JMIR Publications

ABSTRACT

Background:

Serum protein electrophoresis (SPE) is a key diagnostic method in clinical biochemistry. While convolutional neural networks (CNNs) have recently been applied to automate classification, these models treat electrophoretic curves as images, raising concerns about interpretability and computational efficiency, since the underlying data are inherently tabular.

Objective:

This study aimed to evaluate whether machine learning methods designed for tabular data could provide a more accurate, interpretable, and computationally efficient alternative to CNN-based classification of SPE.

Methods:

Electropherogram and gel images from the dataset of Lee et al. were processed to extract tabular features, including grayscale normalization and protein fraction quantification. Several algorithms—XGBoost, TabPFN, and CatBoost—were compared. Performance was evaluated using the same training and test sets as the original CNN study, focusing on sensitivity, specificity, and F1-score.

Results:

CatBoost achieved the best performance among tested models. Compared with CNN-based results, the tabular approach showed higher weighted sensitivity (0.844 vs 0.773), specificity (0.967 vs 0.952), and F1-score (0.768 vs 0.599), with improvements consistent across most pathologies except monoclonal gammopathies. The tabular model required fewer computational resources and no hyperparameter tuning to outperform CNNs.

Conclusions:

Tree-based ensemble methods, particularly CatBoost, represent a robust and interpretable alternative to deep learning for SPE classification. This approach aligns more closely with the intrinsic structure of electrophoresis data, reduces computational demands, and facilitates integration into clinical workflows, offering a promising direction for diagnostic decision support systems.