Introduction

Gastroesophageal Reflux Disease (GERD) is a prevalent gastrointestinal disorder characterized by the backflow of stomach acid into the esophagus, leading to symptoms such as heartburn, regurgitation, and chest pain. It affects a significant portion of the global population and can have a considerable impact on an individual’s quality of life if left untreated1.

The initial step in managing GERD typically involves initiating Proton Pump Inhibitor (PPI) therapies. For those patients with atypical symptoms who do not respond adequately to antisecretory treatments, Classical diagnostic methods such as endoscopy and pH monitoring are used2. Both methods have been mainstays in the field for many years, providing valuable assistance in therapeutic and diagnostic guidance. However, there are still many patients who have normal endoscopic findings and pH monitoring results despite ongoing GERD symptoms and unresponsiveness to PPI therapy3,4.

In the realm of GERD diagnosis and management, Multichannel Intraluminal Impedance pH (MII-pH) monitoring has emerged as a valuable tool for assessing reflux events and their association with symptoms. The integration of MII with pH monitoring, significantly enhances the detection of refluxate presence, proximal extent, and clearance. By combining MII with pH analysis, it becomes possible to accurately differentiate between acid and non-acid GER events3.

MII monitoring measures electrical resistance within the esophagus, offering detailed insights into the movement and properties of refluxate5. It allows for the differentiation between liquid, gas, and mixed reflux events while identifying the direction of bolus movement, differentiating swallows from refluxate. This robust diagnostic tool enables clinicians to correlate symptoms with reflux events accurately. Currently, MII-pH monitoring is the gold standard for diagnosing GERD, offering superior diagnostic capabilities compared to conventional methods6.

While MII-pH monitoring represents a significant advancement in GERD diagnostics, the accurate interpretation of impedance data is paramount to its effectiveness in clinical practice7. Proper analysis of impedance measurements is essential for maximizing the utility of MII-pH monitoring and ensuring optimal patient outcomes in the management of GERD5.

Several clinical studies have highlighted the importance of thorough analysis and interpretation of impedance data5,7,8,9. However, the automatic detection of GER events using MII data has been explored in only a limited number of studies10,11,12,13.

In12 a cascade Multivariate Long Short-Term Memory with Fully Convolutional Network (MLSTM-FCN) system was introduced for GER detection and diagnosis using MII-pH signals. In12, the authors classified 84 acid reflux events versus 141 swallows or artifacts. Their detected events are characterized as either 6, 11, or 21 s long. This study primarily relied on pH sensor data, limiting its scope to identifying acid GER events. Their method relies on an initial preprocessing step based on pH threshold comparison to identify potential acid reflux events, followed by MLSTM-FCN classification to distinguish acid reflux from artifacts or swallows.

A limitation of their approach is that it only provides an approximate duration of a reflux event, categorized as 6, 11, or 21 s, without localizing the precise onset and offset times across the six channels. Another limitation is that they consider only acid reflux events, whereas GERs include both acid and non-acid refluxes. Non-acid refluxes are prevalent in clinical practice14 and should not be excluded.

Additionally, artifacts such as pH readings outside the 0–14 range are omitted in their method, which could lead to missing genuine reflux events, because many reflux events (acid and non-acid) can still be characterized via MII data. This ability to characterize diverse reflux events is precisely why MII-pH monitoring remains the gold standard for diagnosing GERD15.

In contrast13, introduced S4, a state-space sequence model that serves as a versatile building block for modeling signal data. The S4 model is treated as a “black box” with limited details provided. Initially developed for multidimensional signals such as images and videos, the S4 model was adapted into a multiscale architecture capable of handling extremely long audio sequences, including MII data. This innovative approach provided a universal framework for modeling diverse, multidimensional signals. However, the method achieved only 68.7% sensitivity and 80.8% specificity in identifying GER events within a dataset of 45 patients.

A limitation of this study is that in this study, MII data are segmented into 60-second clips labeled as containing reflux or not, but the exact start time of reflux within each segment is not specified. Moreover13, does not specify how they distinguish GER events from swallows or artifacts. Since swallows and peristaltic waves are highly prevalent, it is essential to address the need for their accurate differentiation.

Notably, both the MLSTM-FCN and S4 approaches were limited to event detection and did not address the segmentation of GER events across all six channels.

The method described in11 was an initial exploration aimed at understanding the structure of unfamiliar MII-pH data. In11, a time-domain analysis of MII tracings was conducted to characterize liquid events, deliberately excluding the common occurrence of mixed GERs. A limitation of this approach is that it focuses solely on liquid reflux events, neglecting the more common mixed refluxes.

Additionally10, successfully automated the characterization of both mixed and liquid GER events through sparse representation. This study posited that MII represents a sparse signal comprised of GER and swallow events and extended periods of inactivity, corresponding to isoelectric intervals, as well as noise and interferences. It demonstrated the effectiveness of sparse representations for modeling MII, given that GER or swallow events are infrequent in comparison to isoelectric intervals. However, we believe that the success of deep learning methods in10 has been hampered by the limited availability of a large dataset of MII tracings.

Artificial Intelligence has recently expanded its applications to biomedical signal processing16,17,18,19,20,21,22,23,24,25. Attempts have been made to leverage U-Net26, for waveform segmentation tasks in various biomedical signals, including ECG27,28, plethysmography29, and heart sound signals30. In this study, we aim to investigate the feasibility of employing a semi-U-Net deep learning architecture to MII data. As detailed in10, GER is associated with several variables of clinical importance. Our method accounts for both acid and non-acid, as well as liquid and mixed physical states. It accurately delineates the onset and offset times of each GER event, enabling detailed segmentation rather than simple classification. Furthermore, our approach effectively differentiates GER events from artifacts and swallows through model training, and employs a dedicated post-processing step to separate GERs from swallows. We leverage onset times across multiple channels and exploit the characteristic retrograde propagation pattern of GERs, with specific criteria discussed in31.

To our knowledge, most existing studies only identify the peak or approximate start of events in 1-D signals, often with significant error margins. In contrast, our method uses deep learning to simultaneously localize both the onset and offset of GERs with high precision. We introduce a semi-U-Net architecture combining 2D and 1D CNNs that exploit the relationships between channels for effective segmentation of GERs from MII signals. This compact design efficiently utilizes parameters, enhancing its ability to generalize across diverse GER events.

In the following sections, we will first review the characteristics of our dataset. Next, we will provide an overview of our neural network architecture, inspired by the U-Net. We will then present the statistical analyses and results obtained from our study. Finally, we will conclude with a summary of our findings.

Methods

In our study, we focused on learning GER patterns without directly addressing the extended duration of MII data, which covers 24 h. This challenge can be managed through a preprocessing approach that analyzes the MII signal in non-overlapping 2-minute segments. Candidate intervals can be efficiently identified by calculating the entropy of the 24-hour signal32. Segments with entropy values exceeding a predefined threshold are then selected for further analysis using our proposed models. However, in our dataset, these 2-minute episodes were manually chosen to ensure a diverse representation of GER patterns.

MII dataset

Our dataset is comprised of 202 episodes, including a total of 208 GER events, which is an extended version of the dataset that was initially presented in10. All setups for acquiring this extended dataset mirrors the procedures outlined in the previous study10. The dataset can be accessed at33.

A typical 24-h MII–pH study recording often includes prolonged durations of isoelectric intervals, which may not be optimal for effective learning5. To address this issue, we extracted 2-min episodes from the MII data that are both informative and well-balanced. Each of these episodes spans a duration of 2 min and includes at least one GER event, with the potential inclusion of swallows, as detailed in10.

A thorough review process was conducted with three expert gastroenterologists to ensure accurate and consistent labeling of GER events across six channels. Their insights helped define event characteristics, leading to a standardized labeling system that improved data reliability and coherence.

The dataset overview in Table 1 provides a structural summary. Each episode spans a 2-min interval sampled at 50 Hz across six channels, with GER events labeling approximately 30.77% of the total data.

Figure 1 depicts plots of two samples from the dataset, showcasing two distinct episodes of MII data across the six channels plotted against the time axis, along with their corresponding labels.

Table 1 Dataset description.
Full size table
Fig. 1
figure 1

Sample MII data, all non-zero labelled signal points are emphasized as red, dashed lines are representative of Baseline Impedance (BI) which is extracted using a moving average filter with a 30 s window size. a An episode including two GER events, b another episode that includes one GER event.

Full size image

Signal normalization

Normalization of the amplitude of MII data was done before inputting it into the semi-U-Net structure. Studies have highlighted the importance of considering impedance levels in the interpretation of MII data, with variations in impedance values observed between proximal and distal channels within an empty lumen34. Normalization of the amplitude helps mitigate the impact of these differences, particularly given the differential presence of air and potential variations in cross-sectional areas between electrodes. Moreover, research has shown that esophageal impedance levels can vary among individuals, underscoring the need for standardized normalization procedures to reduce bias in the learning process34.

Amplitude normalization is intended to adjust signal amplitudes to a consistent range. This process helps reduce the impact of differing magnitudes on the convergence speed of the model and promotes the stability of numerical calculations35. The normalization step was implemented using (1):

$$x_{n} = \frac{x}{{\left\| x \right\|_{2} }} = \frac{{\left( {x_{1} ,~~x_{2} ,~ \ldots ,~~x_{n} } \right)}}{{\sqrt {x^{T} x} }}$$
(1)

In (1), x represents the impedance data of a specific channel. The normalization process was carried out individually for each channel.

Proposed approach

In this work we propose a two-stage scheme for the GER segmentation task which includes (1) elementwise classification at the 6th channel and (2) event segmentation across all of the channels. Our approach leverages the concept of making an initial estimation of potential GER events using MII data from all channels. Subsequently, we refine this initial estimation to determine the authenticity of the GER event, as well as to characterize its onset and offset times across all channels. Our proposed scheme, illustrated in Fig. 2 comprises two main components: the element-wise classification, detailed in section C1, and the GER event segmentation, described in section C2.

Fig. 2
figure 2

Block diagram of the proposed method.

Full size image

The neural network architecture A semi-U-Net architecture is Introduced that integrates both 2D and 1D CNNs for effective segmentation of GER events from MII signals. The network takes as input normalized impedance data from all six channels, along with the ground-truth label of the 6th channel. A number of 182 episodes were used for training of the network with holdout cross validation method.

Processing impedance data was based on the utilization of a U-Net like structure and CNN layers to extract feature maps. We proposed to keep the number of network parameters low, in order to be able to effectively learn from the limited input dataset36. The design of the neural network architecture is based on the following principles.

Convolutional layers play a crucial role in extracting meaningful features from raw MII data. By utilizing stacked convolutional layers and multiple feature maps, the model can effectively extract rich and valuable features37. Studies have shown a correlation between input length and segmentation accuracy in models38. Consequently, our proposed architecture featured an initial two-dimensional convolutional layer. This two-dimensional convolution enables the network to monitor changes in impedance throughout the esophagus effectively.

The pooling layer reduced the input length by half and captured relevant MII features. The temporal location of a GER in an episode was not considered a crucial feature to preserve. Pooling layers helped create representations that were almost invariant to translation39.

A sigmoid function was employed to scale output values within the range of 0 to 140. We evaluated with the DICE score, a well-established metric for segmentation accuracy evaluation in the field of computer vision41. The proposed algorithm employs an encoding-decoding architecture, with detailed specifications provided in Table 2 and a corresponding block diagram illustrated in Fig. 3.

The input of the network was a matrix of size 6×l, where l represented the length of the two-minute MII episode, each episode contains \(\:l=2\times\:60\times\:{f}_{s}\) samples for each channel. Despite other frameworks that suggest independent channel analysis42, we proposed simultaneous analysis of all 6 channels. In section III we will compare result of our proposed framework to the case in which each channel is analyzed individually.

The initial convolutional layer is two-dimensional with a kernel size of (211 × 6) to expand the receptive view of the network. During the annotation process, gastroenterologist experts considered all impedance channels together ensuring a thorough understanding of the dynamics and propagation of GER events within the esophageal lumen. This idea was inspired in choosing a 2D convolution at the first layer to gain a comprehensive understanding of GER event dynamics.

Reflux events typically initiate and terminate at the most distal site, making the 6th channel crucial for reflux duration detection. Based on this information, the network was trained to classify each sample point at the distal site as GER or Non-GER. The output was a vector of size \(1 \times l\), compared to the target vector representing the label of the 6th channel. The output of the network is an intermediate result that presents elementwise classification of the 6th channel. To localize a GER event, additional process steps are needed. Post processing generates final segmentation result in terms of GER event onsets and offsets across all 6 impedance sites.

Fig. 3
figure 3

The architecture of the network. The first layer, C1, is special as it uses a 2D CNN to capture features across all channels of MII data, the rest of the encoders and decoders use 1D CNNs to preserve the 1D nature of the signal.

Full size image
Table 2 The proposed NN architecture for elementwise classification of input MII data.
Full size table

Post-Processing To finalize GER segmentation the post processing step was conducted. The output of the previous step is the primary flag of the 6th channel. As described earlier, our final goal is to segment a GER event across all channels, meaning that we should find a final flag, the same size as input MII data that provides element wise classification for each sample at each channel. The post processing step can be considered as three blocks that are described below.

Cross sections with BI According to clinical literature a GER event typically initiates right after the impedance data reaches 50% of its baseline value11. In order to investigate this criterion, Baseline Impedance (BI) is determined by applying a 30-second moving average to each channel individually43. Next, BI is used to identify cross sections, where the impedance value reaches 50% of the BI. Subsequently the output is divided into a vector \(\:{z}_{l}\), specific to channel 6 and a matrix \(\:{Z}_{l5}\) for the remaining channels. The intersection between \(\:{z}_{l}\) and output of the semi-U-Net model, \(\:{y}_{l}\) is computed to ensure the selection of GER events that align with the defined criteria, potentially enhancing the specificity of the detected GERs.

In addition to this, the definition of a GER event has other criteria, which are further investigated in the third block of our post-processing step.

Morphological editing In the second block, a morphological editing process is applied to the outcomes of the preceding step. This stage consists of four morphological sub-blocks that function similarly to their two-dimensional counterparts, known as opening and closing operations44. Each sub-block involves the application of a moving average filter followed by thresholding. The specifics of kernel size and threshold values (θ) for each sub-block are summarized in Table 3. All kernels are rectangular in shape and normalized to have an L2-norm of 1. The result of this block is the determination of the onset and offset points for each GER event at the 6th channel.

Table 3 Details of morphological editing block.
Full size table

GER definition criteria The last block of the post-processing procedure is dedicated to channels 5 through 1. This block takes as input \(\:{Z}_{l5}\) together with the predicted labels from the 6th channel. Its objective is to determine onset and offset points for each assigned GER across channels 5 to 1.

The ground truth labels of channels 5 to 1 from the training dataset were used to regularize parameters of this block with the objective to best adapt with the definition criteria of GER events11,31. Then the optimized block was employed to predict labels for a query MII episode. The output of this block, known as \(\:{R}_{n12}\) is the final result of the post-processing procedure which is a matrix with n rows and 12 columns. With n representing the number of detected GER events, and the number of columns corresponds to 6 couple of onset and offset points belonging to the channels. In reporting values of \({R_{n12}}\), Any empty values were denoted by a “-” mark.

Fig. 4
figure 4

Block diagram of the proposed post-processing procedure.

Full size image

Figure 4 shows details of the post-processing step. Line and bold arrows symbolize the flow of single (vector) and multi-channel (matrix) data among different blocks, respectively. Each matrix is displayed in capitalized bold format, while each vector is in non-capital bold format. The subscript of each matrix or vector denotes its size. The vector \(g{t_l}\) represents the ground truth annotations of the 6th channel which was used to train the semi-U-Net network. The output of the network is denoted by \({y_l}\) which is the predicted label flag of GER at semantic segmentation level.

Figure 5 further elaborates on Fig. 4 and illustrates the roles of each component in the post-processing pipeline and how they collaboratively contribute to the final GER detection for a query MII episode.

Fig. 5
figure 5

a The 133th MII episode, Xl6, along with its baseline impedance, b the 6th channel Impedance data. c The ground truth annotation for the 6th channel, gtl. d The output of the model, yl, e the output of the cross sections with BI block, zl, f the output of the dot product of d and e, g output of the morphological Edit block. h the same episode from a with ground truth annotation highlighted in blue, and detected GER event, Rn12, marked with green crosses indicating the onset points and red crosses indicating the offset points. The horizontal axis in all panels indicates time.

Full size image

In summary, Fig. 6 provides an overview of our proposed model’s entire architecture, illustrating the process from MII episode input to GER event detection output. This figure integrates Figs. 2 and 4, highlighting the input, processing steps, and final detection results of our approach.

Fig. 6
figure 6

The proposed two-stage scheme for the GER segmentation task which includes elementwise classification at the 6th channel and event segmentation across all of the channels. It makes an initial estimation of potential GER events then refines it to determine the authenticity of the GER event, as well as to characterize its onset and offset times across all channels.

Full size image

Results

The proposed semi-U-Net neural network was trained and experiments were conducted using the extended MII dataset. To assess the performance of the method the dataset was randomly divided into train and test sets, comprising 182 and 20 episodes, respectively, as described in Table 1. Within the training episodes, 21 were set aside for validation during the neural network training and later combined for the final fine-tuning step.

To provide more insight on the regularization of the proposed method, Fig. 7 shows the accuracy and loss plots against iteration for both the train and test folds. Based on the analysis of accuracy and loss plots, our proposed method demonstrates effective regularization, indicating a good balance between model complexity and generalization ability. The consistent performance across both training and testing folds suggests that the model is not overfitting to the training data and can reliably segment GER events in unseen data.

Fig. 7
figure 7

Performance of the network versus epochs, for the train and test phases: a MSE loss and b accuracy.

Full size image

The Adam optimization method was employed to minimize loss by adjusting the network weight. During the initial training, the learning rate was set to 0.001, the mini-batch size was 32, and the training was limited to 150 epochs. To prevent overfitting, the maximum epoch was set to 55 and model parameters at epoch 55 were saved for subsequent fine-tuning. The best network weights, determined by the maximum F1-score, were saved at this stage.

To further enhance the network’s performance and capitalize on the knowledge gained during the initial training, a fine-tuning stage was also implemented. During fine-tuning, the training and validation data were combined to create a larger training dataset. The network’s weights with the best F1-score were loaded, and training with the expanded dataset took place for the specified number of epochs for fine-tuning (which was set to 10).

The performance of the proposed method was evaluated with accuracy (Acc), sensitivity (Sen), positive predictive value (PPV), F1 score and Temporal Error (TE) determined as follows45:

$$Acc~\left( \% \right)=100~\frac{{TP+TN}}{{TP+TN+FP+FN}}$$
(2)
$$F1\_Score~\left( \% \right)=100~\frac{{2 \times TP}}{{2 \times TP+FP+FN}}$$
(3)
$$Sen~\left( \% \right)=100~\frac{{TP}}{{TP+FN}}$$
(4)
$$PPV~\left( \% \right)=100~\frac{{TP}}{{TP+FP}}$$
(5)
$$TE=~\left| {actual~GER~duration - detected~GER~duration} \right|$$
(6)

The TP, TN, FP and FN are calculated using output of the network and label annotations of the 6th channel and is described by the confusion matrix of Table 4.

Table 4 Confusion matrix for elementwise classification.
Full size table

Building upon the evaluation metrics outlined above, Table 5 reports performance of our proposed semi-U-Net architecture at elementwise classification level for the 20 test episodes, that contain 21 GER events. Notably, prior research in12, indicated that using a single classifier yields unsatisfactory results across GER events of varying durations. Our analysis, as illustrated in Table 5, reveals that the mean duration of GER events is 17.52 s, accompanied by a standard deviation of 6.39 s. This significant variation underscores the challenges associated with analyzing MII data, highlighting the diverse nature of GER events.

In our investigation, we also evaluated a method that analyzes each channel of the MII data independently, called as 1D semi-U-Net model. The results obtained from this approach are comparable to those of our proposed method. However, the semi-U-Net model, which integrates data from all channels, demonstrates slightly superior performance. Specifically, Table 5 indicates that the total TE for the proposed method is 2.99 ± 3.09, compared to 3.25 ± 3.09 for 1D model, which underscores the effectiveness of our approach in accounting for the complete set of MII channels.

Table 5 Performance of the proposed GER detection method for the 6th channel at elementwise classification level.
Full size table

The results presented in the Table 5 correspond to the output of the trained network before applying the dedicated post-processing steps. In the following, we aim to systematically assess the impact of each post-processing component by analyzing different configurations of the proposed method. Table 6 summarizes the quantitative effects of each component within the framework. To evaluate the contribution of individual blocks, we tested various configurations, each omitting a specific processing step. Excluding the “GER Definition Criteria,” “Morphological Edit,” or “Cross Sections with BI” blocks, or bypassing post-processing entirely—using only the raw network output—allows us to understand how each component influences overall detection performance and robustness. Table 6 reports the results of each configuration on the test set at the element-wise classification level.

Table 6 Quantitative evaluation of component contributions in the proposed framework via ablation study.
Full size table

As shown in Table 6, the elimination of the two blocks ‘GER Definition Criteria’’ and ‘Cross sections with BI’ does not have a significant effect on the final results. This can be justified by the fact that these two blocks are dedicated to other objectives more than just to enhance the performance metrics. The ‘GER Definition Criteria’ and ‘Cross sections with BI’ blocks were added to provide GER segmentation across all channels and to ensure that each detected GER fulfills clinical literature criteria. They might be essential in a real-world setting for ensuring clinical validity but could be optional for purely performance-focused tasks.

The C4 and C2 configurations results in degraded performance, as indicated in Table 6. In C2, the removal of the ‘morphological edit’ block slightly degrades overall performance, indicating its role in improving the quality of results. Interestingly, metrics such as accuracy, sensitivity, and F1 score are slightly higher in C4 (raw network output) than in C2. This may be because, in C4, the raw output isn’t truncated via intersection operations, whereas in C2, some GER areas might be truncated during those intersection procedures with the BI block. Meaning that if we use the output of the model solely we get a higher sensitivity but the PPV in C2 -after some operations of the post processing step, except for the morphological edit-, is higher meaning that each candidate GER has a higher probability of being a true positive, possibly reducing false positives.

The proposed method, C5, outperforms all ablation configurations, confirming that each component contributes positively to the system’s robustness and accuracy. The superior performance of C5 over C1 may be attributed to its ability to effectively distinguish true GER events from swallows and artifacts, thereby reducing false positives and ensuring more precise detections.

At event segmentation level, Table 7 presents the performance metrics for GER localization, comparing the results of our proposed method with those obtained from our previous approaches that utilized sparse coding and discriminative dictionary learning methods10. Additionally, we include a comparison with the 1D convolutional semi-U-Net model. Our algorithm achieved an F1-score of 97.56% for localizing GER events, outperforming the other methods.

Table 7 Comparative analysis of GER event localization between proposed and other Methods.
Full size table

Examples of GER event segmentation using our proposed method are illustrated in Figs. 8 and 9. Figure 8 specifically presents the segmentation results for the 197th episode, which was excluded from the database in10 due to excessive noise. As evident in the figure, this episode exhibits significant noise. Remarkably, our proposed method effectively handles both mixed GER types and challenging noisy episodes. As indicated in Table 5, the proposed method segmented this GER event with a TE of just 1.82 s. This error is relatively minimal compared to the overall duration of the GER, which is 20.36 s, making it negligible for practical purposes.

Fig. 8
figure 8

Example of GER event localization demonstrating a low TE. The blue highlighted areas represent the ground truth, while the green cross marks indicate the onset and red cross marks denote the offset of the localized GER event.

Full size image

Figure 9 shows segmentation results for the 4th episode which has a considerable amount of TE. We believe that the insufficiency of the segmented area is initiated from the post-processing which require refinement to optimize the output results effectively.

Fig. 9
figure 9

An example of GER event localization with a high TE. The blue highlighted areas represent the ground truth, while the green cross marks indicate the onset and red cross marks denote the offset of the localized GER event.

Full size image

Processing of the algorithm takes 26 ms per 2 min of the MII signal on a system equipped with an Intel(R) Xeon(R) E-2176 M CPU, 32 GB of RAM, and an Intel(R) UHD Graphics P630 GPU with 16 GB of VRAM, which makes it suitable for online implementations as well.

Discussion

In this work, we presented a novel semi-U-Net architecture that combines 2D and 1D convolutional neural networks to effectively learn and identify GER events. The model was evaluated using a robust dataset of 202 episodes containing 208 GER events, collected from 26 patients undergoing 24-h MII pH monitoring. After training of the network, a post-processing unit was implemented that accurately segmented GER areas across all 6 channels, ensuring alignment with clinically defined criteria.

Our proposed architecture is characterized by its compact size, low weight, and efficient utilization of a small number of parameters, enhancing its generalizability. Our method demonstrates adaptability to varying durations of input data, making it versatile for application in scenarios with diverse data collection settings.

Our method achieved a sensitivity of 95.24% and a positive predictive value of 100%, demonstrating superior segmentation quality compared to existing methods. Notably, our method outperforms existing approaches in terms of segmentation quality, boasting an F1-score of 97.56%. Its robust generalizability is underscored by its ability to effectively handle challenging episodes that posed difficulties for previous methods.

The potential application of our approach in segmenting GER events holds promise for enhancing the utility of 24-hour MII-pH monitoring devices. The algorithm processes MII signals at a speed of 26 ms per 2 min, making it suitable for online implementations in clinical settings. Healthcare providers can leverage the segmentation results generated by our network to potentially improve decision-making processes, particularly in selecting patients for invasive surgical procedures.

While our method has yielded improvements in GER characterization, several limitations warrant further investigation. Currently, our pipeline relies on a post-processing step to refine initial detections, verify the authenticity of GER events, and determine their onset and offset times across all channels. This adds complexity to the workflow and may introduce additional sources of error. Additionally, there is a need for models that more accurately extract the proximal extents and capture the dynamics of GER across all channels.

To address these challenges, future research could focus on developing end-to-end models that inherently capture the dynamics and spatial correlations of GER events, eliminating the need for post-processing. Recent advances, such as Dense Associative Networks46 or Dense Attention Mechanisms47,48, offer promising solutions for this purpose.

Additionally, our analysis revealed that the statistical distributions of GER and non-GER regions resemble Laplace and normal distributions, respectively. Leveraging this insight, future directions could include the development of novel knowledge distillation techniques49, to transfer knowledge from complex, high-capacity models to smaller, more efficient models suitable for clinical deployment.