Supervised machine learning model for serum protein electrophoresis data interpretation

Serum protein electrophoresis (SPE) is a frequently employed laboratory test in clinical settings, with over 10,000 requests per annum in our centre. It is primarily utilised for the diagnosis and monitoring of paraproteinaemia. Interpretation on SPE is time-consuming and relies on the expertise of pathologists, with potential interobserver variability. Assistance from machine learning algorithms could improve efficiency and objectiveness. Digitised capillary electrophoresis (CE) tracings acquired using the Sebia Minicap Protein(E) 6 kit were extracted from the analyser, and corresponding reports for SPE were obtained from the laboratory information systems of Princess Margaret Hospital (PMH) and Tuen Mun Hospital (TMH). Three artificial neural networks (for fractionation, classification and location plus quantification) were trained and evaluated against reference interpretations from one to two pathologists. Samples from PMH constitute the training datasets. Trained models were subsequently evaluated with samples from TMH. A total of 41,448 and 24,501 CE tracings and corresponding reports for SPE, spanning from October 2014 to November 2022, were obtained from PMH and TMH, respectively; 25,661-41,014 samples from PMH constituted the training datasets. Trained models were subsequently evaluated with 24,238 samples from TMH. The classification model achieved an area under the receiver operating characteristic curve of 0.976 in the testing dataset, with an agreement rate of 93.8%. The fractionation model had mean and standard deviation difference from reported manual fractioning of -0.0884 to 0.155 g/L and 0.315 to 2.04 g/L, respectively, across the six serum protein bands. Peak quantification by the location plus quantification model correlated with manual quantification, with Spearman's r of 0.976. The machine learning models achieved near-human performances. They enabled high-throughput SPE analyses and interpretation and improved objectiveness and reproducibility of results.