The study was published in Frontiers in Artificial Intelligence on 21 October 2026 by S. Vinothini and G. Magesh of the School of Computer Science Engineering and Information Systems at Vellore Institute of Technology in India. It focused on classifying Parkinson’s disease from voice recordings drawn from the University of California, Irvine Machine Learning Repository. The first dataset, designated D1, contains 195 samples and 23 acoustic features, including 48 healthy controls and 147 participants with Parkinson’s disease. The second, D2, contains 756 samples and 752 features, with 192 controls and 564 participants. The source reported no missing values in either dataset, and both were imbalanced toward the Parkinson’s class.

The central problem addressed by IFI-ML is the tension between sensitivity, or the ability to identify participants with Parkinson’s disease, and specificity, or the ability to identify controls. In the baseline 10-fold cross-validation experiments, a support vector machine detected 97.28% of patients in D1 but classified only 45.83% of controls correctly. On D2, the same classifier produced 97.70% recall but just 43.75% specificity. Global resampling methods such as Synthetic Minority Over-sampling Technique, or SMOTE, and Random Over-Sampling, or ROS), can improve class balance, but the study reported that they did not consistently resolve this trade-off. For example, SMOTE raised the support vector machine’s D1 specificity to 66.67% while reducing recall to 81.63%.

IFI-ML instead uses an uncertainty-aware, dual-track architecture. Its first stage selects features through information gain, calculated as the reduction in target entropy when a feature is known, and then uses a Gini-based decision tree to divide samples into leaf regions. The tree’s threshold is adapted to each training fold: it is defined as the mean Gini impurity across multi-label leaf nodes, with reported approximate values of 0.28 for D1 and 0.25 for D2. Regions below that threshold are treated as non-sensitive and assigned to conventional machine-learning classifiers. Regions above it contain greater class overlap and are routed to a Sugeno Type-2 fuzzy inference system, whose rules are optimized by gradient descent to handle local ambiguity. The framework therefore separates decision-boundary uncertainty from clearer regions rather than applying one model or one resampling strategy to every observation.

Evaluation used a nested cross-validation design intended to reduce data leakage. The outer loop consisted of stratified 10-fold cross-validation, while hyperparameters for 15 conventional classifiers were tuned in an inner 10-fold loop restricted to the active training data. Information-gain selection, partitioning-tree training, threshold calculation and classifier fitting were all performed within the training portion, with the test fold held out from both fitting and hyperparameter optimization. Any required class balancing was also restricted to the training partition rather than being applied globally before the data were split. The researchers compared the framework with logistic regression, linear and quadratic discriminant analysis, Gaussian naive Bayes, k-nearest neighbors, decision trees, support vector machines, random forest, extra trees, bagging, gradient boosting, XGBoost, CatBoost, Gaussian process and a multilayer perceptron. Final metrics were calculated across the machine-learning and fuzzy-inference outputs using a weighted average method.

On D1, IFI-ML produced accuracies ranging from 94.87% to 96.98% across the evaluated configurations, with Matthews correlation coefficients above 0.89 for all 15 classifiers. The best result came from bagging with ROS, which achieved 96.98% accuracy, 96.34% specificity and a Matthews correlation coefficient of 0.936. The support vector machine with SMOTE reached 95.90% recall, 95.29% specificity and 96.43% accuracy, compared with its baseline recall of 97.28%, specificity of 45.83% and accuracy of 84.45%. For k-nearest neighbors on D1, IFI-ML with SMOTE increased specificity from 45.83% in the baseline and 85.42% with SMOTE alone to 94.23%, while maintaining 95.90% recall and achieving 95.88% accuracy. Across the D1 results, the study reported false-negative rates of roughly 3% to 7% and false-positive rates in a similarly low range, although the exact values varied by classifier and balancing configuration.

The results were more variable on the larger D2 dataset, which contained 752 features and a more heterogeneous feature space. IFI-ML accuracies ranged from 84.47% to 92.73%, and its Matthews correlation coefficients exceeded 0.66 across the tested configurations. The Gaussian process with ROS reached 92.17% accuracy, 84.46% specificity and 96.42% recall. For k-nearest neighbors, SMOTE alone yielded 81.25% specificity but only 64.89% recall, whereas IFI-ML with SMOTE produced 82.03% specificity, 95.40% recall and a Matthews correlation coefficient of 0.7796. Baseline CatBoost accuracy of 85.45% fell to 81.74% with SMOTE but was 86.85% under IFI-ML, with the coefficient rising from 0.5199 to 0.6944. The broader D2 false-positive range, reported as approximately 12% to 24% in the study’s summary, shows that the framework’s gains were not uniform across the larger dataset.

The authors also used SHAP, or Shapley Additive Explanations, to examine which acoustic variables influenced the model’s predictions. In D1, pitch period entropy, spread1, noise-to-harmonic ratio, fundamental frequency and related measures of vocal instability had the largest absolute SHAP values. High pitch period entropy and spread1 were associated with predictions toward the Parkinson’s class, while lower values pushed predictions toward the healthy-control class. Jitter and shimmer measurements also contributed. The direction was not identical in D2: several high-dimensional entropy, energy and Teager-Kaiser Energy Operator coefficients were associated with more negative SHAP values and control predictions. The authors described this pattern as consistent with dataset- or component-specific differences in irregularity or energy in Parkinson’s phonation, but the analysis remains a model interpretation rather than evidence of a new biological mechanism.

Several limitations temper the study’s clinical significance. The evaluations used two public collections of unequal size, and the experiments measured classification performance through cross-validation rather than deployment in a clinic. D1 contained only 195 observations, while D2 used 752 features for 756 samples, a setting in which the authors noted increased training complexity and potential redundancy. The framework also required initial entropy-based partitioning, and performance depended on the selected classifier, balancing method and dataset. Although the results support IFI-ML as a promising computational approach to vocal classification, they do not establish it as a standalone diagnostic method for Parkinson’s disease. The authors identify longitudinal, multi-institutional validation and integration with motor and cognitive biomarkers as necessary next steps.

Subject of Research: Entropy-guided fuzzy inference and machine learning for voice-based Parkinson’s disease classification

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

Article References: S., V., & G., M. (2026). IFI-ML: a novel framework combining Fuzzy Inference and Machine Learning for Parkinson's disease classification. Frontiers in Artificial Intelligence, 9, Article 1949960. https://doi.org/10.3389/frai.2026.1949960

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1949960

Keywords: Parkinson’s disease, voice biomarkers, fuzzy inference, machine learning, class imbalance, information gain, Gini impurity, cross-validation, SHAP analysis, medical artificial intelligence, IFI-ML, novel