The Construction Of Progression-free Survival Prediction Models For Patients With Colorectal Cancer Based On Real-world Data

Objective: Colorectal cancer is one of the main causes of tumor-related deaths in China and even in the world.And the reduced survival rate of patients with colorectal cancer is mainly related to the disease progression.Therefore,accurate prediction of treatment outcomes is helpful to develop individualized disease management and improve treatment benefits for patients.At present,machine learning models based on algorithm interpretation for progression-free survival of colorectal cancer are still lacking at home and abroad.Based on the real-world data of local patients with colorectal cancer,this thesis intends to use machine learning algorithms to construct localized progression-free survival prediction models for patients with colorectal cancer undergoing chemotherapy.Methods: Firstly,relevant diagnosis and treatment data of colorectal cancer patients hospitalized and received chemotherapy in Sichuan People’s Hospital from August 2018 to April 2021 were collected,and standardized real-world data sets were constructed by using data preprocessing methods.Then machine learning algorithm was used to construct a prediction model of progression-free survival related prognosis for patients with colorectal cancer undergoing chemotherapy from the perspectives of classification prediction model and survival analysis model respectively.The best model was optimized through the performance evaluation.Finally,SHAP algorithm was used to analyze the influence of each feature on outcome prediction,and Streamlit tool was used to complete the online deployment and visualization of the model.Results: A total of 569 patients’ medical data and 127 characteristics were included in this thesis,including demographic characteristics,oncology characteristics,imaging and test results of colorectal cancer patients.Fifteen important features,including chemotherapy completion,platelet count and carcinoembryonic antigen,were obtained by using Light GBM algorithm.Seven machine learning algorithms were applied to the data sets before and after feature screening to construct 14 prediction models for the progression of patients with colorectal cancer chemotherapy,among which the model based on XGBoost algorithm had the best performance(AUC=0.8158).After grid search hyperparameter optimization,the model performance was further improved(AUC=0.8796).Besides,three algorithms were used to construct a progression-free survival analysis model for patients with colorectal cancer undergoing chemotherapy with survival data set.The model based on GBDT algorithm had the best performance(Cindex=0.7035,IBS=0.1524).After feature simplification,the model performance was improved(C-index=0.7201,IBS=0.1495).Finally,the model was interpreted by SHAP,and the influence of various characteristics on the prediction results was expounded combined with the clinical practice,and the online deployment of the progression predicting model was completed.Conclusion: The progression-free survival prediction model for colorectal cancer patients established based on classification prediction model and survival analysis model had good performance,and the model interpretation was realized via SHAP.The methodological exploration of model construction carried out in this thesis would provide certain references for the development of tumor-related clinical aid decision-making system and the individualized treatment management of colorectal cancer.