Prediction of early progression (EP) to CDKIs first line treatment in ER+/HER2- metastatic breast cancer (MBC) using machine learning.

Abstract

e13040 Background: The treatment of luminal MBC has undergone a substantial change with the use of cyclin dependent kinase 4/6 inhibitors (CDKIs). Nevertheless, there is not a clearly defined subgroup of patients who do not initially respond to CDKIs and show EP. Methods: MBC ER+/HER2- patients who have received at least one line of treatment were eligible. The event of interest was disease progression within 6 months of first line treatment according to the type of therapy administered. The first line treatments were categorized in chemotherapy (CT), hormonal therapy (HT), CT plus maintenance HT and HT plus CDKIs. Free text data from clinical visits registered in our Electronic Health Record were obtained until the date of first treatment in order to generate a feature vector composed of the word frequencies for each visit of every patient. Six different machine learning algorithms were evaluated to predict the event of interest and to obtain the risk of EP for every type of therapy. Area under the ROC curve (AUC), True Positive Rate (TPR) and True Negative Rate (TNR) were assessed using 10-fold cross validation. Results: 610 RE+/HER2- MBC treated between November 1991 and August 2019 were included. Median follow up for metastatic disease was 28 months. 17426 clinical visits were analyzed (per patient: range 1-173; median 30). 119 patients received CT as first line treatment, 311 HT, 117 CT plus maintenance HT and 63 HT plus CDKIs. There were 379 patients with disease progression, from which 126 were within 6 months from first line treatment (54 events with CT, 57 with HT, 4 with CT plus maintenance HT and 11 with HT plus CDKIs). The model that yields the best results was the GLMBoost algorithm: AUC 0.72 (95%CI 0.67-0.77), TPR 70.85% (95%CI 70.63%-71.06%), TNR 66.27% (95% 66.08%-66.46%). Conclusions: Our model based on unstructured data from real-world patients predicts EP and establishes the risk for each of the different types of treatment for MBC ER+/HER2-. Obviously an additional validation is needed, but a tool with these characteristics could help to select the best available treatment when that decision has to be made, avoiding those therapies that are probably not to be effective.