Evaluating robotic actions

Introduction: Enhancing medical robot training traditionally relies on explicit feedback from physicians to identify optimal and suboptimal robotic actions during surgery. Passive brain-computer interfaces (BCIs) offer an emerging alternative by enabling implicit brain-based performance evaluations. However, effectively decoding these evaluations of robot performance requires a comprehensive understanding of the spatiotemporal brain dynamics identifying optimal and suboptimal robot actions within realistic settings.
Methods: We conducted an electroencephalographic study with 16 participants who mentally assessed the quality of robotic actions while observing simulated robot-assisted laparoscopic surgery scenarios designed to approximate real-world conditions. We aimed to identify key spatiotemporal dynamics using the surface Laplacian technique and two complementary data-driven methods: a mass-univariate permutation-based clustering and multivariate pattern analysis (MVPA)-based temporal decoding. A second goal was to identify the optimal time interval of evoked brain signatures for single-trial classification.
Results: Our analyses revealed three distinct spatiotemporal brain dynamics differentiating the quality assessment of optimal vs. suboptimal robotic actions during video-based laparoscopic training observations. Specifically, an enhanced left fronto-temporal current source, consistent with P300, LPP, and P600 components, indicated heightened attentional allocation and sustained evaluation processes during suboptimal robot actions. Additionally, amplified current sinks in right frontal and mid-occipito-parietal regions suggested prediction-based processing and conflict detection, consistent with the oERN and interaction-based ERN/N400. Both mass-univariate clustering and MVPA provided convergent evidence supporting these neural distinctions.
Discussion: The identified neural signatures propose that suboptimal robotic actions elicit enhanced, sustained brain dynamics linked to continuous attention allocation, action monitoring, conflict detection, and ongoing evaluative processing. The findings highlight the importance of prioritizing late evaluative brain signatures in BCIs to classify robotic actions reliably. These insights have significant implications for advancing machine-learning-based training paradigms.