A Novel Network for Diagnosis of Heart Failure with High Dimensional SMOTE

Automated and effective detection of heart failure is crucial for timely interventions and managing cardiac conditions. However, current heart failure detection techniques encounter a significant challenge due to data imbalance, with a much higher prevalence of normal heartbeat samples compared to abnormal ones. Electrocardiogram (ECG) data, which captures not only temporal details but also spatial details, plays a pivotal role in diagnosing heart failure due to its ability to reflect the electrical activities of the heart comprehensively. This study introduces an innovative approach for detecting heart failure using ECG signals. We designed a neural network model that generates new minority class samples distinct from the original ones using high-dimensional feature space. This approach aims to rectify the prevalent data imbalance, thereby improving the model’s ability to detect abnormal heartbeats accurately. To tackle the classification of ECG signals, we devised a novel deep learning model based on residual structures and attention mechanisms to extract deep features of ECG. These features enable the model to focus on relevant signal patterns. Experimental validation was conducted using the MIMIC-IV clinical waveform repository. Our model demonstrates superior performance across evaluation metrics: achieving an AUROC of 0.823, a G-mean of 0.743, and a F-measure of 0.524, surpassing other imbalance data handling methods and deep learning classification models. These results indicate that our proposed method is a reliable tool for assisting in the diagnosis of heart failure.