Machine learning-based stratification of prediabetes and type 2 diabetes progression

Area of Science:

• Biomedical Informatics
• Computational Biology
• Data Science in Healthcare

Background:

• Diabetes mellitus presents significant global health challenges, necessitating early detection and accurate staging for effective patient management.
• Machine learning (ML) and bioinformatics offer advanced tools to improve diagnostic accuracy and identify critical biomarkers for diabetes.
• Precise classification of diabetes stages is crucial for timely intervention and preventing severe complications.

Purpose of the Study:

• To develop and evaluate machine learning models for classifying individuals into four distinct health states: healthy, prediabetes, type 2 Diabetes Mellitus (T2DM) without complications, and T2DM with complications.
• To identify key molecular and biochemical markers that are predictive of diabetes progression and severity.
• To assess the performance of different machine learning classifiers in accurately staging diabetes.

Main Methods:

• A multi-class classification framework was implemented using molecular markers, biochemical markers, or a combination of both.
• Five machine learning classifiers were tested: Random Forest, Extra Tree Classifier, Quadratic Discriminant Analysis, Naïve Bayes, and Light Gradient Boosting Machine.
• Recursive Feature Elimination with Cross-Validation (RFECV) and fivefold cross-validation were employed to enhance model robustness and feature selection.

Main Results:

• The Extra Trees Classifier demonstrated superior performance, achieving an Area Under the Curve (AUC) of 0.9985 (95% CI: [0.994-1.000]).
• Key predictive markers identified included molecular markers (miR342, NFKB1, miR636) and biochemical markers (albumin-to-creatinine ratio, HDLc).
• The combined model effectively discriminated between the four diabetes health states, highlighting the synergy of different marker types.

Conclusions:

• Integrating machine learning with molecular and biochemical data provides a powerful approach for accurate diabetes staging.
• The findings support the potential for developing more personalized diabetes management strategies based on individual biomarker profiles.
• This study highlights the utility of advanced computational methods in advancing diabetes diagnostics and care.