Decoding human somatosensory sensitivity through resting EEG and behavioral analysis: a multimodal fusion approach
Hao-Yuan Lin,Congying He,Cheng-Hua Su,Li-Ling Hope Pan,Fu-Jung Hsiao,Yu-Te Wu,Yen-Feng Wang,Shuu-Jiun Wang,Li-Wei Ko
DOI: https://doi.org/10.1109/TNSRE.2024.3434353
2024-07-29
Abstract:In precision medicine and clinical pain management, the creation of quantitative, objective indicators to assess somatosensory sensitivity was essential. This study proposed a fusion approach for decoding human somatosensory sensitivity, which combined multimodal (quantitative sensory test and neurophysiology) features to classify the dataset on individual somatosensory sensitivity and reveal distinct types of brain activation patterns. Sixty healthy participants took part in the experiment on somatosensory sensitivity that implemented cold, heat, mechanical punctate, and pressure stimuli, and the resting-state electroencephalography (EEG) was collected using BrainVision. The quantitative sensory testing (QST) scores of the participants were clustered using the unsupervised k-means algorithm into four subgroups: generally hypersensitive (HS), generally non-sensitive (NS), predominantly thermally sensitive (TS), and predominantly mechanically sensitive (MS). Furthermore, two types of power spectral density (PSD), band-based PSD (BB-PSD) and frequency-based PSD (FB-PSD), and two types of inter-electrode connectivity (IEC), band-based connectivity (BBC) and frequency-based connectivity (FBC), derived from resting-state EEG were subjected to feature selection with a proposed prior-compared minimum-redundancy maximum-relevance (PCMRMR) protocol. Their effectiveness was then tested by the supervised classification tasks using support vector machine (SVM), k-nearest neighbor (kNN), random forest (RF), and Gaussian classifier (GC). Brain networks of four somatosensory types were revealed by decoding fused multimodal data, namely type-averaged connectivity. The data from sixty healthy individuals were divided into training (n = 59) and validation (n = 1) datasets according to leave-one-subject-out (LOSO) criteria. The FBC was identified, which can serve as better brain signatures than BB-PSD, FB-PSD, and BBC to classify subjects as HS, NS, TS, or MS groups. Using the SVM, kNN, RF, and GC models, the best accuracy of 87% was obtained when classifying participants into HS, NS, TS, or MS groups. Moreover, the brain networks were decoded from HS, NS, TS, and MS groups by decoding the type-averaged connectivity fused from somatosensory phenotypes and selected FBC. It indicated that quantified multi-parameter somatosensory sensitivity could be achieved with acceptable accuracy, leading to considerable possibilities for using objective pain perception evaluation in clinical practice.