Construction and validation of a model for predicting the risk of immune checkpoint inhibitor pneumonitis / 中国实用护理杂志

ABSTRACT

Objective:To construct and validate a risk prediction model for immune checkpoint inhibitor-associated pneumonia (CIP) using machine learning algorithms and the nomogram, aiming to provide an accurate and intuitive method to assist nurses in screening people at high risk of developing CIP.Methods:This was a retrospective case -control study. A total of 230 oncology patients treated with immune checkpoint inhibitors attending Zhujiang Hospital of Southern Medical University from January 2019 to February 2022 were collected using the hospital's electronic medical record system. The prediction models were built using five machine learning algorithms and nomogram. The models were then validated on a separate test set, and their differentiation and stability were assessed using evaluation indices like AUC and accuracy rate.Results:Underlying lung disease, smoking history, serum albumin≤35 g/L and radiotherapy history were identified as important influencing factors of CIP in all six models. The AUC of K nearest neighbor, support vetor machines (SVM), naive Bayesian, decision tree and random forest models predicted CIP were 0.647, 0.696, 0.930, 0.870, and 0.934, respectively. The AUC of the model created by the nomogram was 0.813, which was lower than the best random forest model in the machine learning algorithm, but with good predictive performance (AUC=0.934).Conclusions:The nomogram model can assess the patient′s risk more intuitively, but the risk prediction model of CIP based on a machine learning algorithm has a higher diagnostic value. It is suggested that the accuracy and usefulness of the prediction model can be increased by combining the nomogram's foundation with the machine learning algorithm.