Left atrial overload detection in ECG using frequency domain features with machine learning and deep learning algorithms

There are many studies conducted to identify left atrial overload (LAO) biomarkers using ECG however, low specificity remained to be a common problem. Therefore, in this study, we explored specific features and best AI models for a highly accurate diagnosis of LAO using the ECG signal, with a broad comparison from baseline to state-of-the-art methodologies. The frequency domain properties of the ECG signal were obtained by wavelet transform and wavelet scattering methods to detect LAO from the ECG signal. ECG data were obtained from the PTB-XL database. For feature extraction, 10 s ECG waveforms from 403 healthy and 352 LAO-diagnosed in-dividuals were used after carefully filtering. Each signal was then divided into 1.5 s epochs resulting in 4530 ECG signals. Then, four different machine learning (support vector machine, K-nearest neighbor, linear discriminant analysis, random forest) and three different deep learning algorithms (long-short term memory, 1-D convolu-tional neural network, 1D-CNN, and 2D convolutional neural network, 2D-CNN) were tested for the detection of LAO. Our results showed that frequency-domain features are much more capable of detecting LAO than time -domain features. Wavelet scattering features were superior to wavelet transform features such as Shannon en-tropy, variance, and energy, but we achieved the highest success rate with 2D-CNN and continuous wavelet transforms scalogram inputs (accuracy: 92%). Given its high success rate, 2D-CNN may assist clinicians by detecting the pathology with a low-cost and operator-independent method based on a short-term recording.