A Balanced Bagging Classifier machine learning model-based web application to predict risk of febrile neutropenia in cancer patients

Background: Although several models exist to predict risk of febrile neutropenia in cancer patients, there is still need to more accurately quantify this risk to minimize morbidity of and mortality from this treatment toxicity. Material and methods: From previous reports of our group, un updated predictive model had emerged. We refined our algorithm even further by using Balanced Bagging Classifier (BBC) machine learning in the previous model derivation cohort, discarding all the missing data. Moreover, we made a web application to make it accessible for experimental clinical use. Results: We used clinical data from 3439 cycles of chemotherapy obtained from the periods of 2010-2011 and 2015-2019, with 133 episodes of febrile neutropenia observed (after 4% of chemotherapy cycles). BBC resulted in a more efficient model as reflected by an area under curve (AUC) of 0.97, accuracy of 0.95, sensitivity of 0.93, and specificity of 0.95. Permutation importance analysis revealed previous febrile neutropenia, cancer type and receipt of previous radiotherapy as the most important features for the BBC model. The web app that integrates the BBC model with a user-friendly user interface has been found to be clinically useful. Conclusions: Using machine learning with our previous data, we are now able to predict the risk of febrile neutropenia more effectively after chemotherapy in cancer patients. The resultant web application is functional and makes use of the developed machine learning model to predict febrile neutropenia.