Serum electrophoresis (SPEP) is a method used to analyze the distribution of the most important proteins in the blood. The major clinical question is the presence of monoclonal fraction(s) of antibodies (M-protein/paraprotein), which is essential for the diagnosis and follow-up of hematological diseases, such as multiple myeloma. Recent studies have shown that machine learning can be used to assess protein electrophoresis by, for example, examining protein glycan patterns to follow up tumor surgery. In this study we compared 26 different decision tree algorithms to identify the presence of M-proteins in human serum by using numerical data from serum protein capillary electrophoresis. For the automated detection and clustering of data, we used an anonymized data set consisting of 67,073 samples. We found five methods with superior ability to detect M-proteins: Extra Trees (ET), Random Forest (RF), Histogram Grading Boosting Regressor (HGBR), Light Gradient Boosting Method (LGBM), and Extreme Gradient Boosting (XGB). Additionally, we implemented a game theoretic approach to disclose which features in the data set that were indicative of the resulting M-protein diagnosis. The results verified the gamma globulin fraction and part of the beta globulin fraction as the most important features of the electrophoresis analysis, thereby further strengthening the reliability of our approach. Finally, we tested the algorithms for classifying the M-protein isotypes, where ET and XGB showed the best performance out of the five algorithms tested. Our results show that serum capillary electrophoresis combined with decision tree algorithms have great potential in the application of rapid and accurate identification of M-proteins. Moreover, these methods would be applicable for a variety of blood analyses, such as hemoglobinopathies, indicating a wide-range diagnostic use. However, for M-protein isotype classification, combining machine learning solutions for numerical data from capillary electrophoresis with gel electrophoresis image data would be most advantageous.
Read full abstract