What is the clinical evidence from this study?

Study design: Other. Population: Cardiovascular diseases (n=290). Intervention: LSTM model vs. CNN, MLP, CNN-MLP, and ConvLSTM models. Primary outcome: Accuracy.

What question did this study set out to answer?

The aim is to evaluate various deep learning models for ECG-based diagnosis and classification of cardiovascular diseases.

March 9, 2026Open Access

Comparative evaluation of deep learning models for cardiovascular disease diagnosis and classification

Key Result

The LSTM model achieved the best overall performance for cardiovascular disease diagnosis with an accuracy of 99.98%, an F1 score of 98%, and a precision of 100%.

Key Points

The aim is to evaluate various deep learning models for ECG-based diagnosis and classification of cardiovascular diseases.
Compared multiple AI-driven deep learning models including CNN, LSTM, and MLP on the PTB-ECG dataset.
Used discrete wavelet transform for feature extraction and addressed data imbalance to enhance model performance.
Evaluated models based on diagnostic accuracy and computational complexity, measured in Real Multiplications Per Symbol.
The LSTM model achieved the highest accuracy of 99.98% and an F1 score of 98%.
Precision was recorded at 100%, indicating flawless positive predictions.
Models demonstrated that high diagnostic accuracy can coexist with low computational complexity.

Structured PICO

Population

549 ECG records from 290 subjects aged 17 to 87 years from the publicly available PTB-ECG dataset, encompassing various cardiovascular diseases including myocardial infarction, bundle branch block, heart failure, dysrhythmia, valvular heart disease, myocardial hypertrophy, and myocarditis.

Intervention

Low-complexity Artificial Intelligence (AI)-driven Deep Learning (DL) models (CNN, LSTM, MLP, CNN-MLP, and ConvLSTM) using Discrete Wavelet Transform (DWT)-based feature extraction.

Comparator

Comparison among the different deep learning architectures.

Outcome

Diagnostic performance (accuracy, precision, recall, F1 score) and computational complexity measured in Real Multiplications Per Symbol (RMpS).

An LSTM-based deep learning model can achieve near-perfect accuracy (99.98%) in classifying cardiovascular diseases from ECG signals while maintaining low computational complexity suitable for real-time clinical deployment.

Abstract

Cardiovascular diseases (CVDs) remain the leading cause of mortality worldwide, highlighting the need for accurate and computationally efficient diagnostic tools. Electrocardiography (ECG) is widely used for the diagnosis of CVD and is a noninvasive, cost-effective method. However, manual interpretation requires substantial time and can lead to human error. In this study, we propose and systematically compare low-complexity Artificial Intelligence (AI)–driven Deep Learning (DL) models for multi-class ECG-based CVD diagnosis and classification. Using the publicly available PTB Diagnostic ECG (PTB-ECG) dataset, which comprises multiple CVD categories, we evaluate CNN, LSTM, MLP, CNN-MLP, and ConvLSTM architectures, focusing on balancing diagnostic performance and computational complexity. Discrete Wavelet Transform (DWT)–based feature extraction and data-level imbalance handling are employed to enhance efficiency and robustness. Experimental results show that the LSTM model achieves the best overall performance, with an accuracy of 99.98%, an F1 score of 98%, and a precision of 100%, while maintaining low computational complexity, measured in Real Multiplications Per Symbol (RMpS). These findings demonstrate that high diagnostic accuracy can be achieved without high computational complexity, supporting the feasibility of real-time and resource-constrained clinical deployment.

Bookmark

View Full Paper

Cite This Study

Bhia et al. (Sat,) conducted a other in Cardiovascular diseases (n=290). LSTM model vs. CNN, MLP, CNN-MLP, and ConvLSTM models was evaluated on Accuracy. The LSTM model achieved the best overall performance for cardiovascular disease diagnosis with an accuracy of 99.98%, an F1 score of 98%, and a precision of 100%.

synapsesocial.com/papers/69af235f3eac3accde8a1735 https://doi.org/https://doi.org/10.1038/s41598-026-43543-3

Bookmark

View Full Paper