A 2D convolutional neural network using continuous wavelet transform scalogram inputs detected left atrial overload with 92% accuracy, outperforming other machine learning and LSTM models (83%).
Does a 2D-CNN model using frequency-domain features on ECG improve the detection accuracy of left atrial overload compared to other machine learning and deep learning models?
A 2D-CNN deep learning model using frequency-domain features from ECG signals can detect left atrial overload with 92% accuracy, offering a potential low-cost, operator-independent diagnostic tool.
Absolute Event Rate: 92% vs 83%
• Frequency-domain features on ECG were used for LAO/LAE detection. • Frequency-domain features on ECG give better results for the detection of LAE/LAO. • Four ML (SVM, KNN , RF, and LDA) and 3 DL (LSTM, 1D-CNN, and 2D-CNN) models were compared based on frequency-domain features on ECG. • 2D-CNN with CWT gives better results (accuracy: 92%) compared to ML and LSTM models (highest accuracy: 83%). 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 individuals 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 convolutional 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 entropy, 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.
Uslu et al. (Wed,) conducted a other in Left atrial overload (LAO) (n=755). 2D convolutional neural network (2D-CNN) with continuous wavelet transforms (CWT) vs. Machine learning (SVM, KNN, RF, LDA) and LSTM models was evaluated on Accuracy of left atrial overload detection. A 2D convolutional neural network using continuous wavelet transform scalogram inputs detected left atrial overload with 92% accuracy, outperforming other machine learning and LSTM models (83%).
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: