Acute myocardial infarction (AMI) remains the leading cause of death and disability around the world (1, 2). Despite improvements in the pharmacological approaches to improve patients with AMI, their prognosis remains generally poor. Besides, with the global population age and the increase in life expectancy, diagnosis and prevent of AMI are more imperative than ever. Early diagnosis and interventional treatment of AMI significantly contribute to improving the prognosis of AMI patients (3). Currently, the evaluation of cardiac biomarkers, such as cardiac troponin I and cardiac troponin T, is considered one of the gold-standard tools for the diagnosis of AMI. However, the diagnosis of AMI based on these biomarkers is still unsatisfactory due to the limitations in sensitivity and specificity (4, 5).

In recent years, microarray technology has been widely used and improved. Rapid development in the field of bioinformatics provides novel methods for the prediction of AMI. Various biomarkers were identified using integrated bioinformatics, such as MMP9, TLR2, and ILB1, and they can be regarded as predictive and diagnostic tools for AMI (6, 7). Moreover, it has been reported that a multi-biomarker approach may significantly enhance the diagnostic accuracy for AMI (8). However, because of the complex genetic architecture, they may lack powerful ability. In such cases, poor accuracy, low efficiency, and lack of early screening prompt us to establish a novel predictive model for AMI. Huang et al. constructed an optimal prediction model for AMI using multiple machine-learning algorithms (area under the curve [AUC] = 0.794) (4). Obviously, this predictive model is not efficient enough for screening and early detection of AMI. Fortunately, the development of machine-learning techniques, such as random forest (RF) and artificial neural network (ANN), has been successfully applied in biomarker discovery (9, 10).

Thus, in our study, we combined the utilization of RF and ANN to construct a diagnostic model of AMI with microarray data from Gene Expression Omnibus (GEO) database. Firstly, we used the RF classifier to identify AMI-specific genes for classification, and the ANN algorithm was conducted to calculate the weights of each key gene. Then, we established a scoring model named neuralAMI with the cooperation of RF and ANN. Finally, we evaluated the classification performance of our model using the receiver operating characteristic (ROC) curve and validated this model with an independent dataset from GEO. The analysis process of our study is shown in Figure 1.

With the development of machine-learning algorithms and the availability of gene expression information in public databases, we have the opportunity to identify the biomarkers for the diagnosis or prognosis of the disease in various other fields (23, 24). In the field of AMI diagnosis, various attempts have been conducted to explore a diagnostic better method for AMI using diverse machine-learning algorithms (25, 26). In our study, we aimed to construct a diagnostic model for AMI based on the datasets from the GEO database. Firstly, we identified 11 AMI-specific DEGs via an RF classifier, and the ANN was used to calculate the weight of these genes. Then we developed the scoring model neuralAMI and constructed ROC to validate the accuracy and stability of the diagnostic model we established.

The enrichment analysis of DEGs indicates that most of the screened DEGs were mainly involved in the immune response. We found that neutrophil response-related biological processes. In practice, it has been reported that neutrophils are involved in the development of AMI, and we can improve AMI by suppressing the activation of neutrophil (27, 28). In addition, the “IL-17 signaling pathway” is the best-characterized pathway among the KEGG pathways. Recent studies have reported that IL-17 plays a role in the development of AMI (6, 29). Thus, the IL-17 signaling pathway may be a novel therapeutic target for AMI. There have also been reported that the IL-17 level in patients with AMI is markedly higher than in controls, which indicates that IL-17 might be a biomarker for AMI (30). The correct interpretation of DEG enrichment analysis results contributes to reveal the molecular mechanism of AMI and discover novel diagnostic predictors and therapeutic targets.

In this study, the top 11 key genes in DEGs were screened by the RF model based on MeanDecreaseGini. Furthermore, 10 of the 11 genes were also considered as AMI candidate genes in other studies: Acyl-CoA synthetase long-chain family member 1 (ACSL1) (31), nuclear factor interleukin-3-regulated (NFIL3) (32), interleukin-1 receptor-associated kinase 3 (IRAK3) (33), interleukin-1 receptor type 2 (IL1R2) (7), dysferlin (DYSF) (32), S100A12 (34), B-cell lymphoma 6 (BCL-6) (35), mast cell-expressed membrane protein 1 (MCEMP1) (36), serpin family A member 1 (SERPINA1) (37), and intelectin-1 (ITLN1) (38). Besides, we identified for the first time that zinc finger protein-36 (ZFP36) may get involve in the pathogenesis of AMI. It was reported that the expression of ZFP36 was low in the healthy aorta but was high in endothelial cells covered with atherosclerotic lesions in human (39).

The application of RF and ANN in constructing disease models has proven to be sophisticated (21). The highlighted novelty of our diagnostic model was firstly integrated RF and ANN algorithms to improve the predictive ability of the AMI predictive model. In addition, our scoring model was achieved by comprehensively considering the genes and their weight, which is vital to classify AMI and control samples. Gene weight is the actual value associated with each gene, indicating the importance of that gene in predicting the final output value. Besides, we removed the batch effects of training datasets and validation datasets by gene scoring, thus improving the predictive power of the model. As one of the most important machine-learning approaches, the advantages of RF include relatively good accuracy, precision, and ease of use, contributing to recognizing the key genes (40). The ANN algorithm has a good predictive ability, and the ANN model is stable and reliable (41, 42). It has been reported that the combined machine-learning methods of RF and ANN were efficient in many data-generating processes (43). As shown in our study, our model has a superior predictive capacity (AUC = 0.980) when compared to another model built by Chen et al. (AUC = 0.8550) (44). Moreover, the AUC of the predictive model achieved 0.900 in the validation set GSE61144, 0.882 in set GSE34198, and 1.00 in set GSE97320, which suggested that our model is of great applicability. The high AUC score in our predictive model indicates it could distinguish between AMI and control samples with a promising probability in microarray data.

However, there are still some limitations to our study. First, even though our predictive model is equipped with satisfactory performance in training and validation sets, the sample size in each dataset is relatively small. Thus, we combined 3 and 3 small-size datasets to get microarray training and validation sets. Second, the construction of the predictive model was based on the datasets from the GEO database, we should conduct in vitro and in vivo experiments to practice and verify our predictive model. Last but not least, because of the limitation of sample size, we did not conduct a fivefold cross-validation in ANN analyses. However, the expression data in our study are from peripheral blood and circulating endothelial cells, thus the predictive model can be applied to determine the likelihood of AMI through timely blood tests or tissue biopsy, which suggests our model has an outstanding classification performance. There is no doubt that our diagnostic model has a certain clinical value. Our predictive model needs to be investigated further in clinical work.

We established a novel predictive model for AMI based on machine-learning algorithms and a neuralAMI scoring formula that could be used to predict patients with AMI. Besides, we validated the model with an independent dataset from GEO. Our study provides clinicians with a novel diagnostic strategy that shows better prediction performance than using existing biomarkers.

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found in the article/Supplementary Material.

YW and WX conceived and designed the study. HC, LL, LZ, KD, TW, JP, XP, ZZ, and TJ collected the data and assisted in data analysis. TJ, HC, LL, and WX provided the significant suggestions on the methodology. YW conducted the data management and bioinformatics analysis. YW and HC drafted the manuscript. LL, LZ, KD, TW, JP, XP, ZZ, and TJ edited and revised the article. All authors read and approved the final manuscript.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.