Cancer Survival Prediction

Cancer survival prediction estimates outcomes such as overall survival, recurrence, and treatment response while accounting for censored follow-up, where the event time is unknown for some patients. Methods range from Cox models and random survival forests to deep learning on pathology slides, radiology, clinical records, and genomic or other molecular data; multimodal models use attention, graphs, prototypes, or mixture-of-experts to combine these sources. Much work evaluates risk ranking on TCGA cohorts and clinical benchmarks, while handling missing modalities, learning interpretable features, and testing generalization across institutions remain central challenges. Strong concordance does not ensure accurate survival probabilities, so calibration, external validation, fairness, and robustness to data shifts also matter for clinical use.