Multi-Task Multiple Kernel Machines for Personalized Pain Recognition from fNIRS
Date:
I delivered an oral presentation at the International Conference on Pattern Recognition (ICPR 2018) in Beijing, China, presenting our work on personalized pain detection using functional near-infrared spectroscopy (fNIRS) brain signals. The paper received the Best Student Paper Award.
Our study introduced a multi-task machine learning approach for detecting experimentally induced pain from hemodynamic responses measured with fNIRS. Traditional approaches to pain recognition often struggle with significant inter-subject variability, an inherent challenge when modeling a subjective phenomenon such as pain. To address this, we formulated pain detection as a multi-task learning problem, grouping participants into latent profiles based on spectral clustering and training related models jointly.
The method combined multi task learning with multiple kernel learning, allowing the algorithm to automatically estimate the relevance of each fNIRS channel and adapt the classifier to subject specific response patterns. As detailed in the paper, this framework improved the interpretability of the resulting models, highlighting for instance the increased importance of anterior prefrontal cortex channels, and substantially enhanced detection performance compared to single task baselines.
This work demonstrated the feasibility of deriving personalized, brain-based biomarkers of pain using fNIRS and machine learning, with potential applications in clinical contexts where traditional self-report may be unreliable or unavailable.