A deep learning-based approach to enhance accuracy and feasibility of long-term high-resolution manometry examinations

Background High-resolution manometry (HRM) is the gold standard for diagnosing esophageal motility disorders. However, its short-term laboratory setting often fails to capture intermittent abnormalities. Long-term HRM (LTHRM, up to 24h) provides richer insights into swallowing behavior, but the resulting data volume is immense. Manual analysis by medical experts is laborious, time-consuming, and prone to errors, limiting its clinical feasibility. Methods We propose a deep learning-based approach for automatic analysis of LTHRM data. Our method detects both swallow events and secondary non-deglutitive motility disorders with high accuracy. Detected swallows are then clustered into distinct classes of similar events, creating a structured overview of motility patterns and their frequency. This reduces the analytical burden by allowing clinicians to focus on a small number of representative swallows rather than manually reviewing thousands of individual events. We evaluate our pipeline on 25 LTHRMs that were meticulously annotated, resulting in a dataset of more than 23,000 expert-labeled events. Results Our approach is able to detect more than 94% of all relevant events in LTHRM sequences, while the subsequent clustering is able to capture and group all relevant events into distinct swallow groups. To evaluate the overall approach, we conduct a user study with medical experts, demonstrating its effectiveness and positive clinical impact. Conclusions Our findings demonstrate that deep learning-based approaches to analyze LTHRM examinations are capable of providing a more reliable and efficient diagnostic process, ultimately making LTHRM assessments more feasible in clinical care.