The University of Auckland
Browse

File(s) under embargo

Reason: Embargo for 1 year to allow time for publication of results.

8

month(s)

15

day(s)

until file(s) become available

Exploring potential contributors to improved breast cancer patient prognosis over time in Aotearoa New Zealand - BWoodhouse Master's dissertation

thesis
posted on 2024-02-15, 22:21 authored by Braden WoodhouseBraden Woodhouse

This is a dissertation submitted by Braden Woodhouse in partial fulfillment of the requirements for the degree of Master of Professional Studies in Data Science, The University of Auckland, 2023.

Background: In Aotearoa New Zealand (NZ), 1 in 9 women are expected to develop breast cancer during their lifetime. Te Rēhita Mate Ūtaetae (the Breast Cancer Foundation National Register) contains a wealth of clinical, pathological, and outcome data to study prognoses for women with breast cancer in NZ. An initial summary of the data in Te Rēhita Mate Ūtaetae (Te Rēhita) showed, using crude models, that breast cancer outcomes improved over time across all ethnicities. Apart from this recent report, no other study has explored outcomes over time for women with breast cancer in NZ.

Methods: The characteristics of Te Rēhita cohort were described over time to provide the foundation for understanding potential trends in the analysis. K-medians survival clustering was implemented in order to discretise diagnosis years into groups with similar survival. Smoothed hazard rate curves were used to investigate non-linearity and time-dependent relationships of covariates. Univariable and multivariable Cox regression was performed to analyse the relationships between the covariates and breast cancer-specific survival. To account for non-breast cancer related death, competing risk analysis was implemented across ethnicities. A predictive modelling approach considered death from breast cancer by 5 years as a binary classification problem, and feature importance and SHAP values were used to understand how the models made their predictions. Machine learning methods specialised for survival analysis (e.g. Random Survival Forest) were used to account for censoring of patients. Landmark analysis and piecewise Cox regression were used to assess a clinically-interpretable milestone for patients.

Key Findings: An inflection point for breast cancer-specific survival was identified in year of diagnosis at 2009, where women diagnosed before and after had different outcomes. Wāhine Māori had the best outcome in competing risk analysis compared to all other ethnicities. Disease stage and tumour grade were important risk factors for breast cancer outcomes. Despite its aggressiveness, women with triple negative breast cancer who survived 5-years after their breast cancer diagnosis had a decreased risk of distant metastases and breast cancer-specific death, compared to all women with triple negative breast cancer.

Conclusions: Breast cancer outcomes for women in NZ have improved over time. Disease stage and tumour grade were found to be important prognostic factors using Cox proportional hazards and machine learning approaches for survival analysis.

History

Publisher

University of Auckland

Usage metrics

    FMHS

    Licence

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC