Unraveling Functional Dysphagia: A Game-Changing Automated Machine-Learning Diagnostic Approach

Abstract

(1) Background: Dysphagia affects around 16% of the US population. Diagnostic tests like X-ray barium swallow and endoscopy are used initially to diagnose the cause of dysphagia, followed by high-resolution esophageal manometry (HRM). If the above tests are normal, the patient is classified as functional dysphagia (FD), suggesting esophageal sensory dysfunction. HRM records only the contraction phase of peristalsis, not the distension phase. We investigated the utilization of esophageal distension–contraction patterns for the automatic classification of FD, using artificial intelligent shallow learners. (2) Methods: Studies were performed in 30 healthy subjects and 30 patients with FD. Custom-built software (Dplots 1.0) was used to extract relevant esophageal distension–contraction features. Next, we used multiple shallow learners, namely support vector machines, random forest, K-nearest neighbors, and logistic regression, to determine which had the best performance in terms of accuracy, precision, and recall. (3) Results: In the proximal segment, LR produced the best results, with accuracy of 91.7% and precision of 92.86%, using only distension features. In the distal segment, random forest produced accuracy of 90.5% and precision of 91.1% using both pressure and distension features. (4) Conclusions: Findings emphasize the crucial role of abnormality in the distension phase of peristalsis in FD patients.

1. Introduction

In total, 16% of the US population experience dysphagia, only half of whom seek medical care. The remainder manage their symptoms by modifying their diet [1]. X-ray barium esophagogram and endoscopy with biopsy to exclude eosinophilic esophagitis are the initial tests for dysphagia diagnosis. If the above tests are normal, high-resolution esophageal manometry (HRM) is recommended to diagnose the primary and secondary esophageal motility disorder [2]. HRM is the gold standard for the diagnosis of esophageal motility disorders because it is not feasible to obtain muscle tissue (biopsy) for final confirmation. HRM records intraluminal pressures every 1 cm apart and displays these pressures as a heatmap or topographical plots over time, along the entire length of the esophagus and its two sphincters. Studies done before and after the advent of HRM show that primary esophageal motility disorders such as achalasia, diffuse esophageal spasm, and nutcracker esophagus/jackhammer esophagus combined are seen in ~20% of patients. Another primary esophageal motility disorder, esophagogastric junction outflow obstruction (EGJOO), i.e., impaired LES relaxation (integrate relaxation pressure (IRP) >15) in the presence of normal peristalsis, is seen in 5–24% of patients with dysphagia [3]. However, only in a minority it is likely the cause of dysphagia because uncontrolled studies show that therapeutic strategies to address EGJOO (Botox, dilation, and myotomy) do not relieve patients’ symptoms. Therefore, in a significant number of patients, the cause of dysphagia remains obscure, and these patients are currently thought to have functional dysphagia, which implies sensory dysfunction of the esophagus [4,5,6,7].

The current classification system (Chicago Classification 4.0, CC) [8] employs a hierarchical algorithm to assign the HRM diagnosis. The interpretation process involves two stages that utilize feature-based analysis algorithms. In the first stage, outcomes of each swallow are determined using landmark identification and predefined metrics, leading to swallow-level diagnoses, based on esophageal contraction amplitude, pressurization, and esophagogastric junction (EGJ) pressure measurements. The second stage combines the swallow-level diagnoses and outcomes to generate the overall study diagnosis as per the Chicago Classification (CC) algorithm. There are two major problems with the CC approach of assessing esophageal motor function. The first hurdle is that the HRM interpretations are prone to inter-rater variability and inaccuracies due to the subjective nature of feature extraction, as evidenced by studies highlighting inconsistent landmark identification as a significant contributor to inter-rater disagreement [9]. This reliance on subjective experience poses challenges in medical training, as revealed by a recent survey showing a lack of confidence among GI fellows in interpreting esophageal manometry results [10]. A second and more significant issue arises from recent studies indicating that, in many patients, the abnormality lies in the distension rather than contraction phase of esophageal peristalsis [11,12]. Since HRM records only the contraction phase of peristalsis, relying solely on pressure parameters would not be an effective predictor in classifying esophageal motility disorders.

Functional dysphagia is a diagnosis of exclusion; it refers to difficulty in swallowing that is not caused by structural abnormalities or obviously identifiable physiological issues. Patients with functional dysphagia often exhibit symptoms such as difficult swallowing, a sensation of food sticking in the throat, or the regurgitation of food. The dysphagia score as determined by validated questionnaires in patients with functional dysphagia can be similar to patients with identifiable esophageal motility disorders such as achalasia esophagus, distal spasm, and esophagogastric junction outflow obstruction [13,14]. The esophagus has one simple function: to transfer swallowed contents to the stomach. To achieve the above, the esophagus must distend to a larger cross-sectional area than the swallowed bolus, and with adequate contraction strength to propel the bolus towards the stomach. However, determining distension and contraction parameters simultaneously has been a challenge. In the last 10 years, the authors have worked on measuring the esophageal cross-sectional area during peristalsis using the impedance principles [11,15,16,17,18]. We have developed a software called Dplots [19] that displays distension–contraction plots of the esophagus to extract relevant parameters. These parameters differ significantly between functional dysphagia patients and normal subjects [12]. The real significance of our work is that in the current scheme (Chicago Classification), in patients with functional dysphagia and even other esophageal motility disorders, the distension phase of peristalsis is not being considered. If diagnosed correctly, future therapeutic strategies, whether pharmacological or surgical, can be focused on improving the distension phase of peristalsis.

Artificial intelligence (AI) is a broad field, focused on developing systems with human-like intelligence. AI could potentially alleviate some of the issues associated with the clinical interpretation of HRM, as evidenced by its success in other fields of medicine [20,21]. AI is developing rapidly in healthcare [22,23,24,25,26]. In the last two decades, the use of AI has increased substantially in imaging studies in gastroenterology and hepatology [27,28,29]. The current focus of AI in gastroenterology has been on the detection of intestinal malignancies or premalignant lesions from endoscopy videos and X-ray images, e.g., the identification of premalignant or malignant lesions (e.g., the identification of esophageal adenocarcinoma in Barrett’s esophagus [30] and pancreatic malignancies [31], polyp identification and classification [32], small-bowel bleeding lesions on capsule endoscopy [33], pancreatic cystic lesions [34]), predicting disease prognosis or treatment responses, and determining the percentage of patients with inflammatory bowel disease (IBD) who will benefit from biologic therapy [35].

The lack of emphasis on the distension phase of peristalsis has been a significant contributing factor to the limited success of AI methods in the automated diagnosis of esophageal motility disorders [28,36,37,38,39,40,41,42,43,44,45,46]. Shallow learning (SL), a subset of AI, involves algorithms with limited layers and handcrafted features. Shallow learners offer certain advantages over deep learners in AI, particularly in scenarios where simplicity and computational efficiency are essential. Moreover, their straightforward nature makes them easier to interpret and understand, which can be crucial in certain domains where interpretability is essential, such as in regulatory compliance or medical applications. Additionally, shallow learners tend to have a lower risk of overfitting when dealing with small datasets, making them more robust in situations where data availability is constrained.

2. Materials and Methods

2.1. Study Population

We conducted a study involving two distinct groups of participants. The first group consisted of 30 healthy asymptomatic individuals (with an average age of 37 years, ranging from 21 to 65 years, and including 12 males). The second group comprised 30 patients who were diagnosed to have functional dysphagia (FD) (average age of 54 ± 13 years, including 6 males). None of the individuals in the healthy group exhibited any symptoms related to the esophagus. The patients, on the other hand, were referred to the GI function laboratory at the University of California San Diego for an esophageal manometry study, which aimed to determine the underlying cause of their dysphagia. The diagnosis of functional dysphagia was based on a process of exclusion, meaning that patients had dysphagia and met the following criteria: (1) a normal upper endoscopy, (2) normal esophageal biopsy, and (3) normal results on high-resolution impedance manometry (HRMZ). The HRMZ tests were conducted using the Manoscan system developed by Medtronic Inc., located in Minneapolis, MN, USA, and results were interpreted according to the Chicago Classification (version 4.0) for categorizing esophageal motility disorders [47]. The human investigation committee of the University of California San Diego approved the study protocol (IRB # 182156 and 202106X) and subjects signed informed consent forms prior to their participation in the study protocol. Participants filled out a standardized questionnaire, recording details such as age, gender, body mass index (BMI), and symptoms before undergoing the manometry study. The degree of dysphagia was evaluated using a validated and standardized questionnaire known as the brief esophageal dysphagia scoring (BEDS) questionnaire [48], and only patients with a BEDS score of >10 were included in the study.

2.2. HRMZ Recordings

Subjects underwent the examination using an HRMZ catheter (with a diameter of 4.2 mm; manufactured by Medtronic Inc., MN). This catheter was equipped with 36 pressure transducers spaced 1 cm apart and 18 impedance electrodes spaced 2 cm apart. To ensure local anesthesia, viscous lidocaine (2% lidocaine hydrochloride topical solution, USP) was administered both orally and nasally. Subsequently, the catheter was placed through the nasal passage. The assessment proceeded according to a routine clinical manometry study protocol. This protocol involved placing patients in a supine position and having them perform 10 swallows, each involving 5 mL of 0.5 N saline solution at room temperature. Following the above, the examination setup was adjusted by tilting the stretcher to an angle of −15 degrees (Trendelenburg position). In this position, the patients underwent 8 to 10 swallows of 10 mL, 0.5 N saline, which was heated to a temperature of 37 degrees Celsius. The utilization of the Trendelenburg position facilitated the precise measurement of the luminal cross-sectional area (CSA) using the impedance data obtained from the HRMZ recordings, because swallowed air, being lighter than liquid, is separated from saline during the passage of the bolus through the esophagus, due to gravity, which improves the accuracy of the CSA measurement by the impedance technique. All solutions were prepared fresh before the study and warmed to body temperature (37 °C). The conductivity of all solutions was checked in vitro using an Omega model CDH221 conductivity meter (Omega Engineering Inc., Norwalk, CT, USA) prior to each study.

2.3. Data Analysis

For analysis, we selected 5–7 (10 cc) swallows that demonstrated full clearance of bolus, as indicated by impedance recordings. The HRMZ data were displayed using the Manoview program (developed by Medtronic Inc., MN, USA). The relevant data from the chosen swallows were extracted as text files. Utilizing the Dplots software (Motilityviz, La Jolla, CA, USA), which offers user interactivity, the program extracted specific features from the multi-channel pressure and impedance signals obtained from the HRMZ recordings. Dplots is the first commercial software that allows (1) the visualization of distension–contraction color plots of esophageal peristalsis in several formats (Figure 1 and Figure 2) and (2) the quantitation of relevant parameters. For each selected swallow, the region of interest spanned between the lower border of the upper esophageal sphincter (UES) and the contraction deceleration point (CDP) of swallow-induce peristalsis, and the time duration between the onset of UES relaxation and return of LES pressure back to baseline after swallow-induced peristalsis. Out of the many features extracted by Dplots, the following parameters were used in our study.

• Peak distension time (T1): The time duration starting from the initiation of UES relaxation caused by swallowing to the point of lowest impedance (which is the same as peak distension) recorded at each specific location within the esophagus.
• Peak pressure time (T2): The time duration between the commencement of UES relaxation triggered by swallowing and the point of highest pressure reached at that location.
• Amplitude of peak esophageal distension: The largest luminal cross-sectional area (CSA) during the distension phase of peristalsis.
• Amplitude of peak esophageal contraction: The maximal luminal pressure during the contraction phase of peristalsis.
• Sum of distension: Values falling within the distension waveform (same as the area under the curve).
• Sum of contraction: Values falling within the contraction waveform (same as the area under the curve).
• Standard deviation of esophageal contraction wave.
• Standard deviation of esophageal distension wave.

Figure 1. Sample distension–contraction (0.45 N, 10 cc, Trendelenburg) of (a) a normal subject and (b) a functional dysphagia (FD) patient, saline swallow.

Figure 1. Sample distension–contraction (0.45 N, 10 cc, Trendelenburg) of (a) a normal subject and (b) a functional dysphagia (FD) patient, saline swallow.

Figure 2. Transit of the bolus in (a) a normal subject and (b) a functional dysphagia patient (10 cc, TB).

Figure 2. Transit of the bolus in (a) a normal subject and (b) a functional dysphagia patient (10 cc, TB).

2.4. Shallow Learners

We used multiple state-of-the-art shallow learners to achieve our goals, namely SVMs, random forest, K-nearest neighbors (KNN), and logistic regression [49]. Shallow learners, also known as shallow or weak classifiers, are machine learning algorithms that have limited depth or complexity in their decision-making process.

• Support vector machine (SVM) is a shallow learning algorithm that separates data into different classes by creating an optimal hyperplane in a high-dimensional space. Linear SVM aims to find a hyperplane that best separates different classes in the input data by maximizing the margin between them. It works well when the data are linearly separable. On the other hand, radial basis function (RBF) SVM uses a non-linear transformation to map data into a higher-dimensional space, where it seeks to find a hyperplane that effectively separates classes even when they are not linearly separable in the original feature space. The RBF kernel computes the similarity between data points using the Gaussian function, enabling it to capture complex relationships. While linear SVM is simpler and more interpretable, RBF SVM is more flexible and can handle intricate patterns in data, making the choice between them dependent on the nature of the problem and the dataset at hand.
• Random forest is another shallow learning method that constructs multiple decision trees and combines their predictions to make a final classification. It leverages the concept of ensemble learning to improve accuracy and can handle high-dimensional data. It operates by constructing multiple decision trees during training and combines their predictions to make more accurate and robust predictions. Each decision tree is trained on a different subset of data and employs a randomized selection of features. This randomness helps to reduce overfitting and increases the diversity of the individual trees, leading to better generalization. During prediction, the random forest aggregates the outputs of these trees to arrive at a final prediction, whether it is a class label in classification or a numerical value in regression. The above approach makes random forest resilient to noise, outliers, and missing values, and it often performs well on a wide range of tasks without requiring extensive hyperparameter tuning.
• K-nearest neighbors (KNN) is a simple yet effective shallow learning algorithm that classifies new instances based on the majority vote of its K-nearest neighbors in the feature space. It works by finding the K data points in the training dataset that are closest to a given input point and making predictions based on the majority class (in classification) or the average value (in regression) of these K neighbors. KNN’s strength lies in its ability to capture local patterns in the data, which makes it suitable for tasks where data points of similar characteristics tend to have similar outcomes. However, KNN’s performance can be sensitive to the choice of the distance metric and the value of K, and it might struggle with high-dimensional or noisy data. Regularization techniques and preprocessing steps like feature scaling can help to improve KNN’s performance in various scenarios.
• Finally, we utilized logistic regression (LR), another shallow learning algorithm, for our analysis. It operates by modeling the log-odds of the probability of the positive class using a linear combination of input features. The logistic (sigmoid) function is then applied to the linear combination to constrain the output between 0 and 1, representing the probability of the positive class. Model parameters are iteratively optimized through techniques like maximum likelihood estimation, aiming to minimize the log-loss function that quantifies the disparity between predicted probabilities and actual class labels. Regularization techniques such as L1 or L2 regularization can be incorporated to prevent overfitting. Logistic regression is relatively straightforward, computationally efficient, and allows for coefficient interpretation, making it suitable for scenarios with linearly separable classes.

The algorithms were trained on 2 annotated datasets, normal and functional dysphagia patients, optimizing the hyperparameters through cross-validation. Model performance was evaluated using various metrics, including accuracy, precision, and recall.

3. Results

A sample swallow comparing a normal subject to a functional dysphagia patient is shown in Figure 1 and Figure 2. Note the differences in the pattern of distension between the two subjects. The bolus arrives faster in the distal esophagus (a shorter T1) and luminal distension (luminal cross-sectional area) is smaller in the patient as compared to the normal subject. To evaluate the overall performance of the classification models, a hold-out approach was employed. The dataset was divided into two parts, a training set, and a test set, using five-fold cross-validation and an 80–20 train–test split. The five classification models (LR, RBF SVM linear, KNN, and random forest) were compared to determine which had the best performance using different evaluation metrics, such as accuracy, precision, and recall. The results are shown in

Table 1. Classification results of the different AI shallow learners using pressure (P) and distension (CSA) features, showing accuracy (Acc), precision (Pre), and recall (Re).

		LR			KNN			Linear SVM			RBF SVM			Random Forest
Seg 1	%	Acc	Pre	Re	Acc	Pre	Re	Acc	Pre	Re	Acc	Pre	Re	Acc	Pre	Re
	P&D	86.11	87.14	86.11	91.67	93.45	91.67	86.11	86.51	86.11	91.67	92.87	91.67	91.67	92.86	91.67
	P	72.22	75.24	72.22	77.78	83.33	77.78	61.11	53.1	61.11	72.22	73.39	72.22	66.67	71.81	66.67
	D	91.67	92.86	91.67	88.89	91.67	88.89	80.56	80.79	80.56	83.33	86.51	83.33	88.89	91.07	88.89
Seg 2	P&D	72.22	73.93	72.22	77.78	83.33	77.78	77.78	85.4	77.78	77.78	78.47	77.78	88.89	89.68	88.89
	P	63.89	63.44	63.89	80.56	81.43	80.56	63.89	68.52	63.89	77.78	79.64	77.78	77.78	80.49	77.78
	D	80.56	83.57	80.56	80.56	83.57	80.56	86.11	86.51	86.11	83.33	84.72	83.33	83.33	84.72	83.33
Seg 3	P&D	83.33	83.33	83.33	86.11	87.14	86.11	94.44	95.24	94.44	94.44	94.44	94.44	88.89	89.52	88.89
	P	52.78	53.39	52.78	72.22	73.93	72.22	63.89	66.2	63.89	77.78	79.64	77.78	72.22	75.93	72.22
	D	88.89	91.07	88.89	97.22	97.62	97.22	100	100	100	91.67	94.44	91.67	97.22	97.62	97.22
Seg 4	P&D	85.71	86.94	85.71	88.1	88.89	88.1	85.71	86.21	85.71	80.95	81.35	80.95	90.48	91.07	90.48
	P	61.11	61.27	61.11	72.22	74.31	72.22	66.67	67.68	66.67	75	77.32	75	61.11	61.27	61.11
	D	86.11	86.51	86.11	96.11	86.51	86.11	83.33	83.33	83.33	86.11	86.51	86.11	88.89	89.68	88.89

. The significance of our approach lies in its inclusion of functional dysphagia as a distinct group within the classification framework. To obtain a visual insight into the distribution of some of the pressure and distension features, the averages of some features are shown in Figure 3.

In seg 1, using only distension-related features (i.e., T1, peak distension, AUC distention, and STD of distension values), LR had the best overall results using only distension, with accuracy of 91.7%, precision of 92.86%, and recall of 91.7%. Next, using only pressure-related features (i.e., T2, peak pressure, AUC pressure, and STD pressure), KNN had the best overall results, with accuracy of 77.78%, precision of 83.33%, and recall of 77.78%. Finally, using both pressure- and distension-related features, KNN had the best overall results, with accuracy of 91.67%, precision of 93.45%, and recall of 91.67%.

In seg 2, again using only distension features (i.e., CSA), linear SVM produced the best overall results, with accuracy of 86.1%, precision of 86.5%, and recall of 86.1%. Next, using only pressure parameters, KNN had the best overall results, with accuracy of 80.56%, precision of 81.43%, and recall of 80.56%. Finally, using both pressure- and distension-related features, random forest had the best overall results, with accuracy of 88.89%, precision of 89.68%, and recall of 88.89%.

In seg 3, once again using only distension features (i.e., CSA), linear SVM produced the best overall results, with perfect accuracy of 100%, precision of 100%, and recall of 100%. Next, using only pressure-related features, RBF SVM had the best overall results, with accuracy of 77.78%, precision of 79.64%, and recall of 77.78%. Finally, using both pressure- and distension-related features, linear SVM had the best overall results, with accuracy of 94.44%, precision of 95.24%, and recall of 94.44%.

Finally, in seg 4, using both contraction- and distension-related features, random forest produced the best overall results, with accuracy of 90.5%, precision of 91.1%, and recall of 90.5%. Using both pressure and distension features, it produced the highest results in the most distal segment. Feature importance was extracted using the Gini Index (GI) [50]. The two top features were T1 (GI = 0.22) and AUC distension (GI = 0.19) in the distal segment. Using only pressure parameters, RBF SVM had the best overall results, with accuracy of 75%, precision of 77.32%, and recall of 75%. Finally, using distension-related features, KNN had the best overall results, with accuracy of 96.11%, precision of 86.51%, and recall of 86.11%.

4. Discussion

In this study, we utilized AI shallow learners, commonly known as traditional machine learning algorithms, to distinguish between normal subjects and functional dysphagia patients for the first time and achieved high classification accuracies of more than 90%. As mentioned earlier, the diagnosis of functional dysphagia has traditionally been a diagnosis of exclusion. Our data show a prominent distension abnormality in functional dysphagia patients. We achieved sensitivity (recall) of higher than 90% for the FD group, and this separation was mostly observed in the proximal (close the UES) and distal (close to the LES) segments. Our results indicate that adding distension features significantly improves the classification results compared to using only pressure features. It appears that the primary abnormality in functional dysphagia patients lies in the distension phase of peristalsis rather than the contraction phase.

The above being said, a significant issue with shallow learners lies in their incapability to capture intricate patterns and connections within data that are not explicitly presented. These models lack the complexity required to comprehend subtle nuances and hierarchical structures within the data, thus making them less suitable for tasks involving unstructured or nonlinear data. Moreover, shallow learners frequently encounter difficulties with feature engineering, necessitating the manual extraction and selection of pertinent features, which can consume time and is susceptible to errors. Additionally, the challenge of limited sample sizes is worth noting. When dealing with small datasets, as in our study, classifiers might struggle to encompass the full spectrum of variability inherent in the underlying data distribution, which can lead to overfitting, where the model learns from noise and outliers in the training data, resulting in poor adaptation to new, unfamiliar data. Another facet of the generalization problem introduces another hurdle, i.e., a classifier that demonstrates proficiency on the training data may not necessarily perform well on new data points. This predicament arises due to the model’s limited capacity to grasp intricate relationships within the data, particularly in scenarios involving nonlinear boundaries or spaces with high dimensions. Consequently, the classifier might either fail to capture significant patterns (underfitting) or become overly focused on the training data (overfitting), thereby failing to generalize to novel instances. In situations where the sample size is substantial, resorting to more advanced classifiers, like ensemble methods or deep learning architectures, could be considered.

In our study, due to the limited sample size of the two groups, we aimed to maintain a balance and implemented meticulous regular cross-validation and hyperparameter tuning across the classifiers, which are essential measures to ensure that a classifier can effectively extend its learned patterns to unseen data. As a result, we achieved favorable outcomes. Our study did not ascertain the precise origin of the distension anomaly in the groups of patients that we examined. Another noteworthy aspect to highlight is that our findings reveal the significance of certain characteristics, such as the standard deviation of peak pressure, in distinguishing between the two groups. This contrasts with the Chicago Classification, where such differentiation is not applicable. In the Chicago Classification, the pressure is measured as the distal contractile integral (DCI), which is the integrated value of contraction pressure over a 10 cm length of the distal esophagus (above the lower esophageal sphincters), irrespective of the actual length of the esophagus. The latter may differ based on the height of the individual, which is not a consideration in the Chicago Classification. Our approach overcomes the Chicago Classification’s shortcoming by dividing the esophagus into four equal compartments, thus normalizing across different patients of different heights.

5. Conclusions

In this study, we utilized traditional machine learning algorithms, often referred to as AI shallow learners, to achieve novel and accurate differentiation between normal subjects and functional dysphagia patients, obtaining classification accuracies exceeding 90%. Unlike the historical approach of diagnosing functional dysphagia through exclusion, our findings highlight a distinct distension abnormality in these patients, particularly evident in proximal and distal segments. Remarkably, incorporating distension features significantly enhances the classification outcomes compared to using pressure features alone, underscoring the role of the distension phase over the contraction phase in functional dysphagia pathology. However, the limitations of shallow learners lie in their inability to grasp intricate patterns within unstructured or nonlinear data, demanding manual feature engineering and risking overfitting, especially with small datasets. To address this, we emphasize the potential of advanced classifiers like ensemble methods or deep learning in scenarios with larger sample sizes. In our study, careful cross-validation and hyperparameter tuning counteracted limited sample size concerns, resulting in successful classification. Notably, while the origin of the distension anomaly remains unclear, our work underscores the importance of specific characteristics, like peak pressure standard deviation, in distinguishing groups, addressing the limitations of existing classifications. Our method’s division of the esophagus into equal compartments circumvents height-related disparities, providing a valuable alternative to the Chicago Classification.

The results obtained from this study are highly promising and could pave the way for advancements in the diagnosis and management of patients with functional dysphagia. Based on our findings, we observed that a significant problem in patients with dysphagia is impaired distension of the esophagus during peristalsis, while subtle or no contraction abnormalities were noted in these patients. Future studies may explore the use of pharmacologic approaches, esophageal dilation, and possibly surgical interventions targeting the distension phase of peristalsis.