Spectral super-resolution for Parkinson's voice via representation-level methods under mixed-reality acquisition

Area of Science:

• Biomedical signal processing
• Machine learning for healthcare
• Neurological disorder diagnostics

Background:

• Voice analysis is a practical remote biomarker for Parkinson's disease (PD).
• Real-world voice data often has low-resolution features, hindering accurate PD diagnosis.
• Existing methods may require waveform resynthesis for feature enhancement.

Purpose of the Study:

• To evaluate if feature-level spectrogram super-resolution (SR) improves PD vs. healthy control (HC) discrimination.
• To compare SR methods performed inside the model, avoiding waveform resynthesis.
• To assess performance under realistic, low-resolution audio capture constraints.

Main Methods:

• Speech data from 161 participants (75 PD, 86 HC) were recorded using a mixed-reality (MR) protocol.
• Log-mel spectrograms were extracted and fed into ImageNet-pretrained models (ConvNeXt-Tiny, ResNet-50, EfficientNetV2-S).
• Six SR strategies, including a frozen universal feature SR module (AnyUp), were compared using 5-fold cross-validation.

Main Results:

• The AnyUp module consistently outperformed other SR methods across various tasks and backbones.
• Significant improvements in PD vs. HC classification were observed, particularly for sustained vowels and DDK tasks.
• Macro-averaged gains with AnyUp reached +0.045 AUC/+0.045 ACC for ConvNeXt-Tiny and +0.052 AUC/+0.048 ACC for EfficientNetV2-S.

Conclusions:

• Feature-level SR, particularly using the AnyUp module, offers a compute-efficient way to enhance PD voice classification.
• This approach provides consistent improvements without requiring waveform synthesis, making it suitable for low-resolution clinical audio.
• Representation-level SR is a pragmatic alternative or complement to bandwidth extension for PD detection.