Predictive models for chemotherapy-related toxicity in adults with cancer: a systematic review

Chemotherapy-related toxicities remain a major barrier to optimal cancer treatment and can lead to dose reductions, delays, or discontinuation. Prediction models have been developed to estimate individualized toxicity risk, but the landscape of model types, predictors, validation practices, and reporting quality has not been comprehensively synthesized across cancers and toxicity endpoints. We performed a systematic review following PRISMA 2020 guidance. PubMed/MEDLINE, Scopus, Web of Science, and the Cochrane Library were searched from January 1, 1980 through August 2, 2025. Eligible studies included peer-reviewed English-language articles that developed or validated models predicting chemotherapy-related toxicity in adults with cancer. Data were extracted on study characteristics, cancer type, chemotherapy context, modeling approach, predictor domains, outcomes, validation strategy, and performance reporting. Risk of bias and applicability were assessed using a structured PROBAST-based framework. After deduplication, 587 unique records were screened and 166 full texts were assessed, yielding 126 included studies. Publication volume increased substantially after 2015 with broad geographic representation. Most studies were observational (retrospective 39.7%, prospective 34.9%), and most focused on model development with validation (67.5%). Logistic regression was the most frequently used method, and models predominantly relied on routinely available clinical predictors, including treatment exposure, demographics, laboratory values, and cancer characteristics. Internal validation was common (any internal validation 88.9%), whereas external validation was less frequent (any external validation 20.6%). Discrimination was typically reported (AUROC/C-statistic 81.0%), while calibration reporting was less common (any calibration 29.4%); Brier score was rarely reported (3.2%). PROBAST assessments indicated generally low applicability concerns overall (100.0%), but overall risk of bias was most often moderate (82.5%), driven largely by analysis-domain limitations. Chemotherapy toxicity prediction modeling is a rapidly expanding field dominated by regression-based methods and pragmatic clinical predictors. However, limited external validation and incomplete calibration reporting, together with analysis-domain risk of bias, were important barriers to clinical translation across the included literature. Future work should prioritize standardized predictor and outcome definitions, routine calibration assessment, and multi-site external validation.