Enhancing Noise Robustness of Parkinson's Disease Telemonitoring via Contrastive Feature Augmentation

TL;DR

This paper tackles robustness in UPDRS prediction from at-home PD speech data by addressing noise from patient, environment, and transmission channels. It introduces NoRo, which bins a key feature, uses contrastive learning to produce a noise-robust hidden representation, and concatenates it with the original features to form augmented inputs for regression. Across a real PD telemonitoring dataset and multiple downstream models, NoRo delivers consistent improvements under noisy conditions, reducing RMSE by up to $40\%$ and MAE/MedianAE by more than $10\%$ in several cases, especially for non-ensemble methods. The work also provides an evaluative noise-injection framework, analyzes hyperparameter sensitivity, and demonstrates that the augmented feature space preserves discriminative structure under noise, supporting practical deployment in PD telemonitoring.

Abstract

Parkinson's disease (PD) is one of the most common neurodegenerative disorder. PD telemonitoring emerges as a novel assessment modality enabling self-administered at-home tests of Unified Parkinson's Disease Rating Scale (UPDRS) scores, enhancing accessibility for PD patients. However, three types of noise would occur during measurements: (1) patient-induced measurement inaccuracies, (2) environmental noise, and (3) data packet loss during transmission, resulting in higher prediction errors. To address these challenges, NoRo, a noise-robust UPDRS prediction framework is proposed. First, the original speech features are grouped into ordered bins, based on the continuous values of a selected feature, to construct contrastive pairs. Second, the contrastive pairs are employed to train a multilayer perceptron encoder for generating noise-robust features. Finally, these features are concatenated with the original features as the augmented features, which are then fed into the UPDRS prediction models. Notably, we further introduces a novel evaluation approach with customizable noise injection module, and extensive experiments show that NoRo can successfully enhance the noise robustness of UPDRS prediction across various downstream prediction models under different noisy environments.

Figures (6)

• Figure 1: NoRo, a framework of the noise-robust UPDRS prediction process. NoRo enhances prediction robustness through a self-supervised Contrastive Learning (CL) approach that generates noise-robust augmented features. First, a Random Forest algorithm selects the feature dimension $\bm{x}_b$ with the highest importance score across the original speech feature $X$, maximizing correlation with UPDRS scores. Second, $\bm{x}_b$ undergoes equal-width binning to group $X$ into $K$ bins. Third, following the idea of CL, same-bin features are treated as positive pairs and cross-bin features as negative pairs to train a Multilayer Perceptron (MLP) encoder $W$ to project $X$ as hidden states $H=\sigma(WX)$. Then, $H$ is concatenated with $X$ as the augmented feature $X'=\left[X,H\right]$. Finally, augmented features are then fed into downstream prediction models to achieve robust UPDRS prediction.
• Figure 2: Distance Coefficient $\alpha$ for the $m$-th Bin. The curve represents the value of $\alpha$: (1) Same-bin $\alpha$ equals to $1$, which is the highest, e.g., $\alpha_{mm}=1$. (2) $\alpha$ decreases as the distance between the $m$-th bin and other bins increases, e.g., $0<\alpha_{m,m-2}<\alpha_{m,m-1}<\alpha_{m,m}$. (3) $\alpha$ is symmetric around the central bin, e.g., $\alpha_{m,m-1}=\alpha_{m-1,m}$ and $\alpha_{m,m-2}=\alpha_{m-2,m}$.
• Figure 3: Qualitative Analysis Results. The relative errors between the prediction errors (RMSE, MAE, MedianAE) of NoRo and baseline ($\delta=({E}_{x'} - {E}_x)/{{E}_x}$) under different SNR levels of extra noise environments are presented. Relative error $\delta<0$ demonstrates the prediction error using NoRo is better (lower) than baseline where NoRo enhances the robustness of the downstream model under the noisy environments with extra noise at certain SNR levels.
• Figure 4: Results on different bin numbers $K$ settings. The relative errors of RMSE, MAE, and MedianAE between baseline and NoRo of different downstream models with extra noise at SNR=10dB are presented. Here, baseline is obtained from different MLP encoders of different bin numbers $K$ in each plot. Relative error < 0 demonstrates the prediction error using NoRo is better (lower) than baseline, then NoRo with the certain hyperparameter $K$ enhances the robustness of downstream models.
• Figure 5: T-SNE visualization of feature spaces. Points with the same color are the samples of the same bin. One color represents one certain bin. The relative positions between the points indicates the relative positions between them in original feature space or augmented feature space. Corresponding metrics are reported in each subplot, where higher Silhouette score and higher Calinski-Harabasz index indicate better results.