Attention-based deep learning for immunoglobulin typing from electrophoresis and laboratory data

Interpretation of immunotyping results from serum protein electrophoresis (SPE) remains labor-intensive and subject to inter-observer variability. While capillary electrophoresis systems such as those provided by Sebia offer improved reproducibility, yet expert interpretation is still required for detecting monoclonal immunoglobulins.

We developed a deep learning framework based on image data generated from Sebia's capillary immunotyping system, and incorporated clinically relevant laboratory parameters-Creatinine (CREA), Calcium (Ca), Lactate Dehydrogenase (LDH), Erythrocyte Sedimentation Rate (ESR), Red Blood Cell count (RBC), and Globulins (GLO)-to construct a multimodal classification model. Sample imbalance was addressed via hybrid sampling, and attention mechanisms were introduced to enhance model interpretability. External validation was performed using 200 cases from an independent cohort. In addition, model performance was compared to that of physicians across three experience levels.

The multimodal model achieved an overall accuracy of 0.975 on a balanced internal validation set, with a Cohen's Kappa score of 0.955, F1-score of 0.975, and recall of 0.985, substantially outperforming both the image-only and laboratory-only models across all metrics. External validation confirmed generalizability. Comparisons between human readers and AI demonstrated performance comparable to that of senior experts.

This study presents the first multimodal deep learning model designed for interpreting Sebia-based SPE immunotyping images. With diagnostic performance comparable to senior experts and robust external generalization, the model offers significant potential for clinical decision support in laboratory medicine.