Methodological Quality and Health Equity of AI Prediction Models in Rectal Cancer

Abstract

Rectal cancer care has become a sequence of high-consequence, often irreversible decisions, and artificial intelligence (AI) prediction models have now been proposed as aids at nearly every decision point. Whether a patient is spared total mesorectal excision and routed to watch-and-wait organ preservation turns on predicting pathological complete response (pCR); whether occult nodal disease or extramural venous invasion (EMVI) is present, whether the tumor will recur, and whether surgery will leave a permanent stoma or disabling bowel dysfunction are each distinct prognostic problems with their own treatment and quality-of-life stakes. Across these tasks, radiomics, classical machine learning, and deep learning have all been advanced, and internally derived areas under the curve (AUC) routinely sit between 0.80 and 0.92. We conducted a structured, task-stratified appraisal of a documented sample of 41 primary rectal cancer prediction model studies and 10 task-level systematic reviews. Each was read against the Prediction model Risk Of Bias ASsessment Tool (PROBAST) [ 1 ], the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis: AI extension (TRIPOD + AI) [