An efficient neural network-based method for patient-specific information involved arrhythmia detection

As researches on computer-aided arrhythmia detection deepen, the application in clinical practice is still challenging due to weak generalization ability. The utilization of the existing prior patient knowledge can be an effective approach to address the issue and to ultimately architect a network appropriate for clinical use. An inverted residual block-embedded deep neural network (IRBEDNN) is proposed to accurately detect arrhythmias based on processed ECGs. Firstly, continuous ECGs are segmented into individual heartbeats based on R peak information. To simulate the real clinical scenario, these heartbeats are converted into 2-dimensional heartbeat images without any signal processing. Then, selected heartbeats are fed into the proposed IRBEDNN which combines CNN and inverted residual block (IRB) to extract implicit features. Also, patient-specific knowledge is exploited in the network to enhance arrhythmia detection. Finally, 24 data records from the MIT-BIH arrhythmia database are utilized to validate the method’s effectiveness and superiority. The effect on utilizing different duration of patient-specific information to the final experimental results is investigated and analyzed in detail, which demonstrates the effectiveness of using patient-specific information. On the 24 data-segment tests, the highest classification accuracy could reach 100%, while the overall ACC is 96.326%, higher than that of the existing comparison models. The precision of the S class can reach 0.816, also higher than comparison methods. After the boosting method, the recall reaches 0.852 for the S class and 0.921 for the V class, which outperforms the other comparison methods on average. Ablation experiments is conducted to investigate the performance of the IRBEDNN. Results show that the proposed IRBEDNN-based method achieves good generalization ability with promising accuracy on both the MIT-BIH arrhythmia database and the INCARTDB. Therefore, the proposed IRBEDNN could accurately and efficiently detect arrhythmias and has positive significance to clinical applications.