A Novel Machine Learning Framework for Reliable and Accurate Diabetes Detection
Journal of Eye-Innovation in Machine Learning · 12.08.2026 · 1 (20261) · 71
DOI: 10.68990/jieml.202614 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Eye-Innovation in Machine Learning, online erschienen am 12.08.2026. Verfasst von Theyazn Aldhyani und 1 weiteren Autor:innen. Die Arbeit nennt 38 Literaturangaben.
Themen: Künstliche Intelligenz in der Pflege, Diabetes mellitus
Abstract (Originalsprache)
Public, large-scale diabetes data enable reproducible machine-learning comparisons, but high apparent performance can mask uncertain provenance, duplication, class imbalance, diagnostic feature dependence, and leakage risk.
Objective
We reproduced a leakage-aware benchmark on the Kaggle Diabetes Prediction Dataset and measured performance, uncertainty, calibration, threshold behavior, and dependency on HbA1c and blood glucose.
Methods
The downloaded file of 100000 rows was hash-verified. We removed 3,854 exact duplicate full rows, grouped identical predictor vectors to avoid cross-partition overlap, and kept 96,146 records. An 80/20 holdout was created that was fixed and group-aware; all preprocessing and three-fold stratified group cross-validation was done within the training data. Average precision was used as the selection criterion to compare logistic regression, random forest, histogram gradient boosting, and a prevalence dummy. Test metrics received 95% class-stratified bootstrapped intervals.
Results
Histogram gradient boosting achieved the highest training cross-validation PR-AUC of 0.887 +/- 0.010 and test ROC-AUC of 0.977 [0.974, 0.979] and PR-AUC of 0.877 [0.865, 0.889]. At threshold 0.5, sensitivity was 0.898 [0.884, 0.912], precision was 0.479 [0.466, 0.491], and the MCC was 0.614 [0.601, 0.626]. Test PR-AUC dropped to 0.328 [0.309, 0.351] after HbA1c and blood glucose were removed.
Conclusions
The benchmark shows good within-dataset discrimination, but high dependence on variables overlapping diagnostic criteria. Uncertain sampling with potential synthetic or augmented records, incomplete licensing information, questionable person-level independence, miscalibration, and lack of external validation preclude claims of clinical or population generalizability.
Zitieren
APA
Aldhyani, T., & Ahmed, Z. (2026). A Novel Machine Learning Framework for Reliable and Accurate Diabetes Detection. Journal of Eye-Innovation in Machine Learning, 1(20261), 71. https://doi.org/10.68990/jieml.202614
Vancouver
Aldhyani T, Ahmed Z. A Novel Machine Learning Framework for Reliable and Accurate Diabetes Detection. Journal of Eye-Innovation in Machine Learning. 2026;1(20261):71. doi:10.68990/jieml.202614
BibTeX
@article{Aldhyani2026,
author = {Aldhyani, Theyazn and Ahmed, Zeyad},
title = {A Novel Machine Learning Framework for Reliable and Accurate Diabetes Detection},
journal = {Journal of Eye-Innovation in Machine Learning},
year = {2026},
volume = {1},
number = {20261},
pages = {71},
doi = {10.68990/jieml.202614}
}
EUVIMED