Machine learning evaluation for identification of M-proteins in human serum

Area of Science:

• Clinical Chemistry
• Bioinformatics
• Machine Learning in Medicine

Background:

• Serum protein electrophoresis (SPEP) is crucial for diagnosing hematological diseases like multiple myeloma by detecting monoclonal proteins (M-proteins).
• Machine learning (ML) shows potential in analyzing protein electrophoresis data, including glycan patterns for tumor monitoring.

Purpose of the Study:

• To compare 26 decision tree algorithms for identifying M-proteins using serum protein capillary electrophoresis data.
• To utilize a game theoretic approach to identify key diagnostic features in electrophoresis data.
• To evaluate ML algorithms for M-protein isotype classification.

Main Methods:

• Anonymized serum protein capillary electrophoresis data from 67,073 samples were used.
• Twenty-six decision tree algorithms were evaluated for M-protein detection.
• A game theoretic approach was employed to determine feature importance.
• Performance was assessed for M-protein detection and isotype classification.

Main Results:

• Five algorithms demonstrated superior M-protein detection: Extra Trees (ET), Random Forest (RF), Histogram Grading Boosting Regressor (HGBR), Light Gradient Boosting Method (LGBM), and Extreme Gradient Boosting (XGB).
• The gamma and beta globulin fractions were identified as the most significant features.
• ET and XGB algorithms performed best in M-protein isotype classification.

Conclusions:

• Serum capillary electrophoresis combined with decision tree algorithms offers a rapid and accurate method for M-protein identification.
• These ML approaches have broad diagnostic potential for various blood analyses, including hemoglobinopathies.
• Integrating ML with both numerical capillary electrophoresis and gel electrophoresis image data could enhance M-protein isotype classification.