IS THE PERISTALTIC PRESSURE PROFILE IN FUNCTIONAL DYSPHAGIA PATIENTS TRULY NORMAL? UNCOVERING NOVEL PERSPECTIVES THROUGH DEEP LEARNING ARTIFICIAL INTELLIGENCE

Introduction: About 16% of the US population experiences dysphagia symptoms, but only half of them seek medical attention. Standard diagnostic tests involve X-ray barium swallow and endoscopy with biopsy. If these tests yield normal results, high-resolution esophageal manometry (HRM) and the Chicago Classification (CC) method are employed to assess esophageal peristalsis. Surprisingly, 30-50% of patients with dysphagia symptoms exhibit normal HRM studies. Our hypothesis is that the CC method's chosen manometric parameters might unintentionally lead to the misclassification of patients as normal, raising questions about its accuracy in identifying esophageal dysfunction in certain cases. Aims: To assess whether a holistic deep learning approach, capturing comprehensive pressure topograph dynamics may reveal concealed dynamics and pressure differentiations in patients with functional dysphagia, i.e., dysphagia in the setting of normal HRM, as defined by Chicago Classification. Methods: The HRM studies of 20 normal subjects and 20 functional patients with a brief esophageal dysphagia score of > 4, and a normal high resolution esophageal manometry (HRM) study with 8 swallows of 5ml, 0.5N bolus in the supine position were analyzed (Figure 1). Four cutting-edge deep learning methods (DenseNet121, InceptionV3, VGGNet16, ResNet50) analyzed pressure data related to swallow events. DenseNet121 emphasizes dense connections, InceptionV3 uses inception modules, VGGNet16 employs simple convolutional layers, and ResNet50 introduces residual connections to mitigate vanishing gradient problems. These methods were chosen for advanced capabilities in pattern recognition and complex data analysis. Results: All patients were classified as normal based on the CC method. However, using pressure alone, ResNet50 revealed differences between the 2 groups with accuracy, precision, and recall (i.e., sensitivity) of 91%, 91% and 91% respectively. All other neural networks achieved results larger than 87%. Above suggests that the pressure patterns are different in patients with “functional dysphagia” and are overlooked by the parameters used in the CC method of the diagnosis. Conclusion: This is the first time where pressure parameter has been shown to be significantly different between normal and “FD” patients. Feature selection often involves human domain knowledge to identify and choose relevant features. The CC’s feature selection approach may fall short in capturing the true dynamics of FD patients, leading to misclassification. However, DL automatically learns representations without heavy reliance on explicit feature engineering. This study advocates the need for comprehensive, new data-driven artificial intelligence approaches to improve the accuracy of functional dysphagia diagnosis.