Generalised machine learning models outperform personalised models for cognitive load classification in real-life...

Area of Science:

• Human-Computer Interaction
• Wearable Technology
• Cognitive Science

Background:

• Personal assistants could improve societal health and life satisfaction by managing cognitive load.
• Wearable sensors are crucial for real-time physiological data collection, but multi-modal, real-life datasets are scarce.
• Existing studies often lack a bridge between controlled laboratory and uncontrolled real-life settings.

Purpose of the Study:

• To investigate the efficacy of commercially available sensors for building cognitive load assistants.
• To collect multi-modal physiological data in both controlled and uncontrolled environments.
• To create a publicly available dataset for future research on cognitive load monitoring.

Main Methods:

• Collected data from ten university students over eight-hour workdays, split between controlled and uncontrolled settings.
• Utilized commercially available wearable sensors for real-time physiological data acquisition.
• Employed machine learning models, including Logistic Regression, for cognitive load classification.

Main Results:

• No single feature strongly correlated with cognitive load, but smartwatch indices and biomarkers showed differences between low and high load.
• Machine learning models achieved high F1 scores (up to 0.91) for classifying cognitive load levels.
• The study successfully generated a valuable dataset for cognitive load research.

Conclusions:

• Commercially available wearable sensors show promise for developing real-life cognitive load assistants.
• The study design bridges the gap between laboratory and real-world data collection.
• The public release of this dataset facilitates advancements in mental state monitoring technology.