When does artificial intelligence genuinely add value over tried-and-tested statistical methods? A systematic review comparing AI and machine learning (ML) models with logistic regression and Cox models in breast, colorectal and pancreatic cancer concludes that the advantage depends on the task and the data: AI stands out when extracting information from images and modelling non-linear relationships, but conventional regression remains competitive on large, well-structured datasets.
What was already known
Predictive models for risk, complications and survival are central to precision oncology. Logistic regression and Cox proportional hazards regression remain the most widely used, yet they fall short with non-linear interactions, high-dimensional imaging features or combined clinical and metabolic data. AI promises to fill those gaps, but the evidence on when it truly outperforms conventional models had been fragmented.
What this work adds
Studies published between January 2019 and March 2025 were searched in PubMed, Scopus and Web of Science. Two reviewers independently screened titles and abstracts, assessed full texts and applied the PROBAST tool for risk of bias. Sixty-five studies with 907,567 participants were synthesised, classified by cancer site, predictive task, model family, comparator, validation strategy, predictor modality and reporting of calibration or explainability.
Main results
The 65 studies covered breast (35), colorectal (21) and pancreatic (9) cancer. AI superiority was task- and data-dependent. CNN and U-Net models predominated in imaging and body-composition tasks; tree-based ensembles consistently outperformed logistic regression for tabular perioperative complication prediction; and Cox regression remained competitive and, in the largest pancreatic risk study, superior to XGBoost (C-index 0.802 vs 0.723). PROBAST analysis-domain risk was moderate in 54 of 65 studies (83%), driven by limited external validation, sparse calibration reporting (11 of 65) and few decision-curve analyses (7 of 65).
What it means
The practical message is that AI is not better by definition. It adds most when extracting image features or capturing non-linear relationships, but on large, tidy datasets conventional statistics may be preferable and easier to interpret. To reach the clinic, models need external validation, calibration assessment, decision-curve analysis and explainability, in line with the TRIPOD+AI and CLAIM standards.
Reference: Abidin NZ, Shariff NM, Zamri EN, Shamsuddin S. Beyond predictive performance: A systematic review and critical methodological appraisal of AI/ML and conventional modelling strategies in breast, colorectal, and pancreatic Cancer. Artificial intelligence in medicine, 2026. DOI: 10.1016/j.artmed.2026.103506
