Robust Methods to Detect Abnormal Initiation in the Gastric Slow Wave from Cutaneous Recordings
| Autor: | Todd P. Coleman, David C. Kunkel, Alexis B. Allegra, Anjulie S. Agrusa |
|---|---|
| Rok vydání: | 2020 |
| Předmět: |
medicine.medical_specialty
Gastroparesis 0206 medical engineering Population 02 engineering and technology Convolutional neural network 03 medical and health sciences 0302 clinical medicine Smooth muscle Internal medicine Wave frequency medicine Upper gastrointestinal Humans Computer Simulation Dyspepsia education Electrodes education.field_of_study business.industry Stomach medicine.disease 020601 biomedical engineering medicine.anatomical_structure Cardiology 030211 gastroenterology & hepatology Abnormality business |
| Zdroj: | EMBC |
| ISSN: | 2694-0604 |
| Popis: | Upper gastrointestinal (GI) disorders are highly prevalent, with gastroparesis (GP) and functional dyspepsia (FD) affecting 3% and 10% of the US population, respectively. Despite overlapping symptoms, differing etiologies of GP and FD have distinct optimal treatments, thus making their management a challenge. One such cause, that of gastric slow wave abnormalities, affects the electromechanical coordination of pacemaker cells and smooth muscle cells in propelling food through the GI tract. Abnormalities in gastric slow wave initiation location and propagation patterns can be treated with novel pacing technologies but are challenging to identify with traditional spectral analyses from cutaneous recordings due to their occurrence at the normal slow wave frequency. This work advances our previous work in developing a 3D convolutional neural network to process multi-electrode cutaneous recordings and successfully classify, in silico, normal versus abnormal slow wave location and propagation patterns. Here, we use transfer learning to build a method that is robust to heterogeneity in both the location of the abnormal initiation on the stomach surface as well as the recording start times with respect to slow wave cycles. We find that by starting with training lowest-complexity models and building complexity in training sets, transfer learning one model to the next, the final network exhibits, on average, 80% classification accuracy in all but the most challenging spatial abnormality location, and below 5% Type-I error probabilities across all locations. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|