Age Dominates Machine Learning Survival Prediction in Haematological Malignancies: A Multi-Disease Analysis of 192,825 Patients Demonstrates Cox Regression Sufficiency

This study systematically compares machine learning (ML) methods with Cox proportional hazards regression for survival prediction across six haematological malignancies using SEER data (2000-2020). Key Findings: Age accounts for 75–97% of predictive signal across all diseases (mean 89.5%) ML offers no clinically meaningful advantage over Cox regression (mean C-index difference: 0.29 percentage points) Even with extensive hyperparameter optimization (130 configurations), the best ML model exceeded Cox by only 0.13 percentage points Random Survival Forest significantly underperformed Cox in 2 of 6 diseases and produced larger racial performance gaps in 4 of 6 diseases Diseases analyzed: Acute myeloid leukaemia (AML), chronic lymphocytic leukaemia (CLL), chronic myeloid leukaemia (CML), acute lymphoblastic leukaemia (ALL), multiple myeloma (MM), and diffuse large B-cell lymphoma (DLBCL). Implications: When predictor-outcome relationships are dominated by age, simpler models should be preferred for their interpretability, equity, and equivalent performance. These findings challenge enthusiasm for algorithmic complexity in oncology prognostication using routinely available clinical variables. Contents: Main manuscript with embedded figures and tables, supplementary materials (Tables S1–S8, Figures S1–S5), completed STROBE checklist, and README documentation.