Robust U-Nets for Fetal R-Peak Identification in Electrocardiography

Abstract

Accurate fetal R-peak detection from low-SNR fetal electrocardiogram (FECG) signals remains a critical challenge as current NI-FECG methods struggle to extract high SNR FECG signals and conventional algorithms fail when signal quality deteriorates. We proposed a U-Net-based method that enables robust R-peak detection directly from low-SNR FECG signals (0–12 dB), bypassing the need for high-SNR inputs that are clinically difficult to acquire. The method was evaluated on both real (A&D FECG) and synthetic (FECGSYN) databases, comparing against ten state-of-the-art detectors. The proposed method significantly reduces false predictions compared to commonly used detection algorithms, achieving a PPV of 99.81%, an SEN of 100.00%, and an F1-score of 99.91% on the A&D FECG database and a PPV of 99.96%, an SEN of 99.93%, and an F1-score of 99.94% on the FECGSYN database. Further investigation of robustness in low-SNR conditions (0 dB, 5 dB, and 10 dB) achieved 87.38% F1-score at 0 dB SNR on real signals, surpassing the best-performing algorithm implemented in Neurokit by 13.58%. In addition, the algorithm showed ≤2.65% performance variation across tolerance windows (50 reduced to 20 ms), further underscoring its detection accuracy. Overall, this work reduces the reliance on high-SNR FECG signals by reliably extracting R-peaks from suboptimal signals, providing implications for the reliability of fetal heart rate variability analysis in real-world noisy environments.

1. Introduction

Fetal heart rate (FHR) and fetal heart rate variability (FHRV) are crucial indicators derived from fetal heart monitoring (FHM) that help predict potential complications or changes in fetal health, such as intrauterine growth restriction, fetal arrhythmia, and an increased risk of perinatal death [1,2]. Therefore, the ultimate goal of FHM is to accurately obtain FHR and FHRV, which relies on the precise detection of fetal R-peaks. Non-invasive fetal electrocardiography (NI-FECG) is a promising FHM technique that enables continuous and real-time monitoring by capturing fetal electrocardiographic signals from abdominal electrocardiograms (AECGs) [3]. The AECG signals obtained are a mixture of maternal electrocardiogram (MECG), fetal electrocardiogram (FECG), and various artifacts. The typical method for extracting FHR (and subsequently FHRV) from AECG involves estimating and removing the MECG, leaving a noisy FECG signals, to which existing peak detection algorithms (designed for adult) are used to detect the fetal R-peak locations.

Existing peak detection methods face significant limitations when applied to FECG signals due to inherent challenges in morphology, signal overlap, and noise sensitivity. Traditional QRS detection algorithms rely on morphological features such as amplitude, slope, and duration, but these struggle to adapt to fetal signals because of fundamental differences between fetal and adult ECG patterns. Additionally, distortions introduced during FECG extraction further obscure fetal R-peaks, reducing detection reliability. A major complicating factor is the spectral and temporal overlap between maternal and fetal ECG signals, making separation difficult. Since fetal R-peaks (≤60 μV) are much smaller in amplitude than maternal R-peaks (100–150 μV), they are easily masked within the composite signal [4]. Moreover, extracted FECG signals often exhibit a low SNR, exposing a critical weakness in conventional peak detection methods [5]. Techniques like the Pan–Tompkins Algorithm (PTA), which integrate squared signals over a moving window, are prone to misinterpreting high-frequency noise as QRS complexes, leading to false positives. Previous studies [6,7,8] used PTA to detect fetal R-peaks after removing MECG, but the detection accuracy decreased significantly when detecting low signal-to-noise ratio signals. These limitations underscore the need for more robust, fetal-specific detection approaches that account for morphological variability and improve noise resilience.

Agostinelli et al. [9] proposed an adjustment to the traditional PTA to better interpret FECG features. Specifically, their method modified the parameters of the traditional PTA to account for the differences in heart rate, QRS complex amplitude, and QRS duration between fetal and adult ECG signals. The modified algorithm incorporates a bandpass filter with a frequency range of 9–27 Hz, an 80 ms moving window, and an R-peak corrector algorithm, thereby enhancing the detection reliability of noisy FECG signals. However, as a modification of the traditional PTA, it does not leverage advanced signal processing or machine learning techniques. Consequently, its ability to learn complex patterns or dynamically adapt to changes in fetal ECG signals is limited. In contrast, the deep-learning-based method proposed in this research aims to address these limitations by proposing a robust deep-learning-based algorithm to effectively extract and learn meaningful features from noisy data (As another research direction, advanced denoising algorithms can be applied before peak detection. The state-of-art algorithms include a diffusion model [10], forward–backward dynamic mode decomposition (FBDMD) method [11], and retentive network [12]), rendering them more reliable for real-time monitoring applications.

The U-Net architecture was originally developed for 2D image segmentation tasks and has achieved remarkable success across various applications [13,14,15,16]. Unlike sequence modeling approaches (e.g., CNN-LSTMs or Transformers), U-Net’s encoder–decoder structure provides synergistic capabilities critical for fetal ECG analysis. First, the contracting path captures long-range context to distinguish fetal QRS complexes from noise. Second, the incorporation of skip connections mitigates overfitting, facilitating accurate segmentation of small objects even with limited training data [13]. In addition to its success in 2D tasks, this architecture has shown superior noise robustness in analogous 1D applications, achieving a 96.03% F1-score at <6 dB SNR in adult ECG [17] and reducing false-positive predictions by 10% versus E-D CNN without skip connections in Holter ECG signals [18,19]. Critically, U-Net’s inherent design enables effective training with limited annotated samples—a necessity in fetal ECG analysis where large-scale labeled datasets are scarce—while Transformer-based models typically require massive data volumes to achieve comparable performance due to their lack of spatial inductive biases and higher parameter complexity [20,21]. Therefore, the U-Net architecture is uniquely suited to our target of fetal R-peak detection in noise-dominated recordings.

In this work, motivated by these findings, we propose a novel deep-learning-based method that uses the U-Net architecture to accurately detect fetal R-peaks from low SNR FECG signals, thereby addressing the strong demand for extracting high SNR FECG signals and the limitations of PTA-based peak detection methods with FECG signals. The architecture of the proposed fetal R-peak detection method is adapted from the conventional 2D U-Net model to accommodate 1D time-series tasks. The Abdominal and Direct Fetal Electrocardiogram Database (A&D FECGDB) provides the gold standard direct FECG signals, which are treated as the FECG signals extracted by the NIFECG method in this study. In addition, the synthetic fetal ECG signals provided in the Fetal ECG Synthesis Database (FECGSYN) are also used. The robustness of the proposed algorithm is validated using both real recordings and synthetic recordings. The real recordings consist of three datasets with SNRs ranging from 0 to 10 dB, generated by introducing additional noise to the original A&D FECGDB. Additionally, synthetic fetal ECG signals from the Fetal ECG Synthesis Database (FECGSYN) are also utilized. In addition, the relationship between the choice of matching window size and performance deviation is investigated to assess the algorithm’s prediction accuracy. Ten existing peak detection methods are applied and compared with our proposed method to validate its robustness and efficacy.

The remainder of this paper is organized as follows. Section 2 provides a detailed description of the proposed U-Net-based method for fetal R-peak detection. Section 3 outlines the dataset utilized and the evaluation metrics applied and introduces ten existing peak detection methods for comparison. Section 4 offers a comprehensive evaluation of the performance of the proposed method, examining the impact of key thresholds on the performance and the noise robustness of the method. Section 5 compares the proposed method with ten existing peak detection methods and explores the impact of matching window size on the performance. A comprehensive analysis of the robustness and the predicted peak locations accuracy is performed. Finally, Section 6 presents the conclusions.

2. Methods

This section details the proposed U-Net-based fetal R-peak detection method, with the main three stages summarized in Figure 1.

2.1. Data Pre-Processing

This first stage processes and prepares both training and testing data for subsequent stages. First, the fetal ECG signal is downsampled to 250 Hz to mitigate heightened computational costs without substantial benefits. Lower sampling rates are considered sufficient for most clinical applications and are more practical in terms of data storage and processing requirements. Furthermore, Mohebbian et al. found that the effective field of view of units in CNN-based networks can be significantly increased by reducing the sampling rate [22].

The second step is to generate pulse train maps based on true peak annotations, where each pulse shares a uniform width and is centered on the R-peak position. Figure 2 provides an illustrative example of a raw fetal scalp ECG signal (blue) and the corresponding generated pulse train plot (black). The pulse train uses a label of one for each R-peak region, and zeros for other sample points. Chivers et al. reported that the average range of fetal QRS complex duration is between 54.72 ms (manual measurement) and 58.34 ms (algorithmic measurement). At a sampling frequency of 250 Hz, this range corresponds to 13.68 to 14.50 samples [23]. Therefore, a pulse width of 15 samples was chosen. This approach not only highlights the peak region but also helps to achieve data balance, thereby mitigating potential challenges related to label imbalance.

The final step is segmentation, employing a fixed sliding window of 1000 samples (equivalent to four seconds at a frequency of 250 Hz), with a shift of 200 samples. Baseline fetal heart rate typically ranges from 110 to 160 beats per minute (bpm), with an observed decrease with advancing gestational age, resulting in higher heart rate in early pregnancy compared to later stages [24]. Therefore, the selected segment length of 1000 samples used in this study allows for encapsulating approximately 7–10 heartbeats. Furthermore, the chosen shifting length of 200 samples is selected to be short enough to enhance the ergodicity of model training while long enough to reduce training time and ensure ample new information is available in each segment to yield an effective model. The processed segments of fetal ECG data and corresponding pulse train maps were then used in subsequent training and testing stages.

2.2. Model Training

In the training stage, processed segments of fetal ECG data and corresponding pulse train maps from the training data set are used to generate a model for features of the fetal ECG signal. The architecture of the proposed U-Net neural network used is illustrated in Figure 1. The scalp ECG segments are initially fed into the U-Net through the encoding blocks. The output of the 1D-convolutional (1D-CNN) layer is preserved before proceeding through the max-pooling layer. Subsequently, decoding blocks are employed to facilitate the upsampling process via 1D deconvolution (transpose) convolutional layers. The output of the 1D transpose CNN layer is connected through skip connections with the output of the 1D-CNN layer, which was saved during the encoding stage. Finally, the output of the last activation function is directed to the last convolutional layer to generate a predicted pulse train using a sigmoid activation function.

Both the encoder blocks and decoder blocks are composed of three 1D-CNN layers. The bottleneck block consists of two 1D-CNN layers. The filter parameters and kernel size in the convolutional layer in the encoding block were {16, 16, 32} and {9, 9, 6}, respectively. Each convolutional layer is followed by a ReLU activation function. A max-pooling layer with a pool size of 2 is connected after each 1D-CNN layer to achieve dimensionality reduction. In the decoding block, the filter parameters and kernel sizes were {32, 16, 16} and {9, 9, 6} for the 1D-CNN layer and 1D-deconvolutional CNN layer, respectively. The stride of the deconvolution layer is set to 2.

The model values were randomly initialized before training. Training and model parameters, including learning rate, neurons per layer, activation function, batch size, and number of epochs, were evaluated in terms of their effect on model learning and accuracy and selected accordingly. Binary cross entropy was selected as the loss function used by the model. The Adam optimizer was used to minimize the loss between the predicted and the original pulse train map. The learning rate was set to 0.001, with a batch size of 64. Each model was trained for 400 epochs. Five-fold cross-validation was performed using four records as the training set and the remaining record as the test set.

2.3. Fetal R-Peak Detection

In the testing R-peak detection stage, constructed inputs from test signals are applied to the model constructed in the training stage. The output from the model then undergoes additional post-processing steps, aimed at identifying and eliminating false positives, thereby enhancing the accuracy of R-peak detection.

The output map segments are restored to a full length through an overlap-add process, and additional processing steps are executed to finalize the peak screening process. Firstly, a filtering criteria removes potential false-positive predictions where amplitudes of the output pulse train are weak and below the first threshold ($Th{r}_1$). Currently, the threshold $Th{r}_1$ is set to a fixed value of 0.3. This means that predicted values below $Th{r}_1$ will be converted to 0, and the remaining values will be converted to 1.

The middle value of each pulse is then taken to generate a fiducial R-peak series, corresponding to predicted R-peak locations. Note that where a predicted pulse was found to have a width of one (compared to width of 15 samples associated with typical QRS-complex), that pulse is excluded, and no pulse is predicted in that location.

Finally, a second threshold ($Th{r}_2$) is applied to eliminate closely spaced potential beats. As specified in Section 2.1, fetal heart rates are typically within the range of 110–160 bpm [24]. A threshold $Th{r}_2$ of 81 sampling points is employed to filter out potential R-peaks that are too close to be deemed physically possible. When two beats are found to be closer than the defined threshold $Th{r}_2$, the beat further from the midpoint of preceding and subsequent beats is selectively removed. The result is an output signal where each pulse corresponds to the location of a predicted R-peak within the original FECG signal.

3. Experimental Setup

3.1. Database for Training and Evaluation

Two public fetal ECG datasets accessed on PhysioNet [25] were employed to develop the pipeline and test existing methods, including the Abdominal Direct Fetal ECG Database (A&D FECGDB) [26] containing real data and the Fetal ECG Synthetic Database (FECGSYN DB) [27] containing synthetic data.

The original databases were further processed to generate extra training and test datasets with different signal-to-noise ratio (SNR) levels for robustness evaluation. Each dataset is processed according to the details described in Section 2.1.

3.1.1. A&D FECGDB

This database has been considered as gold standard and used as a benchmarking tool due to its inclusion of direct fetal ECG [28]. It contains five multi-lead fetal electrocardiogram (FECG) recordings obtained from five women at different gestational ages between 38 and 41 weeks. Each recording contains four differential signals acquired from the maternal abdomen and a reference direct fetal ECG recorded from the fetal head. These recordings were acquired by the Department of Obstetrics of the Medical University of Silesia using the KOMPOREL Fetal ECG Acquisition and Analysis System (ITAM Institute, Zabrze, Poland) and stored in EDF format. Each recording lasts five minutes and has a sampling rate of 1 kHz and a resolution of 16 bits, numbered as r01, r04, r07, r08, and r10. The provided annotation of the fetal R-wave locations was first automatically determined by the online analysis of the KOMPOREL system and then manually verified by a professional cardiologists. It can be accessed via https://www.physionet.org/content/adfecgdb/1.0.0/ [accessed on 28 July 2025].

To evaluate the noise robustness of our method, we generated calibrated datasets with controlled signal-to-noise ratios (SNRs) from the original A&D database (denoted as AD_origin). Since the native SNR of AD_origin is unspecified due to inherent noise contamination, we applied additive white Gaussian noise (AWGN) with SNR of 0 dB, 5 dB, and 10 dB to AD_origin to create three standardized dataset: AD_0, AD_5, and AD_10. Noise was added to the original signal $x\left(t\right)$ to create the noisy signal $y\left(t\right)$, as given by Equation (1). The noise $n\left(t\right)$ was generated according to a normal distribution with a mean of zero and a standard deviation determined by the desired signal-to-noise ratio (SNR) in decibels, where SNR is calculated as given by Equation (2).

$$y\left(t\right)=x\left(t\right)+n\left(t\right)$$

(1)

$${\mathrm{SNR}}_{\mathrm{dB}}=10·{log}_{10}\left({\displaystyle \frac{\mathrm{Signal}\mathrm{Average}\mathrm{Power}}{\mathrm{Noise}\mathrm{Average}\mathrm{Power}}}\right)$$

(2)

3.1.2. FECGSYN DB

The synthetic NI-FECG signals provided are generated by the open-source FECGSYN toolbox, which synthesizes physiologically plausible fetal and maternal ECG signals by simulating the propagation dynamics of fetal and maternal ECG signals in the maternal abdomen. The simulator adopts a source mixing approach, treating each signal component (fetal ECG, maternal ECG, and noise) as an independent signal source to provide separable waveforms and accurate adult and fetal heartbeat annotations. The database simulates ten different pregnancies, each of which contains seven types of physiological events, including the following:

• Baseline: Abdominal mixture (no noise or events).
• Case 0: Baseline (no events) + noise.
• Case 1: Fetal movement + noise.
• Case 2: MHR/FHR acceleration/decelerations + noise.
• Case 3: Uterine contraction + noise.
• Case 4: Ectopic beats (for both fetus and mother) + noise.
• Case 5: Additional NI-FECG (twin pregnancy) + noise.

Each simulation had a duration of 5 min and was sampled at 250 Hz with a 16-bit resolution (WFDB MIT format) for a total of 34 channels (channels 1–32 are FECG channels, and channels 33–34 are maternal reference ECG channels). By superimposing multiple signal components, a noisy NI-FECG signal can be obtained to support the algorithm verification of this work.

This study extracted two event types: baseline (noise-free abdominal mixture) and Case 0 (baseline + additive noise). From these, four datasets were constructed—each comprising 50 records (10 subjects × 5 records)—with defined noise levels:

• SYN_origin: Pure baseline signals (no noise).
• SYN_6/SYN_9/SYN_12: Case 0 signals at SNR of 6 dB, 9 dB, and 12 dB, respectively.

The full database is publicly accessible at https://physionet.org/content/fecgsyndb/1.0.0/ [accessed on 28 July 2025].

3.2. Evaluation Metrics

In accordance with the ANSI/AAMI guideline [29], sensitivity (SEN), positive predictability value (PPV), and F1-score allow for the assessment of the presence of R-peaks within a tolerance window. Each of these metrics have been used to evaluate the performance of the proposed method.

PPV is a measure of accuracy, which shows the proportion of the algorithm’s predictions that correspond to the true annotations. SEN is a measure of completeness, which shows the proportion of the true annotations that were correctly detected by the algorithm F1-score is the harmonic average of PPV and SEN. In this study, a match window size of 50 milliseconds (ms) (see Figure 3) is applied to determine the true positives (TP), true negatives (TN), and false positives (FP), which is consistent with previous studies in the literature [30,31,32]. Each performance metric is then calculated according to Equations (3)–(5).

$$\mathrm{PPV}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}}\times 100\%$$

(3)

$$\mathrm{SEN}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}}\times 100\%$$

(4)

$$\mathrm{F}1\text{-Score}={\displaystyle \frac{2\times \mathrm{PPV}\times \mathrm{SEN}}{\mathrm{PPV}+\mathrm{SEN}}}\times 100\%$$

(5)

3.3. Existing Detection Approaches

To compare the performance with the proposed method in this work, 10 open-source pre-existing methods were applied to Datasets A&D FECGDB and FECGSYN DB, as summarized in

Table 1. Summary of the ten existing detection methods compared in this study.

Source	Method	Filtering Approach	Feature Enhancement	Detection Mechanism
Neurokit	Neurokit2, 2021 [33]	0.5 Hz HPF and 50 Hz notch	Gradient analysis to obtain QRS complexes	R-peaks are detected as local maxima in the QRS complexes
	Pan–Tompkins, 1985 [34]	5–15 Hz BPF	Derivative and Squaring and Integration	Multi-stage thresholding
	Hamilton, 2002 [35]	8–16 Hz BPF	Rectification	Modified Pan–Tompkins (smaller integration window)
	Christov, 2004 [36]	Multi-moving averaging filters	Multi-head analysis and Proposed Threshold System	Dual algorithm collaboration and adaptive threshold
	EngZee, 2012 [37]	48–52 Hz notch and Multiple LPF	Differentiation	Adaptive threshold (Christov-inspired)
	Kalidas, 2017 [38]	Resampling (80 Hz) and SWT (db3)	Squaring and Moving Window Average	Thresholding
	Nabian, 2018 [39]	–	Sliding window for liberal initial R-peak list detection (To obtain more potential r-peaks and reduce missed detection)	Modified Pan–Tompkins
	Elgendi, 2010 [40]	8–20 Hz BPF	Moving Window Integration	Thresholding
	Rodrigues, 2021 [41]	Double derivative and Squaring and Moving window integration	FSM refinement	Exponential decaying and threshold-based
ECGPUWAVE	ECGPUWAVE, 2000 [25]	0.5–40 Hz BPF and Notch (50/60 Hz)	WT and Slope analysis	Multi-lead correlation and Adaptive search window
Our Methods	Proposed U-Nets	None	Pulse-Train map to enhance R-peak regain for model training	Thresholding for false-prediction removal

LPF: low-pass filter, HPF: high-pass filter, BPF: band-pass filter, WT: wavelet transform, SWT: stationary wavelet transform, and FSM: finite-state machine.

. These approaches represented the available open-source methods from Python toolbox: Neurokit [33] and ECGPUWAVE (denoted PTA_E) [25] for R-peak extraction.

4. Evaluation Results of Proposed Method

This section states the evaluation results of the proposed method. We first analyzed whether the bias between SEN and PPV could be managed by using the post-processing threshold. Moreover, we evaluated the robustness on both datasets with different SNR conditions.

4.1. Effect of Threshold $Th{r}_1$ on Performance

The R-peak detection stage of the proposed method incorporates a tunable threshold, $Th{r}_1$ to filter out weak or false-positive predictions. This threshold acts as a critical decision boundary, directly controlling the TP, FP, and FN counts and influencing trade-off between SEN and PPV. To quantify this relationship and determine the optimal threshold value, we evaluated $Th{r}_1$ across a range of $0.1$ to $0.9$ in steps of $0.2$ on the eight datasets introduced in Section 3.1.

The results show that a higher threshold discards marginal predictions, thereby reducing the risk of FP but increasing the risk of FN. Lower thresholds retain weak predictions and classify them as R-peaks, resulting in admitting more FPs. Figure 4 clearly shows that PPV is positively correlated with $Th{r}_1$, and SEN is negatively correlated with $Th{r}_1$. The F1-score, as a compromise between PPV and SEN, peaked at $Th{r}_1$ = 0.3 across all SNRs, indicating that this threshold optimally balances PPV and SEN. Therefore, this work sets $Th{r}_1$ to 0.3 to prioritize clinical reliability by minimizing missed R-peaks (FN reduction), which is critical for applications such as heart rate monitoring and arrhythmia detection.

4.2. Robustness

4.2.1. Evaluation on Real Recordings

To evaluate the proposed method, we first performed 5-fold cross-validation on the origin of A&D FECG database and three derived datasets with varying noise levels: SNR 0 dB, 5 dB, and 10 dB. These datasets are denoted as AD_Origin, AD_0, AD_5, and AD_10, respectively. With the post-processing threshold fixed to $Th{r}_1$ = 0.3, the average performance metrics are summarized in

Table 2. PPV, SEN, and F1-score on A&D FECG database with different SNR conditions. Five-fold cross-validation is applied for each dataset.

Dataset	PPV (%)	SEN (%)	F1-Score (%)
AD_0	88.53	86.31	87.38
AD_5	98.77	97.95	98.35
AD_10	99.84	99.62	99.73
AD_Origin	100.00	99.84	99.92

. TP, FP, and FN, as well as PPV, SEN, and F1 scores for each test record are listed in

Table 3. Experimental results of the proposed U-Net across different SNRs. The results for the 0 dB case are highlighted in bold.

Test Record	Dataset	TP	FP	FN	PPV (%)	SEN (%)	F1-Score (%)
r01	AD_0	627	10	17	98.43	97.36	97.89
	AD_5	643	1	1	99.84	99.84	99.84
	AD_10	644	0	0	100.00	100.00	100.00
	AD_Origin	644	0	0	100.00	100.00	100.00
r04	AD_0	469	132	163	78.04	74.21	76.07
	AD_5	600	4	32	99.34	94.94	97.09
	AD_10	629	0	3	100.00	99.53	99.76
	AD_Origin	632	0	0	100.00	100.00	100.00
r07	AD_0	624	1	3	99.84	99.52	99.68
	AD_5	627	0	0	100.00	100.00	100.00
	AD_10	627	0	0	100.00	100.00	100.00
	AD_Origin	627	0	0	100.00	100.00	100.00
r08	AD_0	643	7	8	98.92	98.77	98.85
	AD_5	650	1	1	99.85	99.85	99.85
	AD_10	649	1	2	99.85	99.69	99.77
	AD_Origin	651	0	0	100.00	100.00	100.00
r10	AD_0	393	190	244	67.41	61.70	64.43
	AD_5	606	33	31	94.84	95.13	94.98
	AD_10	630	4	7	99.37	98.90	99.13
	AD_Origin	632	0	5	100.00	99.22	99.61

.

Our method achieves high detection accuracy on real recordings from the origin A&D FECG database, with a PPV of 100.00%, SEN of 99.84%, and an F1-score of 99.92%. The sole exception was record r10, where five FP predictions occurred in regions lacking both fetal ECG activity and expert annotations, likely due to electrode disconnection When subjected to extreme 0 dB noise conditions, the U-Net maintains robust detection capability, surpassing 97% F1-score in three of five test records (r01: 97.89%, r07: 99.68%, and r08: 98.85%). The performance degradation observed in records r04 and r10 under noisy conditions due to their characteristically challenging waveforms: these recordings already contained inherent high-amplitude noise prior to the addition of artificial noise, and the subsequent 0 dB SNR amplification further obscured the true R-peaks beyond typical detection thresholds, as evidenced by the morphological analysis in Figure 5.

The ability of our method to accurately predict the peak location and the role of the post-processing step on the method performance are demonstrated in Figure 6. The signal condition is significantly worse at 0 dB (left sub-figure) than 10 dB (right sub-figure), which leads to the generation of more weak pulse predictions that are corrupted by noise. The false weak pulses caused by noise are excluded from potential beats by applying $Th{r}_1$, which helps to reduce the generation of FP predictions. However, it may also cause missed heartbeats (causing FNs), which are highlighted in blue boxes.

4.2.2. Evaluation on Synthetic Recordings

To further evaluate the method’s robustness, we conducted experiments on synthetic data from the FECGSYN database, including the original dataset and three noise-corrupted versions with signal-to-noise ratios (SNRs) of 6 dB, 9 dB, and 12 dB (denoted as SYN_origin, SYN_6, SYN_9, and SYN_12, respectively). For each dataset, we performed 5-fold cross-validation with 40 records from two patients are used for training and 10 records from two additional patients for testing. With the post-processing threshold fixed at $Th{r}_1$ = 0.3, the average performance metrics of the U-Net model across different SNR levels are presented in

Table 4. PPV, SEN, and F1-score on FECGSYN database with different SNR conditions. Five-fold cross-validation is applied for each database.

Dataset	PPV (%)	SEN (%)	F1-Score (%)
SYN_6	63.77	54.61	58.12
SYN_9	87.16	83.67	85.05
SYN_12	89.87	87.30	88.44
SYN_Origin	99.96	99.93	99.94

. The proposed method demonstrates strong detection accuracy on SYN_origin, obtaining 99.96% PPV, 99.93% SEN, and 99.94% F1-score. It also maintains a score of over 85% on the SNR of 9dB and 12 dB conditions. However, performance degrades significantly under high noise conditions (6 dB SNR), with the average F1-score dropping to 58.12% across all 50 test records.

4.3. Runtime

The proposed model contains trainable parameters of 41,889, and the computational efficiency was evaluated on an NVIDIA Titan RTX (24 GB) GPU using FECG segments from the FECGSYN database. The average inference time per 4-s segment (1000 samples at 250 Hz) was 40.39 ± 11.06 ms. For continuous monitoring applications, processing one minute of ECG data required only 605.82 ms, which is equivalent to processing 1.7 min of FECG data processed per second. For future implementations, we will further explore its real-time deployment potential by applying Google’s TensorFlow Lite (TF Lite Micro) [42].

5. Comparative Assessment of Existing R-Peak Detection Methods

5.1. Performance Relative to Robustness

To further assess the performance of the proposed method, we now compare the proposed U-Net with ten existing detection methods in the literature, which are presented in Table 1 in Section 3.3. The evaluation uses the same datasets, including the original A&D FECG (AD_Origin) data and datasets with signal-to-noise ratio (SNR) levels of 0 dB and 10 dB (denoted as AD_0 and AD_10) and FECGSYN (SYN_Origin) data and datasets with SNR levels of 6 dB and 12 dB (SYN_6 and SYN_12). The same evaluation metrics are used: PPV, SEN, and F1-score, and a tolerance range of 50 ms is set. Figure 7a,b illustrate the performance of compared methods on the A&D FECG and FECGSYN databases, respectively. The results demonstrate the efficacy of the proposed method across datasets with varying SNR conditions.

For the A&D FECG database, under baseline conditions, the violin plot distribution of the U-Net(ours) across the three metrics is concentrated, with the median close to 1, indicating the highest overall performance. Neurokit2021, Rodrigues2021, Engzeemod2012, and Kalidas2017 are identified as comparable methods as their median values exceed 0.9, accompanied by narrow interquartile ranges (IQRs), indicating high detection accuracy. As the noise level increases, a general decline in performance is observed across all methods. When the SNR decreases to 0 dB, indicative of a high-noise environment, the distribution of the U-Net remains relatively compact, with the median close to 1, suggesting greater robustness to noise. However, the PPV, SEN, and F1-scores for half of the methods in the comparison experience significant drops, with the median falling below 0.5. Notably, only Neurokit2021 maintains a median exceeding 0.8 across the three metrics, although its performance exhibits considerable fluctuation, as evidenced by a wide IQR, indicating sensitivity to noise.

In the synthetic database, the superior performance of the U-Net and Rodrigues2021 under baseline conditions is evident, characterized by a high median and a narrow IQR. As with the A&D FECG database, the performance of all methods declines with increasing noise levels; however, U-Net consistently maintains the highest performance. Specifically, on the database with an SNR of 6 dB, the median F1-score for U-Net decreases to approximately 0.6, which is notably twice as high as the median scores of the other compared methods, all of which fall below 0.3. This significant difference underscores the robustness of the U-Net in challenging noise conditions compared to existing methods.

Under ideal conditions (10 dB and 12 dB SNR), U-Net’s F1-score approaches the theoretical upper limit of 1.0, with the distribution of prediction results being more concentrated. In contrast, traditional methods exhibit limited performance and significant fluctuations in extreme values, highlighting the advantages of the U-Net in achieving consistent and reliable r-peak detections. In low SNR scenarios (0 dB and 6 dB), U-Net shows a small IQR and a high median on both databases, indicating that it is more robust to both physiological noise (ADFECGDB) and synthetic noise (FECGSYN) than other existing methods.

Additionally,

Table 5. Comparative analysis of fetal R-peak detection methodologies with explicit low-SNR evaluation.

Method	Preprocessing Requirement	Dataset	F1 (%) (Performance on Origin Database)	F1 (%) (Performance on Low-SNR (0–12 dB))	Robustness to Low-SNR
Agostinelli et al., 2017 [4] * Derivative and Squaring and Integration and Multi-stage thresholding	9–27 Hz BPF	ADFECG DB	99.4	Not reported	-
Neurokit2 et al., 2021 [33] (Toolbox Method) (Gradient Analysis: R-peaks are detected as local maxima in the QRS complexes)	Double derivative and Squaring and Moving window integration	ADFECG DB	98.93	73.80 (AD_0)	Medium
		FECGSYN	77.87	42.42 (SYN_12)	Low
Rodrigues et al., 2021 [41] (Toolbox Method) (Modified Pan–Tompkins)	8–16 Hz BPF	ADFECG DB	99.47	36.18 (AD_0)	Low
		FECGSYN	81.6	66.09 (SYN_12)	Medium
Lampros et al., 2023 [43] * (Pan–Tompkins’ algorithm)	Decomposed by EMD and Denoised by Wavelet soft thresholding and An algorithm based on correlation analysis produces the optimal IMF subset	ADFECG DB	93.24	Not reported	-
Qiao et al., 2023 [44] * (Variance-based fetal R-peak seed selection, time-varying coarse prediction, and adaptive probability mask calibration)	None	ADFECG DB	97.6	Not reported	-
Proposed U-Nets	None	ADFECG DB	99.92	87.38 (AD_0)	High
		FECGSYN	99.94	88.44 (SYN_12)	High

BPF: band-pass filter. EMD: Empirical mode decomposition. IMF: Intrinsic mode function. *: The reported F1-score is quoted from the original work. All toolbox methods evaluated on identical low-SNR datasets in this study. Bold values highlight superior performance in target condition.

provides a rigorous comparative analysis with explicit low-SNR evaluation. While existing methods achieve competitive F1-scores (>93%) on high-quality data, their performance degrades substantially under low-SNR conditions (36.18–73.80%). In contrast, our U-Net maintains robust detection accuracy (87.38–88.44% F1-score) without preprocessing, demonstrating 23.5–51.2% absolute improvement in challenging environments.

5.2. Performance Relative to Match Window Tolerance

This section systematically evaluates the influence of matching window tolerance on FECG peak detection accuracy across multiple methods. We examined three symmetric window widths centered on true R-peak annotations: ±25 ms (50 ms total), ±15 ms (30 ms), and ±10 ms (20 ms), corresponding to progressively stricter localization criteria. Building on Section 5.1, we analyzed the top three methods by F1-score from the A&D FECG (AD_origin) and FECGSYN (SYN_origin) databases. For broader relevance, we included PTA_E2020, a widely adopted approach based on the ECGPUWAVE peak detector, despite its suboptimal ranking.

Figure 8 reveals that PPV, SEN, and F1-score improve with wider windows. Figure 8a shows that Rodrigues2021 and PTA_E2020 exhibit significant performance degradation as windows narrow (e.g., 16.11–16.84% F1-score drop from 50 ms to 20 ms), reflecting inherent localization limitations. In contrast, our U-Net and Neurokit202 maintain stable metrics (±2.65% variation) across all window width. Notably, U-Net achieves superior performance (F1 = 83.79% at 20 ms and 0 dB), demonstrating superior detection accuracy and noise robustness. Figure 8b further highlights the exceptional accuracy of U-Net under tightened window width, reinforcing its efficacy in precise peak detection.

The matching window tolerance in R-peak detection affects the determination of TP. While most methods perform well over a wider tolerance range, U-Net’s consistent accuracy under tight windows makes it uniquely valuable in fetal HRV analysis and other applications that require precise peak localization.

6. Conclusions

This study presents a U-Net-based architecture for robust fetal R-peak detection in low-SNR FECG signals, addressing the critical limitation of conventional algorithms in noisy environments. The proposed method demonstrates high-performance metrics and exhibits excellent robustness by minimizing FP and FN across most recordings. Notably, impressive results are achieved on fetal ECG signals from the original A&D fetal ECG database, with a PPV of 100.00%, an SEN of 99.84%, and an F1-score of 99.92%. Similarly, on the original FECGSYN database, the method achieves a PPV of 99.96%, an SEN of 99.93%, and an F1-score of 99.94%.

The robustness of the U-Net model is further evaluated using real datasets with SNRs of 0 dB, 5 dB, and 10 dB, as well as synthetic datasets with SNRs of 6 dB, 9 dB, and 12 dB. Additionally, the performance of the proposed method is compared with ten existing peak detection methods, revealing that U-Net outperforms all of them even in low SNR scenarios (0 dB and 6 dB). Further analyzing the impact of matching window tolerance, U-Net’s consistent performance under tightened tolerances (20 ms) underscores its detection precision.

The proposed U-Net method effectively solves the problem of inaccurate detection of traditional peak detection algorithms in low SNR fetal electrocardiogram signals. Its strong robustness and high accuracy highlight its potential as a reliable solution for R-peak detection in low SNR fetal electrocardiogram signals.