VitaDetect: Deep Learning Based Vitamin Deficiency Detection System

Abstract: VitaDetect is a preliminary, non-invasive vitamin-deficiency screening prototype combining EfficientNetV2S image classification with a Multi-Layer Perceptron (MLP) trained on questionnaire data. It predicts Normal, Vitamin A, Vitamin B-12, Vitamin C, and Vitamin D classes. The image dataset was audited for duplicates, conflicting labels, and corrupted images; the symptom branch uses Scikit-Learn preprocessing and SMOTE. Since the modalities have no shared patient identifiers, recall-weighted decision-level fusion with disagreement arbitration is used. On independent test sets, image and symptom models achieved 75.61% accuracy on 783 photographs and 78.00% on 600 records. A React/Vite frontend and Flask backend support up to four photos per assessment.

Keywords: vitamin deficiency detection, deep learning, EfficientNetV2, symptom classification, multimodal learning, late fusion, uncertainty handling, data quality auditing