Artificial intelligence aided serum protein electrophoresis analysis of Finnish patient samples: Retrospective validation

Background and aims: Serum protein electrophoresis interpretation requires a substantial amount of manual work. In 2020, Chabrun et al. created a machine learning method called SPECTR for the task. We aimed to validate and test the SPECTR method against our results of more precise immunofixation electrophoresis. Materials and methods: We gathered 34 625 patients and their first serum protein electrophoresis sample in Helsinki University Hospital. We trained three neural network models: (1) a fractionation model to fractionate electropherograms; (2) a classification model to classify samples to normal, ambiguous, and abnormal (i.e. containing paraprotein); (3) an integration model to predict concentration and location of paraproteins. Results: The fractionation model demonstrated an error rate of ≤0.33 g/L in 95 % samples. The classification model achieved an area under the curve of 97 % in receiver operating characteristic analysis. The integration model demonstrated a coefficient of determination (R²) of 0.991 and a root-mean-square error of 1.37 g/L in linear regression. Conclusion: The neural network models proved to be suitable for partial automation in serum protein electrophoresis reporting, i.e. classification of normal electropherograms. Furthermore, the models can accurately suggest the location and concentration of paraproteins.