| Popis: |
BACKGROUND In recent years, heart diseases cause more than 18 million deaths every year. Heart failure (HF) prediction is essential to slow disease progression by changing lifestyle and pharmacologic interventions before heart diseases occur. Various researches have been proposed recently to predict heart failure. However, these methods did not combine different data sources with high-dimensional for heart failure prediction. In addition, the existing methods failed to consider the coexisting risk factors for heart failure and the complex relationships among them. OBJECTIVE Our goal is to make early warning and prediction of heart failure, which can offer the opportunity to test and ultimately develop effective lifestyle and pharmacologic interventions. In this paper, both electronic medical records and physiological data are considered, so as to provide enough source information to identify valuable risk factors of heart failure and make HF prediction. METHODS In this paper, an early warning and prediction method for heart failure is proposed using deep learning and trend similarity measure approaches. First, we present the data fusion and feature extraction method to merge different sources of data and get several important risk factors, which contain relevant and valuable information for HF. Second, an ensemble deep learning model for HF prediction is proposed based on gradient algorithms and back propagation techniques. In addition, we present an anomaly detection method to eliminate abnormal data caused by mood changes or environmental factors. Finally, evaluated by the Haar wavelet decomposition strategy, a data sequence trend similarity measure method is proposed aiming at prediction and early warning of heart failure in massive medical data. RESULTS The proposed method is evaluated based on our real research project HeartCarer, which includes risk factor information and physiological data. We combine these datasets from 2015 to 2020 to make a better performance evaluation for the proposed deep learning model and similarity measure method. The combined dataset totally involves 2,976 HF patients, 18,203 family members closely related to patients, and 295,801 healthy people. By comparing with other state-of-the-art methods and our prior work in [2] (90%), the proposed method can obtain a higher accuracy of 98.5% in heart disease prediction. CONCLUSIONS Heart failure (HF) prediction is essential to slow disease progression by changing lifestyle and pharmacologic interventions before heart diseases occur. An early warning and prediction method for heart failure is proposed using deep learning and trend similarity measure approaches in this paper. The proposed method is evaluated based on our real research project HeartCarer and obtain a high accuracy in heart disease prediction. |
