C77-13 Construction of a Risk Prediction Model for Serious Infection Following Immune Checkpoint Blockade Therapy in Lung Cancer Patients Based on the Machine Learning Algorithm

Rationale Immune checkpoint inhibitors (ICIs) therapy has revolutionized anti-cancer therapy, with lung cancer exhibiting sustained clinical responses to it. However, infectious complications have emerged as a growing concern in this population, highlighting the need for effective risk prediction tools. This study aimed to develop and validate a machine learning (ML)-based model for predicting severe infections in lung cancer patients treated with immune checkpoint inhibitors (ICIs) to enhance clinical applicability. Methods A retrospective cohort of lung cancer patients receiving ICIs at Zhongshan Hospital, Fudan University, was analyzed. Baseline demographic, clinical, and laboratory data were collected from electronic medical records. The dataset was divided into training (80%) and testing (20%) cohorts. To address class imbalance, the Synthetic Minority Oversampling Technique for Nominal and Continuous variables (SMOTENC) was applied. Recursive Feature Elimination (RFE) was used for dimensionality reduction and feature selection. Seven key predictors were retained based on clinical relevance and model performance, including albumin, tumor necrosis factor-alpha (TNF-α), interleukin-8 (IL-8), lymphocyte count, CD4/CD8 ratio, corticosteroid use, and COPD. Nine ML algorithms were trained and evaluated by 5-fold cross-validation, including Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Naïve Bayes (NB), Extreme Gradient Boosting (XGB), Stochastic Gradient Boosting Tree (SGBT), and Neural Network (NNET). Model performance was assessed by accuracy, sensitivity, specificity, F1 score, negative predictive value, positive predictive value and area under the receiver operating characteristic curve (AUC).Results Among all models, the Random Forest (RF) algorithm showed the best overall performance, achieving an AUC of 0.86 (95% CI (0.83-0.89)), a sensitivity of 0.85, and a negative predictive value of 0.80, indicating a strong ability to identify high-risk patients while minimizing missed infections. The impact of the top seven ranked features on model prediction outcomes is illustrated in Figure 1. Based on these influential variables, a simplified linear risk score model was developed to enhance interpretability and facilitate clinical application. Conclusion The ML-based infection risk prediction model, particularly the RF algorithm, effectively identified lung cancer patients at high risk of severe infections following ICIs therapy. Integrating immune and clinical features may improve early infection surveillance and guide preventive strategies in clinical practice. This abstract is funded by: None