Two-Level Ensemble with Four Meta-Features for Diabetes Classification on Clinical Tabular Data
Journal of Applied Informatics and Computing · 21.04.2026 · 10 (2) · 1694-1706
DOI: 10.30871/jaic.v10i2.12239 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Applied Informatics and Computing, online erschienen am 21.04.2026. Verfasst von Farid Ma'ruf und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-SA 4.0).
Themen: Diabetes mellitus
Abstract (Originalsprache)
Diabetes mellitus remains a major global public health challenge due to its increasing prevalence, high risk of chronic complications, and growing burden on healthcare systems. In this context, early detection supported by artificial intelligence has become increasingly important, particularly for large-scale clinical tabular data. However, no single model consistently performs best across all clinical tabular datasets, and models with strong discriminative ability do not always provide reliable probability estimates or sufficient interpretability. This study proposes a two-level ensemble model with four meta-features for diabetes classification on clinical tabular data. At the first level, XGBoost and a baseline Deep Neural Network (DNN) were used as heterogeneous base learners. Their prediction probabilities were then transformed into four meta-features, namely the XGBoost probability, the DNN probability, the absolute difference between the two probabilities, and their product, which were subsequently modeled using Logistic Regression at the second level. The proposed model was evaluated against XGBoost, Random Forest, and Baseline DNN using Stratified 5-Fold Cross-Validation and an independent hold-out test. Performance was assessed using ROC-AUC, accuracy, precision, recall, F1-score, specificity, Brier score, confusion matrix, threshold optimization for screening mode, isotonic probability calibration, SHAP, LIME, and DeLong statistical testing. On the hold-out test, the proposed Meta Level-2 LR (4 features) achieved a ROC-AUC of 0.979451, accuracy of 0.97170, F1-score of 0.806562, precision of 0.962480, specificity of 0.997486, and the best Brier score of 0.022476. Although XGBoost obtained the highest ROC-AUC (0.979969), the proposed model demonstrated the most balanced overall performance, particularly in terms of F1-score, precision, specificity, calibration quality, and suitability for clinical decision support. SHAP and LIME further indicated that the most influential features were clinically plausible, especially HbA1c_level, blood_glucose_level, age, and BMI. These findings indicate that the proposed two-level ensemble provides a strong balance among discriminative performance, probability reliability, and interpretability, and therefore has strong potential for clinical decision support in diabetes classification.
Zitieren
APA
Ma'ruf, F., Prasetya, I. W. D., & Sa’ida, I. A. (2026). Two-Level Ensemble with Four Meta-Features for Diabetes Classification on Clinical Tabular Data. Journal of Applied Informatics and Computing, 10(2), 1694-1706. https://doi.org/10.30871/jaic.v10i2.12239
Vancouver
Ma'ruf F, Prasetya IWD, Sa’ida IA. Two-Level Ensemble with Four Meta-Features for Diabetes Classification on Clinical Tabular Data. Journal of Applied Informatics and Computing. 2026;10(2):1694-1706. doi:10.30871/jaic.v10i2.12239
BibTeX
@article{Maruf2026,
author = {Ma'ruf, Farid and Prasetya, Ifnu Wisma Dwi and Sa’ida, Ita Aristia},
title = {Two-Level Ensemble with Four Meta-Features for Diabetes Classification on Clinical Tabular Data},
journal = {Journal of Applied Informatics and Computing},
year = {2026},
volume = {10},
number = {2},
pages = {1694--1706},
doi = {10.30871/jaic.v10i2.12239}
}
EUVIMED