Prediction of Lymphoma Aggressiveness Using Machine Learning Algorithms

Introduction: Lymph nodes are essential to diagnose lymphoid neoplasms, metastases, and infections. Some lymphomas, particularly aggressive non-Hodgkin lymphomas (NHL), need urgent diagnosis. Combining lymph node cytology (LNC) and flow cytometry (FC) with other rapidly available parameters through multivariable predictive models could offer valuable diagnostic information while waiting for anatomopathological results. Materials and Methods: Results of 196 lymph node specimens were retrospectively analyzed for parameters like age, sex, LNC, FC, positron emission tomography scan, lymphocytosis, leukocytosis, lactate dehydrogenase (LDH) levels, and hemoglobin. We constructed five multivariable models predicting the aggressive nature of lymphoma as defined by the anatomopathological diagnostic. The first three were logistic regression models based on two (model 1), four (model 2), and up to 16 independent variables (model 3). The last two models were based on ensemble learning algorithms, bagging (model 4) and boosting (model 5), respectively. The performance of these five models was compared after 10-fold cross-validation, evaluating metrics such as sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC). Results: Compared to individual variables associated with the aggressive nature of the lymphoma (AUCs from 0.69 to 0.87), the multivariable models achieved better AUCs, ranging from 0.88 to 0.94. The best model (model 5) achieved a sensitivity and a specificity of 77% and 94%, respectively. Conclusion: LNC, FC, and other rapidly available parameters are associated with the aggressive nature of the lymphomas. It is possible to combine them in multivariable models to obtain a valuable diagnostic information and to initiate a prompt treatment.