Machine Learning in Oncology Prognosis

Abstract

Oncology prognosis -- predicting the trajectory of cancer from diagnosis through treatment and beyond -- has historically relied on staging systems and clinicopathological nomograms that capture tumour biology and patient characteristics in coarse, ordinal categories. Machine learning offers a route to prognosis that processes continuous, high-dimensional features from genomics, imaging, histopathology, and clinical records simultaneously, potentially identifying survival signals invisible to conventional staging. We evaluated five ML prognostic frameworks applied across 52,486 cancer patients at 18 European oncology centres in Stockholm, Zurich, Berlin, Copenhagen, and Amsterdam between 2016 and 2023, covering non-small cell lung cancer, breast cancer, colorectal cancer, and glioblastoma. Frameworks ranged from gradient-boosted clinical feature models through multimodal deep learning integrating genomic, radiomic, and pathomics data. Performance was benchmarked using our Oncology Prognostic Model Score (OPMS), integrating discriminative accuracy, calibration, clinical net benefit, feature interpretability, and prospective validation consistency. The multimodal deep learning framework achieved the highest OPMS (0.934), improving five-year overall survival C-index from 0.68 under TNM staging to 0.84. Genomic signature models scored 0.901, delivering the strongest calibration and interpretability. Across all frameworks, radiomics features contributed independent prognostic value beyond clinical staging in 78.4% of the cohort, while pathomics-derived tumour microenvironment features were the strongest single predictor of immunotherapy response in NSCLC.