Quantitative abnormalities in the β-region of the electrophoretic profile of serum proteins as predictive markers of monoclonality: Machine learning for monoclonality prediction

Objective of the study

To develop a machine learning algorithm aimed at predicting the presence of a monoclonal (M-) protein when the β-globulin fraction is elevated.

Materials and method

Patients were selected as part of the University Hospital Laboratory of Brussels routine diagnostic procedures from October 2021 to April 2022. Adult patients with serum protein electrophoresis showing elevated β1 and/or β2 fractions were included. The selection was done following strict exclusion criteria such as acute inflammation, iron deficiency anemia signs or supposed liver disease. To construct a predictive model for prediction of positive immunofixation (IFE) for monoclonality, the following factors were used: age, sex, β1 and β2 concentration (g/L), total proteins (g/L), IgA, IgM, IgG values (g/L) and hypogammaglobulinemia. The dataset underwent a random split, divided into a foundational training set (80%, 247 samples) and a foundational test set (20%, 62 samples). The training sets were subjected to five different algorithms: logistic regression, decision tree, random forest, gradient boosting, and support vector.

Results

309 patients were selected; 149 exhibited a negative IFE and 160 a positive IFE for monoclonality. The evaluation of the five tested models demonstrated very good performance, the chosen model was Random Forest for its high sensitivity (85%) and area under the receiver operating characteristic curve (91%).

Conclusion

An accurate algorithm was achieved for predicting the presence of M protein when the β-globulin fraction is elevated which enables early and improved diagnosis of monoclonal gammopathy.