心肾代谢综合征冠状动脉狭窄中西医结合智能评估模型OA

Objective To develop an optimized assessment model for coronary artery stenosis in patients with cardiovascular-kidney-metabolic(CKM)syndrome using the Tabular Prior-Data Fitted Network(TabPFN),thereby providing support for clinical decision-making.Methods Based on a cross-sectional study design,229 hospitalized patients were enrolled from two campuses of Jiangsu Province Hospital of Chinese Medicine,including 178 cases from the Main Campus and 51 cases from the Zidong Branch.Data from the Main Campus were randomly assigned to a training set(123 cases)and an internal validation set(55 cases)at a 7∶3 ratio,while data from the Zidong Branch served as the external validation set(51 cases).Risk factors were screened through integrated regression analyses,and models were constructed using the TabPFN and eight conventional machine learning techniques.Performance metrics,calibration curves,decision curves,and learning curves were compared in a multidimensional manner to conduct model selection,validation,and generalizability assessment.Finally,a local Shiny calculator was developed.Results Through stepwise regression and ensemble least absolute shrinkage and selection operator(LASSO)regression,a total of 8 risk factors were identified,including Chinese Medicine pathogenic elements(stagnation,phlegm,stasis,deficiency),coronary artery computed tomography angiography(CTA)findings,duration of diabetes,lymphocyte count,and low-density lipoprotein cholesterol(LDL-C).After multi-dimensional comparisons,TabPFN demonstrated superior comprehensive performance in small-sample settings compared to traditional models,and was selected as the optimal model.The model achieved area under the receiver operating characteristic curve(AUC)of 0.994(95％CI:0.984～1.000)in the training set,0.933(95％CI:0.862～1.000)in internal validation,and 0.902(95％CI:0.820～0.984)in external validation.Calibration curves indicated high predictive consistency,decision curve analysis confirmed clinical utility across all threshold probabilities,and the learning curve suggested strong feature learning and generalization capabilities.Compared with assessment based on CTA alone,the optimal model reduced misdiagnosis rate by 28.0％and missed diagnoses rate by 14.5％.Permutation feature importance analysis ranked the predictors in descending order of importance as follows:stasis,CTAfindings,lymphocyte count,LDL-C,stagnation,deficiency,duration of diabetes,phlegm.Conclusions TabPFN exhibits promising potential for small-sample medical data analysis.The developed Chinese-Western medicine multimodal assessment model demonstrates robust efficacy,offering an optimized assessment tool for coronary stenosis in CKM patients.