Comparing machine learning model accuracy for predicting Vitamin D deficiency at different cut-offs in Saudi men

Abstract

Background

Despite Saudi Arabia's sunny climate, vitamin D deficiency is prevalent. We aimed to utilize several machine learning algorithms to predict Vitamin D deficiency among Saudi men and identify correlated features.

Methods

We collected data from the records of King Khalid University Hospital in Riyadh between 2019 and 2021. Variables included health conditions, race, blood pressure, BMI, blood glucose, lipid profile, and total 25-hydroxyvitamin D (25OHD) levels. Six ML algorithms (Support Vector Machine, decision tree, linear regression, gradient boosting, XGBoost, and random forest) were employed to construct models for both cutoff points of <75 nmol/L and <50 nmol/L. The dataset was randomly divided into the training set and validation set at a ratio of 8:2. The accuracy of the algorithm was tested using the ten-fold cross-validation method. Performance metrics such as accuracy, precision, recall, F1 score, and area under the precision-recall curve (PRC, AUC) were assessed and compared between the models for both categories.

Results

Among 1700 Saudi men, 76.1% exhibited vitamin D deficiency (<75 nmol/l), with 52.5% deficient using an alternative cutoff (<50 nmol/l). Gradient boosting algorithm exhibited superior predictive accuracy, especially at the <75 nmol/L cutoff, outperforming the discrimination power of other algorithms for both classes. BMI emerged as the strongest predictor of vitamin D deficiency, followed by age and blood sugar results. The PRC, AUC for the <50 cutoff varies between 0.59 and 0.63 for the four classifiers (SVM, Linear regression, XGBoost, and Gradient boosting), whereas it reached 0.83 for the <75 cutoff.

Conclusion

Vitamin D deficiency is prevalent among healthy Saudi Arabian men. ML algorithms have proven effective in identifying correlated features within this highly impacted population, enabling tailored interventions and appropriate preventive strategies to address vitamin D deficiency and insufficiency. This model's high accuracy can help identify individuals at risk without the need for costly and time-consuming blood tests.