Supervised machine learning model for serum protein electrophoresis data interpretation.

Area of Science:

• Clinical Chemistry
• Artificial Intelligence in Medicine
• Bioinformatics

Background:

• Serum protein electrophoresis (SPE) is crucial for diagnosing paraproteinaemia but is time-consuming and prone to interobserver variability.
• Over 10,000 SPE tests are requested annually, highlighting the need for more efficient and objective interpretation methods.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for automated interpretation of digitised capillary electrophoresis (CE) tracings in SPE.
• To assess the performance of ML models in fractionation, classification, and quantification of serum proteins compared to expert pathologists.

Main Methods:

• Three artificial neural networks were trained using CE tracings and SPE reports from Princess Margaret Hospital (PMH).
• Models were evaluated on independent datasets from Tuen Mun Hospital (TMH), comprising over 24,000 samples.
• Performance metrics included area under the ROC curve, agreement rates, and correlation coefficients (Spearman's r).

Main Results:

• The classification model achieved an AUC of 0.976 and 93.8% agreement in the testing dataset.
• Fractionation models showed minimal differences from manual methods (mean difference -0.0884 to 0.155 g/L).
• Peak quantification demonstrated strong correlation with manual methods (Spearman's r = 0.976).

Conclusions:

• ML models achieve near-human performance in SPE interpretation, offering improved objectivity and reproducibility.
• These AI tools facilitate high-throughput analysis, potentially revolutionizing clinical laboratory workflows for paraproteinaemia detection.