Abstract

Acute Myocardial Infarction (AMI) is a vital public health concern, because it is the primary factor of death globally. Therefore, timely identification is crucial, especially in resource-constrained situations without centralized testing. (1) Background: Assessment, risk assessment, and treatment all depend on electrocardiograms (ECGs). ECG segments are artificially corrupted with various noise types (e.g., Gaussian noise, baseline wander) to create noisy training data.; (2) Methods: In this paper, signal denoising with an Optimized Variational Stacked Autoencoder (OVSAE) model which involves training a Neural Network (NN) to reconstruct clean ECGs from noisy versions, effectively learning to separate signal from noises and then decompressing the noise removed signals. OVSAE is introduced to adaptively remove noisy signals from ECG signals. VSAE learns hierarchical latent representations of clean signal structures through multiple nonlinear encoding layers. To get improved parameter initialization and optimum results, layer-wise prior training and modification are implemented. To enhance training stability and reconstruction accuracy, a layer-wise greedy pre-training strategy is adopted, followed by global fine-tuning of the entire network. Self-Attention Long Short-Term Memory (SALSTM) classifier is designed for AMI detection. To adaptively weight the significance of various temporal aspects and numerous ECG signals, the LSTM model incorporates a self-attention mechanism. A gating mechanism is added by LSTM, a gated network, to regulate the NN's information transfer; (3) Results: Clinical Parameters in Risk Stratification, and PTB-XL ECG diagnostic dataset includes of 18885 patients' 10-second clinical 12-lead ECGs, that of 21837.Furthermore, the results are quantified using measures like Mean Square Error (MSE), Mean Absolute Error (MAE), Structured Similarity Index (SSIM), and Peak Signal to Noise Ratio (PSNR). AMI results have been evaluated using metrics such as ERRor (ERR), ACCuracy (ACC), SPEcificity (SPE), and SENsitivity (SEN) against k-fold cross-validation.