Can we develop prognostic models for estimating life expectancy following a cancer diagnosis? Evaluating model complexity, data requirements, and optimal modelling frameworks.
AI interpretation is pending for this paper.
Open original publication →What the AI sees
Not AI summarized yet.
Research significance
Pending deeper interpretation.
Source abstract
As the burden of global cancer diagnoses rise, there is growing importance to provide accessible survival statistics to patients. Although life expectancy (LE) is often used to assess economic benefit of new treatments, it is rarely provided to help patients comprehend their diagnosis as this metric often requires extrapolation. We sought to determine the best modelling framework for providing this extrapolation, the minimum follow-up required, and circumstances where reliable estimates can be obtained. We analysed United States cancer registration data collected via the Surveillance, Epidemiology and End Results (SEER) Program. 122,703 patients aged 18-89 diagnosed between 1988 and 1991 with breast, colorectal, lung, or stomach cancer at localised, regional, or distant stages were included (follow-up until 2021). Mortality risk factors included age, sex, and stage at diagnosis. All-cause mortality was modelled and extrapolated using all-cause, cause-specific, and relative survival frameworks across 2-, 3-, 5-, 10-, and 20-year follow-up. Flexible parametric models were built to assess the importance of model complexity (Simple Model(s): Main-effects, Complex Model(s): Main-effects, interactions and time-dependent-effects). Timescales for other-cause mortality within the cause-specific framework were also compared (Time-since-diagnosis, or Attained-age). For evaluation, patients were categorised into 30 risk groups (5 age-groups, 3 stages, and 2 sexes; females only for breast cancer). Using at least 10-years of follow-up, LE/30-year restricted mean survival time (RMST) can be predicted to within 1-year/10% of observed values, with at least 80% of covariate groups predicted to within this relatively small difference for breast, colorectal, and lung cancer patients of varied risks. The relative and cause-specific survival frameworks produced reasonable extrapolated estimates with attained age the recommended timescale for other-cause mortality in the cause-specific setting. Increasing model complexity improves accuracy, particularly among lower-risk patients, typically younger, localised individuals. While the study demonstrated reasonable estimates for 50% of stomach cancer patients primarily those at high risk, further research is required to calculate LE/30-year RMST for lower-risk stomach cancer patients.