Prediction of lymphedema occurrence in patients with breast cancer using the optimized combination of ensemble learning algorithm and feature selection

Abstract

Background Breast cancer-related lymphedema is one of the most important complications that adversely affect patients' quality of life. Lymphedema can be managed if its risk factors are known and can be modified. This study aimed to select an appropriate model to predict the risk of lymphedema and determine the factors affecting lymphedema. Method This study was conducted on data of 970 breast cancer patients with lymphedema referred to a lymphedema clinic. This study was designed in two phases: developing an appropriate model to predict the risk of lymphedema and identifying the risk factors. The first phase included data preprocessing, optimizing feature selection for each base learner by the Genetic algorithm, optimizing the combined ensemble learning method, and estimating fitness function for evaluating an appropriate model. In the second phase, the influential variables were assessed and introduced based on the average number of variables in the output of the proposed algorithm. Result Once the sensitivity and accuracy of the algorithms were evaluated and compared, the Support Vector Machine algorithm showed the highest sensitivity and was found to be the superior model for predicting lymphedema. Meanwhile, the combined method had an accuracy coefficient of 91%. The extracted significant features in the proposed model were the number of lymph nodes to the number of removed lymph nodes ratio (68%), feeling of heaviness (67%), limited range of motion in the affected limb (65%), the number of the removed lymph nodes ( 64%), receiving radiotherapy (63%), misalignment of the dominant and the involved limb (62%), presence of fibrotic tissue (62%), type of surgery (62%), tingling sensation (62%), the number of the involved lymph nodes (61%), body mass index (61%), the number of chemotherapy sessions (60%), age (58%), limb injury (53%), chemotherapy regimen (53%), and occupation (50%). Conclusion Applying a combination of ensemble learning approach with the selected classification algorithms, feature selection, and optimization by Genetic algorithm, Lymphedema can be predicted with appropriate accuracy. Developing applications by effective variables to determine the risk of lymphedema can help lymphedema clinics choose the proper preventive and therapeutic method.

Lymphedema and Axillary Sentinel Lymph Node Biopsy (SLNB) in Node-Negative Patients with Breast Cancer

“…Isik et al reported in 2022, in a retrospective study of the BC database from 2013 to 2017, that patients with removal of more than 5 LNs were more likely to develop lymphedema (25). A 2022 study by Yaghoobi Notash et al, aiming to determine factors influencing lymphedema which was done on 970 patients with BC, the percentage of influencing variables was reported as follows: The number of LNs/the number of LNs isolated ratio (68%), the feeling of heaviness (67%), number of LNs isolated (64%), receiving radiotherapy (63%), surgery type (62%), number of involved LNs (61%), BMI (61%) (31). So both of these studies showed a significant association between the number of isolated LN-negatives and Lymphedema in patients same our study.…”

Comparison of machine learning models to predict the risk of breast cancer-related lymphedema among breast cancer survivors: a cross-sectional study in China

“…Notash et al. ( 37 ) used six classification algorithms including C5.0’s decision tree, KNN, SVM, LDA, Bayesian, and MLP to construct the BCRL prediction model, of which the SVM algorithm showed the highest sensitivity and was found to be the best model for predicting lymphedema, based on the accuracy obtained, the algorithm correctly detected the presence or absence of lymphedema in newly diagnosed patients in 88% of cases, which was slightly higher than the AUC of the LR in this study (0.87).Wei et al. ( 38 ) derived and evaluated six machine learning models, and the results showed that the LR model performed the best in the early detection of lymphedema with the best performance, AUC = 0.889 (0.840–0.938), sensitivity = 0.771, specificity = 0.883, accuracy = 0.825, and Brier score = 0.141, which is slightly lower than the sensitivity of LR (0.79) in this study.…”