On Estimating Age and Gender from Parkinson's Disease Diagnostic-Oriented Recordings Using Wav2Vec 2.0

This study demonstrates that a pretrained Wav2Vec 2.0 model can robustly estimate gender and preserve age-related patterns in pathological speech across multilingual datasets, achieving high accuracy for gender and significant age correlations in connected speech while revealing task-dependent limitations for sustained vowel phonation.

The Gist

Imagine you have a super-smart robot that has listened to millions of hours of radio, podcasts, and conversations from around the world. This robot, called Wav2Vec 2.0, is like a musical prodigy who has never been taught to speak a specific language but can instantly recognize the shape of a voice, the rhythm of speech, and the unique "fingerprint" of a speaker.

Now, imagine this robot is handed a collection of voice recordings from people with Parkinson's disease. These people might have shaky voices, speak slowly, or struggle to pronounce words. The big question the researchers asked was: "Can this robot, which was trained on 'normal' healthy voices, still figure out how old these people are and whether they are men or women, even when their voices are affected by disease?"

Here is the story of what they found, explained simply:

1. The Robot's Superpower: Gender

Think of a person's gender (male or female) in their voice like the color of a shirt. Even if the person is wearing a dirty, torn, or wet shirt (representing the disease), the color is still usually obvious.

• The Result: The robot was incredibly good at this. It guessed the gender correctly 94% to 100% of the time, no matter what the person was saying or how sick they were.
• The Analogy: It's like looking at a blurry, black-and-white photo of a person and still being able to tell if they are wearing a red shirt or a blue shirt. The "gender signal" is so strong in the voice that the disease couldn't hide it.

2. The Robot's Struggle: Age

Now, imagine trying to guess someone's age. This is like trying to guess how many rings are on an old tree just by looking at a single, tiny leaf.

• The Good News: When the person was reading a story or speaking in sentences (connected speech), the robot did a decent job. It could tell the difference between a 40-year-old and a 70-year-old reasonably well. It was like looking at a whole branch of the tree; you could see the texture and guess the age.
• The Bad News: When the person was just holding a single note (saying "Ahhh" for a few seconds), the robot got completely confused. It guessed that 60-year-olds were actually 30-year-olds!
• The Analogy: Asking the robot to guess age from a single vowel sound is like asking a chef to guess the age of a cow just by tasting a single drop of milk. There just isn't enough information in that tiny drop to tell you how old the animal is. The robot "hallucinated" that everyone was much younger than they actually were.

3. Why This Matters (The "Why Should I Care?" Part)

You might wonder, "Why do we need a robot to guess age and gender if we already have medical records?"

• The "Missing Label" Problem: Imagine you find a box of old voice recordings from the internet. There are no names, no ages, and no genders attached. If you want to study these voices to help Parkinson's patients, you need to know: "Are these voices mostly from old men? Or young women?" If you don't know, your study might be biased.
• The Robot as a Detective: This study shows that this "off-the-shelf" robot can act as a detective. It can look at a messy, unlabeled pile of voice data and say, "Hey, 80% of these speakers are men, and they are mostly over 60." This helps scientists clean up their data and make sure their studies are fair.
• The "Quality Control" Check: Sometimes, medical records have mistakes. Maybe a 20-year-old's voice was accidentally labeled as a 70-year-old. The robot can spot these errors. If the robot hears a voice and says, "That sounds 20," but the file says "70," you know to double-check the file.

4. The Big Takeaway

The researchers found that self-supervised models (robots that learn by listening to the world) are like Swiss Army Knives.

• They are amazing at gender (the main blade is very sharp).
• They are okay at age when people are talking normally (the screwdriver works fine).
• They are useless at age when people are just humming a single note (the bottle opener is broken).

In short: We don't need to build a new, expensive robot from scratch to guess gender and age from Parkinson's voices. We can use a powerful, pre-trained robot that already exists. It works great for gender and is helpful for age, as long as we ask it to listen to full sentences rather than single sounds. This saves time, money, and helps scientists understand their data better without needing to know every detail about the patients beforehand.

Technical Summary

1. Problem Statement

The rapid adoption of Self-Supervised Foundation Models (SFMs) like Wav2Vec 2.0 in biomedical speech research has raised a critical question: To what extent are demographic attributes (age and gender) encoded in pathological speech?

• The Gap: While SFMs excel at disease detection (e.g., Parkinson's Disease), they are often trained on healthy speech corpora. It is unclear if these models retain the ability to estimate demographics when applied to disordered speech (e.g., dysarthria) without task-specific fine-tuning.
• The Risk: In clinical datasets lacking demographic metadata, models might inadvertently learn to classify disease based on age or gender correlations (shortcut learning) rather than true pathological markers. Conversely, if metadata is missing, researchers lack tools to characterize their datasets or validate that their models are not biased by demographic confounds.
• The Objective: To evaluate whether a single, pre-trained SFM can robustly estimate age and gender across heterogeneous Parkinson's Disease (PD) and related syndromes datasets, using various diagnostic speech tasks, without any exposure to the target data during training.

2. Methodology

Datasets

The study utilized three independent, multilingual datasets comprising 244 unique subjects (Healthy Controls [HC], PD patients, and related parkinsonian syndromes):

1. PC-GITA (Spanish): 50 HC and 50 PD subjects. Tasks: Read text, diadochokinesis ("pataka"), and sustained vowel /a/.
2. Italian Dataset: 15 Young HC, 18 Elderly HC, and 28 PD subjects. Task: Read text.
3. PD/Parkinsonism Dataset (Synthesized): 22 HC, 22 PD, 21 Multiple System Atrophy (MSA), and 18 Progressive Supranuclear Palsy (PSP). Task: Synthesized sustained vowel /A/.

Models and Approaches

The study compared two primary approaches:

1. Primary Approach (Zero-Shot Inference):
   • Model: A 24-layer Wav2Vec 2.0 model specifically fine-tuned for age and gender recognition (trained on general corpora like Common Voice, VoxCeleb2, etc., but not on neurological disorder data).
   • Mechanism: Direct inference. The model takes raw audio and outputs continuous age predictions (normalized 0–100) and gender probabilities (Child/Female/Male). No downstream training was performed on the target datasets.
2. Baseline Approach (Feature Extraction):
   • Model: Wav2Vec 2.0 XLSR-53 used strictly as a fixed feature extractor.
   • Mechanism: Features were extracted from multiple Transformer layers (1, 4, 8, 24), aggregated via mean pooling, and fed into downstream Random Forest classifiers (for gender) and regressors (for age).
   • Validation: Evaluated using Leave-One-Subject-Out Cross-Validation (LOSO-CV).

Evaluation Metrics

• Gender: Classification accuracy.
• Age: Spearman's rank correlation ($\rho$) between predicted and true age, Chi-square goodness-of-fit tests for distribution alignment, and Median Absolute Deviation (MAD).
• Visualization: t-SNE for embedding separability and Bland-Altman plots for bias analysis.

3. Key Results

Gender Estimation

• Performance: Exceptionally robust across all datasets, tasks, and diagnostic groups.
• Accuracy: Ranged from 94% to 100%.
  • PC-GITA: 100% (HC, Read Text), 98% (PD, Read Text).
  • Italian Dataset: 100% across all groups.
• Comparison: The primary zero-shot approach outperformed the feature-extraction baseline by 8–12% in accuracy.
• Insight: Gender information is deeply and robustly encoded in the SFM's representations, even in pathological speech.

Age Estimation

• Performance: Highly dependent on the speech task.
  • Connected Speech (Read Text): Showed statistically significant correlations with true age.
    • PC-GITA (HC): $\rho = 0.52$ ($p < 0.001$).
    • PC-GITA (PD): $\rho = 0.44$ ($p = 0.001$).
    • Italian (EHC): $\rho = 0.49$ ($p = 0.04$).
  • Sustained Vowels (/a/): Failed completely.
    • Correlations were non-significant (e.g., $\rho = 0.05$ for HC in PC-GITA).
    • Systematic Bias: The model severely underestimated age (predicting ~30–36 years for subjects who were actually 60–67 years old) across all diagnostic groups (HC, PD, MSA, PSP).
• Baseline Comparison: The feature-extraction baseline performed significantly worse than the primary approach, particularly failing to capture age structure in read-text tasks.
• t-SNE Visualization: Embeddings showed clear gender separability but no clear age-related clustering, confirming that age is weakly encoded in the raw feature space without task-specific adaptation.

LLM Exploratory Analysis

A small-scale experiment using a multimodal LLM (GPT-5.2) to "listen" to sustained vowels showed slightly better age estimates than the SFM for some samples, suggesting future potential for hybrid approaches, though SFMs remain preferred for reproducibility.

4. Key Contributions

1. Comprehensive Evaluation: Demonstrated that a single, off-the-shelf SFM can estimate gender with near-perfect accuracy and age with moderate accuracy in pathological speech without retraining.
2. Task-Dependent Limitations: Identified a critical limitation: SFMs fail to estimate age from isolated sustained vowels, likely due to the loss of prosodic and contextual cues required for age inference.
3. Baseline Benchmarking: Established that a standard feature-extraction pipeline (XLSR + Random Forest) is inferior to a model specifically pre-trained for demographic tasks, even when the latter has not seen pathological data.
4. Bias Detection: Showed that SFMs can serve as a "quality control" tool to detect metadata inconsistencies or demographic imbalances in clinical datasets.

5. Significance and Implications

• Clinical Utility:
  • Metadata Recovery: Researchers can use these models to automatically annotate large, unlabelled clinical speech datasets with demographic metadata.
  • Bias Mitigation: By estimating age and gender, researchers can verify that disease detection models are not exploiting demographic shortcuts (e.g., classifying "older" as "diseased").
  • Quality Control: Discrepancies between predicted and recorded demographics can flag mislabeled data in pipelines.
• Theoretical Insight:
  • Generalization: SFMs trained on healthy speech generalize well to pathological speech for gender, suggesting these acoustic features are robust to neurological degradation.
  • Task Specificity: The failure on sustained vowels highlights that while SFMs capture global structure, they may lose specific discriminative cues when speech is constrained to isolated phonation.
• Future Directions:
  • The study suggests that while zero-shot inference is useful for characterization, task-specific fine-tuning on pathological data is likely necessary for high-precision age estimation in clinical settings.
  • There is a need for more diverse pretraining data (specifically older adults) to improve age estimation performance.

Conclusion: The paper validates Wav2Vec 2.0 as a powerful tool for demographic metadata extraction in Parkinson's research, particularly for gender and connected speech age estimation, while cautioning against its use for age estimation in isolated vowel tasks without further adaptation.