B-153 Automation of serum protein immunofixation electrophoresis interpretation with artificial intelligence (AI)

Abstract Background Using machine learning and artificial intelligence (AI) techniques, we have developed a computer vision tool for the automated interpretation of electrophoresis gel scans. Serum protein electrophoresis (SPE) is a well-established routine laboratory diagnostic technique used to quantify different types of blood proteins, whose composition can be altered by various pathologies. Serum immunofixation electrophoresis (IFE), a type of SPE, can help determine disease type, severity, prognosis, and is useful for disease monitoring. Serum IFE is primarily indicated for patients suspected of having plasma cell dyscrasias/disorders, such as multiple myeloma, a type of blood cell cancer. Anti-sera is added into distinct gel lanes for identification of antibody heavy chains (IgG, IgA, IgM) and light chains (kappa and lambda) which correspond to specific monoclonal immunoglobulins and have clinical significance. Interpretation of IFE gels requires significant time and effort from highly skilled experts in the field of clinical pathology and/or laboratory medicine. Methods Over 4,000 archived serum IFE scan images from UofL Health have been collected and processed for model development, with 45% positive for monoclonal proteins and the remainder negative. Model parameters were tuned using 70% of the data for model training, with the remainder reserved as validation data. Additional data was collected as internal (N = 906) and external (N = 120) test data, both of which were excluded from the training process. Training, validation, and internal test scans were labeled according to routine clinical laboratory reports which were interpreted by board-certified clinical pathologists. External test images were labeled by consensus agreement among 3 independent visual interpreters. Results When predicting if a band corresponding to a monoclonal protein was present or not present in IFE gel scans, the model achieved a sensitivity of 0.96, specificity of 0.98, and AUROC of 0.99, on internal test data, and sensitivity of 0.89, specificity of 1.00, and AUROC of 0.98, on external test data, compared to routine visual interpretation. AUROC was 0.95 and 0.96 for identifying IgG heavy chains and 0.98 and 0.99 for kappa light chains within internal and external test data, respectively. IgM heavy chains had an AUROC of 0.95 and 0.73 within internal and external test data, respectively. Further, heavy chain and light chain combinations were evaluated for monoclonal protein identification. For instance, IgG-kappa monoclonal proteins were detected with a sensitivity of 1.0 in both internal and external test scans. The model pipeline is also practical to implement, with an estimated average computation time of 16 seconds per sample using personal computer class desktop hardware and less than 0.5 seconds per sample with professional workstation class hardware. Conclusion The computer vision tool is poised to be integrated with routine clinical laboratory workflows where it can assist the interpretive efforts of medical staff by triaging patients with potential plasma cell disorders.