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

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Pain assessment traditionally relies on subjective methods, complicating objective diagnosis, especially after analgesic administration.
• Functional near-infrared spectroscopy (fNIRS) offers a non-invasive approach to measure brain activity, specifically cortical hemodynamic responses.

Purpose of the Study:

• To propose a deep learning (DL) based transfer learning (TL) methodology for objective classification of fNIRS-derived cortical oxygenated hemoglobin responses.
• To differentiate between painful and non-painful stimuli under various timings post-analgesic and placebo drug administration.

Main Methods:

• Utilized a publicly available fNIRS dataset from painful/non-painful stimuli experiments.
• Developed a base DL model from pre-drug fNIRS data and applied TL to six post-drug conditions (30, 60, 90 min post-morphine/placebo).
• Employed DeepSHAP to analyze the contribution of nine regions of interest to the decoding models.

Main Results:

• The pre-drug model and all post-drug models achieved over 90% accuracy, sensitivity, specificity, and AUC.
• Post-placebo models showed higher decoding accuracy than post-morphine models.
• Cortical region contributions to classification varied across post-drug conditions.

Conclusions:

• The proposed DL-based TL methodology can create effective pain decoding models without extensive condition-specific training data.
• This approach minimizes computational costs and the need for data collection in impractical settings.
• Understanding regional brain contributions can inform the design of efficient fNIRS-based brain-computer interface (BCI) systems.