Abstract:Aim To investigate the predictive value of a backpropagation artificial neural network (BPNN) model based on novel inflammatory factors for the severity grading of coronary artery lesions in patients with acute ST-segment elevation myocardial infarction (STEMI). Methods A total of 234 patients with acute STEMI admitted to the Cardiac Center of First Affiliated Hospital of Xinjiang Medical University from January 2022 to October 2024 were enrolled as study subjects. They were randomly divided into a training set (164 cases) and a test set (70 cases) at a 7∶3 ratio.Based on the severity grading of coronary artery lesions (SYNTAX score), the training set was further classified into a low-risk group (0~22 points) with 54 cases (32.9%), a medium-risk group (23~32 points) with 62 cases (37.8%), and a high-risk group (≥33 points) with 48 cases (29.3%). General clinical data, serum inflammatory factors (high sensitivity C-reactive protein (hs-CRP), interleukin-6(IL-6), tumor necrosis factor-α(TNF-α), and novel inflammatory factors (neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), pan-immune inflammation value (PIV), systemic immune-inflammation index (SII) were compared among three groups of acute STEMI patients in the training set. Multivariate ordinal Logistic regression analysis was used to identify independent predictors of coronary artery lesion severity grading in acute STEMI patients. Random forest models and BPNN models were constructed, and the predictive performance of the models was evaluated using the area under the curve (AUC) from ROC curve analysis, as well as accuracy, sensitivity, and specificity calculated from the confusion matrix. Results Multivariate ordinal Logistic regression analysis showed that the triglyceride-glucose index (TyG), hs-CRP, and PIV were independent predictors of coronary artery lesion severity grading in acute STEMI patients (P＜0.05). Based on these three independent predictors, random forest models and BPNN models were constructed to predict the severity grading of coronary artery lesions in acute STEMI patients. The random forest model showed that when 2 654 decision trees were generated, the error rates for predicting high-risk, medium-risk, low-risk, and overall out-of-bag data stabilized at 15.3%, 19.2%, 30.5% and 21.4%, respectively, indicating stable accuracy of the model. Hyperparameter tuning of the BPNN model was performed using 5-fold cross-validation and grid search, and the optimal parameter combination was identified as follows:the network topology of the BPNN was 3-4-2-3, the maximum number of iterations was 107, and the learning rate was 0.5, at which the average accuracy reached a maximum of 84.3%. Finally, after 201 935 iterations of weight updates, the loss function reached a minimum value of 20.659 917. In the training set, the AUC of the BPNN model for predicting high-risk, medium-risk, and low-risk was increased by 6.5%, 8.9% and 8.3%, respectively, compared with the random forest model (all P＜0.05). In the test set, the AUC of the BPNN model for predicting high-risk, medium-risk, and low-risk was increased by 5.2%, 9.4% and 13.4%, respectively, compared with the random forest model (all P＜0.05). In both the training and test sets, the sensitivity, specificity, and accuracy of the BPNN model for predicting high-risk, medium-risk, and low-risk were higher than those of the random forest model. Conclusion Both the BPNN model and the random forest model constructed based on TyG, hs-CRP, and PIV exhibit good predictive efficacy for the severity grading of coronary artery lesions in acute STEMI patients, and the BPNN model outperforms the random forest model.
