Time-dynamic versus time-static prediction models for impaired health-related quality of life in breast cancer: A methodological comparison using international harmonized data from the EORTC BALANCE database

Fatigue and impaired physical functioning are common and often persistent issues in breast cancer, significantly impacting HRQOL. Existing prediction models target HRQoL at fixed timepoints (static horizon), failing to reflect the temporal variability of a patient’s HRQoL trajectory. A time-dynamic approach incorporating repeated HrQOL assessments may increase clinical usability, yet it remains unclear if they can match or outperform time-static models. This study develops and compares time-dynamic and time-static regression and ensemble models (machine learning algorithms that combine multiple models) to predict impaired HRQoL in breast cancer patients. Data from the international multi-cohort EORTC BALANCE dataset ( N = 5,299) with repeated EORTC QLQ-C30 assessments (2001–2024) of female early breast cancer patients were used. Time-dynamic and -static training sets were constructed to predict impaired Physical functioning and clinically-relevant fatigue at 12 and 24 months. Sixteen logistic regression and stacked ensemble models were developed and internally validated. Calibration, discrimination (AUC), Brier score were assessed in a common test set. Model AUCs ranged from 0.703 to 0.785, indicating good discrimination. No significant differences were found between time-static and time-dynamic models or between logistic regression and ensemble algorithms. Calibration was good for all ensemble models and for time-static regression models, but poor for time-dynamic regression models. In conclusion, no statistically significant differences in predictive performance were detected between the time-static or time-dynamic, regression or ensemble models in patients with early breast cancer. The ensemble-based time-dynamic approach allows a more flexible application in the clinic by not using fixed follow-up intervals. This approach still requires evaluation in a clinical setting.