A Predictive Model for Qualitative Evaluation of PG-SGA in Tumor Patients Through Machine Learning

Objective: Patient-Generated Subjective Global Assessment (PG-SGA) was a nutritional status assessment technique specifically tailored for patients with oncology. The goal of this study was to develop a machine learning (ML) prediction model for predicting PGSGA categorization of patients with tumor. Methods: From 2014 to 2020, patients at the First Hospital of Jilin University performed laboratory testing, bioelectrical impedance, physical measures, and the PG-SGA scale. A total of 8230 patients were involved in the study. Patients with missing or partial data were removed, leaving 7287 patients, of which 3743 were males and 3544 were females. ML was used to design a clinical prediction model for PG-SGA categories. Results: Through the least absolute shrinkage and selection operator (LASSO) and the correlation matrix, 135 variables were screened and 6 variables were retained; ML was performed among the remaining variables. The accuracy of neural network prediction models was 70.3% and 70.4% for males and females in the training cohort, respectively, and 74.4% and 73.2% for males and females in the validation cohort, respectively. The area under curve (AUC) of males was 0.87 for PG-SGA scores "0-3", 0.70 for PG-SGA scores "4- 8" and 0.74 for PG-SGA scores ">8". As for females, the AUC was 0.85 for PG-SGA scores "0-3", 0.65 for PG-SGA scores "4-8" and 0.76 for PG-SGA scores ">8". The results of confusion matrix showed that the models were of good predictive validity. The prediction model was nearly 90% accurate for predictions that do not require nutritional support. Conclusion: We demonstrated that neural network learning is the best clinical prediction model using ML. The model can work as a prediction for the PG-SGA classification of patients with cancer and can be promoted further in the clinic.