Statistical models of breast cancer tumour progression for mammography screening data

Abstract

In this thesis we propose a novel statistical natural history model and illustrate how it can be applied to epidemiological breast cancer screening data to increase knowledge about how breast cancers progress over time and how likely they are to be detected by both screening and by symptoms. The model may be useful in helping to design future individualised screening programmes for breast cancer.

In Study I a continuous tumour growth model for jointly estimating tumour growth, time to symptomatic detection and mammography screening sensitivity as a function of percentage mammographic density, PD, is presented. The model is applied to data extracted from Swedish postmenopausal breast cancer cases (the same study base is used in Studies I-III). PD is significantly associated with screening sensitivity. Growth rates are found to have a high individual-to-individual variability. In Study II the continuous tumour growth model is extended to allow for covariates in all submodels (tumour growth, symptomatic detection and screening sensitivity). A previously described positive association between body size and tumour size is found to be mainly caused by difficulties in symptomatic detectability/delay in visiting health care.

In Study III we compare the statistical powers of detecting image markers related to masking between the continuous tumour growth model and logistic regression using interval vs. screen-detected cancer as the dependent variable. Based on simulated data, we show that statistical power can be higher when tests are based on the continuous tumour growth model. Using observational data, we study an image marker of scatteredness of mammographically dense tissues in terms of screening sensitivity. PD did not include any additional information regarding sensitivity once SI’s role in sensitivity was accounted for. In Study IV, using our continuous tumour growth model framework, we derive individual (conditional) lead time distributions, based on a woman’s tumour size, screening history and percentage mammographic density. We propose a lead time bias correction that can be used in survival comparisons between e.g. screendetected and interval cases. In a simulation study, we explore the length-biased sampling. Results showed that the sampling should be viewed in the light of the tumour growth rate and the tumour size at which the tumour would have become symptomatically detected in absence of screening.