# Single- and Multimodal Deep Learning of EEG and EDA Responses to Construction Noise: Performance and Ablation Analyses

## TL;DR

This study uses deep learning to analyze EEG and EDA responses to construction noise, showing that EEG is more effective and that combining data improves results.

## Contribution

The study introduces a deep learning framework for analyzing multimodal physiological responses to construction noise with ablation analyses for optimization.

## Key findings

- EEG-based models outperformed EDA-based models in detecting physiological responses to construction noise.
- Decision-level fusion of EEG and EDA data improved robustness and evaluation metrics.
- Optimal model performance was achieved with specific window sizes, batch sizes, and weight decay settings.

## Abstract

The purpose of the study is to investigate human physiological responses to construction noise exposure using deep learning, applying electroencephalography (EEG) and electro-dermal activity (EDA) sensors. Construction noise is a pervasive occupational stressor that affects physiological states and impairs cognitive performance. EEG sensors capture neural activity related to perception and attention, and EDA reflects autonomic arousal and stress. In this study, twenty-five participants were exposed to impulsive noise from pile drivers and tonal noise from earth augers at three intensity levels (40, 60, and 80 dB), while EEG and EDA signals were recorded simultaneously. Convolutional neural networks (CNN) were utilized for EEG and long short-term memory networks (LSTM) for EDA. The results depict that EEG-based models consistently outperformed EDA-based models, establishing EEG as the dominant modality. In addition, decision-level fusion enhanced robustness across evaluation metrics by employing complementary information from EDA sensors. Ablation analyses presented that model performance was sensitive to design choices, with medium EEG windows (6 s), medium EDA windows (5–10 s), smaller batch sizes, and moderate weight decay yielding the most stable results. Further, retraining with ablation-informed hyperparameters confirmed that this configuration improved overall accuracy and maintained stable generalization across folds. The outcome of this study demonstrates the potential of deep learning to capture multimodal physiological responses when subjected to construction noise and emphasizes the critical role of modality-specific design and systematic hyperparameter optimization in achieving reliable annoyance detection.