Natural history models of breast cancer for screening and treatment evaluation

<jats:p>&lt;p dir="ltr"&gt;Natural history models provide a framework for describing latent disease pro- cesses that cannot be observed directly, such as tumour onset, growth, progres- sion, and detection. In this thesis, we focused primarily on biologically inspired continuous-growth models, in which tumour size evolves continuously over time and rates of growth vary between individuals through random effects. This frame- work was used to investigate breast cancer screening, subtype-specific tumour growth, interval cancers, metastatic progression, and the time-varying effect of hormonal treatment, and was also compared with conventional multi-state modelling approaches.&lt;/p&gt;&lt;p dir="ltr"&gt;&lt;b&gt;Paper I&lt;/b&gt; - We modelled the probability of symptomatic detection among women who regularly attend screening and used it to jointly model tumour size and mode of detection conditional on screening history. Simulation studies were used to confirm the theoretical results. We showed that the joint model improved pa- rameter identifiability compared with a model based on tumour size alone. The estimated mean asymptomatic time was 9.9 years, and the mean tumour volume doubling time was 226 days. The predicted proportions of interval cancers were 12.0%, 28.5%, and 42.5% for screening intervals of one, two, and three years, re- spectively, in line with proportions observed in real-world screening populations. The model provides a tool for understanding how tumour growth, screening sen- sitivity, and screening interval jointly influence the occurrence of interval cancers.&lt;/p&gt;&lt;p dir="ltr"&gt;&lt;b&gt;Paper II&lt;/b&gt; - We extended the continuous-growth model to estimate subtype- specific tumour growth and symptomatic detection processes by incorporating the interval cancer proportion developed in Paper I. The model was applied to data from 7,818 women with invasive breast cancer and was used to estimate mean tumour volume doubling times of 332 days for Luminal A-like, 203 days for Luminal B-like, and 126 days for triple-negative tumours. The corresponding mean asymptomatic times were 11.8, 7.6, and 4.5 years. The model reproduced the observed subtype-specific proportions of interval cancers. These findings suggest that the higher proportion of interval cancers among more aggressive subtypes can be largely explained by faster tumour growth and shorter asymptomatic pe- riods.&lt;/p&gt;&lt;p dir="ltr"&gt;&lt;b&gt;Paper III&lt;/b&gt; - We extended an existing continuous-growth natural history model of metastatic progression to incorporate the time-varying effect of adjuvant hor- monal treatment. The model was applied to data from 9,716 women with invasive oestrogen receptor-positive breast cancer, among whom 299 developed distant metastases during follow-up. We estimated/quantified how hormonal treatment slows the growth of latent distant metastases. For a symptomatically diagnosed patient with a 20 mm primary tumour, the predicted 10-year metastasis-free survival increased from 92.8% with five years of treatment to 96.1% with ten years. These findings confirm that extended hormonal treatment may provide greater benefit for patients with larger primary tumours.&lt;/p&gt;&lt;p dir="ltr"&gt;&lt;b&gt;Paper IV&lt;/b&gt; - We compared four natural history models for breast cancer screening: a three-state Markov model, a three-state semi-Markov model, a seven-state Markov model incorporating tumour size, and a continuous tumour growth model. The models were fitted to data from 65,532 women and evaluated using diag- noses observed during a subsequent 26-month period. The continuous-growth model provided the best overall balance between calibration, discrimination, and biological interpretability, predicting 275.7 compared to 286 observed cancers. The semi-Markov model showed similar discrimination but overpredicted screen- detected cancers, while the Markov models showed limited discrimination. These findings suggest that explicitly modelling tumour growth may provide a more realistic representation of the latent disease process than conventional multi-state approaches.&lt;/p&gt;&lt;h3 dir="ltr"&gt;List of scientific papers&lt;/h3&gt;&lt;p dir="ltr"&gt;I. &lt;b&gt;Orsini, L.&lt;/b&gt;, Czene, K., &amp; Humphreys, K. Random effects models of tumour growth for investigating interval breast cancer. Statistics in Medicine. 2024; 43(15): 2957-2971. &lt;a href="https://doi.org/10.1002/sim.10105" target="_blank" rel="noreferrer"&gt;https://doi.org/10.1002/sim.10105&lt;/a&gt;&lt;/p&gt;&lt;p dir="ltr"&gt;II. &lt;b&gt;Orsini, L.&lt;/b&gt;, Zhang, Y., Czene, K., &amp; Humphreys, K. Estimation of subtype-specific tumour growth rate distributions and interval cancer proportions in breast cancer using biology-inspired natural history models. International Journal of Cancer. [Accepted]&lt;/p&gt;&lt;p dir="ltr"&gt;III. &lt;b&gt;Orsini, L.&lt;/b&gt;, Gasparini, A., Czene, K., &amp; Humphreys, K. Time-Varying Hormonal Treatment and Metastasis-Free Survival Among ER+ Breast Cancer Patients: A Natural History Modelling Approach, Statistics in Medicine. 2026; 45, no. 8-9: e70504. &lt;a href="https://doi.org/10.1002/sim.70504" target="_blank" rel="noreferrer"&gt;https://doi.org/10.1002/sim.70504&lt;/a&gt;&lt;/p&gt;&lt;p dir="ltr"&gt;IV. &lt;b&gt;Orsini, L.&lt;/b&gt;, Vinattieri, V., Kapanidis, E., Gjesvik, J., Strandberg, R., &amp; Humphreys, K. A Comparison of Multi-State and Biologically-Inspired Natural History Mod- els for Breast Cancer Screening Cohorts. [Manuscript]&lt;/p&gt;</jats:p>