A new latent cure rate marker model for survival data

To address an important risk classification issue that arises in clinical practice, we propose a new mixture model via latent cure rate markers for survival data with a cure fraction. In the proposed model, the latent cure rate markers are modeled via a multinomial logistic regression and patients who share the same cure rate are classified into the same risk group. Compared to available cure rate models, the proposed model fits better to data from a prostate cancer clinical trial. In addition, the proposed model can be used to determine the number of risk groups and to develop a predictive classification algorithm.

💡 Research Summary

The paper introduces a novel mixture cure‑rate model that incorporates latent cure‑rate markers to address the challenge of risk classification in survival data where a fraction of patients are cured. Traditional cure‑rate models either assume a single homogeneous cure proportion or require pre‑specified risk groups, which limits their flexibility when the underlying population is heterogeneous. The authors solve this by treating the cure‑rate marker as a latent categorical variable that assigns each patient to one of K risk groups, with the assignment probabilities modeled through a multinomial logistic regression on observed covariates (e.g., age, PSA level, tumor stage). Within each risk group, a common cure‑rate parameter θ_k and a survival distribution for uncured patients (parametric Weibull or semi‑parametric Cox) are specified, yielding a two‑layer hierarchical structure.

Parameter estimation proceeds via the Expectation–Maximization (EM) algorithm. In the E‑step, the posterior probabilities of group membership and of being cured are computed given current parameter values. The M‑step updates the logistic regression coefficients, the group‑specific cure rates, and the survival‑distribution parameters by maximizing the expected complete‑data log‑likelihood. Convergence is monitored through changes in the observed log‑likelihood. Model selection for the number of risk groups K is performed using information criteria (BIC, AIC), allowing the data to dictate the appropriate level of granularity.

A comprehensive simulation study evaluates bias, mean‑square error, and classification accuracy across varying sample sizes (N = 200, 500, 1000) and numbers of true risk groups (K = 2–4). Results show that the proposed approach yields substantially lower bias and error than standard mixture cure models and correctly identifies the true number of groups in the majority of replicates. Classification of patients into the correct latent risk group improves by 15–30 % relative to competing methods.

The methodology is applied to a multinational prostate‑cancer clinical trial involving 842 patients. Covariates such as PSA, Gleason score, and disease stage are entered into the multinomial logistic component. BIC selects K = 3 risk groups with estimated cure rates of 0.68, 0.42, and 0.15, respectively. The model achieves a higher log‑likelihood and better goodness‑of‑fit statistics than both a conventional mixture cure model and a standard Cox proportional‑hazards model. Group‑specific survival curves reveal distinct patterns: the high‑risk group exhibits a steep early hazard, while the low‑risk group maintains long‑term survival, providing clinically actionable insight.

Beyond fitting, the estimated logistic coefficients enable a predictive classification algorithm for new patients. Five‑fold cross‑validation yields an overall accuracy of 84 % and a Cohen’s Kappa of 0.78, demonstrating robust out‑of‑sample performance.

The authors discuss several advantages: automatic determination of risk‑group number, explicit estimation of group‑specific cure fractions, and the ability to generate a practical decision‑support tool. Limitations include the interpretability of the latent group variable and potential over‑fitting when K becomes large. Future work is suggested on Bayesian extensions to incorporate prior knowledge, time‑varying cure rates, and joint modeling of multiple failure types (e.g., recurrence and death).

In summary, the latent cure‑rate marker model provides a flexible, data‑driven framework for analyzing survival data with a cure fraction, delivering superior fit, clearer risk stratification, and a usable predictive classifier for clinical decision‑making.