PREDICTING MYOCARDIAL INFARCTION COMPLICATIONS AND OUTCOMES WITH DEEP LEARNING

Abstract

Early diagnosis of cardiovascular diseases, which have high mortality rates all over the world, can save many lives. Various clinical findings and past histories of patients play an important role in diagnosing these diseases. These days, the prediction of cardiovascular diseases has gained great importance in the medical field. Pathological studies are prone to misinterpretation because too many findings are studied. For this reason, many automatic models that work with machine learning methods on patients' findings have been proposed. In this study, a model that predicts twelve myocardial infarction complications based on clinical findings is proposed. The proposed model is a deep learning model with three hidden layers with dropouts and a skip connection. A binary accuracy metric is used for measuring the performance of the proposed method. Rectified Linear Unit is set to the hidden layers and sigmoid function to the output layer as an activation function. Experiments were performed on a real dataset with 1700 patient records and carried out on two main scenarios; training on original data and training on augmented data with 100 epochs. As a result of the experiments, a total accuracy rate of 92% was achieved which is the best accuracy rate that has been proposed on this dataset.Early diagnosis of cardiovascular diseases, which have high mortality rates all over the world, can save many lives. Various clinical findings and past histories of patients play an important role in diagnosing these diseases. These days, the prediction of cardiovascular diseases has gained great importance in the medical field. Pathological studies are prone to misinterpretation because too many findings are studied. For this reason, many automatic models that work with machine learning methods on patients' findings have been proposed. In this study, a model that predicts twelve myocardial infarction complications based on clinical findings is proposed. The proposed model is a deep learning model with three hidden layers with dropouts and a skip connection. A binary accuracy metric is used for measuring the performance of the proposed method. Rectified Linear Unit is set to the hidden layers and sigmoid function to the output layer as an activation function. Experiments were performed on a real dataset with 1700 patient records and carried out on two main scenarios; training on original data and training on augmented data with 100 epochs. As a result of the experiments, a total accuracy rate of 92% was achieved which is the best accuracy rate that has been proposed on this dataset.