Leveraging Kappa-Lambda Signatures in a Multistage Machine Learning Pipeline for B-Cell Lymphoma Detection by Flow Cytometry

Flow cytometry immunophenotyping is essential for diagnosing B-cell lymphomas, but manual interpretation of high-dimensional data remains subjective, time-consuming, and prone to interoperator variability. Previous computational approaches often overlook clinically relevant principles, such as Ig light chain restriction. To address this gap, a biologically informed, three-stage machine learning pipeline that integrates Ig κ (IGK) and Ig λ (IGL) signatures to improve B-cell lymphoma detection was developed. A total of 200 peripheral blood samples (100 normal, 100 abnormal) were analyzed, comprising >15 million single-cell events characterized by 21 immunophenotypic markers. Three XGBoost models were trained sequentially: the first classified light chain expression (IGK, IGL, or nuisance), the second identified cell phenotypes using marker intensities and IGK/IGL-based neighborhood enrichment, and the third produced sample-level predictions based on aggregated cell features. The IGK/IGL classifier achieved 88.0% test accuracy [area under the receiver operating characteristic curve (AUC), 0.957], whereas the cell-level classification reached 92.9% accuracy (AUC, 0.983), with IGK/IGL enrichment as the most informative feature. Similarly, sample-level classification achieved 94.7% accuracy (AUC, 0.976), with improved performance when IGK/IGL enrichment was included. These findings demonstrate that incorporating biologically grounded features enhances both the accuracy and interpretability of automated flow cytometry analysis. This approach offers a scalable, reproducible, and clinically aligned alternative to the manual review of flow cytometry data for B-cell lymphomas.