Evaluating the performances of various machine-learning solutions to manage and utilize the interleukin-6 test...

This study aims to develop and evaluate machine-learning (ML) models that can predict interleukin-6 (IL-6) test outcomes and support the management of inappropriate IL-6 test requests.

This retrospective study included all IL-6 test results from the Health Science University, Ankara Education and Research Center, between January 1, 2022, and June 30, 2024. Patients without accompanying C-reactive protein (CRP) and hemogram results were excluded. A total of 4142 patient records were analyzed, categorizing patients into two groups based on an IL-6 cut-off value of 7 pg/mL. The dataset initially contained 22 variables, which were reduced after feature selection using Pearson correlation, variance inflation factor (VIF), Boruta, and recursive feature elimination (RFE) methods. In the end, nine features were identified as the most significant. CRP, hemoglobin (HGB), and neutrophil count (NEU) were the most critical variables. Six machine learning models were trained using standardized data and the Synthetic Minority Over-sampling Technique (SMOTE) to address the class imbalance. The models were evaluated based on different metrics, such as accuracy. Statistical analyses were performed using R and Python.

According to the IL-6 cut-off value, 17.5 % of patients were classified as negative and 82.5 % as positive in the dataset. Random Forest performed the most robustly, achieving an accuracy of 0.87 and a sensitivity of 0.87. Support Vector Machine and XGBoost models also performed well, with accuracies of 0.83 and 0.87, respectively, but with varying trade-offs in specificity and NPV.

The ML models developed based on low-cost and frequently requested first-line routine tests could predict IL-6 positivity or negativity and might be used to decrease the rate of inappropriate test requests.