A comparison of performance of mathematical predictive methods for medical diagnosis: identifying acute cardiac ischemia among emergency department patients.

摘要: BACKGROUND There is increasing interest in mathematical methods for the prediction of medical outcomes. Three have attracted particular attention: logistic regression, classification trees (such as ID3 and CART), neural networks. To compare their relative performance, we used a large clinical database to develop models using these methods. METHODS Each modeling method was generate predictive instruments acute cardiac ischemia (which includes myocardial infarction unstable angina pectoris), prospectivel-collected data on 5773 patients, who presented over two year period six hospitals' emergency departments with chest pain or symptoms suggesting ischemia. This set then split into training (n = 3453) test 2320) sets. Of 200 available variables, restricted those within first 10 minutes department care (history, physical exam, electrocardiogram). RESULTS When number variables limited eight, representing practical input real-time setting, regression's receiver-operating characteristic (ROC) curve area, measure diagnostic 0.887; tree model's ROC area 0.858, network's 0.902. by model not limited, 0.905, 0.861, 0.923. Among networks had noticeably poorer calibration. outputs from each unrestricted were other an additional independent variable, areas new "hybrid" significantly better than original unlimited (ROC 0.858 0.920). CONCLUSIONS Logistic tree, network all can provide excellent performance outcomes decision aids policy models. Their ultimate limitations seem due availability information (a "data barrier") rather respective intrinsic properties. Choices between would be most appropriately based needs specific application, premise that any one intrinsically more powerful.