Classifying chronic pain conditions using deep learning and resting-state fMRI

Introduction: Due to its comorbidities with other symptoms and lack of effective treatments, chronic pain is considered a complex disease. At the same time, advancements in brain imaging, the spread of machine learning methods, and the creation of new diagnostic tools based on these technologies have demonstrated that these tools may be an option for assisting healthcare professionals in making decisions. However, a growing body of research in neuroimaging suggests that functional networks show dynamic changes in connection strength as well as variable phase difference (nonzero time-lag) between regions. Our goal is to compare performances in developing and evaluating deep learning (DL) and traditional machine learnings (MLs) in predicting chronic pain.Methods: A resting state functional MRI (rsfMRI) data with pain status were obtained from UK BioBank (UKBB) 6500 dataset. We computed functional brain activity to evaluate how well the 6 ML models (i.e., K-Nearest Neighbors (KNN), Logistic Regression (LR), Random Forest (RF), XGBoost (XGB), AdaBoost (Ada), LightGBM (GBM) and convolutional neural network (CNN), categorized chronic pain patients and pain-free controls in this study. They were compared using Area Under the Curve (AUC).Results: Training a CNN with preprocessed images produced the best results. CNN has an AUC of 0.69 compared to KNN (AUC = 0.52), LR (AUC = 0.55), RF (AUC = 0.54), XGB (AUC = 0.54), Ada (AUC = 0.52), and GBM (AUC = 0.54).Conclusion: The idea that resting-state fMRI data could be used as a biomarker for chronic pain conditions is supported by these findings.