Machine Learning-Driven Multi-Class Prediction of Vitamin Deficiency Diseases: A Clinically Interpretable Approach Integrating Socioeconomic Determinants and Biomarker Profiles

Abstract

Deficiency disease conditions due to micronutrients represent serious public health issues in low-resource settings wherein central laboratory facilities are not accessible. Even though earlier studies using digital health approaches have concentrated on detection of single diseases in a binary fashion, multi-class prediction involving multiple domains like clinical signs and symptoms, biomarkers and socio-economic factors is relatively unexplored.  A proof-of-concept dataset using parameterization (N = 4,000) with 33 parameters from demographic, lifestyle, nutritional, biological markers and pathognomonic sign and symptom domains was used for evaluation. Random Forest, Gradient Boosting and SVM classifiers were trained using stratified splitting in 80/20 ratios and validated by 5-fold cross-validation. Gradient Boosting gave the best multi-class discrimination, with 94.6% global accuracy, a macro-averaged F1-score of 91.5%, and a Matthews Correlation Coefficient of 0.926. Class-specific F1-scores were 0.989 for healthy controls, 0.959 for anemia, 0.937 for night blindness, 0.950 for rickets or osteomalacia, and 0.742 for the minority scurvy cohort. Discrete classification of scurvy was limited by low sample support (4.2% prevalence), but probabilistic risk ranking still performed well (ROC-AUC: 0.9737). The most important features were serum vitamin D (weight: 0.246), night blindness symptom (0.187), serum vitamin B12 (0.149), composite symptom burden (0.112), and hemoglobin (0.089); these ranking matches known physiological pathways. Sequential gradient boosting gives a robust, interpretable screening model for multi-class vitamin deficiency risk. These are proof-of-concept results with internal construct validity; prospective registry validation and threshold calibration are needed before clinical use.