Prediction of Lymphoma Aggressiveness Using Machine Learning Algorithms.

Area of Science:

• Hematology
• Oncology
• Diagnostic Pathology

Background:

• Lymph node examination is crucial for diagnosing lymphoid neoplasms, metastases, and infections.
• Aggressive lymphomas, such as aggressive non-Hodgkin lymphomas (NHL), require urgent diagnosis.
• Integrating various diagnostic parameters can provide timely information for patient management.

Purpose of the Study:

• To develop and evaluate multivariable predictive models for identifying aggressive lymphomas.
• To assess the diagnostic value of combining lymph node cytology (LNC) and flow cytometry (FC) with other clinical parameters.
• To improve diagnostic efficiency in cases requiring urgent lymphoma assessment.

Main Methods:

• Retrospective analysis of 196 lymph node specimens.
• Inclusion of parameters: age, sex, LNC, FC, positron emission tomography scan, lymphocytosis, leukocytosis, lactate dehydrogenase (LDH), and hemoglobin.
• Construction of five multivariable models: three logistic regression models and two ensemble learning models (bagging and boosting).
• Performance evaluation using 10-fold cross-validation, including sensitivity, specificity, and AUC.

Main Results:

• Multivariable models demonstrated superior performance (AUCs 0.88–0.94) compared to individual variables (AUCs 0.69–0.87).
• The best-performing model (boosting) achieved a sensitivity of 77% and a specificity of 94%.
• Rapidly available parameters, including LNC and FC, are significant predictors of lymphoma aggressiveness.

Conclusions:

• Lymph node cytology, flow cytometry, and other readily available clinical data are associated with the aggressive nature of lymphomas.
• Multivariable models integrating these parameters provide valuable diagnostic insights.
• This approach facilitates prompt treatment initiation for aggressive lymphomas.