Abstract

Parkinson’s disease (PD) is the second most common neurodegenerative pathology in the world. It has been shown that 70% to 75% of PD patients would suffer from speech disorders at some stage. PD affects men more frequently than women, which has resulted in previous works using unbalanced datasets that potentially introduce sex-related biases in the models. In this paper, we investigate the sex differences in the detection of PD from speech using various models. We extract features using WavLM, Wav2vec2.0, Whisper, and FBanks and feed them into ECAPA-TDNN, deep neural network architecture, and then a binary classification layer. We use ComParE2016 features with Random Forest as well.
We also conducted sex-specific experiments to assess generalization across sexes. In our main experiments, we use a large subset of the mPower open dataset, where sex and disease status (PD vs Healthy Controls (HC)) are balanced. In subsequent experiments evaluating dataset size and sex-specific trainings, we used 3 additional datasets, mPower matched, mPower small, and PC-GITA, where age was also matched between PD and HC groups.
All the datasets include sustained phonation task. We observe that on the large dataset, female speakers are more easily detected than males. However, on smaller datasets, none of the classification results were found to exceed chance performance. Sex-specific results
show that when the models are trained on one sex, they fail to generalize to the other one especially when using traditional Random Forest pipeline. Our results highlight the importance of including sex as a variable in the development of fair PD detection systems.