Electrophysiological biomarkers for foundational signal markers in chronic pancreatitis using machine learning models

Most patients with chronic pancreatitis (CP) experience abdominal pain, but the pathophysiology differs. We aimed to investigate whether using only electroencephalography (EEG) and electrocardiography (ECG) could be used to identify distinct differences between healthy controls and people with CP with and without pain.

This study utilised cross-sectional data from a multicentre research project, including healthy controls (n = 35) and patients with CP (n = 124), with (n = 94) and without (n = 30) pain. EEG and ECG recordings were analysed while resting and in experimental pain states. Analysis was performed using the machine learning models: decision tree classifier, random forest classifier, logistic regression, and support vector machine. The features used in the machine learning models were statistical-, entropy-, and fractal-feature extraction.

Classifying differences between all CP patients and controls yielded 78% F1-score (both EEG and ECG) using a support vector machine. Classifying differences between CP patients with and without pain was less precise, with a maximum 63% F1-score (both EEG and ECG) using a support vector machine. Overall, EEG and ECG improved classifications by up to 10 percentage points. Adding the cold pressor (tonic pain) condition did not reliably change the separation between patients with and without pain, which remained near chance.

The classification helps to show the importance of potential EEG and ECG measures in characterization of chronic pancreatitis and pain. Future work should focus on incorporating other disease-specific features in the model.