Machine learning approach for the detection of vitamin D level: a comparative study

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Nutritional Science

Background:

• Vitamin D's critical role in health, amplified by the COVID-19 pandemic.
• Limitations of 25-hydroxy vitamin D (25-OH-D) blood testing, including feasibility issues.
• The challenge of multicollinearity in health datasets, potentially leading to inaccurate machine learning models.

Purpose of the Study:

• To develop a clinically acceptable machine learning model for detecting vitamin D status in North Cyprus adults.
• To accurately assess vitamin D levels without requiring 25-OH-D blood tests.
• To mitigate the impact of multicollinearity on machine learning model performance.

Main Methods:

• Comparative analysis of four supervised machine learning models: Ordinal Logistic Regression (OLR), Elastic-Net Ordinal Regression (ENOR), Support Vector Machine (SVM), and Random Forest (RF).
• Evaluation of models based on sensitivity to metabolic syndrome status, hyper-parameter tuning, and training data size.
• Assessment of classification performance metrics including accuracy, specificity, sensitivity, precision, F1-score, and Cohen's kappa.

Main Results:

• Support Vector Machine (SVM) performance was negatively impacted by multicollinearity.
• Random Forest (RF) exhibited greater robustness to variations in training data size and multicollinearity.
• RF and ENOR models outperformed OLR and SVM, especially with reduced sample sizes, indicating resilience to multicollinearity.

Conclusions:

• Random Forest (RF) demonstrated superior accuracy (0.94), specificity (0.96), sensitivity (0.94), precision (0.95), F1-score (0.95), and Cohen's kappa (0.90) compared to other models.
• The findings support the development of an intelligent vitamin D detection system using RF.
• This system could reduce the cost and time associated with traditional vitamin D level detection methods.