Artificial intelligence as a precision tool for predicting risk of in-hospital death after aortic valve replacement

Abstract

Abstract Background/Introduction The prediction of risk of in-hospital death associated with cardiac surgery still has important gaps. In this scenario, the computational tools and mathematical techniques, the pillars of artificial intelligence, can represent an effective solution to this problem. Purpose To develop an in-hospital death prediction model for isolated aortic valve replacement (AVR) based on an artificial intelligence constituted by an artificial neural network (ANN). Methods 352 patients consecutively submitted to isolated AVR between 2010 and 2020 were included. Altogether, 30 baseline variables were evaluated. Initially, the Extra Tree Classifier machine learning algorithm was used to select the attributes with the highest association with death. With the application of the algorithm, it was possible to identify the 11 variables with the greatest weight associated with in-hospital death. After selecting the variables and dividing the dataset into training (70%) and testing (30%), a risk prediction model was structured through an ANN with multiple layers. The ReLU activation function was used in the hidden layers and the SoftMax activation function was used in the output layer. As an optimizing function of the ANN, the Nadam function was used. In addition, a thousand cycles of propagation and data return (Epochs) were performed to induce machine learning based on the cyclic adjustment of the weights of each of the independent variables included in the model. Accuracy assessments were performed using the ROC curve in the test dataset. The model was developed using the Python programming language. Results A predictive accuracy of 93,6% (AUC 0,936) was observed for the occurrence of in-hospital death in the test dataset to the ANN. When comparing the performance of traditional risk scores, also tested only in the test dataset, we found that the ANN-based model was significantly superior to the scores (EuroScore I = 84,0% (AUC 0,840); EuroScore II = 84,4% (AUC 0,844), STS Score = 74,0% (AUC 0,740). The area under the curve of the model based on the ANN was significantly higher when compared to the areas of the scores using the DeLong test (p<0.05). When applying the same model only to patients aged 75 and over, the results were as follows: ANN AUC 0,877; ES1 AUC 0,652; ES2 AUC 0,590; STS AUC 0,663 (p<0,05). Conclusion The application of artificial intelligence modelling is feasible for the creation of prediction models in the health area. In this study, the accuracy of the ANN was significantly higher than that of the other traditional risk scores in the general sample and for patients with more advanced age. These findings demonstrate the great potential that representative datasets have when accessed through artificial intelligence techniques. The demand for massive volumes of information is mitigated when well-structured datasets with extreme data quality is used. Funding Acknowledgement Type of funding sources: None.