Pain decoding under analgesic conditions using functional near infrared spectroscopy and transfer learning

Abstract

Objective: Pain decoding using hemodynamic responses is an objective but challenging approach due to the variable nature of hemodynamic response. Moreover, the effects of different analgesic conditions increase the complexity of this problem.In this study, we aimed to decode the intensity level of nociceptive stimuli under analgesic conditions by utilizing fNIRS derived hemodynamic responses and a deep transfer learning approach. Methods: A previously collected fNIRS dataset collected from 14 healthy male volunteers was utilized. Each subject had two site visits where they were orally administered with a morphine or a placebo pill. At each site visit, subjects had 4 fNIRS scans which were taken during a nociceptive stimuli protocol a)before and b) after 30,60,90 minutes of drug administration. 6 noxious and 6 innocuous stimuli were given to left thumb. After data preprocessing, a deep learning model was trained on the pre-drug dataset to classify painful and non-painful stimuli. Then, the knowledge obtained in this model was then transferred to classify post-drug dataset. Results: Accuracy performance of the pre-drug model was 0.97. Accuracy of post morphine drug models were 0.91 after 30 min, 0.90 after 60 min and 0.91, after 90 min. For placebo administration, they were found as 0.92 after 30 min, 0.92 after 60 min, 0.91 after 90 min respectively. Statistical comparison of performance metrics showed that accuracy values were significantly higher in pre-drug models compared to post-morphine and post-placebo models. Conclusion: Our deep transfer learning approach showed that knowledge obtained from a pre-drug model trained by using hemodynamic responses can be used to decode pain level after drug administration. We demonstrate the potential of fNIRS derived signals for transferring information from a model trained with baseline data to models built for different clinical or daily life conditions where collection of training data may not be feasible/practical to build novel ML or DL models.