Steering Vectors for Parkinson's Disease Voice Detection

Figure 1: Steering vector approach for PD detection. Left: In the learned embedding space, PD and healthy control (HC) samples cluster separately. The steering vector (green arrow) captures the direction from HC to PD. Right: Classification pipeline projecting audio embeddings onto the steering vector.

Methods

Datasets

We combined multiple publicly available PD voice datasets for training and evaluation:

For disease specificity testing, we used the TORGO database [7]:

Model Architecture

We use wav2vec2 [8] (facebook/wav2vec2-base-960h) as a frozen feature extractor, followed by a lightweight adaptive classifier:

Audio → wav2vec2 → Mean Pooling → 2-Layer Encoder → Classification Head
                                       ↓
                              768-dim embeddings

Steering Vector Computation

The steering vector is computed as the difference between mean PD and healthy control embeddings:

steering_vector = mean(embeddings_PD) - mean(embeddings_HC)

Classification proceeds by projecting new samples onto this vector:

• Positive projection → PD-like
• Negative projection → HC-like

This provides a transparent, interpretable decision boundary compared to black-box classifiers.

Results

Classification Performance

Figure 2: Classification performance on the held-out test set (n=172). Left: Bar chart showing accuracy, F1 score, and AUC for both the full classifier and steering vector approach. Right: Confusion matrix showing true/false positives and negatives.

The steering vector approach achieves comparable performance to the full classifier, with the advantage of complete interpretability.

Steering Vector Statistics:

• PD mean projection: +13.9
• HC mean projection: -12.5
• Separation: t=12.5, p < 10⁻²⁵

Disease Specificity

A critical test: does our model detect PD specifically, or just dysarthria in general?

We evaluated on TORGO database [7] speakers with non-PD dysarthria (ALS and cerebral palsy):

Figure 3: Disease specificity analysis on TORGO database. Left: Mean P(PD) scores showing non-PD dysarthric speakers receive lower PD probability than healthy controls. Right: Per-speaker false positive rates, showing high variability across individuals. Green bars indicate good specificity (<40% FP), orange moderate (40-50%), red poor (>50%).

Key Finding: The model assigns lower PD probability to non-PD dysarthric speakers than to healthy controls (t=-2.5, p=0.012). This suggests the model is not simply detecting "abnormal speech" but capturing something more specific to PD.

Specificity varies considerably by speaker, with some dysarthric speakers (M04, M05) showing good specificity while others (M03) are frequently confused with PD.

Acoustic Interpretability

What acoustic features correlate with the steering vector direction?

Figure 4: Acoustic feature analysis. Left: Spearman correlations between acoustic features and steering vector projection. Higher-order MFCCs (5-12) show the strongest positive correlations, while zero-crossing rate shows negative correlation. Right: Cohen's d effect sizes for PD vs HC differences. Features with d > 0.8 (large effect) are clinically meaningful.

Top Correlations with Steering Vector Projection:

The strong correlation with higher-order MFCCs (7-9) is clinically meaningful. These coefficients capture subtle spectral shape variations associated with articulatory precision, vocal tract configuration, and phonatory stability [9].

PD vs HC Feature Differences:

Effect sizes above 0.8 are considered "large" in clinical research [10]. Our top features show Cohen's d values approaching or exceeding 1.0, indicating robust discrimination.

Feature Importance Convergence

Multiple analysis methods converge on the same key features:

Figure 5: Feature importance rankings across four analysis methods. Numbers indicate rank (1 = most important). The final column shows aggregate scores combining all methods. mfcc_7_mean consistently ranks as the most important feature across methods.

Aggregate Top Features (combining all methods):

1. mfcc_7_mean (score: 94.75)
2. onset_rate (score: 91.5)
3. mfcc_8_mean (score: 91.25)
4. spectral_contrast_6_mean (score: 90.5)
5. mfcc_6_mean (score: 90.0)

Clinical Feature Correlations

Using UCI telemonitoring data [1] (5,057 samples with UPDRS scores):

These correlations align with established PD voice research: increased aperiodicity (higher NHR, shimmer) and reduced complexity (DFA changes) are documented PD markers [11].

Discussion

Comparison to Related Work

Our approach draws inspiration from Goodfire AI's work on finding Alzheimer's disease directions in neural language model embeddings [4]. However, key differences exist:

Figure 6: Comparison between our approach and Goodfire AI's Alzheimer's detection work. Both use steering vectors in neural network representations, but differ in modality and disease target. Our work adds explicit disease specificity testing.

The Specificity Challenge

Our 53.5% correct rejection rate on non-PD dysarthria is meaningful but imperfect. This moderate specificity suggests:

1. Partial success: The model captures something PD-specific, not just general dysarthria
2. Shared features: Some acoustic manifestations of different motor speech disorders overlap
3. Room for improvement: Multi-task learning or explicit confound modeling could improve specificity

The fact that P(PD) is actually lower for non-PD dysarthric speakers than controls (0.468 vs 0.550) is encouraging. A naive dysarthria detector would show the opposite pattern.

Clinical Implications

The acoustic features our model relies on (MFCCs 5-12, onset rate, spectral contrast) correspond to:

• Articulatory precision: Fine spectral structure reflects precise tongue/lip positioning
• Speech timing: Onset rate captures the characteristic bradykinesia in PD speech [12]
• Phonatory quality: Spectral contrast relates to harmonic structure and vocal fold vibration

These are exactly the features speech-language pathologists assess when evaluating PD patients, providing face validity for our learned representations.

Limitations

1. Dataset size: 172 test samples limits statistical power for subgroup analyses
2. Cross-language generalization: Training data spans Italian, Spanish, and English; generalization to other languages is untested
3. Clinical validation: Prospective studies needed before screening deployment
4. Confounds: Age, sex, and recording conditions may influence results despite attempts at matching

Conclusion

We demonstrate that steering vectors in wav2vec2 embeddings can detect Parkinson's Disease with 91.5% AUC while maintaining moderate specificity against confounding dysarthrias. The learned direction correlates strongly with established acoustic biomarkers, providing interpretability absent from black-box approaches.

Figure 7: Summary of key results. Our approach achieves strong classification performance (91.5% AUC), comparable steering vector performance (89.4% AUC), moderate disease specificity (53.5% correct rejection of non-PD dysarthria), and clear interpretability through acoustic feature correlations.

This work suggests a promising path toward voice-based PD screening: neural audio embeddings capture disease-relevant information that can be extracted and interpreted. Future work should focus on improving disease specificity through multi-task learning and validating these findings in prospective clinical studies.

Reproducibility

All code and experiment details are available at: https://github.com/thin-signal/research

Key Results Summary:

• Classification AUC: 0.915
• Steering Vector AUC: 0.894
• Disease Specificity: 53.5% correct rejection of non-PD dysarthria
• Top Feature: mfcc_7_mean (r=0.52 with steering projection)

References

[1] Tsanas, A., Little, M.A., McSharry, P.E., & Ramig, L.O. (2010). Accurate telemonitoring of Parkinson's disease progression by noninvasive speech tests. IEEE Transactions on Biomedical Engineering, 57(4), 884-893.

[2] Postuma, R.B., & Berg, D. (2016). Advances in markers of prodromal Parkinson disease. Nature Reviews Neurology, 12(11), 622-634.

[3] Turner, A., Thiergart, L., Udell, D., Leech, G., Mini, U., & MacDiarmid, M. (2023). Steering language models with activation engineering. arXiv preprint arXiv:2308.10248.

[4] Wang, N., Fang, C., Bissell, M., Jain, A., Balsam, D., & Byun, M. (2024). Using interpretability to identify a novel class of Alzheimer's biomarkers. Goodfire Research. https://www.goodfire.ai/research/interpretability-for-alzheimers-detection

[5] Dimauro, G., Di Nicola, V., Bevilacqua, V., Caivano, D., & Girardi, F. (2017). Assessment of speech intelligibility in Parkinson's disease using a speech-to-text system. IEEE Access, 5, 22199-22208.

[6] Orozco-Arroyave, J.R., Arias-Londoño, J.D., Vargas-Bonilla, J.F., Gonzalez-Rátiva, M.C., & Nöth, E. (2014). New Spanish speech corpus database for the analysis of people suffering from Parkinson's disease. Proceedings of LREC, 342-347.

[7] Rudzicz, F., Namasivayam, A.K., & Wolff, T. (2012). The TORGO database of acoustic and articulatory speech from speakers with dysarthria. Language Resources and Evaluation, 46(4), 523-541.

[8] Baevski, A., Zhou, Y., Mohamed, A., & Auli, M. (2020). wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in Neural Information Processing Systems, 33, 12449-12460.

[9] Rusz, J., Cmejla, R., Ruzickova, H., & Ruzicka, E. (2011). Quantitative acoustic measurements for characterization of speech and voice disorders in early untreated Parkinson's disease. The Journal of the Acoustical Society of America, 129(1), 350-367.

[10] Cohen, J. (1988). Statistical Power Analysis for the Behavioral Sciences (2nd ed.). Lawrence Erlbaum Associates.

[11] Little, M.A., McSharry, P.E., Hunter, E.J., Spielman, J., & Ramig, L.O. (2009). Suitability of dysphonia measurements for telemonitoring of Parkinson's disease. IEEE Transactions on Biomedical Engineering, 56(4), 1015-1022.

[12] Skodda, S., Visser, W., & Schlegel, U. (2011). Vowel articulation in Parkinson's disease. Journal of Voice, 25(4), 467-472.