Toward robust AI-based ECG R-peak detection for real-time physiological monitoring in operational environments

Figures

• 

  Figure 1. Standard ECG waveform morphology. Illustration of a typical ECG cycle showing the P wave, QRS complex, and T wave (adapted with permission from Ergun_Pinar/Shutterstock.com). The R-peak corresponds to the maximum amplitude within the QRS complex and is used for beat-to-beat timing and HRV analysis.

• 

  Figure 2. Example of noisy ECG signal illustrating the challenge of R-peak detection in operational environments. A representative segment of monitoring-grade ECG data is shown, characterized by motion artifacts, amplitude variability, and nonstationary noise. Several R-peaks are difficult to identify due to signal corruption, particularly in the central portion of the recording, highlighting the limitations of classical threshold-based detection methods under realistic conditions. Dots indicate the annotated R-peaks.

• 

  Figure 3. Framework for AI-based real-time physiological monitoring using ECG. The figure illustrates the end-to-end pipeline for robust R-peak detection and HRV analysis in operational environments. Monitoring-grade ECG signals are acquired under realistic conditions characterized by noise and motion artifacts. During evaluation of operational ECG data, physiologically consistent annotations are generated to preserve cardiac timing while handling artifacts conservatively. AI-based models, including LSTM and foundation-model architectures, are trained using benchmark ECG data and subsequently evaluated under operational conditions. While the proposed architecture is designed to support future autonomous, real-time operation for R-peak detection without manual intervention, the present study focuses on the development and evaluation framework, including data acquisition, annotation methodology, and AI-based model training and evaluation. Fully autonomous deployment and downstream physiological state assessment therefore remain subjects for future work.

• 

  Figure 4. Examples of common noise artifacts in monitoring-grade ECG signals. Representative ECG segments illustrating two typical sources of noise encountered in operational environments. The left panel shows baseline wander, characterized by low-frequency drift often caused by respiration, movement, or electrode instability. The right panel shows muscle artifact, characterized by high-frequency irregular fluctuations resulting from muscle activity. Both types of noise can distort ECG morphology and complicate accurate R-peak detection in monitoring-grade recordings acquired under non-controlled conditions.

• 

  Figure 5. LSTM-based architecture for R-peak detection. The model processes fixed-length ECG segments and outputs a binary sequence indicating the presence of R-peaks at each time step. Bidirectional LSTM layers capture temporal dependencies across cardiac cycles, enabling discrimination between physiological R-peaks and noise-related artifacts. The architecture is based on the R-peak detection approach described by Laitala et al. (2020). Tensor dimensions are indicated for each stage, where B denotes batch size.

• 

  Figure 6. Architectures of the Mantis-8M-based R-peak detection models with task-specific output layers. Two variants of the foundation model approach are illustrated. In both cases, input ECG segments are first processed by the pretrained Mantis-8M backbone, which extracts high-dimensional feature representations. These features are concatenated with the original input signal to preserve temporal alignment and signal-level information. Left panel: Mantis-8M + LSTM architecture. The concatenated features are passed to a bidirectional LSTM network (two layers), which captures temporal dependencies across cardiac cycles before a dense layer produces a binary R-peak prediction sequence. Right panel: Mantis-8M + CNN architecture. The concatenated features are processed by a series of convolutional layers with ReLU activations, progressively reducing the feature dimension and mapping the representation to a single-channel output corresponding to R-peak predictions. Tensor dimensions are indicated for each stage, where B denotes batch size and the final output corresponds to a binary sequence indicating the presence of R-peaks at each time step.