Paper #3

Explainable AI-Assisted Parkinson's Disease Diagnosis Using Machine Learning and Deep Neural Networks

Ferdaus Ibne Aziz, Daniel Ojeda Rosales, Becky Firomssa Gudeta and Jia Uddin

Abstract: Timely and specific interventions can substantially help in managing the disease, provided that the PD is diagnosed at an early stage. This paper compares machine learning (ML) and deep learning (DL) methods of PD detection with the help of vocal characteristics of a canonical sample (197 samples with 22 voice attributes pre-extracted). To reduce the issue of class imbalance, the Synthetic Minority Over-Sampling Technique (SMOTE) was used on the training data, which enhanced the strength of the models. The classical Machine Learning (ML) classifiers, such as Logistic Regression, Support Vector Machine, Random Forest, Extra Trees, Decision Tree, AdaBoost, and K-Nearest Neighbors (KNN) were evaluated, and KNN produced the best accuracy of 85% as well as competitive accuracy, recall, F1 score, and AUC ROC. In the case of deep learning, 1D CNN, 2D CNN, and LSTM were used, and 2D CNN and LSTM performed better than 1D CNN, with test accuracy of 89.7% and 84.6%, respectively, indicating their capability to learn both time-based and spatial patterns in data. Interpretability was added through Local Interpretable Model Agnostic Explanations (LIME) to ML models that point to Spread2, Recurrence Period Density Entropy, and MDVP-related frequency measures as significant vocal biomarkers. Although the framework has constraints relating to data volume and single modality, it offers a reproducible, interpretable baseline of PD detection and highlights the possibilities of explainable AI and neural networks as an assistant clinical decision-making tool. In the future, larger, multi-modal datasets are needed to offer better generalizability.

Keywords: Parkinson's Disease Diagnosis; Tabular Data; Machine Learning; Deep Neural Network.