A clinical decision support framework for automated M-protein isotype classification from capillary zone electrophoresis immunotyping signals.

BACKGROUND

Machine learning applications in laboratory medicine rarely translate into clinical practice, largely due to the absence of clinical decision support (CDS) frameworks that address uncertainty and workflow integration. No existing system classifies monoclonal protein (M-protein) isotype directly from capillary electrophoresis signals or integrates classification within a deployment-ready CDS framework. We aimed to develop and externally validate such a framework.

METHODS

We developed a 3-level hierarchical cascade classifier for 9-class M-protein isotype classification from 6-channel capillary zone electrophoresis with immunotyping (CZE-IT) signals, integrated within a CDS framework comprising probability calibration, conformal prediction, confidence-based triaging, automated reflex test recommendations, and explainability. The system was trained on 2219 samples using 399 peak-based features with XGBoost and externally validated on 498 independent samples with frozen model weights.

RESULTS

The cascade achieved 9-class accuracy of 0.872 (95% CI 0.858-0.885), outperforming flat baselines by 13.8 percentage points in macro F1 (95% CI 7.8-19.2). External validation accuracy was 0.871. Conformal prediction achieved 95.9% empirical coverage. The confidence triaging system assigned 62.2% of samples to a high-confidence zone at 96.2% accuracy. External high-zone accuracy was 98.8%. At 10% clinical prevalence, negative predictive value was 0.981. A tuning-free cascade achieved an out-of-fold accuracy of 0.866, indicating negligible hyperparameter overfitting.

CONCLUSIONS

This study presents the first ML-based CDS framework for M-protein isotype classification from CZE-IT signals, demonstrating that uncertainty-aware design can identify 62% of samples as eligible for automation-assisted verification while routing lower-confidence cases to expert review through confidence-based triaging and per-sample explainability. The validated task is agreement with expert CZE-IT interpretation within a supervised laboratory workflow.