Features from Pupillary Reflex Dilation and neural networks classification of remifentanil effect

Abstract

Background: Pupillary Reflex Dilation (PRD) is a physiological response characterized by pupil dilation following noxious stimuli. PRD is attenuated by opioids, thus serving as an indicator of nociception-antinociception balance. Although various metrics derived from PRD have been proposed, there is no consensus on the optimal measure correlating with opioid pharmacological effects. Neural network modeling has shown potential in interpreting complex physiological signals; however, no previous studies have systematically evaluated multiple PRD-derived features using advanced machine learning techniques to classify opioid effects. In this study we analyze multiple PRD-derived metrics using neural network modeling to classify remifentanil effect-site concentrations during general anesthesia with propofol and remifentanil. Methods: This prospective observational study enrolled 86 patients undergoing elective general surgery or gynecological procedures under anesthesia exclusively with propofol and remifentanil. PRD was measured using a portable infrared pupillometer during standardized noxious stimulation at varying estimated effect-site concentrations of propofol and remifentanil. Data were modeled using an ordinal differential equation incorporating stimulus characteristics. Extracted features included baseline pupil diameter, maximum dilation amplitude, dilation delays, percentage dilation, maximum and minimum pupil diameter variations, and area under the curve (AUC). Correlation analysis reduced feature dimensionality before training a neural network classifier to categorize remifentanil concentrations into three groups: 0 ng/mL, 0-2 ng/mL, and >2 ng/mL. Results: From an initial dataset of 1,294 PRD records, 1,264 were analyzed after exclusions. The developed model accurately captured dynamic PRD characteristics across anesthetic conditions. Higher remifentanil concentrations correlated with reduced peak pupil size, smaller AUC values, decreased percentage and absolute changes in pupil diameter, and shorter dilation delays. The neural network achieved an overall classification accuracy of 72.3% on the test set, performing well at distinguishing extreme groups (80.0% accuracy for 0 ng/mL and 88.1% for >2 ng/mL) but less effectively classifying intermediate concentrations (40.6% accuracy for 0-2 ng/mL). Discussion: The study demonstrates that PRD-derived metrics analyzed via neural networks can effectively classify opioid effect-site concentrations during anesthesia. Lower accuracy in intermediate concentration classification likely reflects interpatient variability inherent to population-based pharmacokinetic-pharmacodynamic modeling rather than limitations of PRD measurement itself. Standardized automated pupillometry provided consistent stimulus intensity and reduced operator-dependent variability.