Addressing Incomplete Data in Survival and Quality of Life Prediction: A Pancreatic Cancer Case Study

Models that predict the survival and quality of life (QoL) in patients with malignancies can support prognostic counselling and patient-centered evaluation of disease trajectories. Challenges in the development of such models include incomplete data in real-world registries, including missing baseline measurements and irregular, sparse availability of longitudinal QoL outcomes during treatment. These limitations complicate the incorporation of QoL into survival prediction and the reliable modeling of QoL over time. We address these limitations with a novel framework to predict survival and quality of life in the face of sparse data. The framework consists of a regularized missingness-avoiding random survival forest (MA-RSF) for survival prediction, and a conditional diffusion-based approach to predict QoL at various follow-up intervals. We apply the approach to a pancreatic cancer use-case, and evaluate it on a nationwide Dutch cohort. We find that our approach maintains high predictive accuracy (C-index 0.77) for survival prediction and minimizes reliance on sparse pre-treatment data, including QoL PROMs. In addition, we demonstrate that the QoL prediction expressed as the Static-Dynamic diffusion architecture reduces trajectory prediction error (RMSE) at key follow-up intervals compared to standard conditional generation (e.g. 12.9 vs 14.0 RMSE at 7 months). The proposed framework facilitates the prediction of survival and QoL outcomes in settings characterized by incomplete data and could enhance prognostic counseling in patients with malignancies, including pancreatic cancer.