A layer-wise analysis of nine SSL speech backbones using a low-capacity logistic regression probe reveals that the transferred discriminative signal lacks pathological specificity, highlighting critical limitations that must be addressed before speech-based pathology recognition models can be reliably deployed in clinical settings.
Abstract
Self-supervised learning (SSL) speech representations achieve strong performance for Parkinson's disease (PD) detection within individual corpora. However, it remains unclear whether these models capture disease-related characteristics or exploit dataset-specific confounds, particularly since most SSL backbones are pretrained exclusively on healthy speech. To investigate this question, we perform a layer-wise analysis of nine SSL speech backbones using a low-capacity logistic regression probe across three languages. We structure the evaluation as multiple scenarios that progressively introduce distribution shifts in participant identity, recording conditions, language, and pathology. Our results reveal two key findings. First, layer selection is highly corpus-dependent: the optimal representation layer is determined primarily by the source dataset rather than by the SSL architecture itself. Second, the transferred discriminative signal lacks pathological specificity: classifiers trained to detect PD assign similarly high probabilities to both PD and dementia speech in the target corpus. These results highlight critical limitations that must be addressed before speech-based pathology recognition models can be reliably deployed in clinical settings.
Self-supervised speech models (S3Ms) provide powerful representations for Parkinson's disease (PD) detection, making cross-lingual transfer attractive for languages lacking labeled patient speech. However, these representations also encode language identity, which can confound this transfer: without target-language PD speech, classifiers may separate languages rather than pathology, yielding high specificity but low sensitivity on target patients. We propose \emph{language orthogonalization}, a closed-form ridge residualization of S3M features against external VoxLingua107 language embeddings, fitted using only healthy-control (HC) speech. By removing language-predictable components while retaining pathology-related variation, it produces a less language-dependent geometry in which HC representations concentrate while PD representations disperse. Across five S3M backbones, three speech tasks, and three target languages, our method consistently improves cross-lingual PD-detection performance while correcting the high-specificity/low-sensitivity failure.
Minu Kim, Eunjung Yeo, Kwanghee Choi et al.· 0 citations
Parkinson’s disease (PD) is a progressive neurodegenerative disorder that severely impairs motor control and quality of life. Conventional diagnostic approaches such as clinical motor assessment and neuroimaging are expensive, invasive, and lack sensitivity for early-stage detection. However, subtle alterations in vocal characteristics, including tremors, pitch instability, and hoarseness, can manifest years before motor symptoms become evident, offering a promising non-invasive biomarker. To address these challenges, we propose a novel ensemble-driven machine learning framework that leverages acoustic features from speech signals for early PD detection. Using an open-access voice dataset, 22 clinically validated vocal biomarkers were extracted. Six supervised models were trained and evaluated, with a custom Stacking Classifier achieving superior performance with an accuracy of 97.77%, precision of 98.39%, recall of 95.31%, and an F1-score of 96.83%. This ensemble method integrates the strengths of diverse learners to ensure stable and generalizable predictions. Furthermore, the system incorporates a post-diagnostic chatbot powered by a Large Language Model (LLM), enabling personalised support and guidance. The proposed framework is non-invasive, scalable, and well-suited for telemedicine, representing a significant step toward accessible early-stage PD screening.
Acoustic biomarkers show promise for detecting Alzheimer's Disease (AD), yet whether the cues driving diagnostic AI align with those salient to human listeners is underexplored across languages and genders, where pathological markers and perceptual strategies differ. We train models to predict clinical AD status (pathology) and human perceptual scores across Mandarin and Greek, male and female speakers. Using SHAP for interpretability and statistical models for validation, we compare feature importance by subgroup. Results reveal a context-dependent divergence: pathological-perceptual alignment is significant for Mandarin and female speakers but disappears for Greek and male speakers, where pathology models did not exceed chance; this is a failure mode that population-specific auditing surfaces. Global Explainable AI (XAI) explanations can mask critical demographic divergences, highlighting the need for population-specific explainability auditing for equitable deployment of clinical speech AI.
Liu He, Yuanchao Li, Yinlong Liu et al.· arXiv.org· 0 citations
Parkinsons Disease (PD) is a progressive neurodegenerative disorder which significantly affects motor function, daily coordination and verbal communication. Speech-based biomarkers provide a non-invasive and scalable approach to early detection, as dysphonia is one of the earliest and most consistent clinical markers of PD. The dataset used in this study is publicly available and consists of 756 voice recordings from 252 subjects (188 with PD and 64 neurologically healthy controls) with a wide range of acoustic parameters such as Mel-Frequency Cepstral Coefficients (MFCCs), energy-based parameters, and higher-order statistical derivatives. After systematic preprocessing and z-score normalisation, five machine learning classifiers were tested: K-Nearest Neighbors (KNN), Extreme Gradient Boosting (XGBoost), Random Forest (RF), Support Vector Machine (SVM), and Naive Bayes (NB) under a subject-independent, GroupKFold cross-validation protocol. The KNN classifier performed best overall with an accuracy of 92.10%, F1 score of 94.50%, and a precision rate of 98.09%, reducing the number of false positive diagnoses. To overcome the lack of interpretability of black-box predictive models, SHapley Additive exPlanations (SHAP) were used to explain the contribution of each feature to the prediction of an individual. The most diagnostically salient acoustic biomarkers were identified as features from the SHAP analysis: std delta delta log energy, the first Mel-Frequency Cepstral Coefficient, and Tunable Q-Factor Wavelet Transform (TQWT). This work introduces a machine learning framework that is both reproducible and clinically interpretable, combining high predictive accuracy with transparent, physiologically grounded decision logic.
D. Howlader, T. Ahmed, M. M. Rahman· medRxiv· 0 citations
Parkinson’s disease (PD) diagnosis relies on subjective clinical examination of motor signs that can be mild, intermittent, or absent early in the disease course, motivating objective, low-cost, non-invasive markers for earlier, more consistent detection. Voice recordings, acquirable with nothing more than a microphone, are a strong candidate, and this study develops a machine learning pipeline for voice-based PD screening built on the Competitive Swarm Optimizer (CSO), which jointly searches the acoustic feature subset, classifier configuration, and binarization transfer function, instead of optimizing the feature subset alone as most prior pipelines do. Evaluated on two public, subject-grouped voice datasets, Oxford and Naranjo, against five baselines under an identical protocol across 20 runs per method, our proposed pipeline attains the highest mean balanced accuracy on Naranjo with 0.847 and the second-highest on Oxford with 0.810, accuracies consistent with the wider voice-based PD screening literature; because this evidence comes from two small, single-recording-protocol, retrospective public datasets of 31 and 80 subjects each, we present it as an initial, encouraging step toward a first-pass triage or between-visit monitoring tool, pending external validation on a prospectively collected, multi-site cohort, not as a standalone diagnostic instrument. As an initial step toward clinical interpretability, we check which acoustic features are selected most consistently across 20 repeated runs of the proposed pipeline against established physiological correlates of Parkinsonian dysphonia; agreement between any two runs’ complete feature subsets is weak, but pitch period entropy, a nonlinear-dynamical measure of aperiodic pitch period variability, is selected far more often than chance on both datasets, consistent with the underlying pathophysiology and not merely predictive. These results support voice-based, metaheuristic-optimized screening as a plausible, interpretable, low-burden tool for telemedicine and home monitoring.
Speech impairment affects the large majority of Parkinson’s disease (PD) patients and can appear before motor signs become clinically evident. Two decades of computational research have produced a heterogeneous literature in which reported accuracy varies enormously across studies. This review argues that the spread is governed less by algorithmic capability than by three evaluation choices: subject-level data leakage, leave-one-out cross-validation on very small datasets, and the near-absence of genuinely independent validation. Building on the Ngo et al. 2022 systematic review of the 2010–2021 literature as a verified foundation, this work extends coverage through 2025 with emphasis on self-supervised foundation models and federated learning, both absent from prior surveys. A verified benchmark spanning the validation-rigour spectrum illustrates the effect directly: the foundational sound-booth study of 31 participants reports 91.4% accuracy, whereas the largest telephone-quality study, using a considerably richer feature set and a larger cohort, reports 66.4% balanced accuracy—a 25-point gap attributable to acquisition conditions rather than to method. A structured comparison between handcrafted acoustic features and self-supervised speech embeddings is provided, alongside a three-tier gap analysis linking ten research deficits to prioritised future directions, and a privacy and ethics analysis not present in prior PD voice biomarker surveys. Fine-tuned transformer models achieve the strongest independently validated result to date. The field’s central unsolved problem remains generalisation, not algorithmic sophistication.
