ARISE ensemble tops 15 biomedical benchmarks with 0.793 balanced accuracy

A new arXiv preprint introduces ARISE, a feature selection framework built for the tough reality of small-sample biomedical omics studies, where the number of genes or proteins far exceeds the number of patients. Developed by Zardad Khan, Amjad Ali and colleagues, ARISE stands for Adaptive Residual-Informed Stability Ensemble. It fuses seven percentile-normalized relevance components through 15 predefined profiles, using nested inner cross-validation to adaptively weight each profile. This lets the method balance predictive relevance, class-balanced stability, and residual-informed redundancy control within a single transparent pipeline.

The team validated ARISE on five molecular datasets, testing eight feature-set sizes with three classifiers (k-nearest neighbours, support vector machine, random forest) and six filter-based comparators. Using five-fold outer cross-validation repeated 50 times, they generated 210,000 held-out assessments. ARISE achieved the top rank in every dataset-metric combination, with equal-dataset means of 0.793 balanced accuracy, 0.776 macro-F1, and 0.725 Cohen's kappa—beating the strongest aggregate comparator by 0.022, 0.023, and 0.028 respectively. It maintained strong performance even with compact feature sets, though optimal budgets varied by dataset, suggesting ARISE robustly adapts to different biological signals.