关键词: 疼痛, 认知偏向, 事件相关电位, 漂移扩散模型, 机器学习

Abstract: Previous studies have shown that chronic pain is associated with pain-related cognitive biases. However, how a painful state reshapes attentional, interpretive, and memory processing, and whether these biases can represent individual pain state and pain intensity, remains unclear. The present study combined drift-diffusion modeling (DDM), event-related potentials (ERPs), and machine-learning approaches to investigate the behavioral and neural mechanisms of cognitive biases during a painful state, and to examine their value in pain decoding. Ninety healthy participants were assigned to either a pain group or a control group. Sustained experimental pain was induced using topical capsaicin in the pain group, whereas the control group received a matched control cream. Participants completed three cognitive-bias tasks during EEG recording: a modified spatial cueing task assessing attentional bias, an incidental learning task assessing interpretation bias for ambiguous faces, and a recognition task assessing memory bias for pain-related pictures. DDM was used to characterize latent decision processes underlying behavioral performance. ERP analyses were conducted to examine the temporal dynamics of cognitive-bias processing, including early attentional processing and later evaluative stages. Cross-task DDM and ERP features were further entered into machine-learning models for pain-state classification and pain-intensity prediction. The results showed that pain state did not globally enhance all pain-related cognitive biases. Instead, it mainly altered the interpretation of ambiguous information and the attentional processing pattern associated with pain-related cues. Specifically, pain state weakened the happy interpretation advantage for ambiguous faces, making their evidence accumulation, pre-decisional bias, and VPP and LPP neural representations more similar to pain-related interpretations. Pain state also altered the disengagement-difficulty pattern under invalid threat-cue conditions, indicating that ongoing pain changed the processing of pain-related cues when attention had to be redirected. In contrast, pain pictures showed a stable recognition advantage over non-pain pictures, but this memory bias was not further amplified by pain state. Machine-learning analyses showed that cross-task cognitive-bias features and their neural signatures could distinguish pain from non-pain states to some extent, with the best model achieving an AUC of 0.691. Within the pain group, these features also predicted individual pain intensity, with the best model achieving an out-of-sample R^2 of 0.182. Feature attribution analyses indicated that interpretation-bias features contributed more to pain-state classification, whereas attention- and memory-bias features were more closely associated with pain-intensity differences. Together, by integrating cross-task cognitive-bias paradigms, ERPs, DDM, and machine-learning analysis, the present study systematically revealed the computational and neural mechanisms of cognitive biases during a pain state. These findings suggest that painful state selectively reshapes different forms of cognitive bias, and that cognitive biases and their neural signatures may provide important complementary indicators for individualized pain decoding and assessment beyond traditional subjective pain ratings.

Key words: pain, cognitive bias, event-related potentials, drift-diffusion model, machine learning