JMIR Publications

ABSTRACT

Background:

Breast cancer patients often experience health-related quality of life (HRQoL) impairments that remain difficult to predict on an individual level. Prediction models can aid to understand individual survivorship trajectories. However, current prognostic models are based on fixed intervals, limiting their utility in clinical follow-up schedules.

Objective:

This study aimed to develop and externally validate time-dynamic machine learning (ML) models that predict clinically relevant HRQoL impairments in non-metastatic breast cancer patients.

Methods:

Using the pooled multi-cohort EORTC BALANCE (big data in patients with breast cancer) dataset (N=6,316) containing repeated HRQoL measurements (EORTC QLQ-C30), we constructed over 70,000 patient assessment pairs. ML algorithms were trained using the earlier HRQoL assessment and clinical data to predict dichotomized impairments in QLQ-C30 domains at the later assessment between two weeks and five years ahead. The best performing model was determined via the Area Under the Receiver Operating Characteristic Curve (AUC) in the internal validation and externally validated in an independent cohort of the BALANCE dataset, in which the calibration and predictive performance in risk groups (patients: post menopause, with financial difficulties, with obesity, with 2 or more comorbidities, with lower educational status, with frailty) were also evaluated.

Results:

ML models showed good discrimination (AUC 0.64-0.84) across most domains, especially for persistent symptoms like fatigue, financial difficulties, or functioning scales. Gradient boosting models performed best, but tended to be overconfident, with poor calibration for low-prevalence symptoms like diarrhoea or constipation. Model performance varied by risk group (e.g., lower education, frailty), though no group consistently performed poorly. Performance remained stable across time windows, with prior HRQoL being the strongest predictor at the respective scale level, while clinical variables such as the type of treatment were less important for prediction.

Conclusions:

Time-dynamic ML models can support personalized HRQoL prediction in breast cancer care. Future improvements should focus on calibration and fairness to enable equitable, clinically meaningful implementation.