To evaluate the application of decision tree method and Logistic regression in the prediction of acute myocardial infarction (AMI) events. The clinical data of 295 patients, who underwent coronary angiography due to angina or chest pain with unidentified causes in Zhejiang provincial People's Hospital during October 2018 and April 2019, were retrospectively analyzed. Fifty five patients were identified as AMI. Logistic regression and decision tree methods were performed to establish predictive models for the occurrence of AMI, respectively; and the models created by decision tree analysis were divided into Logistic regression-independent model (Tree 1) and Logistic regression-dependent model (Tree 2). The performance of Logistic regression and decision tree models were compared using the area under the receiver operating characteristic (ROC) curve. Logistic regression analysis showed that history of coronary artery disease, multi-vessel coronary artery disease, statin use and apolipoprotein (ApoA1) level were independent influencing factors of AMI events (all P<0.05). Logistic regression-independent decision tree model (Tree 1) showed that multi-vessel coronary artery disease was the root node, and history of coronary artery disease, ApoA1 level (the cutoff value:1.314 g/L) and anti-platelet drug use were descendant nodes. In Logistic regression-dependent decision tree model (Tree 2), multi-vessel coronary artery disease was still the root node, but only followed by two descendant nodes including history of coronary artery disease and ApoA1 level. The area under the curve (AUC) of ROC of Logistic regression model was 0.826, and AUCs of decision tree models were 0.765 and 0.726, respectively. AUC of Logistic regression model was significantly higher than that of Tree 2 (95% CI=0.041-0.145, Z=3.534, P<0.001), but was not higher than that of Tree 1 (95% CI=-0.014-0.121, Z=-1.173, P>0.05). The predictive value for AMI event was comparable between Logistic regression-independent decision tree model and Logistic regression model, implying the data mining methods are feasible and effective in AMI prevention and control.
Read full abstract