REA Press null REA Press3042-01803042-0180 REA Press https://doi.org/10.22105/scfa.v3i2.102 Research Article Active learning, Label efficiency, Class imbalance, Label noise, Query-by-committee, Clinical tabular data Label-Efficient Active Learning on Clinical-Style Tabular Data: Uncertainty, Margin, and Committee Strategies Across Imbalance and Noise RegimesLabel-Efficient Active Learning on Clinical-Style Tabular Data: Uncertainty, Margin, and Committee Strategies Across Imbalance and Noise Regimes Alimoradi Mahmoud Department of Computer Engineering and Information Technology, La.C., Islamic Azad University, Iran. Ahmed Shaban Awaz Department of Information Technology, Technical College of Informatics–Akre, Akre University for Applied Sciences, Duhok, Iraq. Madina Esdauletova Isametova Satbayev University, 050000 Almaty, Kazakhstan. 06 2026 07 06 2026 3 2 © 2026 REA Press 2026

This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

 Label-Efficient Active Learning on Clinical-Style Tabular Data: Uncertainty, Margin, and Committee Strategies Across Imbalance and Noise Regimes

This study compares uncertainty, margin, entropy and Query-by-Committee (QbC) acquisition with random sampling on the Wisconsin Diagnostic Breast Cancer dataset and two derived stress regimes. The reported design uses thirty seeds, three learners, twenty initial labels and batches of ten up to 150 labels. The retained summaries report favorable Area-Under-Learning-Curve (AULC) differences in all 36 strategy-regime-learner cells and Holm-adjusted paired Wilcoxon p-values at most 0.0039. Under the noisy regime, logistic-regression AULC is reported as 0.9587 for random and 0.9837 for uncertainty; the corresponding mean retrospective threshold budgets are 32.0 and 24.7 labels, approximately 23% lower. QbC has the largest reported mean AULC gain under imbalance for each learner. Favorable cell medians coexist with individual seeds below random sampling. AULC integration, binary-score tie handling, ANOVA units and execution provenance require original code and outputs to validate these comparisons. The conclusions concern one retrospective benchmark with controlled fixed-count label corruption, rather than clinical annotation policies or a general benefit of noisy labels.

Null