Self-Supervised Learning for Near-Wild Cognitive Workload Estimation

Area of Science:

• Neuroscience
• Machine Learning
• Human-Computer Interaction

Background:

• Cognitive workload estimation is crucial for reducing decision-making errors.
• Machine learning models using physiological data (EEG, ECG) show promise but require extensive labeled data.
• Commercial devices offer low-cost data collection but suffer from artifacts in real-world settings.

Purpose of the Study:

• To develop a hybrid machine learning model for estimating cognitive-physical workloads outside controlled laboratory settings.
• To identify the most relevant physiological modalities for approximating cognitive workload.
• To reduce the need for costly and time-consuming data labeling in machine learning models.

Main Methods:

• A hybrid approach combining feature selection and self-supervised machine learning techniques was implemented.
• Physiological data from seven modalities were collected outside laboratory settings.
• The model was used to identify relevant modalities and approximate six levels of cognitive-physical workload.

Main Results:

• The study successfully identified key physiological modalities for approximating cognitive workload.
• The hybrid model demonstrated the ability to estimate cognitive-physical workload levels using self-supervised learning.
• The approach proved effective in real-world settings, overcoming challenges of artifact contamination.

Conclusions:

• A novel hybrid machine learning framework effectively estimates cognitive workload using real-world physiological data.
• Self-supervised learning significantly reduces the burden of data labeling for cognitive workload approximation.
• This approach enhances the feasibility of using physiological feedback to improve decision-making in everyday environments.