Explainable Artificial Intelligence to Diagnose Early Parkinson′s Disease via Voice Analysis

Background: Parkinson′s disease (PD) is a progressive neurodegenerative disorder that affects motor control, leading to symptoms such as tremors or impaired balance. Early diagnosis of PD is crucial for effective treatment, yet traditional diagnostic models are often costly and lengthy. This study explores the use of Artificial Intelligence (AI) and Machine Learning (ML) techniques, particularly voice analysis, to identify early signs of PD and make a precise diagnosis. Objectives: This paper aims to create an automatic detection and prediction of PD binary classification using vocal biomarkers. We will also use explainability to identify latent and important patterns in the input data in retrospect to the target to inform the definition of Parkinson′s through voice characteristics. Finally, a probability generation will be generated to create a scoring system of a patient′s odds of PD as a spectrum. Methods: We utilized a dataset comprising 81 voice recordings from both healthy control (HC) and PD patients, applying a hybrid AI model combining Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Multiple Kernel Learning (MKL), and Multilayer Perceptron (MLP). The model′s architecture was designed to extract and analyze acoustic features such as Mel-Frequency Cepstral Coefficients (MFCCs), local jitter, and local shimmer, which are all indicative of PD-related voice impairments. Once features are extracted, the AI model will generate prediction labels for HC or PD files. Then, a scoring system will assign a number ranging from 0-1 to each file, indicating the stage of PD development. Results: Our champion model yielded the following results: diagnostic accuracy of 91.11%, recall of 92.50%, precision of 89.84%, an F1 score of 0.9113, and an area under curve (AUC) of 0.9125. Furthermore, the use of SHapley Additive exPlanations (SHAP) provided detailed insight into the model′s decision-making process, highlighting the most influential features contributing to a PD diagnosis. The outcomes of the implemented scoring system demonstrate a distinct separation in the probability assessments for PD across the 81 analyzed audio samples, validating our scoring system by confirming that the vocal biomarkers in the audio files accurately correspond with their assigned scores. Conclusion: This study highlights the efficacy of AI, particularly a hybrid model combining CNN, RNN, MKL, and Deep Learning in diagnosing early PD through voice analysis. The model demonstrated a robust ability to distinguish between HC and PD patients with significant accuracy by leveraging key vocal biomarkers such as MFCCs, jitter, and shimmer.