Problem Statement: Pain has a crucial function in the human body acting as an early warning signal to protect against tissue damage. However, both assessment of pain experience and its clinical diagnosis rely on highly subjective methods. Objective evaluation of the presence of pain under analgesic drug administrations becomes even more complicated.

Objectives: The aim of this study was to propose a transfer learning (TL) based deep learning (DL) methodology for accurate detection and objective classification of the neural processing of painful and non-painful stimuli that were presented under different levels of analgesia.

Method: A publicly available fNIRS dataset of 14 participants was obtained during an experimental protocol that involved painful and non-painful events. Separate fNIRS scans were taken under the same nociceptive protocol before analgesic drug (Morphine and Placebo) administration and at three different times (30,60 and 90 min) post-administration. By utilizing data from pre-drug fNIRS scans, a DL architecture for classifying painful and non-painful stimuli was constructed as a base model. Knowledge generated in pre-drug base model was transferred to 6 distinct post-drug conditions by adapting a TL approach. The DeepSHAP method was utilized to unveil the contribution weights of nine R OIs for each of the pre-drug and post-drug models.

Results: Mean performance of pre-drug base model was above 95% for accuracy, sensitivity, specificity and AUC metrics. Each of the post-drug models had mean accuracy, sensitivity, specificity and AUC performance above 90%. No statistically significant difference across post-drug models were found for classification performance of any of the performance metrics. Post-placebo models had higher decoding accuracy than post-morphine models.

Conclusion: Knowledge obtained from a pre-drug base model could be successfully utilized to build pain decoding models for six distinct brain states that were altered with either analgesic or placebo intervention. Contribution of different cortical regions to classification performance varied across the post-drug models.

Importance: The proposed methodology may remove the necessity to build new DL models for data collected at clinical or daily life conditions for which obtaining training data is not practical or building a new decoding model will have a computational cost. Unveiling the explanation power of different cortical regions may aid the design of more computationally efficient fNIRS based BCI system designs that target other application areas.