MULTI-MODEL COMPARISON AND UNCERTAINTY-BASED RELIABLE DECISION ANALYSIS FOR DIABETES PREDICTION IN LARGE-SCALE HEALTH DATA

Doğan L., Teke O., İnaç T.

INTERNATIONAL BLACK SEA SCIENTIFIC STUDIES AND RESEARCH CONGRES, Rize, Turkey, 4 - 06 May 2026, pp.208-216, (Full Text)

  • Publication Type: Conference Paper / Full Text
  • City: Rize
  • Country: Turkey
  • Page Numbers: pp.208-216
  • Kütahya Health Sciences University Affiliated: Yes

Abstract

ABSTRACT

Diabetes is a significant chronic condition that is rapidly increasing worldwide and can lead to serious complications if not diagnosed early. In this study, the Diabetes Health Indicators Dataset was used for the purpose of predicting diabetes. The dataset was obtained from the Kaggle platform and contains a total of 70,692 individual records, comprising variables related to demographics, lifestyle and health status. In the study, a multi-model comparison was conducted using Logistic Regression, Decision Tree, Random Forest, Extra Trees and Gradient Boosting algorithms. Model performance was evaluated using the Accuracy, Precision, Recall, F1-score and ROC-AUC metrics. According to the results obtained, the Gradient Boosting algorithm was found to deliver the highest performance, with an F1-score of 0.762 and an ROC-AUC value of 0.833. To improve model performance, a data-driven optimal threshold value was determined instead of a fixed decision threshold. As a result of this approach, the optimal threshold value was found to be 0.38, and the model’s recall value rose from 0.79 to 0.88, whilst the F1-score reached 0.775. The main contribution of this study is the proposal of an uncertainty-based decision mechanism in addition to the classical binary classification approach. In this context, model prediction probabilities were analysed, and examples close to the decision boundary were defined as uncertain and excluded from the evaluation process. The findings indicate that approximately 15.48% of the data was classified as uncertain, whilst model performance improved significantly on the remaining reliable examples. With this approach, the model’s recall rose to 0.92, whilst the F1-score reached 0.813. Consequently, the proposed approach demonstrates that machine learning models should be evaluated not only in terms of accuracy but also in terms of decision reliability. Particularly in healthcare applications where misdiagnosis is critical, the separation of uncertain cases contributes to the development of more reliable and clinically applicable decision support systems.

Keywords: Machine Learning, Healthcare Data, Diabetes Prediction, Uncertainty Analysis, Decision Threshold Optimisation