Stable Survival Extrapolation via Transfer Learning

The mean survival is the key ingredient of the decision process in several applications, notably in health economic evaluations. It is defined as the area under the complete survival curve, thus necessitating extrapolation of the observed data. This may be achieved in a more stable manner by borrowing long term evidence from registry and demographic data. Such borrowing can be seen as an implicit bias-variance trade-off in unseen data. In this article we employ a Bayesian mortality model and transfer its projections in order to construct the baseline population that acts as an anchor of the survival model. We then propose extrapolation methods based on flexible parametric polyhazard models which can naturally accommodate diverse shapes, including non-proportional hazards and crossing survival curves, while typically maintaining a natural interpretation. We estimate the mean survival and related estimands in three cases, namely breast cancer, cardiac arrhythmia and advanced melanoma. Specifically, we evaluate the survival disadvantage of triple-negative breast cancer cases, the efficacy of combining immunotherapy with mRNA cancer therapeutics for melanoma treatment and the suitability of implantable cardioverter defibrilators for cardiac arrhythmia. The latter is conducted in a competing risks context illustrating how working on the cause-specific hazard alone minimizes potential instability. The results suggest that the proposed approach offers a flexible, interpretable and robust approach when survival extrapolation is required.

💡 Research Summary

The paper addresses the problem of estimating mean survival—a key metric in health‑economic evaluations—when the observed follow‑up period is insufficient to capture the full survival curve. Traditional extrapolation methods that simply fit naïve parametric models or rely on the restricted mean survival (RMS) often produce biased results, especially for diseases with long‑term outcomes. To improve stability and interpretability, the authors propose a two‑pronged approach that combines Bayesian transfer learning of external mortality data with flexible poly‑hazard survival models.

First, they construct a synthetic baseline population that mirrors the age‑sex composition of the clinical cohort. This is done by fitting a Lee‑Carter mortality model to historical death rates from the Human Mortality Database, then projecting these rates forward using a fully Bayesian framework (following Pedroza, 2006). The resulting projected mortality rates generate individual synthetic lifetimes for the external population, providing an up‑to‑date “anchor” for extrapolation.

Second, both the disease cohort and the external population are modeled with poly‑hazard processes, where the overall hazard is expressed as a sum of M component hazards (typically M = 3). Each component can follow Weibull, log‑normal, log‑logistic, or other flexible families, allowing the model to capture increasing, decreasing, bathtub‑shaped, or crossing hazards. The authors often impose structural relationships between the disease and external hazards—e.g., one component is proportional (C·h₁ᵖ) and another is identical (h₃ᵖ)—thereby borrowing strength from the external data while retaining disease‑specific features.

Likelihood contributions from the clinical and external datasets are multiplied to form a joint likelihood, and inference is performed in Stan via MCMC, yielding posterior distributions for all parameters. For extrapolation beyond the observed horizon, two strategies are offered: (1) “baseline” projection, which simply extends the disease hazard using the estimated parameters; and (2) “constant difference/ratio” projection, which assumes that the last k observed differences (or ratios) between disease and external hazards remain constant into the future, i.e., h_d(t) = h_p(t)+D or h_d(t) = R·h_p(t). Moreover, the framework accommodates cause‑specific hazards: the external hazard is decomposed into a disease‑specific component and an “other‑causes” component, the latter being shared across groups. This is particularly useful in competing‑risk settings, where only the disease‑specific hazard needs explicit modeling, enhancing long‑term stability.

The methodology is illustrated with three real‑world case studies.

1. Breast Cancer (METABRIC dataset) – The authors compare triple‑negative versus non‑triple‑negative patients. The poly‑hazard model captures crossing survival curves, and the estimated mean survival difference is about three years, highlighting the clinical disadvantage of triple‑negative disease.

2. Advanced Melanoma – Using published hazard ratios (HR ≈ 0.56) for the combination of an mRNA vaccine (V940) with pembrolizumab versus pembrolizumab alone, the authors construct a synthetic survival curve for the combination therapy. The extrapolated mean survival gain is roughly 1.8 years, and the 5‑year survival probability improves by about 12 percentage points.

3. Cardiac Arrhythmia – The benefit of implantable cardioverter‑defibrillators (ICD) versus anti‑arrhythmic drugs (AAD) is evaluated in a competing‑risk framework. External mortality provides the “other‑cause” hazard, while the arrhythmia‑specific hazard is modeled with a proportional relationship to the external population. The analysis yields an estimated life‑years‑gained (LYG) of 2.3 for ICD recipients, consistent with meta‑analytic hazard ratios (HR ≈ 0.5).

Across all examples, the proposed approach reduces extrapolation bias, yields realistic uncertainty intervals, and retains interpretability through identifiable hazard components (e.g., early disease effect vs. aging effect).

Limitations include reliance on the accuracy of the Lee‑Carter projections (which may be sensitive to long‑term mortality trends), the somewhat subjective choice of the number of hazard components and change‑points, and the computational burden of Bayesian MCMC for high‑dimensional models. The authors suggest future work on multi‑national mortality projections, dynamic Bayesian updating as new data arrive, and automated model‑selection criteria to streamline practical implementation.

In summary, by anchoring disease survival models to up‑to‑date external mortality projections and employing flexible poly‑hazard structures, the paper offers a robust, interpretable, and statistically sound framework for mean survival extrapolation in health‑economic and clinical decision‑making contexts.