Explainable fNIRS-based pain decoding under pharmacological conditions via deep transfer learning approach.

Significance: Assessment of pain and its clinical diagnosis rely on subjective methods which become even more complicated under analgesic drug administrations.

Aim: We aim to propose a deep learning (DL)-based transfer learning (TL) methodology for objective classification of functional near-infrared spectroscopy (fNIRS)-derived cortical oxygenated hemoglobin responses to painful and non-painful stimuli presented under different timings post-analgesic and placebo drug administration.

Approach: A publicly available fNIRS dataset obtained during painful/non-painful stimuli was used. Separate fNIRS scans were taken under the same protocol before drug (morphine and placebo) administration and at three different timings (30, 60, and 90 min) post-administration. Data from pre-drug fNIRS scans were utilized for constructing a base DL model. Knowledge generated from the pre-drug model was transferred to six distinct post-drug conditions by following a TL approach. The DeepSHAP method was utilized to unveil the contribution weights of nine regions of interest for each of the pre-drug and post-drug decoding models.

Results: Accuracy, sensitivity, specificity, and area under curve (AUC) metrics of the pre-drug model were above 90%, whereas each of the post-drug models demonstrated a performance above 90% for the same metrics. Post-placebo models had higher decoding accuracy than post-morphine models. Knowledge obtained from a pre-drug base model could be successfully utilized to build pain decoding models for six distinct brain states that were scanned at three different timings after either analgesic or placebo drug administration. The contribution of different cortical regions to classification performance varied across the post-drug models.

Conclusions: The proposed DL-based TL methodology may remove the necessity to build 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 brain-computer interface (BCI) system designs that target other application areas.