Accurate Diabetes Risk Stratification Using Machine Learning: Role of Missing Value and Outliers

Area of Science:

• Computational biology and bioinformatics
• Medical informatics and machine learning

Background:

• Diabetes mellitus is a global metabolic disease with rising prevalence, posing challenges for accurate risk stratification.
• Existing machine learning (ML) classifiers often exhibit suboptimal performance on diabetes datasets due to issues like missing values and outliers.

Purpose of the Study:

• To develop an optimized and robust machine learning system for improved diabetes risk stratification.
• To investigate the hypothesis that imputing missing values and outliers with median configurations enhances classification accuracy.

Main Methods:

• Feature selection was performed using six techniques (random forest, logistic regression, mutual information, PCA, ANOVA, FDA).
• Ten classifiers (LDA, QDA, Naive Bayes, GPC, SVM, ANN, Adaboost, LR, Decision Tree, Random Forest) were evaluated.
• The Pima Indian diabetic dataset was utilized, with missing values and outliers imputed using median values.

Main Results:

• The combination of Random Forest feature selection and Random Forest classification, with median imputation, achieved 92.26% accuracy.
• Key performance metrics included 95.96% sensitivity, 79.72% specificity, and 0.93 AUC.
• This approach demonstrated a significant 10% improvement over previously published methods.

Conclusions:

• Replacing missing values and outliers with median values, coupled with Random Forest feature selection and classification, is an effective strategy for diabetes risk stratification.
• The developed ML system is robust, reliable, and offers superior performance compared to existing techniques.