Abstract
Accurate diagnosis of B-cell chronic lymphoproliferative disorders (B-CLPDs) remains challenging due to overlapping phenotypes across subtypes. Machine learning (ML) offers promising tools to improve marker evaluation and refine flow cytometry analysis. We investigated the use of ML algorithms to evaluate the diagnostic value of incorporating CD148, CD180, and CD200 into standard B-CLPD phenotyping panel and to develop a diagnosis decision tree. We trained models with flow cytometry data from 480 patients with B-CLPDs using XGBoost and DecisionTree algorithms. The final models integrated 2 categorical markers (CD5 and CD10) and quantiles of fluorescence intensity of 4 quantitative markers (CD20, CD180, and CD200) to classify 6 B-CLPD subtypes. These trained models were applied to an independent cohort of 433 patients with B-CLPD analyzed on a different flow cytometer platform. DecisionTree models achieved the highest classification accuracy (mean accuracy, 0.88) in the validation cohort. The overall specificity ranged from 0.95 lymphoplasmacytic lymphoma (LPL) to 1 hairy cell leukemia (HCL), whereas sensitivity varied from 0.75 (LPL) to 1 (HCL). The DecisionTree model demonstrated superior identification of chronic lymphocytic leukemia compared to a Matutes score of 4 or 5 (P = .029). In more than half of the cases, a diagnosis was determined with near certainty using only the cytometry data. For the remaining cases, a hierarchical approach incorporating additional tests was proposed. For practical implementation, an interactive interface provides diagnostic predictions, positive predictive values, and Gini index scores. This study establishes a ML-optimized strategy for B-CLPD classification, combining phenotypic, cytogenetic, and molecular data to enhance diagnostic accuracy of leukemic B-CLPD cells. This trial was registered at www.ClinicalTrials.gov as #NCT04952974.