GSR Signals, Extracted Features and ML Trained Models For Startle Events Detection While Walking with a Smart Cane

This dataset accompanies the following publication, which is currently under review: Villalba-Bravo, Trujillo-León, A., & Vidal-Verdú, F.Startle Event Detection Using a Smart Assistive Walking Cane and Machine Learning.IEEE RAS/EMBS 11th International Conference on Biomedical Robotics and Biomechatronics.(Under review) Description This dataset contains physiological signals collected from 18 participants using a wireless smart walking cane equipped with a custom Galvanic Skin Response (GSR) sensor. The participants walked through common areas of the university while being exposed to sudden and loud auditory stimuli designed to induce startle responses. The aim of the study was to identify statistical and temporal features of the GSR signals that are robust to motion artifacts (MAs) and feasible for real-time detection of startle events in vulnerable populations who rely on mobility aids. The Empatica EmbracePlus device was used as a motion artifacts free reference system, which recorded the same physiological signals and features, comparison purposes. Preprocessing included filtering, downsampling, and extraction of windows corresponding to startle and non-startle events. Features were calculated for each window and used to train multiple machine learning models. The structure of the dataset is discussed later. All participants provided written informed consent both to take part in the experiment and to allow their anonymized data to be publicly shared for research purposes. Furthermore, the experiment was approved by the Ethical Committee of the Universidad de Málaga (reference 46-2024-H). Folder Structure recorded_signals: Contains the signals recorded during the experiment from both the smart cane and the reference Empatica device. The signals have been preprocessed as described in Section III, Materials and Methods. The low-pass cutoff frquency for GSR and phasic signal in this folder is set to 1 Hz. features_data: Contains the data from all 10 trials for the proposed features of the selected participants (i.e., those who exhibited a startle response following the auditory stimulus). The data are organized hierarchically, combining different low-pass filter cutoff frequencies for the GSR/phasic signals (0.45 Hz and 1 Hz) and different window lengths (5, 10, and 15 seconds). classification_data: For each of the 10 trials, this folder contains the dataset used for training the ML models, as well as the trained machine learning models with 5-fold cross-validation, saved in .mat format. All models were created and trained usingusing version 25.2 (R2025b) of MATLAB Statistics and Machine Learning Toolbox. Also, this folder includes two CSV tables: one containing the classification metrics for each model in each trial (see classification_metrics_per_trial.csv), and another with the mean and standard deviation of the metrics for the proposed models across the 10 trials (see classification_metrics_summary_10trials.csv). startle_audio_experiment.mp3: Audio track used during the experiment to elicit startle responses to the participants. Contact Information For any questions or further information regarding this dataset, please contact [e-mail] and [e-mail].