Mohammed Ramadhan
Department of Computer Science, University of Zakho, Zakho, Kurdistan Region-Iraq

Published : 1 Documents Claim Missing Document
Claim Missing Document
Check
Articles

Found 1 Documents
Search

Hybrid Oversampling and FT-Transformer-Based Classification for Imbalanced Clinical Tabular Data: A Multi-Dataset Statistical Evaluation Asaad Khaleel; Mohammed Ramadhan; Masood Abdulrahman
International Journal of Research and Applied Technology Vol. 6 No. 2 (2026): December, 2026 (Issue in Progress)
Publisher : Universitas Komputer Indonesia

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.34010/injuratech.v6i2.21230

Abstract

Class imbalance is a problem which is not new on the clinical data, but is not well dealt with in practice. A model trained on a skewed distribution will often claim high accuracy on the skewed dataset, but will miss almost all of the minority class, typically the class that is the most important for detection — such as the class of patients who require detection the most. The three synthetic oversampling methods tested were SMOTE and ADASYN and Borderline SMOTE, alongside the three classifiers Random Forest and Gradient Boosted Trees and a Bayesian-tuned FT-Transformer, applied to three datasets: Pima Diabetes (768 records); a cardiovascular dataset (3,000 records); and Heart Failure Clinical Records (299 records, 203 survived, 96 died). All 14 resulting configurations were tested using 5-fold cross validation repeated 5 times per fold (i.e., 25-fold-pairs), and Wilcoxon Signed-Rank tests were performed to determine if the improvements are statistically significant when compared with the baseline (RF). The answer is: Yes, they were — in many instances very much so. On Pima, ADASYN together with the tuned FT-Transformer obtained F1=0.6677 and Recall=0.8209, while for the unaugmented RF it was 0.5145 and 0.4563, respectively. In the Pima ablation study specifically, the large portion of that gain (+15.6%) was due to oversampling, with hyperparameter tuning contributing another +9.2%; these percentages are specific to the Pima/RF ablation experiment and should not be interpreted as dataset-agnostic general findings. Synthetic augmentation was tested using SHAP on the RF branch only, and did not alter the variables that the model uses.