Rapid identification of tumor patients with PG-SGA ≥ 4 based on machine learning: a prospective study

• Published in: BMC cancer Vol. 25; no. 1; pp. 902 - 12
• Main Authors: Qian, Gui, Jiaxin, Huang, minghua, Cong, beijia, Liu, Yinfeng, Li, Guiyu, Huang, Mingxue, Yang, Xiaoli, Tang, Hongyan, Yan
• Format: Journal Article
• Language: English
• Published: London BioMed Central 20.05.2025; BioMed Central Ltd; Springer Nature B.V; BMC
• Subjects: Accuracy; Adult; Aged; Algorithms; Automation; Biomedical and Life Sciences; Biomedicine; Body Mass Index; Cancer; Cancer Research; Cancer therapies; Care and treatment; Chemotherapy; Data collection; Datasets; Diagnosis; Efficiency; Fat-free body mass; Female; Generalized linear models; Hand Strength; Health Promotion and Disease Prevention; Hospital patients; Hospitals; Humans; Immobilization; Immunotherapy; Laboratories; Learning algorithms; Machine Learning; Male; Malnutrition; Malnutrition - diagnosis; Malnutrition - etiology; Medicine/Public Health; Methods; Middle Aged; Neoplasms - complications; Neoplasms - diagnosis; Nutrition; Nutrition Assessment; Nutritional Assessment; Nutritional status; Oncology; Optimization; Patient-Generated Subjective Global Assessment; Patients; Performance evaluation; Prevention; Product/Service Evaluations; Prognosis; Prospective Studies; Radiation therapy; Regression analysis; Risk factors; Sample size; Surgical Oncology; Tumors; China; Machine Learning; Patient-Generated Subjective Global Assessment; Cancer; Nutritional Assessment
• ISSN: 1471-2407; 1471-2407
• DOI: 10.1186/s12885-025-14222-9

Background Malnutrition is common in cancer patients and worsens treatment and prognosis. The Patient-Generated Subjective Global Assessment (PG-SGA) is the best tool to evaluate malnutrition, but it is complicated has limited its routine clinical use. Methods We reviewed 798 records from 416 cancer patients treated at our hospital from July 2022 to March 2024. We used machine learning methods like XGBoost and Random Forest to find important factors linked to PG-SGA scores of 4 or higher. We confirmed the most important factors with logistic regression analysis. Results Among all models, XGBoost and Random Forest models perform the best, with the area under the curve (AUC) reaching of 0.75 and 0.77. Multivariate logistic regression analysis identified body mass index (BMI) (OR = 0.82, 95%CI 0.66–0.99; P  = 0.045), handgrip strength (HGS) (OR = 0.89, 95%CI 0.82–0.96; P  = 0.004), fat-free mass index (FFMI) (OR = 1.36, 95%CI 1.01–1.88; P  = 0.045), and bedridden status (OR = 3.16, 95%CI 1.17–9.14; P  = 0.026) as key predictors for PG-SGA scores of ≥ 4. Conclusion BMI, HGS, FFMI, and bedridden status were identified as practical indicators to efficiently screen patients likely to have PG-SGA scores ≥ 4.