UC San Diego

One of the primary components of GI motility is the gastric slow wave, which travels along the surface of the stomach and locally actuates smooth muscle contractions, known as peristalsis. Pathological slow wave patterns, detected through intrasurgical recordings, have been shown in high incidence with GI disease and symptoms. The ability to non-invasively detect pathological slow wave patterns is of high clinical interest, as this could significantly enhance diagnostics and lead to tailored treatment regimens. In four parts, we develop statistical analysis methods to infer spatial slow wave patterns from cutaneous high-resolution electrogastrogram (HR-EGG) waveforms.

We first establish a deep three-dimensional convolutional neural network (3D-CNN) framework to classify normal and abnormally initiated slow waves from in silico multi-electrode waveforms. We develop a linear discriminant classifier using spatial features of wave propagation to compare performance. We report the first successful classification study and show robustness across recording and anatomical variability.

Second, we develop a transfer learning framework, in silico, to build a 3D-CNN classifier robust to heterogeneity in both the location of the slow wave abnormal initiation as well as the recording start times with respect to slow wave cycles. We find that by building complexity in training sets, transfer learning one model to the next, the final network exhibits high classification accuracy on data with closer resemblance to human physiologic recordings, suggesting downstream potential for use as widely deployable clinical screening tools for GI disorders.

Third, we establish an optimal transport theory-based philosophical framework for the detection of anomalous probability distributions. We demonstrate a rigorous mathematical proof and show an example algorithm within the framework of our methodology to detect an anomaly in a set of probability distributions that model slow wave direction estimates from HR-EGG recordings on human subjects.

Finally, we implement robust regression and clustering methods on optimal transport-based features of probability density estimates of slow wave directions from HR-EGG recordings on human subjects. We report highly accurate robust regression and suggest the ability to predict treatment response and underlying disease etiology. This methodology may provide clinicians with an opportunity to screen patients and optimize treatment regimens.