Pain assessment relies heavily on a patient describing how they feel. Because this is highly subjective, there is a need of objective, biological way to measure pain using protein biomarkers. A laboratory technique called 2D-GE can map out these proteins visually, but having humans analyze these complex image maps is slow and inconsistent. To solve this, this thesis evaluates how deep learning can automatically scan 2D-GE images of different bodily fluids to find objective, physical evidence of pain.

To achieve this, the study systematically evaluated three distinct neural network architectures: convolutional neural networks (CNNs), specifically ResNet18 and ResNet34, and a vision transformer (ViT). The models were trained and validated on a proprietary dataset of 2D-GE images derived from heterogeneous bodily fluids (saliva, cerebrospinal fluid [CSF], and blood plasma). The methodology evaluated both global (mixed-fluid) and fluid-specific classification tasks. Monte Carlo cross-validation across 50 iterations was employed to ensure evaluation, comparing metrics such as accuracy, balanced accuracy, and Matthews Correlation Coefficient (MCC). Furthermore, the necessity of transfer learning was tested against training from scratch. To confirm the significance of the observed performance disparities, two-sided pairwise permutation tests (10,000 resamples) and Wilcoxon signed-rank tests (α = 0.05) were executed. Finally, visual explainability algorithms specially, GradCAM++ for CNNs and attention heatmap for the ViT were utilized to interpret the spatial reasoning behind the models’ predictions.The study established that pain classification is feasible but fluid-dependent. Models trained on heterogeneous global data struggled to surpass a random-chance baseline due to confounding factor. However, for fluid-specific classification, saliva emerged as a significant diagnostic medium, achieving a good predictive accuracy of 82.49%±5.63%. Conversely, blood plasma proved unsuitable for this task, largely due to high variations in sample quality and background noise.

Computationally, transfer learning was proven necessary; without pre-trained weights, the deeper ResNet34 model suffered considerable performance degradation by performing at 50%. With transfer learning, the CNN architectures demonstrated a significant statistical superiority over the ViT. For the overall binary classification task, ResNet34 achieved the highest overall accuracy of 65.2%±5.8%. Whereas, ResNet18 achieved an accuracy of 62.5%±4.4% and ViT achieved an accuracy of 59.4%±5.2%. For the complex 6-class multiclass task, ResNet34 achieved 63.0%±4.0% accuracy against a 16.67% baseline, again establishing significant superiority over the ViT (p = 0.0002). Despite the predictive shortcomings of the ViT, explainability analyses revealed a crucial trade-off: GradCAM++ showed that ResNet models relied on broad, diffuse regional patterns, while the ViT attention heatmaps acted as a more localized focus, successfully pinpointing discrete, individual protein spots.Ultimately, this research demonstrates that while deeper CNN (ResNet34) are the superior diagnostic classifiers for evaluating complex 2D-GE images, the ViT provides the precise higher localized focus required for downstream biological identification. This thesis lays a foundation for the algorithmic pain-related protein biomarkers detection. Future efforts must focus on mitigating data scarcity within the underrepresented fluid cohorts and exploring multimodal data fusion to further enhance diagnostic capabilities.