Introduction
Parkinson’s Disease (PD) often manifests early through subtle changes in voice quality. While machine learning models are excellent at detecting these patterns, training them requires large amounts of sensitive medical data. The core challenge is the "privacy-utility trade-off": sharing medical records for research risks exposing patient identities, yet standard anonymization is often insufficient against sophisticated attacks.
This research evaluates how Differential Privacy (DP) affects the performance of machine learning models used to detect PD from vocal data. The authors demonstrate that by using specialized dimensionality reduction techniques, it is possible to maintain high diagnostic accuracy while providing mathematical guarantees for patient data privacy.
Methodology
The study utilized a publicly available dataset from Kaggle (Sakar et al., 2019) containing 756 vocal attributes from 252 individuals (188 with PD, 64 healthy controls). The researchers compared a suite of standard machine learning models—including Logistic Regression (RL), Random Forest (RF), Support Vector Machines (SVM), and Gaussian Processes (PG)—across two scenarios:
- Non-Private Approach (ANP): Standard Principal Component Analysis (PCA) was used to reduce the 756 vocal features down to the 100 most significant ones, retaining approximately 88% of the variance.
- Private Approach (AP): The Mod_SULQ algorithm was implemented. This method achieves dimensionality reduction while satisfying Differential Privacy by adding Gaussian noise to a symmetric matrix generated from the data before extracting eigenvalues. This ensures the reduced dataset does not leak specific patient details.
All models were optimized using grid search to find the best hyperparameters for both the private and non-private versions of the data.
Key Results and Insights
The evaluation focused on several metrics, including Accuracy (A), Precision (Pr), Recall (R), F1-Score, and Area Under the Curve (AUC).
The results indicate that the "privacy tax"—the drop in performance when adding noise—is remarkably low. In some cases, the private models performed nearly as well as, or even slightly better than, their non-private counterparts.
- Non-Private Winner: The Gaussian Process (PG) model achieved the best results with an Accuracy of 0.81 and an AUC of 0.83.
- Private Winner: The Random Forest (RF) model took the lead in the private scenario, maintaining an Accuracy of 0.81 and an AUC of 0.81.
- Model Robustness: While PG led in the non-private setting, RF proved more robust when dealing with the noise introduced by the Differential Privacy mechanism.
- Non-Linearity: Logistic Regression consistently performed the worst, suggesting that the relationship between vocal features and a Parkinson's diagnosis is complex and non-linear.
As shown in the confusion matrix, most models maintained consistent performance levels. The Random Forest model in the private approach (AP) achieved 137 True Positives (VP) and 17 True Negatives (VN), demonstrating its effectiveness even under privacy constraints.
Conclusion and Future Work
The study proves the viability of the Mod_SULQ approach for medical diagnostics, showing that privacy does not necessarily come at the cost of utility. However, the authors acknowledge limitations, such as the small dataset size (252 patients). Future work should involve validation on larger cohorts and the use of "Leave-One-Subject-Out" cross-validation to better simulate real-world diagnostic scenarios. Additionally, exploring the impact of varying privacy budgets ($\epsilon$) and the application of deep learning techniques could further refine these privacy-preserving diagnostic tools.
Reference: Filho, M. E. B., et al. (2020). Detectando Doença de Parkinson - Uma Comparação de Modelos de Aprendizagem de Máquina com Redução de Dimensionalidade Diferencialmente Privada. 35º Simpósio Brasileiro de Bancos de Dados (SBBD).
Summary updated