Validation of death prediction after breast cancer relapses using joint models

Audrey Mauguen*, Bernard Rachet, Simone Mathoulin-Pélissier, Gill M. Lawrence, Sabine Siesling, Gaëtan MacGrogan, Alexandre Laurent, Virginie Rondeau

*Corresponding author for this work

    Research output: Contribution to journalArticleAcademicpeer-review

    5 Citations (Scopus)
    49 Downloads (Pure)

    Abstract

    Background: Cancer relapses may be useful to predict the risk of death. To take into account relapse information, the Landmark approach is popular. As an alternative, we propose the joint frailty model for a recurrent event and a terminal event to derive dynamic predictions of the risk of death. Methods: The proposed prediction settings can account for relapse history or not. In this work, predictions developed on a French hospital series of patients with breast cancer are externally validated on UK and Netherlands registry data. The performances in terms of prediction error and calibration are compared to those from a Landmark Cox model. Results: The error of prediction was reduced when relapse information was taken into account. The prediction was well-calibrated, although it was developed and validated on very different populations. Joint modelling and Landmark approaches had similar performances. Conclusions: When predicting the risk of death, accounting for relapses led to better prediction performance. Joint modelling appeared to be suitable for such prediction. Performance was similar to the landmark Cox model, while directly quantifying the correlation between relapses and death.

    Original languageEnglish
    Article number27
    JournalBMC medical research methodology
    Volume15
    Issue number1
    DOIs
    Publication statusPublished - 1 Apr 2015

    Keywords

    • Breast cancer
    • Joint frailty model
    • Landmark
    • Prediction
    • Relapse history
    • Survival

    Fingerprint Dive into the research topics of 'Validation of death prediction after breast cancer relapses using joint models'. Together they form a unique fingerprint.

    Cite this