F-Chain: Personalized overall survival prediction based on incremental adaptive indicators and multi-source clinical records

The abundance of biomarkers across histology, imaging, and clinical endpoints poses a challenge in selecting indicators for personalized clinical decision support. Patient heterogeneity necessitates an adaptive and incremental approach to indicator selection, leading to complex demands due to missing data. To address these challenges, we propose Forest Chain (F-Chain), a learning framework that incrementally selects prognostic indicators for each patient. Using a proposed surrogate preference function, F-Chain achieves consistent evaluations across multiple doctors and data sources. We introduce an indicator selection strategy that integrates data information, gradually adding relevant indicators. Additionally, we develop a missingness-incorporated decision tree for predicting outcomes on multi-source datasets with substantial missing values. We validate the F-Chain model using the SEER database and real clinical data from a hospital, demonstrating superior OS prediction results compared to state-of-the-art methods.