Development of machine learning models to predict cancer-related fatigue in Dutch breast cancer survivors up to 15 years after diagnosis

Abstract

Purpose: To prevent (chronic) cancer-related fatigue (CRF) after breast cancer, it is important to identify survivors at risk on time. In literature, factors related to CRF are identified, but not often linked to individual risks. Therefore, our aim was to predict individual risks for developing CRF. Methods: Two pre-existing datasets were used. The Nivel-Primary Care Database and the Netherlands Cancer Registry (NCR) formed the Primary Secondary Cancer Care Registry (PSCCR). NCR data with Patient Reported Outcomes Following Initial treatment and Long-term Evaluation of Survivorship (PROFILES) data resulted in the PSCCR-PROFILES dataset. Predictors were patient, tumor and treatment characteristics, and pre-diagnosis health. Fatigue was GP-reported (PSCCR) or patient-reported (PSCCR-PROFILES). Machine learning models were developed, and performances compared using the C-statistic. Results: In PSCCR, 2224/12813 (17%) experienced fatigue up to 7.6 ± 4.4 years after diagnosis. In PSCCR-PROFILES, 254 (65%) of 390 patients reported fatigue 3.4 ± 1.4 years after diagnosis. For both, models predicted fatigue poorly with best C-statistics of 0.561 ± 0.006 (PSCCR) and 0.669 ± 0.040 (PSCCR-PROFILES). Conclusion: Fatigue (GP-reported or patient-reported) could not be predicted accurately using available data of the PSCCR and PSCCR-PROFILES datasets. Implications for Cancer Survivors: CRF is a common but underreported problem after breast cancer. We aimed to develop a model that could identify individuals with a high risk of developing CRF, ideally to help them prevent (chronic) CRF. As our models had poor predictive abilities, they cannot be used for this purpose yet. Adding patient-reported data as predictor could lead to improved results. Until then, awareness for CRF stays crucial.