparison,
  author = {Juan Camilo Vásquez-Correa and Tobias Bocklet and Juan Rafael Orozco-Arroyave and Elmar Nöth},
  title = {{Comparison of User Models Based on GMM-UBM and I-Vectors for Speech, Handwriting, and Gait Assessment of Parkinson's Disease Patients}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2020},
  doi = {10.1109/icassp40776.2020.9054348},
}
Figure