Predicting pupillary response and identifying lighting thresholds in road tunnels using bacterial foraging optimization-based stacking ensemble model

The lighting conditions of road tunnels vary sharply across threshold, interior, and exit zones, imposing substantial visual adaptation demands on drivers and directly affecting pupillary response. To improve the prediction accuracy of pupil diameter and identify the nonlinear threshold characteristics of lighting parameters, this study proposed a bacterial foraging optimization-based stacking ensemble model (BFO-OSEM). Real-vehicle experiments were conducted in four tunnels with different lighting environments to collect pupil diameter and five lighting parameters. The proposed model employs XGBoost, LightGBM, and multilayer perceptron as base learners, with hyperparameters independently optimized using the BFO algorithm, and support vector regression as the meta-learner. The generalization capability of model was rigorously evaluated using Leave-One-Subject-Out (LOSO) and tunnel-scenario cross-validations, and its interpretability was enhanced through SHAP and permutation importance analyses. The results demonstrate that optimization of BFO significantly improves base learner performance, with MAE of XGBoost decreasing from 1.411 to 0.759 and RMSE of LightGBM decreasing from 1.473 to 1.115. The stacking ensemble further enhances prediction accuracy, achieving the R 2 of 0.959. The model maintains robust performance under rigorous validation (R 2 = 0.913 of LOSO and R 2 = 0.925 of tunnel-scenario), confirming its generalization ability. Based on the predicted results and statistically validated piecewise regression, three pupillary response zones were identified, revealing dual-threshold nonlinear response characteristics with critical thresholds for road surface illuminance at approximately 800 lx and 1315 lx. These findings provide a predictive tool for visual load assessment, establish a quantitative basis for synergistic optimization of road tunnel lighting, and offer insights for future intelligent lighting control systems.