B-145 An R Shiny App for Automated Peak Deconvolution, Interpretation, and Quantitation of Monoclonal Proteins Using Capillary Electrophoresis Immunotyping Data

Abstract Background Machine learning methods have been utilized to train algorithms for interpretation of immunofixation electrophoresis. This work developed an R Shiny app for immunotyping (IT) data capable of deconvoluting discrete peaks in overlapping backgrounds and automatic quantitation of monoclonal proteins. Unlike machine learning which obfuscates decision making, each step of this deconvolution and quantitation process is displayed to the end user. Methods Individual traces for 1315 previously interpreted IT results were used in the study. A modified Levenberg-Marquardt algorithm using 50 sets of randomized start parameters was used for deconvolution. Primitive peak features such as height, FWHM, AUC and retention time were used to match heavy and light chain deconvoluted peaks for interpretation. The matched peak with the largest AUC was compared to reviewer interpretation and peak quantitation. Repeatability of quantitation was evaluated through 10 consecutive rounds of deconvolution. Code was written for end user visualization in Shiny or batch processing in base R. Results Deconvolution was successful for 97.2% (1279/1315) of specimens. Agreement between the software and reviewer interpretation was 90.6% (11 927/13 150) across all deconvolutions and was consistent for 84.7% of samples in all 10 rounds of deconvolution. Discordant interpretation in 15.3% of samples was often attributed to splitting of large peaks with sharp maximums, large percentage (but small absolute) changes in calculated AUC for monoclonal proteins under 0.5 g/dL, or drifting baseline signal between channels. Mean and median percent CV for consecutive rounds of monoclonal quantitation was 5.9% and 1.16% respectively (protein range 0.02–4.5 g/dL). Deming regression of reviewer vs software quantitation had a calculated slope of 0.728 and Pearson’s r coefficient of 0.933. Conclusion Deconvolution and automatic quantitation of monoclonal protein peaks is possible using the Levenberg-Marquardt algorithm for immunotyping data. Future work aims to improve interpretation agreement with reviewers using more advanced ranking of matched peaks.