Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CALM: Cognitive Assessment using Light-insensitive Model

Akhil Meethal, Anita Paas, Nerea Urrestilla Anguiozar, David St-Onge

TL;DR

CALM investigates how ambient light biases pupillometry-based cognitive load estimation and demonstrates that fusing pupillometry with HRV substantially reduces light sensitivity and improves accuracy and reliability. The study collects synchronized pupillometry and HRV data from 10 participants performing n-back tasks under light and dark conditions, comparing clinical-grade BioPac MP35 with fitness-grade Polar H10 devices. Across RF, MLP, and Transformer classifiers, multimodal models yield notable gains (roughly 12–15 percentage points) over single-modality approaches, with Polars often performing best in mobile settings. The findings support using affordable HRV sensing in real-world CLE and indicate transformer-based learned features can approach or match traditional hand-crafted features, offering a path toward robust, light-insensitive cognitive load assessment.

Abstract

The demand for cognitive load assessment with low-cost easy-to-use equipment is increasing, with applications ranging from safety-critical industries to entertainment. Though pupillometry is an attractive solution for cognitive load estimation in such applications, its sensitivity to light makes it less robust under varying lighting conditions. Multimodal data acquisition provides a viable alternative, where pupillometry is combined with electrocardiography (ECG) or electroencephalography (EEG). In this work, we study the sensitivity of pupillometry-based cognitive load estimation to light. By collecting heart rate variability (HRV) data during the same experimental sessions, we analyze how the multimodal data reduces this sensitivity and increases robustness to light conditions. In addition to this, we compared the performance in multimodal settings using the HRV data obtained from low-cost fitness-grade equipment to that from clinical-grade equipment by synchronously collecting data from both devices for all task conditions. Our results indicate that multimodal data improves the robustness of cognitive load estimation under changes in light conditions and improves the accuracy by more than 20% points over assessment based on pupillometry alone. In addition to that, the fitness grade device is observed to be a potential alternative to the clinical grade one, even in controlled laboratory settings.

CALM: Cognitive Assessment using Light-insensitive Model

TL;DR

CALM investigates how ambient light biases pupillometry-based cognitive load estimation and demonstrates that fusing pupillometry with HRV substantially reduces light sensitivity and improves accuracy and reliability. The study collects synchronized pupillometry and HRV data from 10 participants performing n-back tasks under light and dark conditions, comparing clinical-grade BioPac MP35 with fitness-grade Polar H10 devices. Across RF, MLP, and Transformer classifiers, multimodal models yield notable gains (roughly 12–15 percentage points) over single-modality approaches, with Polars often performing best in mobile settings. The findings support using affordable HRV sensing in real-world CLE and indicate transformer-based learned features can approach or match traditional hand-crafted features, offering a path toward robust, light-insensitive cognitive load assessment.

Abstract

The demand for cognitive load assessment with low-cost easy-to-use equipment is increasing, with applications ranging from safety-critical industries to entertainment. Though pupillometry is an attractive solution for cognitive load estimation in such applications, its sensitivity to light makes it less robust under varying lighting conditions. Multimodal data acquisition provides a viable alternative, where pupillometry is combined with electrocardiography (ECG) or electroencephalography (EEG). In this work, we study the sensitivity of pupillometry-based cognitive load estimation to light. By collecting heart rate variability (HRV) data during the same experimental sessions, we analyze how the multimodal data reduces this sensitivity and increases robustness to light conditions. In addition to this, we compared the performance in multimodal settings using the HRV data obtained from low-cost fitness-grade equipment to that from clinical-grade equipment by synchronously collecting data from both devices for all task conditions. Our results indicate that multimodal data improves the robustness of cognitive load estimation under changes in light conditions and improves the accuracy by more than 20% points over assessment based on pupillometry alone. In addition to that, the fitness grade device is observed to be a potential alternative to the clinical grade one, even in controlled laboratory settings.
Paper Structure (24 sections, 5 figures, 6 tables)

This paper contains 24 sections, 5 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Pupillometry and the light condition
  3. Multimodal CLE with Pupillometry and HRV
  4. Methods and Materials
  5. Participants
  6. Experimental Design
  7. Sensors
  8. Procedure
  9. Data Pre-processing and Feature Extraction
  10. Pupillometry
  11. Heart Rate Variability
  12. Classifiers
  13. Results
  14. Pupillometry and change in light conditions
  15. Multimodal models
  16. ...and 9 more sections

Figures (5)

  • Figure 1: HRV devices: Left, BioPac M35 with the electrodes position, Right, Polar H10 chest band used only in the second experiment.
  • Figure 2: Pupillometry device: Tobii Pro 3 glasses
  • Figure 3: Left: Experiment setup for the Light condition. Right: Experiment setup for the Dark condition.
  • Figure 4: Comparison of confusion matrices between pupillometry only and multimodal models.
  • Figure 5: Left: Density plots of pupillometry and HRV features under light and dark conditions. The top row shows pupillometry features (mean pupil diameter, and pupil diameter variance), while the bottom row shows HRV features (RSMMD, and pNN50). Higher sensitivity to light of pupillometry features can be observed from their complete distributions' change. The HRV features are relatively less sensitive to light. Right: Box plot of the Pupil mean and RMSSD values for all cognitive load conditions using the Polar device. CL1 and CL2 stand for cognitive load for the 1-back (low) and 2-back (high) tasks respectively.