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

Area of Science:

• Medical diagnostics
• Artificial intelligence in healthcare
• Laboratory medicine

Background:

• Serum protein electrophoresis (SPE) interpretation is complex and prone to variability.
• Sebia's capillary electrophoresis improves reproducibility but still requires expert analysis.
• Monoclonal immunoglobulins detection is crucial for diagnosing various conditions.

Purpose of the Study:

• To develop a multimodal deep learning model for interpreting Sebia-based SPE immunotyping.
• To integrate laboratory parameters with image data for enhanced classification.
• To assess the model's performance against human experts and validate its generalizability.

Main Methods:

• A deep learning framework using Sebia capillary immunotyping images and lab data (Creatinine, Calcium, LDH, ESR, RBC, Globulins).
• Multimodal classification model construction with hybrid sampling for data imbalance and attention mechanisms for interpretability.
• External validation on an independent cohort and comparison with physicians of varying experience levels.

Main Results:

• The multimodal model achieved high accuracy (0.975), Cohen's Kappa (0.955), F1-score (0.975), and recall (0.985).
• Performance significantly surpassed image-only and lab-only models.
• External validation confirmed generalizability, with AI performance comparable to senior experts.

Conclusions:

• The study introduces the first multimodal deep learning model for Sebia SPE immunotyping image interpretation.
• The model demonstrates diagnostic performance on par with senior experts.
• It shows potential as a clinical decision support tool in laboratory medicine due to its accuracy and generalizability.