IFI-ML: a novel framework combining Fuzzy Inference and Machine Learning for Parkinson's disease classification

Early detection of Parkinson's Disease (PD) remains challenging due to vocal signal complexity and severe data imbalance, which necessitates that machine learning (ML) models deliver higher classification accuracy. This study aims to break the systematic, inverse trade-off between sensitivity and specificity in voice-based PD classification. We present the Intelligent Fuzzy Inference and Machine Learning (IFI-ML) framework, an uncertainty-aware architecture utilizing entropy-based partitioning, hybrid classifiers, and a Sugeno Type-2 Fuzzy Inference System (FIS). While traditional frameworks apply global resampling techniques such as the Synthetic Minority Over-sampling Technique (SMOTE) or Random Over-Sampling (ROS) across entire datasets, the proposed IFI-ML framework utilizes information gain-driven clustering to segment samples into sensitive (uncertain boundary) and non-sensitive partitions. The framework optimizes conventional ML classifiers on the stable, non-sensitive clusters, while routing critical boundary samples to the Type-2 FIS to handle local class overlap without distorting the overall data distribution. We evaluated our framework on two public, imbalanced datasets from the University of California, Irvine (UCI) Machine Learning repository: Parkinson's voice Dataset 1 (D ₁) and Dataset 2 (D ₂). The proposed framework achieves an accuracy of 97% on D ₁ and 92.17% on D ₂, while simultaneously raising specificity to 94-95% on D ₁ and 75-88% on D ₂. The Matthews Correlation Coefficient (MCC) improves to 0.93 and 0.77, respectively. Furthermore, IFI-ML reduces false-positive and false-negative rates to 3-7% on D ₁ across all classifiers, and limits false-positive rates to 11-12% for the K-Nearest Neighbors (KNN) model on D ₂. Under a rigorous 10-fold cross-validation (10 CV) scheme, the framework consistently outperforms 15 baseline machine learning models. These benchmark results show the clinical reliability and robustness of the proposed approach, establishing a state-of-the-art benchmark for non-invasive vocal biomarkers.