Machine Learning Techniques for Identifying Folate and Vitamin B12 Deficiency Anemia Using Blood Biomarkers

This study explores the classification of vitamin B12 deficiency anemia (B12DA) and folate deficiency anemia (FDA) using machine learning (ML) and standard blood parameters derived from complete blood count (CBC) data. With appropriate ethical clearance, a dataset of 267 patient samples with nearly equal class distribution was gathered from Kasturba Medical College in Karnataka, India. Stratified k-fold cross-validation was employed to reduce the hazards associated with the small dataset, and oversampling was utilized sparingly to balance classes. K-Nearest Neighbours, AdaBoost, and LightGBM were among the ML models that performed well in the evaluation; cross-validation results revealed that KNN had the greatest accuracy (97%). Although our results imply that ML can improve diagnostic precision and facilitate quicker, non-invasive classification between FDA and B12DA. These results should be regarded as proof-of-concept due to the limited dataset size and single-institution sampling. Unlike prior work focused on broader anemia subtypes, this study specifically targets the underexplored diagnostic ambiguity between FDA and B12DA. Furthermore, the integration of explainable AI for enhancing clinical interpretability is a key component of the broader research, though it has not been implemented in the present study. Future incorporation of such techniques, along with larger multi-center datasets and advanced methodologies, could help overcome current limitations. Nevertheless, the findings highlight the potential of ML to improve diagnostic precision, thereby supporting medical professionals in early detection and personalized treatment planning.