Detecting and clustering swallow events in esophageal long-term high-resolution manometry

TL;DR

This paper tackles the challenge of long-term esophageal HRM data analysis by introducing a CNN-based swallow detector that operates across the full sensor field, followed by clustering of detected swallows into clinically meaningful classes. On 25 long-duration HRM recordings, the ML detectors—especially MobileNet—achieve high recall (~94%) and strong precision, outperforming a threshold-based baseline and a commercial tool. The clustering step reduces manual review time and significantly improves inter-rater reliability among clinicians (κ rising from 0.53 to 0.73). Collectively, the approach makes long-term HRM more practical for clinical care and sets the stage for automatic swallow-type classification in future work.

Abstract

High-resolution manometry (HRM) is the gold standard in diagnosing esophageal motility disorders. As HRM is typically conducted under short-term laboratory settings, intermittently occurring disorders are likely to be missed. Therefore, long-term (up to 24h) HRM (LTHRM) is used to gain detailed insights into the swallowing behavior. However, analyzing the extensive data from LTHRM is challenging and time consuming as medical experts have to analyze the data manually, which is slow and prone to errors. To address this challenge, we propose a Deep Learning based swallowing detection method to accurately identify swallowing events and secondary non-deglutitive-induced esophageal motility disorders in LTHRM data. We then proceed with clustering the identified swallows into distinct classes, which are analyzed by highly experienced clinicians to validate the different swallowing patterns. We evaluate our computational pipeline on a total of 25 LTHRMs, which were meticulously annotated by medical experts. By detecting more than 94% of all relevant swallow events and providing all relevant clusters for a more reliable diagnostic process among experienced clinicians, we are able to demonstrate the effectiveness as well as positive clinical impact of our approach to make LTHRM feasible in clinical care.

Figures (6)

• Figure 1: Overview of our computational pipeline, consisting of a swallow detector and a subsequent clustering of the detected swallows.
• Figure 2: The training and inference procedure of our swallow detection method.
• Figure 3: Comparison of different clustering options. Left: Difference between using the pure manometry values and applying the change filter. Right: Comparison if different clustering methods - agglomerative clustering and k-means achieve similarly distinctive clusters, while DTW based k-means results in slightly less distinctive clusters.
• Figure 4: Clustering of all detected swallows for a single patient. Special classes with an occurrence $<15\%$ are clustered more granularly on the right side.
• Figure 5: The distance between the detected swallow and the correct swallow start for 5 patients. It can be observed, that the distances of the MobileNet outputs are centered around -35, indicating that our approach is typically predicting the start of a swallow slightly earlier compared to the labelled start. The distances of the ViMeDat outputs are centered around 0, indicating that the software is typically predicting the swallow exactly at the true swallow start. The baseline, as it is looking at high pressure events which mostly occur during the swallow, is predicting the swallow event after the swallow start. It can be observed that most of the times a high pressure event is either occurring rather soon after the labeled swallow start (50 measurements or 1 second), or after around 270 measurements or 5-6 seconds after the swallow start.