A Survival Analysis Approach to Predicting Severe Leukopenia in Advanced Breast Cancer Patients

Modern therapies for advanced breast cancer can significantly prolong survival and improve patient’s quality of life. Yet, in some patients they can be associated with different, sometimes severe toxicities. Predicting these toxicities in individual patients would enable more targeted monitoring, toxicity prevention and could help saving scarce healthcare resources. Previously we have shown that individual prediction of a cardio toxicity called QT prolongation using survival modelling algorithms is feasible. In this work we target the individual prediction of a blood toxicity called severe leukopenia using the same modelling approach. Seven statistical and machine learning survival modelling algorithms are trained, optimized, validated and compared on both clinical trial and real-world data. The statistical Cox proportional hazards model reached rather modest performance of about 0.6 (Uno’s concordance index) and slightly outperformed its machine learning counterparts. The strongest predictors include variables related to prior therapies, location of metastasis, histologic grade as well as several quality of life scores.