Predictive Value of a Machine Learning Model Based on CT and Clinical Features in Occult Peritoneal Metastasis of Locally Advanced Gastric Cancer
Objective To construct and evaluate a machine learning model using CT multimodal parameters to predict the status of occult peritoneal metastasis(OPM)in patients with locally advanced gastric cancer(LAGC).Methods Clinical information and imaging data from 310 LAGC patients were collected and randomly divided into a training group(217 cases,of which 201 were OPM-negative and 16 were OPM-positive)and a validation group(93 cases,of which 84 were OPM-negative and 9 were OPM-positive)in a 7∶3 ratio.Three models were constructed:a radiomics score model,a clinical fea-tures model,and a combined model.Model performance was evaluated using the area under the receiver operating characteristic curve(AUC).DeLong test and the integrated discriminant improvement index(IDI)were used to assess model generalization performance and diagnostic efficacy.Results Significant features,including geometric feature GeoFd2 and gray-level histogram feature Mean,were identified through multivariate logistic regression analysis to calculate the radiomics score(Rad-score).The clinical features model incorporated the serum tumor marker carbohydrate antigen(CA125),endoscopic tumor site and CT-N staging.The combined model integrated the radiomics score and the clinical features and was constructed using the Bayesian Network(BN)algorithm.In the training and validation groups,the com-bined model demonstrated superior predictive performance,significantly outperforming the radiomics score model(IDI:0.237 and 0.200,P＜0.05)and the clinical features model(IDI:0.177 and 0.278,P＜0.05)alone.Conclusion The joint BN model combining endoscopic tumor site,CA125,and imaging features can predict OPM status in LAGC patients with high predictive accuracy and reliability.
NETL
NSTL
万方数据
