AI-supported assessment of Sebia serum immunofixation electrophoreses with a deep neural network

Objective: The manual assessment of serum immunofixation electrophoreses and non-automatic result transfer to laboratory information systems by laboratory specialists is a time-consuming and resource-intensive task. Additionally, the procedure can lead to variability in inter-rater agreement. To address these challenges, our objective was to develop an artificial intelligence-supported, digital assessment pathway for Sebia serum immunofixation electrophoreses.

AI solution: In cooperation with an Information Technology company, we developed a deep neural network using image-based analysis to assess serum immunofixation electrophoreses.

The system performed densitometric evaluations of individual lanes and provided a confidence score for each assessment. The solution included user interfaces for both laboratory technicians and specialists, with the capability for results to be transferred directly to the laboratory information system after human verification. A proof-of-concept experiment demonstrated promising results, including reduced turnaround times, an improved workflow, and positive user acceptance.

The system could classify results as: negative, IgA-kappa, IgA-lambda, IgG-kappa, IgG-lambda, IgM-kappa, IgM-lambda, biclonal or oligoclonal gammopathies, free light chain gammopathy and unclear. The solution was also designed to be adaptable for other applications, such as urine immunofixation electrophoresis, sodium dodecyl sulfate-polyacrylamide gel electrophoresis and isoelectric focusing.

Lessons learned: Before proceeding with further development, we consulted a specialized law firm to review the project against current European legislation. Under the European Union's Artificial Intelligence (AI) Act and the in vitro diagnostic medical devices regulation (IVDR), our proposed solution was classified as "high-risk" (AI Act) and "risk category C" (IVDR).

This classification would have mandated a conformity assessment by a notified body. The significant time and financial investment required for this assessment made a timely return on investment unfeasible, even when considering potential revenue from distributing the solution to several other laboratories. Consequently, the project was discontinued.

Our primary takeaway is that for similar projects to be financially viable in the future, they would likely need to be implemented at the scale of large national or international laboratory groups to justify the development and regulatory costs.