[Paper Review] 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+2 more|arXiv (Cornell University)|Feb 13, 2020

Voice and Speech Disorders25 references4 citations

TL;DR

This study proposes using GMM-UBM and i-vector user models to assess Parkinson’s disease (PD) severity via multimodal signals—speech, handwriting, and gait—by modeling patient-specific traits against healthy control references. GMM-UBM outperformed i-vectors, achieving a Spearman correlation of 0.634 with the MDS-UPDRS-III scale when fusing harmonic gait, prosody, and articulation features.

ABSTRACT

Parkinson's disease is a neurodegenerative disorder characterized by the presence of different motor impairments. Information from speech, handwriting, and gait signals have been considered to evaluate the neurological state of the patients. On the other hand, user models based on Gaussian mixture models - universal background models (GMM-UBM) and i-vectors are considered the state-of-the-art in biometric applications like speaker verification because they are able to model specific speaker traits. This study introduces the use of GMM-UBM and i-vectors to evaluate the neurological state of Parkinson's patients using information from speech, handwriting, and gait. The results show the importance of different feature sets from each type of signal in the assessment of the neurological state of the patients.

Motivation & Objective

• To evaluate the neurological state of Parkinson’s disease (PD) patients using multimodal biometric signals—speech, handwriting, and gait.
• To investigate whether user models based on GMM-UBM and i-vectors can effectively capture PD-related motor impairments across different modalities.
• To compare the performance of GMM-UBM and i-vector systems in predicting MDS-UPDRS-III scores using diverse feature sets from each modality.
• To explore the contribution of individual feature sets and their fusion in improving neurological state prediction accuracy.
• To identify the most informative features across modalities for assessing PD severity objectively and non-invasively.

Proposed method

• Extracted modality-specific features: phonation, articulation, and prosody from speech; kinematic features from handwriting; harmonic and non-linear features from gait.
• Trained GMM-UBM systems by adapting universal background models (UBMs) to individual PD patients using their signal data.
• Constructed reference i-vector models from healthy control subjects matched by age and gender, then computed i-vector distances from PD patients.
• Computed distance metrics between patient models (GMM or i-vector) and reference models (UBM or control i-vector) as proxies for neurological deviation.
• Fused multimodal distances via concatenation and applied linear regression to predict MDS-UPDRS-III scores using leave-one-out cross-validation.
• Evaluated performance using Pearson and Spearman correlation coefficients, median absolute error (MAE), and p-values.

Fig. 1 : General methodology followed in this study.

Experimental results

Research questions

• RQ1 Can GMM-UBM and i-vector user models effectively represent PD-related motor impairments across speech, handwriting, and gait signals?
• RQ2 Which modality—speech, handwriting, or gait—provides the most informative features for predicting MDS-UPDRS-III scores?
• RQ3 How does feature fusion across modalities improve the prediction of neurological state compared to unimodal analysis?
• RQ4 Does GMM-UBM outperform i-vector modeling in assessing PD severity, and if so, under what conditions?
• RQ5 What is the relative contribution of each feature set (e.g., harmonic gait, prosody, kinematic handwriting) to the final multimodal prediction model?

Key findings

• GMM-UBM models achieved a Spearman correlation of 0.634 with the MDS-UPDRS-III scale when fusing harmonic gait, prosody, and articulation features, outperforming individual modality and i-vector systems.
• Harmonic gait features contributed most significantly to the multimodal model, followed by prosody and articulation features, indicating strong relevance to dysarthria in PD.
• The i-vector system showed lower performance, with a maximum Spearman correlation of 0.381 using non-linear gait features, suggesting limitations in modeling PD-specific traits with current reference data.
• Handwriting kinematic features contributed less than expected, likely due to mismatch between standard kinematic features and PD-specific motor impairments.
• The multimodal fusion improved the Spearman correlation by 2.4 percentage points compared to the best unimodal result (harmonic gait features).
• Random forest and SVR models overfitted the test set, predicting only the mean MDS-UPDRS-III value, indicating that linear regression with distance features was more robust for this task.

Fig. 2 : Prediction of the MDS-UPDRS-III score using multimodal user models based on GMM-UBM systems.