Deep learning model for diagnosing lupus erythematosus in cardiac patients using ECG and audio spectrograms.

Individuals with both Lupus Erythematosus and pre-existing heart conditions are more likely to develop severe symptoms, emphasizing the complex and not fully understood interaction between the disease and cardiovascular health. A universal diagnostic model based on fixed rules has proven ineffective, as demonstrated in the experimental section of this study. To address this challenge, we propose an efficient and novel approach. Our model consists of two complementary subsystems. The first leverages Residual Network (ResNet) to capture complex patterns within ECG datasets, capitalizing on its ability to identify complex patterns in sequential data. The captured features are subsequently processed through Long Short-Term Memory (LSTM) networks. The second subsystem takes an alternative approach, we introduce a novel pipeline that converts ECG images into audio, enabling Mel-spectrogram generation and deep analysis via a fine-tuned Audio Spectrogram Transformer (AST). This audio-based representation reveals richer temporal and spectral features, leading to more accurate and interpretable classification than traditional methods. Experimental findings indicate that our hybrid approach achieves exceptional performance, with accuracy, sensitivity, specificity, and AUC scores of 99%, 99.2%, 96.8%, and 97%, respectively. Furthermore, we validate our model's effectiveness through an explainable deep learning framework using a heatmap algorithm. The results suggest that Lupus Erythematosus may contribute to ventricular hypertrophy, as indicated by the model's emphasis on the QRS region in ECG images from the test dataset.