A Novel Survey-Driven Multimodal Deep Learning Framework for Anemia Detection Using Clinical, Ocular, and Geospatial Data

Roughly a quarter of the global population is anemic, yet laboratory hemoglobin measurement remains hard to obtain in low-resource settings. In this pilot work we introduce AnemiaFusionNet, a deep model that detects anemia by jointly using clinical survey responses, palpebral conjunctiva photographs and location information. A hybrid dataset of 114 participants (74 anemic, 40 non-anemic) from 16 Indian states and union territories was collected through an online survey distributed via WhatsApp and Facebook. Each conjunctival photograph is described in two ways: by convolutional (CNN) features, and by a conjunctiva color-and-pallor-aware (CCP) block that pairs color statistics with GLCM-based texture descriptors. Clinical and geospatial features are encoded separately, and all modalities are fused by a transformer with multi-head self- and cross-attention. Trained on the original 114 samples, the model reached 100% validation accuracy within three epochs, indicating overfitting. After image augmentation and a 70:15:15 split of the augmented dataset, AnemiaFusionNet achieved 92.0% accuracy, 93.5% precision, 89.9% sensitivity, 94.0% specificity and a 91.7% F1-score on 3000 test samples. For comparison, geospatial-only and clinical-only logistic regression reached 70.0% and 88.8% accuracy, and an image-only ResNet-50 reached 100%. Because the augmented samples were split at the sample level, these estimates are likely optimistic; participant-level validation on a larger cohort is required. The study demonstrates the feasibility of survey-based multimodal data collection for anemia screening research.