Leveraging Transfer Learning for Accurate CRP Level Prediction in Diabetic Patients

This study addresses the challenge of predicting C-reactive protein (CRP) levels in patients with type 2 diabetes mellitus by integrating tabular-to-image transformation techniques with transfer learning architectures. Utilizing a dataset of 838 clinical records, three encoding methods (Zero-padded Grid, Recurrence Plot or Gramian Angular Field) were applied to convert tabular clinical data into two-dimensional images. ResNet50 and Vision Transformer, two pre-trained models, were employed. Notably, the findings of this study suggested that predictive accuracy is highly influenced by both the model architecture and image encoding technique. The integration of spatial encoding with Vision Transformer-based models presents a promising direction for non-invasive, data-driven inflammation monitoring in diabetic care. The Vision Transformer model paired with Zero-Padded Grid achieved the highest performance, with a test accuracy of 97.62% and an F1-score of 0.90, demonstrating strong discriminative capacity with AUC values exceeding 0.99 across all classes, the Gramian Angular Field encoding image technique with the Vision Transformer model combination also demonstrated strong results, with an average AUC of 0.98, slightly higher than Zero-Padded Grid with Vision Transformer model.