Skip to main content
PLOS One logoLink to PLOS One
. 2023 Jul 27;18(7):e0289076. doi: 10.1371/journal.pone.0289076

Prediction of gastrointestinal functional state based on myoelectric recordings utilizing a deep neural network architecture

Mahmoud Elkhadrawi 1,*, Murat Akcakaya 1, Stephanie Fulton 2, Bill J Yates 3,4, Lee E Fisher 5,6,7, Charles C Horn 2,4,8,9
Editor: Zhishun Wang10
PMCID: PMC10374095  PMID: 37498882

Abstract

Functional and motility-related gastrointestinal (GI) disorders affect nearly 40% percent of the population. Disturbances of GI myoelectric activity have been proposed to play a significant role in these disorders. A significant barrier to usage of these signals in diagnosis and treatment is the lack of consistent relationships between GI myoelectric features and function. A potential cause of this issue is the use of arbitrary classification criteria, such as percentage of power in tachygastric and bradygastric frequency bands. Here we applied automatic feature extraction using a deep neural network architecture on GI myoelectric signals from free-moving ferrets. For each animal, we recorded during baseline control and feeding conditions lasting for 1 h. Data were trained on a 1-dimensional residual convolutional network, followed by a fully connected layer, with a decision based on a sigmoidal output. For this 2-class problem, accuracy was 90%, sensitivity (feeding detection) was 90%, and specificity (baseline detection) was 89%. By comparison, approaches using hand-crafted features (e.g., SVM, random forest, and logistic regression) produced an accuracy from 54% to 82%, sensitivity from 46% to 84% and specificity from 66% to 80%. These results suggest that automatic feature extraction and deep neural networks could be useful to assess GI function for comparing baseline to an active functional GI state, such as feeding. In future testing, the current approach could be applied to determine normal and disease-related GI myoelectric patterns to diagnosis and assess patients with GI disease.

Introduction

Functional and motility-related gastrointestinal (GI) disorders affect nearly 40% percent of the population, with annual costs of several billion USD. Myoelectric signals have the potential to contribute to clinical diagnosis and treatment of GI disorders; however, current methods of electrogastrography (EGG), based on abdominal surface recordings, and arbitrarily-defined analytic measures, such as dominant frequency, have provided limited insight into GI functional states [13]. This has led to little usage of GI signaling features in the diagnosis and treatment of patients.

Machine learning approaches have been used to increase the predictive power of GI signals. Human studies have used recordings of EGG signals from the abdominal skin surface, which limited signal-to-noise resolution because of the distance of electrodes from the stomach. In this approach, subjects must also remain seated and limit movements to provide a low noise signal. Using non-invasive EGG, a support vector machine (SVM) classifier performed at 88% accuracy to predict the occurrence of motion sickness in one study [4], although EGG was only one of the inputs to the model. SVM also performed well when classifying individuals with functional nausea, with an F1-score of 0.85 [5]. A neural network approach performed at 95% accuracy to predict un-fed vs. fed states in a population of 1000 human subjects [6]. Lastly, in our prior ferret study, we achieved > 75% accuracy in a binary or 3-state classification for predicting baseline and early and late periods before emesis; importantly, in our study, electrodes were placed directly on the surface of the stomach [7].

To more fully determine the clinical potential of the above approaches, we now combine three methods: (1) free-moving/awake testing, because this will permit application in diagnostic testing in different environments, e.g., clinic and at-home; (2) chronic implantation of electrodes on multiple stomach sites (four locations), to increase signal resolution and degrees of freedom in feature engineering; and (3) deep learning with neural networks. We also compare the performance of the deep learning approach with other machine learning approaches including support vector machine (SVM), discriminant analysis, logistic regression, and random forest. We used ferrets for the current study because they are a gold-standard model for GI research, including studies to determine efficacy of anti-emetic medications now used in the clinic [8, 9].

Materials and methods

Animals

This study included four adult purpose-bred influenza-free male ferrets (Mustela putorius furo; Marshall BioResources, North Rose, NY, USA; body weights 1.4 ± 0.2 kg and age of 5.1 ± 1.2 months, mean ± SD). Animals were adapted to the housing facility for at least 7 days before surgery. Ferrets were housed in wire cages (62 × 74 × 46 cm) under a 12 h standard light cycle (lights on at 0700 h), in a temperature (20–24°C) and humidity (30–70%) controlled environment. Food (ferret kibble; Mazuri Exotic Animal Nutrition, St. Louis, MO) and drinking water were freely available; however, food was removed 3 h before experimentation to assure the stomach was empty. At the end of the study, ferrets were euthanized with an intracardiac injection of euthanasia solution (390 mg/ml sodium pentobarbital; SomnaSol EUTHANASIA-III Solution, Covetrus, Dublin, Ohio, USA) under isoflurane general anesthesia (5%). The University of Pittsburgh Institutional Animal Care and Use Committee (IACUC) approved all experimental procedures.

Chronic electrode implant surgery

Anesthesia was induced using an intramuscular injection of ketamine (20 mg/kg), all surgical sites were shaved, and animals were endotracheally intubated with a 3.0 or 3.5 French endotracheal tube. Surgery was conducted in an aseptic surgical suite and animals were maintained under general anesthesia using isoflurane (1–2%) vaporized in O2. Subcutaneous injections of sterile saline were used to replace fluid loss. Body temperature was maintained at 36–40°C using a heating pad and monitored with a rectal probe.

Surgical procedures were similar to those in our prior study [7]. Animals were implanted with four gastric electrodes on the ventral surface of the stomach, including one on the fundus, two on the body, and one on the antrum (Fig 1), as well as two intestinal electrodes, and one abdominal vagus nerve cuff electrode (animal 103–21 was not implanted with a cuff electrode). Connectors from the electrodes were placed on the skull. Data collected from the intestine and the vagus nerve electrodes are not reported because they are not relevant to the current analysis. Each gastric planar electrode was attached to the GI serosal surface using 8–0 surgical silk with eight suture locations, four around each contact point. This was followed by placing a single suture around all electrode leads to anchor them to the left abdominal muscle wall. Before closing, the abdominal cavity was flushed with Cefazolin (1g/L), an antibiotic. The abdominal muscle was closed with 4–0 resorbable suture material and the skin with 3–0 monofilament nylon suture material and followed by applying surgical glue to the incision site.

Fig 1. Gastric myoelectric electrodes and behavioral testing.

Fig 1

A) Geometry of the two-contact electrodes. B) Surgical placement of the electrodes on quadrants of the ventral stomach surface of the ferret relative to the fat pad in the lesser curvature. C) Recording chamber and tether system.

Behavioral testing and gastric myoelectric recording

Animals were acclimated to the test chambers for three days prior to testing. GI myoelectric activity was recorded using a cable tethered to the head connector. Baseline GI myoelectric activity was recorded for 1 h for the first testing session, during which no food was provided. In subsequent sessions, all animals were presented with food (Ensure Original Vanilla Flavor nutritional shake, Abbott Laboratories, Lake Bluff, Illinois, USA) for 15 min. For feeding trials, baseline myoelectric activity was recorded for 15 min prior to food presentation and 30 min after food removal. There were 1 to 2 days between each test. Myoelectric activity was recorded with differential amplifiers attached to a Ripple Grapevine acquisition system (EMG headstage, Ripple Neuro LLC, Salt Lake City, UT). Data were acquired at 2 KHz sample rate using a band pass of DC to 2 KHz. The input range was ±187.5 mV to permit large signal changes due to potential movement artifacts.

Data analysis

Preprocessing

Data were initially imported into Spike 2 software (Version 9; Cambridge Electronic Design Ltd, Cambridge, England) to assess data quality and then files were exported in CSV file format. A custom workflow in Python 3 was written to process these data prior to application of machine learning algorithms, including: (1) down sampling; (2) removing dropped/saturated signals; (3) applying a band-pass filter; (4) removing movement artifacts; and (5) assessing signal quality using power spectral density (PSD) and spectrogram plots. Down sampling to 200 Hz sampling rate was accomplished with an anti-aliasing filter (https://docs.scipy.org/doc/scipy/reference/generated/scipy.signal.decimate.html). Data that exceeded the input range of the amplifier were then removed and replaced with zeros. A bandpass Butterworth filter at 0.05 to 0.7 Hz was applied, forward and backward (https://docs.scipy.org/doc/scipy/reference/generated/scipy.signal.butter.html, https://docs.scipy.org/doc/scipy//reference/generated/scipy.signal.filtfilt.html). This was followed by linear interpolation of data that exceed the physiological range of the signals, using ± 2 mV as the threshold (https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.interpolate.html). Fig 2 demonstrates the quality of our data with examples of raw and processed signal. A GitHub repository is available for this preprocessing pipeline. For machine learning, we only included trials that included < 20% removal of data based on dropped signals, < 20% data interpolated based on movement artifacts, and > 1000 V2/Hz amplitude in the total PSD.

Fig 2. Representative samples of myoelectric signals.

Fig 2

A) Slow wave signals from the four gastric myoelectric channels after preprocessing. B) Full length of the recorded signals for baseline and feeding conditions after preprocessing. C) Power spectral density (PSD) and spectrogram plots of the slow wave signals from channel 4, left = baseline and right = feeding. The Y-axis of the PSD plots is in E (scientific) notation. cpm = cycles per minute.

Machine learning

In this study, we were interested in classifying EGG signals between feeding and baseline control. To achieve this, we applied a deep neural network approach for automatic feature extraction. We also compared the neural network approach to other common machine learning approaches, including support vector machine (SVM), discriminant analysis (DA), logistic regression (LR), and random forest (RF) classification, which use hand-crafted features. We used Pytorch (v 1.12.1, CUDA 11.6) on Python (v 3.8.2) for the deep learning approach, and MATLAB (R2021a) for the other techniques. For all approaches, 1-hour data from each of the 4 animals was segmented into 1-min windows, thus obtaining a total of 480 data samples (240 feeding, 240 baseline). All approaches were validated using 3-fold cross validation.

Support Vector Machines (SVMs) are a type of supervised machine learning algorithm used for classification and regression analysis. They find a hyperplane in a high-dimensional space that maximally separates different classes of data or approximates the relationship between input and output variables (here we use it to classify). SVMs are useful for non-linearly separable data as they can use a kernel function to transform input data non-linearly into a different high-dimensional space like the radial basis function (RBF) kernel. Discriminant Analysis is another supervised machine learning algorithm used for classification. There are two main types of discriminant analysis: Linear Discriminant Analysis (LDA) and Quadratic Discriminant Analysis (QDA). LDA assumes that the covariance matrix of the input data is the same for all classes and that the decision boundary between classes is linear. QDA, on the other hand, assumes that each class has its own covariance matrix and that the decision boundary between classes is quadratic. Logistic Regression is a supervised machine learning algorithm used for binary classification, where the goal is to predict the probability of an input belonging to one of two possible classes. It uses the logistic function (sigmoid function) to transform the output of a linear regression model into a probability score between 0 and 1. Random Forests are another type of machine learning algorithm used for classification and regression. They combine multiple decision trees to improve accuracy and reduce overfitting. To reduce the risk of overfitting, each decision tree is trained using a bootstrap sample of the input data and a random subset of features. Additional details about these methods can be found elsewhere [10].

For deep learning, the network was a 1D residual convolutional neural network (CNN) to extract features from the channels of gastric myoelectric signals (Fig 3). A 1D CNN is a type of neural network architecture commonly used for processing and analyzing one-dimensional signals, such as audio and time-series data [1113]. In a 1D CNN, input data are processed through a series of convolutional layers, which learn to extract relevant features from the input signal. Each convolutional layer consists of a set of filters, which slide over the input data and perform element-wise multiplication and sum operations to produce a feature map. The resulting feature maps are then passed through a nonlinear activation function, such as a rectified linear unit (ReLU) to introduce nonlinearity into the model. These features are then passed to a fully connected neural network followed by a sigmoid activation function, that outputs a value between zero and one to classify the input signal. A residual neural network includes a layer where the output of a layer is summed with the input to that layer, effectively creating a "shortcut" connection that bypasses one or more layers. This allows the network to learn residual functions, which are easier to optimize than the original functions [15].

Fig 3. Architecture of the convolutional neural network.

Fig 3

The convolutional layer parameters are described as follows: Conv [kernel size], [Number of output channels], [/2 means the input is downsampled by 2]. The skip connection of a block that downsamples the input is made of a convolutional layer with kernel size of 1 and stride of 2 followed by a batch normalization layer. The convolutional network is then followed by a fully connected layer, then a sigmoid activation function.

The input myoelectric signal was one minute long and had a sample rate of 200 Hz, containing 12,000 data points. The CNN included 6 layers, and each layer downsampled the signal by a factor of 2, except for the sixth layer which was a convolutional layer with kernel size of one, followed by a batch normalization layer. This last layer is used to achieve an output of only one channel, which corresponded to the features that were passed to a fully connected layer. After these down-sampling procedures, the original 12,000 point signal was converted to a vector of output features with a length of 375. The number of output channels for the first five layers was 6, 12, 24, 48 and 96 successively. Each layer was made of one residual block, such that each block was a convolutional layer (kernel size = 121) followed by a batch normalization layer. Then the input of the block was summed with the output of its batch normalization layer using a skip connection, and the output was passed to a ReLU activation function. The skip connection downsampled the input for the residual block to achieve a convolutional layer of stride equal to 2 and a kernel size equal to 1. The features were then fed to a fully connected layer, which passed its output to a sigmoid activation function with an output range of 0 to 1, which correspond to posterior probabilities, with 0 corresponding to the baseline and 1 corresponding to feeding. The loss function used to train this network was a binary cross-entropy loss. The performance was evaluated using ROC curves. The hyperparameters of the network (number of layers, kernel sizes, etc.) were chosen heuristically by trial and error.

In order to reduce the generalization error, we used the dropout technique [14] after each convolutional layer during training. The network was trained using early stopping to reduce overfitting [15]. Early stopping saves the model that achieves the lowest validation loss and stops the training of the network if it cannot achieve a lower validation loss during the following 50 epochs. The maximum number of training epochs was set to 200. The training batch size was set to 20 with a learning rate of 0.001. The Adam optimizer with weight decay (AdamW) was used [16].

For other classification approaches (i.e., SVM, DA, DA, and RF), we extracted frequency-based features (i.e., the signal power from narrow frequency bands) from each of the four channels of 1-min signals. We calculated the PSD for each channel using Welch’s method [17] (using 16384(214) FFT points). The PSD was normalized (dividing by total power) for each channel and for each animal individually. Prior work from our lab and others has shown that gastric slow wave activity in the ferret typically spans 0.05 to 0.3 Hz, and that activities such as feeding and emesis can increase signal power in the tachygastric (i.e., higher frequency) range of the power spectrum [7, 18]. As such, we included PSD values ranging from 0.05 to 0.7 Hz. To capture variability in different frequency bands across this range, we divided the PSD of each channel into 10 bands and summed the PSD bins for each band. Thus, we obtained 30 features from each of the 1-min 4-channel samples. We applied forward and backward sequential feature selection [19] to select the most relevant features. The feature selection algorithm minimized the misclassification rate as an objective function and performed 3-fold cross-validation with 100 Monte-Carlo simulations to choose the best features. The performance of the selected features was measured performing 3-fold cross-validation, then plotting ROC curves and computing area under the curve (AUC) values. Note that during cross-validation, the training sets features were standardized using z-scores. The computed mean and standard deviation for each training set were then used to standardize the corresponding validation set. Like the NN approach, the output from all the models ranges from 0 to 1, with 0 corresponding to the baseline and 1 corresponding to feeding, except for the SVM models, where the output score values ∈ℝ. Thus, posterior probabilities were fitted on the SVM scores to produce output values of range 0 to 1.

Results

As mentioned above, we collected 1 hour of data with sampling frequency of 2000 Hz (downsampled to 200 Hz) from each animal, for both baseline and feeding conditions. We segmented those signals into smaller 1-minute trials to generate multiple samples for each condition, which enabled us to train the CNN and other machine learning models to classify between feeding and baseline conditions. We compared the performance of the machine learning methods for this classification. Note that our method did not investigate the changes across baseline and feeding based of the short signal duration during this period.

Preprocessing

Gastric channels 1, 2, and 4 from the four animals had < 20% dropped signals, < 20% artifacts removed, and power > 1000 V2/Hz in the frequency range of interest (Table 1). In contrast, channel 3 from two of the animals did not achieve these criteria (Table 1); therefore, our machine learning predictive models were focused on using gastric channels 1, 2, and 4.

Table 1. Assessment of channel quality.

% signal dropped % artifact removed Power amplitude (V2/Hz)
channel 1 2 3 4 1 2 3 4 1 2 3 4
58–21
baseline 0.5 0.5 0.5 0.5 2.0 6.1 1.9 1.2 7.6e+07 8.3e+08 1.4e+08 1.2+e08
feeding 18.3 18.4 17.9 18.2 6.5 2.8 7.3 5.4 2.0e+08 3.1e+08 4.0e+08 2.3+e08
60–21
baseline 0.5 0.5 1.3 0.5 3.0 0.6 51.2 7.9 3.0e+08 7.2e+07 8.7e+08 6.7e+08
feeding 4.4 0.9 1.6 0.9 9.6 4.8 57.2 8.6 4.6e+08 1.4e+08 8.7e+08 4.1e+08
87–21
baseline 2.8 0.7 0.2 0.1 2.2 4.6 3.1 0.4 1.9e+08 9.1e+07 7.7e+07 2.0e+07
feeding 2.6 6.4 3.9 0.3 2.4 12.7 40.9 0.9 1.2e+08 4.7e+08 9.8e+08 9.5e+07
103–21
baseline 0.0 0.0 0.0 0.0 1.4 0.0 0.0 0.0 2.0e+08 1.5e+08 6.4e+07 4.9e+07
eeding 0.0 0.0 0.0 0.0 4.5 0.0 0.1 0.0 5.7e+08 2.0e+08 1.4e+08 7.4e+07

Red indicates channels with low quality based on the criterion of > 20% removal; Power amplitude is in E (scientific) notation

Machine learning

All approaches were validated using 3-fold cross-validation, and the prediction scores were used to plot ROC (receiver operating characteristic) curves and calculate the area under the curve (AUC). In Fig 4A and 4B, we show the results for the methods that use hand-crafted frequency-based features, where we used forward feature selection (Fig 4A) and backward feature selection (Fig 4B). For both, we observe that the SVM model with Radial Basis Function (RBF) kernel achieved the highest performance, with AUC values of 0.88 and 0.87 for forward and backward feature selection respectively. The average AUC value for the forward feature selection approaches was 0.77, which is lower than the average AUC of the backward feature selection methods, which was 0.82.

Fig 4. Classifiers performance-ROC curves and AUC values.

Fig 4

A) ROC curves and AUC values for the frequency-based methods using forward feature selection. B) ROC curves and AUC values for the frequency-based methods using backward feature selection. C) ROC curve and AUC for the deep learning approach.

In Fig 4C, we show the results for the deep learning approach. It outperformed the best frequency-based approach, with an AUC value of 0.96.

To characterize the variability of the performance of the deep learning approach, we repeated the 3-fold cross-validation process 10 times, with a detection threshold of 0.5. This threshold was selected because the classifier outcomes were transformed to represent probabilities from 0 to 1. Then, we calculated the average accuracy, sensitivity, and specificity and their standard deviation, which were 90.4% (1.7%), 89.4% (2.7%) and 91.3% (1.4%) respectively. For comparison, Fig 5A shows accuracy, sensitivity, and specificity for SVM, LDA, QDA, LR, and RF classifiers using 3-fold cross-validation. Various frequency features contributed to the overall performance of SVM, LDA, QDA, LR, and RF classifiers, with no specific pattern observed (Fig 5B). Fig 5B shows the contribution of each frequency bin (power) for each classifier. The performance was consistent between all the animals for all the methods used (results are shown in S1 Table).

Fig 5. Performance and feature selection of classifiers using features for each signal frequency.

Fig 5

A) Performance of classifiers using hand-crafted features. A heat map is used to indicate average performance across the folds for each classifier (red = high percentage to green = low percentage). SVM = Support Vector Machine; LDA = Linear discriminant analysis; QDA = Quadratic discriminant analysis. B) Contribution of power frequency features to classifier performance. Grey shading indicates that the feature was selected by the classifier. The bottom rows indicate the overall percentage that each feature was selected across all classifiers using a heat map (red = high percentage to green = low percentage).

Discussion

Our results showed the NN approach outperforms other classifiers in distinguishing baseline and feeding. A close second place performance occurred with the SVM classifier using a RBF kernel at 0.88 AUC compared to 0.97 AUC for the NN approach. Moreover, backwards feature selection performed better than forward feature selection. There are several strengths of the current study, which could have positive impacts on clinical applications. Unlike hand-crafted features, we used an agnostic approach for feature selection with deep learning methods; this could capture unknown signaling information between gastric states. We also carefully assessed the quality of signals and removed data that were likely non-physiological (i.e., signals above 2 mV). These artifacts likely occur in real world data acquisition of gastric myoelectric signals from patients equipped with internal or external electrodes, especially during movement. While this study incorporated a relatively small sample size of subjects (N = 4) and a short duration of recording, it could serve as a useful template to assess gastric function in individual patients with short test sessions.

Traditional machine learning approaches like support vector machines or discriminant analysis often use hand-crafted features, designed to capture relevant information in the data. In contrast, convolutional neural networks (e.g., the NN approach used in this paper) can learn to extract relevant features from the raw input signal by capturing complex temporal and spatial dependencies in the data, without being restricted to hand-crafted features. This could lead the neural network approach to outperform the other approaches. However, an important trade-off is reduced model interpretability and higher risk of bias and overfitting with the NN approach.

The limitations of the study include model complexity, use of invasive electrodes, and the gastric states used for testing. In general, the NN model provided no insight into what data features produced the model performance. It is also unclear if this classifier performance can be achieved using a less invasive approach to record signals, such as the typical surface electrodes used in clinical EGG. Finally, the states used here (i.e., baseline and feeding) are potentially extreme differences in physiological state and more subtle changes, such as reduced gastric function in gastroparesis patients compared to normal controls, might not be detectable.

Our results suggest several future investigations using preclinical models, followed by testing in humans. These studies could include the use of larger samples with more animals and potentially longer recording sessions under different testing conditions to determine whether the NN approach could be applied to predict gastric physiological states in individual animals. Importantly, this could be useful for personalized diagnostic medicine. Changes to the recording methods should also be investigated, including the use of fewer internal electrodes and dermal surface techniques coupled with the NN approach. In human studies, the non-invasive collection of data and subsequent NN approach could be readily investigated, although there may be limitations on the ability to record low-noise signals non-invasively during unconstrained behavior or in medical conditions such as gastroparesis, where signal amplitudes may be highly suppressed, necessitating implanted approaches [20]. Fortunately, recent evidence from our lab suggests that it should be possible to implant recording electrodes on the serosal surface of the stomach via minimally invasive surgical approaches [21].

Supporting information

S1 Table. Performance of classifiers per animal.

In the table, a heat map is used to highlight performance (red = high percentage to green = low percentage). Results were obtained with a detection threshold of 0.5. We show the average and standard deviation of the forward and backward methods (8 methods for each) to summarize them.

(DOCX)

Acknowledgments

We thank Michael Sciullo for assistance with data collection and preliminary analysis.

Data Availability

Please find the DOI for our data https://doi.org/10.5281/zenodo.8148338. It contains the original data and the code used for preprocessing and classification.

Funding Statement

"Please add BJY along with CCH and LEF as a funding receiver for grant number grant 5R01DK121703. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript".

References

  • 1.Levinthal DJ. Slow Wave(s) of Enthusiasm: Electrogastrography as an Electrodiagnostic Tool in Clinical Gastroenterology. Dig Dis Sci. 2021; s10620-021-07029–8. doi: 10.1007/s10620-021-07029-8 [DOI] [PubMed] [Google Scholar]
  • 2.Abid S. Electrogastrography: Poor correlation with antro-duodenal manometry and doubtful clinical usefulness in adults. WJG. 2007;13: 5101. doi: 10.3748/wjg.v13.i38.5101 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Sha W, Pasricha PJ, Chen JDZ. Correlations among electrogastrogram, gastric dysmotility, and duodenal dysmotility in patients with functional dyspepsia. J Clin Gastroenterol. 2009;43: 716–722. doi: 10.1097/MCG.0b013e31818b8ed9 [DOI] [PubMed] [Google Scholar]
  • 4.Li C, Zhang Z, Liu Y, Zhang T, Zhang X, Wang H, et al. Multi-Dimensional and Objective Assessment of Motion Sickness Susceptibility Based on Machine Learning. Front Neurol. 2022;13: 824670. doi: 10.3389/fneur.2022.824670 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Olson JD, Somarajan S, Muszynski ND, Comstock AH, Hendrickson KE, Scott L, et al. Automated Machine Learning Pipeline Framework for Classification of Pediatric Functional Nausea Using High-Resolution Electrogastrogram. IEEE Trans Biomed Eng. 2022;69: 1717–1725. doi: 10.1109/TBME.2021.3129175 [DOI] [PubMed] [Google Scholar]
  • 6.Gandhi C, Ahmad SS, Mehbodniya A, Webber JL, Hemalatha S, Elwahsh H, et al. Biosensor-Assisted Method for Abdominal Syndrome Classification Using Machine Learning Algorithm. Al qaness MAA, editor. Computational Intelligence and Neuroscience. 2022;2022: 1–14. doi: 10.1155/2022/4454226 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Nanivadekar AC, Miller DM, Fulton S, Wong L, Ogren J, Chitnis G, et al. Machine learning prediction of emesis and gastrointestinal state in ferrets. PLoS One. 2019;14: e0223279. doi: 10.1371/journal.pone.0223279 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Andrews PLR, Rudd JA. The Role of Tachykinins and the Tachykinin NK1 Receptor in Nausea and Emesis. In: Holzer P, editor. Tachykinins. Springer Berlin Heidelberg; 2004. pp. 359–440. [Google Scholar]
  • 9.Reynolds DJM, Andrews PLR, Davis CJ. Serotonin and the Scientific Basis of Anti-Emetic Therapy. Philadelphia: Oxford; 1995. [Google Scholar]
  • 10.Pattern Classification, 2nd Edition | Wiley. In: Wiley.com [Internet]. [cited 18 Apr 2023]. Available: https://www.wiley.com/en-us/Pattern+Classification%2C+2nd+Edition-p-9780471056690
  • 11.LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521: 436–444. doi: 10.1038/nature14539 [DOI] [PubMed] [Google Scholar]
  • 12.Schmidhuber J. Deep learning in neural networks: An overview. Neural Networks. 2015;61: 85–117. doi: 10.1016/j.neunet.2014.09.003 [DOI] [PubMed] [Google Scholar]
  • 13.He K, Zhang X, Ren S, Sun J. Deep Residual Learning for Image Recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2016. pp. 770–778. doi: 10.1109/CVPR.2016.90 [DOI] [Google Scholar]
  • 14.Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: A Simple Way to Prevent Neural Networks from Overfitting. Journal of Machine Learning Research. 2014;15: 1929–1958. Available: http://jmlr.org/papers/v15/srivastava14a.html [Google Scholar]
  • 15.Prechelt L. Early Stopping—But When? In: Montavon G, Orr GB, Müller K-R, editors. Neural Networks: Tricks of the Trade: Second Edition. Berlin, Heidelberg: Springer; 2012. pp. 53–67. doi: 10.1007/978-3-642-35289-8_5 [DOI] [Google Scholar]
  • 16.Loshchilov I, Hutter F. Decoupled Weight Decay Regularization. arXiv; 2019. doi: 10.48550/arXiv.1711.05101 [DOI] [Google Scholar]
  • 17.Hayes MH. Statistical Digital Signal Processing and Modeling. 1st ed. USA: John Wiley & Sons, Inc.; 1996. [Google Scholar]
  • 18.Percie du Sert N, Chu KM, Wai MK, Rudd JA, Andrews PL. Reduced normogastric electrical activity associated with emesis: a telemetric study in ferrets. World J Gastroenterol. 2009;15: 6034–43. doi: 10.3748/wjg.15.6034 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Aha DW, Bankert RL. A Comparative Evaluation of Sequential Feature Selection Algorithms. Pre-proceedings of the Fifth International Workshop on Artificial Intelligence and Statistics. PMLR; 1995. pp. 1–7. Available: https://proceedings.mlr.press/r0/aha95a.html
  • 20.Yin J, Chen JDZ. Electrogastrography: Methodology, Validation and Applications. J Neurogastroenterol Motil. 2013;19: 5–17. doi: 10.5056/jnm.2013.19.1.5 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21.Horn CC, Wong L, Shepard BS, Gourash WF, McLaughlin BL, Fisher LE, et al. Surgical placement of customized abdominal vagus nerve stimulating and gastrointestinal serosal surface recording electrodes. Journal of Surgical Case Reports. 2021;2021: rjab463. doi: 10.1093/jscr/rjab463 [DOI] [PMC free article] [PubMed] [Google Scholar]

Decision Letter 0

Zhishun Wang

8 Mar 2023

PONE-D-22-33411Prediction of gastrointestinal functional state based on myoelectric recordings utilizing a deep neural network architecturePLOS ONE

Dear Dr. Elkhadrawi,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

Please submit your revised manuscript by Apr 22 2023 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

Please include the following items when submitting your revised manuscript:

  • A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.

  • A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.

  • An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols.

We look forward to receiving your revised manuscript.

Kind regards,

Zhishun Wang, Ph.D.

Academic Editor

PLOS ONE

Journal Requirements:

When submitting your revision, we need you to address these additional requirements.

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and 

https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf

2. We note that you have stated that you will provide repository information for your data at acceptance. Should your manuscript be accepted for publication, we will hold it until you provide the relevant accession numbers or DOIs necessary to access your data. If you wish to make changes to your Data Availability statement, please describe these changes in your cover letter and we will update your Data Availability statement to reflect the information you provide

Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice.

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: Partly

**********

2. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: N/A

**********

3. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: Yes

**********

4. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

**********

5. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: Summary:

In this study, the authors aimed to classify gastric myoelectric signals into feeding and baseline control states using machine learning techniques. They compared the performance of a deep neural network approach with other traditional machine learning approaches that use hand-crafted features. The deep learning approach outperformed other classifiers, with an AUC value of 0.96, while SVM with RBF kernel achieved a close second place performance, with an AUC value of 0.88. The study's strengths include the use of an agnostic approach for feature selection with deep learning methods, careful signal quality assessment, and a relatively small sample size of subjects with short test sessions. However, the limitations included model complexity, the use of invasive electrodes, and potentially extreme physiological states used for testing. Future investigations could include the use of potentially longer recording sessions and different testing conditions to determine whether the NN approach could be applied to predict gastric physiological states in individual animals and humans.

The motivation behind this research is to develop an automated method for detecting gastric myoelectric activity and to compare the performance of machine learning approaches, including a deep neural network, with traditional approaches that use handcrafted features.

Strengths:

• The study proposes a new approach for gastric myoelectric signals analysis using deep learning, which shows better performance than traditional machine learning algorithms.

• The authors carefully assessed the quality of signals and removed data that were likely non-physiological.

• The study provides a useful template to assess gastric function in individual patients with short test sessions.

• The findings could have positive impacts on clinical applications in personalized diagnostic medicine.

Weaknesses:

• Small sample size: The study was conducted on a small sample size of only four ferrets, which limits the generalizability of the findings to other animal models or humans.

• Short duration of recordings: The study used only 1-hour recordings, which might not be sufficient to capture subtle changes in gastric myoelectric signals and their classification.

• Invasive electrode placement: The use of invasive electrodes in ferrets may not be applicable to humans, where non-invasive techniques, such as surface electrodes, are preferred.

• Limited gastric states used for testing: The study only tested the classification performance between feeding and baseline control states, which are potentially extreme differences in physiological state. More subtle changes, such as reduced gastric function in gastroparesis patients compared to normal controls, might not be detectable.

• Complexity of deep learning models: The deep learning models used in the study are complex and provide no insight into what data features produced the model performance. This lack of interpretability could be a limitation in clinical applications where interpretability and transparency are crucial.

Questions:

1. Can you provide more insight into the features that the deep neural network approach used for classification?

2. Have you considered investigating the performance of the deep learning approach on a larger sample size of subjects or recordings?

3. Is it possible to achieve similar performance using less invasive methods for signal acquisition, such as surface electrodes commonly used in electrogastrography?

4. How would you address the potential challenge of detecting more subtle changes in gastric function in gastroparesis patients compared to normal controls using the current approach?

Comments:

The authors should clarify some of the technical details of their approach, such as the specifics of the deep neural network architecture used, the hyperparameters selected, and the normalization method applied to the data.

The manuscript could benefit from additional proofreading to address any spelling or grammatical errors.

The authors should include more detailed information on the data collection process, such as the electrode placement and recording parameters used.

Reviewer #2: The presented method to detect the GI functional state based on myoelectrical recordings using deep neural networks. The approach is interesting however, the following points should be addressed.

1- The features used for training and predicting the output should be explained in detail i.e., frequency ranges of the 10 bands. Why were those bands used and how they differ theoretically during various GI functions?

2- False positives were not addressed in any case. Please explain how they effect the overall performance.

3-Please introduce all the ML approaches briefly and discuss their performance in more details. Why do you think NN outperformed others

4- The figure 5 is not legible, hence the numbers could not be verified. Please provide clear figure to verify the results.

5- Check the table number of the supplementary table

6- Please explain in more details about how the small signal duration can affect the results and how you solved with this issue.

**********

6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: Yes: M. Khawar Ali

**********

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step.

PLoS One. 2023 Jul 27;18(7):e0289076. doi: 10.1371/journal.pone.0289076.r002

Author response to Decision Letter 0


9 May 2023

Reviewer #1

Questions

1. Can you provide more insight into the features that the deep neural network approach used for classification?

We thank you for this question and we understand the interest for a deeper insight in the interpretability of neural network features. However, machine learning approaches are often referred to as “black box” models” due to the difficulty in understanding how they work. The goal of deep learning approaches is to capture complex non-linear patterns and relationships between the input and the output, but at the expense of a lack of understanding of those underlying relationships. In this paper a 1-minute signal is fed to a convolutional neural network to be classified by extracting non-trivial patterns within the raw signal. In conclusion, the trade-off for better model performance is having a more complex model with features that are difficult to interpret and explain. We describe this limitation in the third paragraph of the Discussion.

2. Have you considered investigating the performance of the deep learning approach on a larger sample size of subjects or recordings?

This is an excellent question, but we have not attempted to apply this approach to a larger cohort of animals. This will be the subject of future work. However, please note that even though the study was limited to four animals, we collected a large amount of data from each animal, which made it possible to apply deep learning methods. Moreover, in order to ensure that our approaches generalize across animals and do not overfit, all machine learning approaches were validated using 3-fold cross-validation. We have, however, added a statement in the fourth paragraph of the Discussion suggesting that future studies should include a larger sample size with more animals and longer recordings.

3. Is it possible to achieve similar performance using less invasive methods for signal acquisition, such as surface electrodes commonly used in electrogastrography?

This is an important question that we do not currently have data to address. However, there is good reason to believe that there would be major challenges in achieving the levels of performance reported here with non-invasive electrogastrogram recordings. As described by Yin and Chen (J Neurogastroenterol Motil, 2013), to achieve a low-noise reliable non-invasive electrograstrogram recording, patients “should be asked not to talk, move, read or make phone calls during the procedure.” In our study, animals were freely moving and active throughout the experiment. As such, while non-invasive approaches may be appropriate for diagnostic testing, an implanted approach may be necessary for applications that involve real-time classification of electrograstrogram signals under noisy conditions. We have added a statement at the end of the fourth paragraph of the Discussion addressing this issue.

4. How would you address the potential challenge of detecting more subtle changes in gastric function in gastroparesis patients compared to normal controls using the current approach?

As described above, we believe that there are multiple advantages to recording signals from implanted electrodes. For subtle changes in signals that might occur with patients with gastroparesis, we believe the low-noise capabilities of an implantable approach may be important, though we do not currently have evidence to confirm this hypothesis. We have added a statement at the end of the fourth paragraph of the Discussion, describing this potential advantage of an implanted recording system.

Comments:

The authors should clarify some of the technical details of their approach, such as the specifics of the deep neural network architecture used, the hyperparameters selected, and the normalization method applied to the data.

The manuscript could benefit from additional proofreading to address any spelling or grammatical errors.

The authors should include more detailed information on the data collection process, such as the electrode placement and recording parameters used.

We appreciate the suggestion regarding the addition of detail around the approach and have added a brief explanation about all the machine learning approaches used in this paper. These details are found in the “Machine learning” subsection (Methods and material> Data Analysis>Machine learning). The details of the convolutional neural network architecture, layers, hyperparameters, the kernels and optimizer are also provided.

We have also carefully proofread the paper and corrected grammatical mistakes where appropriate.

Details of the electrode placement are included in the second paragraph of the “Chronic electrode implant surgery” section of the Materials and Methods, and recording parameters are described at the end of the paragraph “Behavioral testing and gastric myoelectric recording” in the Material and Methods section. We have made edits to both sections to add additional clarity.

Reviewer #2

Questions

1- The features used for training and predicting the output should be explained in detail i.e., frequency ranges of the 10 bands. Why were those bands used and how they differ theoretically during various GI functions?

Slow wave activity in the ferret typically spans 3-18 cycles per minute (i.e., 0.05 to 0.3 Hz), and our recent work (Nanivadekar, et. al, 2019) has shown that interventions such as feeding or infusion of an emetic drug can increase signal power content at higher frequencies. As such, for machine learning approaches that used hand-crafted features, we chose to include a range of frequencies spanning those typically observed during behavior as well as higher frequencies (i.e. up to 0.7 Hz) to ensure we captured the frequencies most likely to change during feeding. Each of the 10 bands spanned 0.065 Hz to ensure sufficient resolution of individual features in the frequency domain to capture variability of different portions of the power spectrum. We have added additional detail to the sixth paragraph of the section titled “Machine Learning” in the Materials and Methods describing the justification for this choice of parameters.

2- False positives were not addressed in any case. Please explain how they effect the overall performance.

We investigated the overall performance using 3-fold cross-validation and showed the average sensitivity and specificity in Figure 5. Also, we plotted ROC curves for all the classification approaches in Figure 4. We noticed that the neural network (NN) approach achieved better results than the other approaches, thus the NN approach is more robust to false positives.

3-Please introduce all the ML approaches briefly and discuss their performance in more details. Why do you think NN outperformed others

Thank you for your recommendation. A brief introduction of all the applied machine learning methods was added to the machine learning subsection (Methods and material> Data Analysis>Machine learning).

Traditional machine learning approaches like support vector machines or discriminant analysis often use hand-engineered features, designed to capture relevant information in the data. In contrast, convolutional neural networks (the NN approach used in this paper) can learn to extract relevant features from the raw input signal by capturing complex temporal and spatial dependencies in the data, without being restricted to engineered features. This could lead the neural network approach to outperform the other approaches. However, the trade-off is reduced model interpretability and higher risk of bias and overfitting.

4- The figure 5 is not legible, hence the numbers could not be verified. Please provide clear figure to verify the results.

A higher quality version of the figure was uploaded. Please let us know if it is not clear enough.

5- Check the table number of the supplementary table

This has been addressed in our resubmission. Thank you.

6- Please explain in more details about how the small signal duration can affect the results and how you solved with this issue.

We interpret this as a question about the size of the trials used for classification. Please note that we are not investigating changes in sample signal duration. We have collected a large amount of data (1 hour with sampling frequency 2000 Hz) from each animal, for both baseline and feeding conditions. We segmented those signals into smaller 1-minute chunks to generate multiple samples for each condition, which enabled us to train the CNN model. Then, we compared feeding and baseline conditions based on these observations through ML. This doesn’t mean that we investigate the changes across conditions based on small signal duration.

Attachment

Submitted filename: Response to Reviewers.pdf

Decision Letter 1

Zhishun Wang

11 Jul 2023

Prediction of gastrointestinal functional state based on myoelectric recordings utilizing a deep neural network architecture

PONE-D-22-33411R1

Dear Dr. Elkhadrawi,

We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements.

Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication.

An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at http://www.editorialmanager.com/pone/, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to-date. If you have any billing related questions, please contact our Author Billing department directly at authorbilling@plos.org.

If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

Kind regards,

Zhishun Wang, Ph.D.

Academic Editor

PLOS ONE

Additional Editor Comments (optional):

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.

Reviewer #1: All comments have been addressed

Reviewer #2: All comments have been addressed

**********

2. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Partly

Reviewer #2: Yes

**********

3. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: N/A

**********

4. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: Yes

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

**********

6. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: 1. I understand the 'black-box' nature of your model. However, I am asking for the details of using the model. Consider exploring interpretability techniques in future

2. I appreciate the idea to test a larger cohort in future work. Expanding to other species could also be an interesting addition.

3. Your thoughts on non-invasive approaches are insightful. Despite challenges, pursuing such methods would be a valuable direction for future studies.

4. Investigating whether implanted electrodes better detect subtle gastroparesis changes is a solid plan. Empirical evidence will be key.

Your clarifications on the deep learning approach, grammar corrections, and details on electrode placement and recording are helpful.

Reviewer #2: (No Response)

**********

7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: Yes: M. Khawar Ali

**********

Acceptance letter

Zhishun Wang

19 Jul 2023

PONE-D-22-33411R1

Prediction of gastrointestinal functional state based on myoelectric recordings utilizing a deep neural network architecture

Dear Dr. Elkhadrawi:

I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department.

If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact onepress@plos.org.

If we can help with anything else, please email us at plosone@plos.org.

Thank you for submitting your work to PLOS ONE and supporting open access.

Kind regards,

PLOS ONE Editorial Office Staff

on behalf of

Dr. Zhishun Wang

Academic Editor

PLOS ONE

Associated Data

    This section collects any data citations, data availability statements, or supplementary materials included in this article.

    Supplementary Materials

    S1 Table. Performance of classifiers per animal.

    In the table, a heat map is used to highlight performance (red = high percentage to green = low percentage). Results were obtained with a detection threshold of 0.5. We show the average and standard deviation of the forward and backward methods (8 methods for each) to summarize them.

    (DOCX)

    Attachment

    Submitted filename: Response to Reviewers.pdf

    Data Availability Statement

    Please find the DOI for our data https://doi.org/10.5281/zenodo.8148338. It contains the original data and the code used for preprocessing and classification.


    Articles from PLOS ONE are provided here courtesy of PLOS

    RESOURCES