Pilot Study Employing a Machine Learning Approach as a Potential Method for Predicting Parkinson’s Disease Using Voice as a Digital Biomarker and the SHAP Approach for Feature Engineering

Introduction: Voice-based digital biomarkers have emerged as a promising approach for distinguishing individuals with neurodegenerative disorders, particularly Parkinson’s disease (PD), from healthy subjects (HS). With the increasing availability of smartphone and web-based recording tools, voice data can be collected efficiently in both clinical and remote settings. However, further validation is required before such approaches can be translated into routine clinical practice. Methods: This study used a cross-sectional analysis at the recording level, treating repeated recordings from the same participant as separate observations collected at Vito Fazzi Hospital in Lecce, Italy. Speech recordings from individuals with Parkinson’s disease (PD) and healthy controls were collected using the dedicated Talia smartphone and web application. Sustained vowel phonation (/a/) was analyzed as the primary speech task. Following data acquisition, feature extraction was performed as a crucial step in the speech analysis pipeline, as the quality and relevance of the extracted features directly influence the ability of machine learning models to discriminate between Parkinson’s disease (PD) patients and healthy controls. To capture various aspects of speech impairment associated with PD, a comprehensive set of acoustic features was extracted, including long-term features (pitch, jitter, and shimmer), nonlinear descriptors such as Recurrence Period Density Entropy (RPDE), and short-term feature based on Mel-Frequency Cepstral Coefficients (MFCCs). These features were subsequently used to develop and evaluate machine learning models for the classification of Parkinson’s disease and healthy subjects. Feature selection was performed using SHAP to identify the most informative vocal biomarkers. Model performance was assessed using five independent random train–test splits (70% training and 30% testing), supported by an internal five-fold cross-validation procedure within the training data. Multiple machine learning models were developed and evaluated, including Random Forest, Logistic Regression, Support Vector Machine, Naive Bayes, K-Nearest Neighbors, Decision Tree, Artificial Neural Network, and Gradient Boosting. Results: The evaluated models demonstrated strong recording-level classification performance. Artificial Neural Networks (ANN) and K-Nearest Neighbors (KNN) achieved the highest accuracy scores (0.9545 and 0.9494, respectively), along with superior recall (up to 0.9500), precision (up to 0.9551), and F1-score (up to 0.9525). Both models also exhibited excellent discriminative ability, with ROC-AUC values reaching 0.9882 (ANN) and 0.9893 (KNN). In contrast, Naive Bayes and Decision Tree showed comparatively lower performance across all metrics. Log-loss analysis further confirmed the robustness of ANN and KNN, which achieved the lowest values (0.2552 and 0.2510, respectively), indicating well-calibrated predictions. Overall, the findings highlight the consistency and generalizability of ANN and KNN across cross-validation splits. Conclusions: This study demonstrates that machine learning models, particularly ANN and KNN, can effectively differentiate Parkinson’s disease from healthy conditions using voice recordings. The integration of explainable AI for feature selection enhances model transparency and clinical relevance. However, the reported performance estimates were obtained from a recording-level analysis and should be interpreted as preliminary findings. Further studies involving larger cohorts and participant-level validation strategies are required to determine the generalizability and clinical applicability of these approaches.