Study On Machine Learning-based Colorectal Cancer Prognosis Model And Its Generalization

Colorectal cancer is one of the leading causes of tumor-related death worldwide,which has seriously threatened people's life and health and caused a huge disease burden,so it is of great significance to provide clinicians with tools for the prognosis analysis and treatment plan formulation of colorectal cancer.Data-driven prognostic model is considered as an important tool to improve the accuracy of prognostic prediction of colorectal cancer.In addition,the generalizability of the model,which means that the prognostic model obtained by learning from training data can perform well in the real clinical data in more medical institutions,is the premise of the application of the prognostic model in clinical practice.However,the current data-driven prognostic model research is still in the stage of model development,lacking of the model generalizability validation and clinical application.Therefore,it is of great significance to establish a more accurate prognostic model for colorectal cancer,improve the generalizability and clinical usefulness of the prognostic model,and enable it to be applied in more medical institutions and perform well to assist clinicians in formulating treatment plans based on the real clinical data and machine learning methods.This paper verified the time dependent and the nonlinear effects of prognostic factors in colorectal cancer,proposed a deep learning-based semi-supervised multitask survival analysis method to effectively improve the accuracy of the prognostic model,and used a semi-supervised logistic regression method to improve the generalizability of risk prediction model in order to promote its application in clinical practice.The main innovation points are as follows:The time-dependent and nonlinear effects of prognostic factors in non-metastatic colorectal cancer were discovered and visualized.A more accurate prognostic model for colorectal cancer was established,which predicted the prognostic risk of patients with non-metastatic colorectal cancer more accurately than the model that ignored the time-dependent and nonlinear effects.An effective decision-making support tool for the prognosis analysis and treatment plan formulation of colorectal cancer.was provided for clinicians.A deep learning-based semi-supervised multitask survival analysis method was proposed.Based on the deep neural network,the survival analysis problem was transformed into a multitask learning model composed of the semi-supervised learning problem with the survival probability prediction of multiple time sequence points.We proposed the semi-supervised loss and ranking loss to deal with the data censoring and non-increasing trend of survival probability in survival analysis,respectively.At the same time,a method for evaluating the importance of predictors was provided,and the ability of the model to capture the effects of predictors was visualized.An effective method based on deep learning for survival analysis of clinical complex structured data was provided.We used a semi-supervised learning method to improve the generalizability of risk prediction models by using unlabeled data.The semi-supervised logistic regression model was compared with supervised learning methods in terms of discrimination,calibration,generalizability,interpretability and clinical usefulness,which increases our current understanding of the generalizability of different models and provides a reference for predictive model development for clinical decision-making.The validity of this method is verified by external data,which provides technical support for constructing a prognostic model based on multi-center real clinical data.