Explainable Machine Learning for Vitamin D-Related Chronic Disease Risk Prediction Using MIMIC-III

The prevalence of vitamin D deficiency has been noted to be of great concern to the population due to its association with multiple chronic ailments, such as those that affect the cardiovascular system, diabetes, bone diseases, and the immune system, and even some cancerous forms. The underlying mechanics are not well understood, but epidemiological and clinical research demonstrates notable correlations, correlated to different degrees by population, based on genetic, environmental, and lifestyle differences. The use of advanced machine learning methods in this study aimed to examine the association between serum vitamin D levels and chronic disease risk using the de-identified electronic health records from Intensive Care Unit (ICU) patients in the MIMIC-III database. A final cohort of 38,712 unique patients was selected after applying predefined inclusion and exclusion criteria, the data being used for machine learning model development and evaluation. To predict the status of chronic diseases based on vitamin D levels, supervised machine learning models were used: Logistic Regression (LR), Support Vector Machines (SVM), Random Forest (RF), and K-Nearest Neighbors (KNN). Data preprocessing involved strategies such as missing data imputation, normalization, and feature engineering to improve model training. The accuracy, precision, recall, F1-score, and receiver operating characteristic curve (ROC-AUC) were used to assess model performance. LR performed the best with the ROC-AUC of 92.5% and accuracy of 85.2%, and SVM was close behind the ROC-AUC: 90.9 percent. The chronic disease was associated with 2.1 times higher risk of chronic disease among people with severe vitamin D deficiency (p < 0.001), which was in agreement with earlier observational data. RF did not work as well (ROC-AUC: 87.7%) because it required higher dimensionality data, e.g., genetic and lifestyle parameters. The study shows the possible use of AI-based diagnostics in preventive medicine. As an example, presented vitamin D testing of high-risk population (e.g., increasing tests in old age, obesity, or dark-skinned population) adapted to ML-based risk assessment as well as supplementation plans could enhance this prevention.