COSMOS: a dataset for Classification Of Stress and workload using multiMOdal wearable Sensors

Prolonged stress and high mental workload can have deteriorating long-term effects developing several stress-related diseases. The existing stress detection techniques are often uni-modal and limited to controlled setups. One sensing modality could be unobtrusive but mostly results in unreliable sensor readings, especially in uncontrolled environments. Our study recorded multi-modal physiological signals from twenty-five participants in controlled and uncontrolled environments by performing given and self-chosen tasks of high and low mental demand. In this version, we processed and published a subset of the dataset from six participants while working on the rest. The subset of the data is used to check the feasibility of our study by engineering features from electroencephalography (EEG), photoplethysmography (PPG), electrodermal activity (EDA), and temperature sensor data. Machine learning methods were used for the binary classification of the tasks. Personalized models in the uncontrolled environment achieved a mean classification accuracy of up to 83% while using one of the four labels, unveiling some unintentional mislabeling by participants. In controlled environments, multi-modality improved the accuracy by at least 7%. Generalized machine learning models achieved close to chance-level performances. This work underlines the importance of multi-modal recordings and provides the research community with an experimental paradigm to take studies of mental workload and stress workload and stress out of controlled into uncontrolled environments长期压力与高精神负荷可诱发多种压力相关性疾病，造成进行性加重的长期健康损害。现有压力检测技术多为单模态（uni-modal），且仅局限于受控实验环境。单一感知模态虽可实现无侵扰式信号采集，但在非受控环境下往往难以获得可靠的传感器读数。本研究招募25名受试者，在受控与非受控环境中分别完成指定任务与自主选择任务，涵盖高、低两种精神负荷水平，同步采集多模态生理信号。本版本已完成处理并发布其中6名受试者的数据集子集，其余受试者的数据仍在处理中。该子集数据用于验证本研究的可行性：我们从脑电图（electroencephalography, EEG）、光电容积描记法（photoplethysmography, PPG）、皮肤电活动（electrodermal activity, EDA）及温度传感器数据中提取特征，并采用机器学习方法完成任务的二分类任务。在非受控环境下，仅使用四类标签中的一类时，个性化模型的平均分类准确率可达83%，同时该结果也揭示了受试者存在部分无意识的标注错误。在受控环境中，多模态数据可将分类准确率提升至少7%。通用型机器学习模型的性能则接近随机猜测水平。本研究凸显了多模态生理信号采集的重要性，并为科研社区提供了一套可将精神负荷与压力研究从受控环境拓展至非受控环境的实验范式。