On the redundancy of denoising in electrocardiogram reconstruction

Figures

• 

  Figure 1. Overview of the work done in this paper. Left (Blue): Single model pipeline (SMP) which takes in ECG with various noise levels and outputs denoised synthesised leads. Middle (Green): Transfer learning continuation of the SMP. It fine-tunes the generic model of the SMP to a patient-specific form to improve the performance of the model on each patient irrespective of heart condition. Right (Red): Multiple model pipeline which classifies the noise level of the ECG before transferring it to the model designed for that noise level for synthesising denoised leads.

• 

  Figure 2. (A) Model architecture used in the pipelines. Each block shows the type of layer while the arrow to the right shows the activation function and number of neurons. (B) Flow diagram for training and testing the pipelines, showing how various noise levels are added to the denoised/cleaned ECG signal prior to inputting it to the model for either model training or model testing. (C) Multiple Model Pipeline (MMP) showing how a classification algorithm is required to decide the noise level before passing it on to the relevant model (Model 1 to Model 5) for reconstruction. (D) Single Model Pipeline (SMP) in general use showing the input signal directly applied to the model irrespective of noise level.

• 

  Figure 3. Sample reconstruction of a randomly selected test record under input SNR conditions of 0 dB and 15 dB, where the black trace denotes the baseline denoised ECG signal prior to noise injection, the red trace denotes the reconstruction produced by the multiple model pipeline (MMP), the blue trace denotes the single model pipeline (SMP), the green trace denotes the SMP trained on an external dataset without adaptation, and the yellow trace denotes the personalised SMP following transfer learning on the external dataset; the left panel presents cross-validation comparisons between MMP and SMP. The far-left sub-panel showing overlaid original and reconstructed signals where increased overlap indicates higher reconstruction similarity, and the sub-panel next to it repeats the same signals with vertical offsets for clearer visualisation. The right panel illustrates the effect of transfer learning on SMP performance, where its left sub-panel corresponds to a rhythmic patient and the right sub-panel corresponds to an arrhythmic patient, comparing the generic SMP with its personalised variant. The reconstructed signals are vertically offset (0.2 mV and 0.4 mV) to improve morphological…

• 

  Figure 4. Figure showing the trend of the average correlation and average RMSE across all leads of the single model pipeline (SMP) and multiple model pipeline (MMP) with relation to the input SNR. The SMP is denoted in orange, while the MMP is denoted in blue. The white or grey vertical shade denotes the different input SNR segments specialised on by the different MMP models.

• 

  Figure 5. Trend of the average correlation across all leads for the fully trained single model pipeline (SMP) with relation to the input SNR. The SMP is evaluated on 10 normal and 10 myocardial infarction patient records from the PTB-XL database. Tests were conducted on the model in its generic form and in its fine-tuned (through transfer learning) form. In the legend, ‘Norm’ indicates evaluation on normal records, ‘Infarc’ denotes evaluation on myocardial infarction records, ‘Gen’ denotes evaluation with generic model, and ‘Tran’ denotes evaluation with fine-tuned model.

• 

  Figure 6. Trend of the average correlation and average RMSE across all leads for each patient after transfer learning with relation to the input SNR. To the left, a table of the time interval between the initial signal used to fine-tune the model and final signal used to test the model. “Normal” represents patients with sinus rhythm while “Infarction” represents patients with myocardial infarction.