Predictive models for detecting patients more likely to develop acute myocardial infarctions

Acute myocardial infarction (AMI) is a major cause of death worldwide. In the USA, there are approximately 0.8 million persons suffering from AMI annually with a death rate of 27%. The risk factors of AMI include hypertension, family history, smoking habits, diabetes, serenity, obesity, cholesterol, alcoholism, coronary artery disease, and so forth. In this study, data acquired from a subset of the National Health Insurance Research Database (NHIRD) of Taiwan were used to develop a clinical decision support system (CDSS) to predict AMI. The integrated genetic algorithm and support vector machine (IGS) and deep neural network (DNN) were both applied to design the predictive models. A balanced dataset (6087 AMI patients and 6087 non-AMI patients) and an imbalanced dataset (6,087 AMI patients and 12,174 non-AMI patients) with each patient record including 74 features were retrieved to design the predictive models. Tenfold cross-validation was used to obtain the optimal model with best prediction performance during training. The experimental results showed that the CDSSs reached a prediction performance with accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) of 79.75–84.4%, 68.29–83.7%, 82.45–92.07%, and 0.8424–0.9089, respectively, for models designed based on the balanced dataset, as well as 81.86–86.27%, 52.65–81.22%, 84.29–96.47%, and 0.8503–0.9098, respectively, for models implemented based on the imbalanced dataset. The IGS and DNN algorithms and a combination of age, presence of related comorbidities, and other comorbidity-related features, including diagnosed age and annual physician visits of individual comorbidities, have been shown to be promising in designing strong predictive models in detecting patients who are more likely to develop AMI in the near future as well as for realizing mobile-health (m-Health) systems in managing their comorbidities to prevent occurrence of AMI events. Future work will focus on realizing an ensemble model by combining the model designed based on the long-term NHIRD dataset and the model based on the short-term EMR data and the real-time IoT sensor data, as well as implementing a transfer learning model by transferring the knowledge learned from the long-term model for training the short-term model, so that the predictive performance can be enhanced.