Confidence‑Based Misdiagnosis Analysis in Machine Learning Models for Diabetes Prediction

This paper presents a confidence‑based framework for examining diagnostic errors in machine learning models for diabetes prediction. Instead of looking only at accuracy, the framework focuses on how each model represents and handles uncertainty. To evaluate this idea, three models with different probabilistic characteristics—Gaussian Naive Bayes, LogitBoost, and Logistic Regression—were trained on the Pima Indian Diabetes dataset. Using a margin‑based criterion, errors were categorized into Severe Misdiagnosis and Near‑Miss cases. The analysis shows that although the models achieve similar accuracy (LogitBoost ۰.۷۱۹, Logistic Regression ۰.۷۵۸, Naive Bayes ۰.۷۳۹), the three models produce fundamentally different misdiagnosis patterns. LogitBoost yields the highest number of severe errors (۴۲ vs. ۶ and ۱۴), while Logistic Regression and Naive Bayes generate substantially more near‑miss errors (۳۱ and ۲۶ vs. ۱). These results suggest that trust‑oriented error analysis can reveal hidden weaknesses in predictive models and support the development of safer medical diagnosis systems.

کلیدواژه ها:

Diabetes Diagnosis ، Naïve Bayes ، LogitBoost ، Logistic Regression ، Error Analysis ، Model Confidence