Machine learning-based prediction models for malnutrition risk in patients with cancer: a scoping review

Background Machine learning (ML)-based prediction models have increasingly been developed to identify malnutrition risk among patients with cancer; however, their predictive performance, methodological quality, and clinical applicability have not been comprehensively mapped. Methods Following the JBI methodology for scoping reviews, we systematically searched PubMed, Embase, Web of Science, Scopus, the Cochrane Library, CNKI, WanFang Data, VIP Database, and Chinese Biomedical Literature Database (CBM) from inception to March 16, 2026. Studies that developed, validated, or compared ML-based prediction models for malnutrition or nutritional risk in adult patients with cancer were eligible. Data regarding study characteristics, predictors, algorithms, validation strategies, and model performance were extracted. Methodological quality was assessed using PROBAST+AI. Results Nineteen studies published between 2021 and 2026 were included. The models incorporated multidimensional predictors spanning demographic, anthropometric, laboratory, tumor-related, treatment-related, and functional domains. Reported AUC values ranged from 0.770 to 1.000. Among the 16 studies that compared ML approaches with logistic regression within the same dataset, 15 reported higher discrimination for at least one ML model. However, this finding was based on within-study comparisons and does not establish the overall superiority of ML over conventional regression. Calibration and clinical utility were inconsistently evaluated, and only three studies conducted external validation. Conclusion ML-based prediction models show potential for identifying malnutrition, nutritional risk, and related nutritional outcomes in patients with cancer. However, current evidence is insufficient to establish consistent superiority of ML over conventional regression or to support widespread clinical implementation. Limited external validation, heterogeneous nutritional outcome definitions, and incomplete evaluation of calibration and clinical utility remain important challenges. Future studies should prioritize standardized outcomes, rigorous external validation, and transparent reporting. Systematic review registration 10.17605/OSF.IO/2V3UA .