Challenge Big Data in Biomedicine: Long-term High-Resolution Impedance Manometry (LTHRIM) of the Esophagus
Dr. Alissa Jell,
Klinikum rechts der Isar der TUM
High-resolution manometry of the esophagus is the gold standard in the diagnosis of esophageal motility disorders . As the measurements take place in clinic for a very limited period of time only little is known about circadian changes and there is no recording of intermittent symptoms in esophageal motility. However, disorders of esophageal motility can lead to severe symptoms even if they occur (only) sporadically.
For this purpose, a combined impedance-manometry probe with 36 pressure sensors and 17 impedance fields (Unisensor AG, Swiss) with a mobile data logger (MALT, Standard Instruments, Karlsruhe) were used. With a sampling rate of 50 hertz and a prolonged examination time of a full of 24 hours, very large amounts of data (8GB, respectively about 212 mio. entries per patient) are quickly produced. In the case of manual evaluation, an experienced physician needs three working days to evaluate a long-term data recording containing about 800 to 1500 acts of swallowing as well as a various number of esophageal motility phenomena (e.g. diffuse esophageal spasm, tertiary peristalsis, …).
Therefore, we implemented a combined matlab and python scripting for extracting every pressure event of the whole examination. In the first step an automated swallowing detection algorithm based on applying Bayes’ theorem  combined with a loss-and-risk-analysis was developed. After cross-validation and minimizing the cv-risk we were able to establish a risk of 0,11 to our dataset compared to the risk of naive bayes with 0,61. A typical long-term measurement generates around 2000 images which were standardized extracted in 640x480 pixel2 images (fig. 1) and used for clustering. Orange, a python library, commonly used for image embedding, was modified and trained with characteristic volunteer data sets as well as defined diseases (e.g. achalasia, Zenker’s diverticula, fig. 2).
We validated the most common imaging neuronal networks (Inception v3, VGG16, VGG 19, Squeeze Net, Painters, DeepLoc) for inter- and intra-observer accordance.
1. Roman S, Pandolfino J, Mion F. High-resolution manometry: a new gold standard to diagnose esophageal dysmotility? Gastroenterologie clinique et biologique 2009; 33: 1061-1067
2. Rish, I. An empirical study of the naive Bayes classifier. IJCAI Workshop on Empirical Methods in AI. 2001