Mental Workload Prediction Using Physiological Signals: Balancing Performance and Interpretability

Paper Title

Mental Workload Prediction Using Physiological Signals: Balancing Performance and Interpretability

Publication Info

• Topic area: Mental workload prediction using physiological signals and machine learning.
• Keywords: Mental workload, physiological signals, machine learning, interpretability, respiratory signals, cardiac signals, oculomotor signals, logistic regression, decision tree, safety-critical systems.

Background and Problem

• Problem / challenge: Existing machine learning models for mental workload prediction prioritize performance but often neglect interpretability and error analysis, which are critical for safety-critical systems.
• Significance: Accurate and interpretable workload prediction can prevent overload and underload states in operators, reducing errors and improving safety in critical environments.
• Motivation and related work: Prior studies have explored physiological indicators for workload prediction but lack focus on interpretability and systematic error analysis. This study aims to address these gaps by developing models that balance accuracy and interpretability.

Solution

• Proposed approach: Development of interpretable machine learning models (logistic regression and decision tree) for predicting mental workload using physiological signals.
• Novelty:
  1. Creation of interpretable models that achieve competitive accuracy using minimal features.
  2. Systematic optimization of preprocessing strategies alongside model hyperparameters.
  3. Comprehensive error analysis to identify scenario-, time-, and individual-dependent prediction errors.
• Procedure and key techniques:
  • Collection of respiratory, cardiac, and oculomotor data from participants under varying task demands.
  • Preprocessing steps including outlier handling, scaling, feature selection, and balancing.
  • Validation using 5-fold cross-validation and random search for hyperparameter optimization.
  • Selection of logistic regression for binary classification and decision tree for multi-class classification based on performance and interpretability.
  • Analysis of prediction errors across scenarios and participants.

Results

• Concrete findings:
  • Logistic regression achieved an f1-score of 90.5% (accuracy: 92.2%) for binary classification.
  • Decision tree achieved an f1-score of 72.0% (accuracy: 72.3%) for multi-class classification.
  • Respiratory and pupil features were most predictive, while cardiac features showed higher variability.
• Advantage over baselines:
  • Logistic regression outperformed black-box models like random forests and support vector machines in binary classification, achieving higher accuracy with fewer features.
  • Decision tree performed competitively with complex models in multi-class classification.
• Experiments / evaluation:
  • Dataset included 30 participants for respiratory and cardiac data, and 17 participants for oculomotor data.
  • Models were tested on unseen data, with separate training and test sets to ensure generalizability.
  • Error analysis revealed systematic misclassification patterns influenced by task dynamics, physiological response properties, and individual differences.
• Limitations and future work:
  • Small sample size may limit generalizability; larger datasets are needed.
  • Task load as ground truth oversimplifies the multidimensional nature of mental workload.
  • Future research should explore continuous modeling, incorporate diverse workload states, and investigate person-specific models.

Summary

This study developed interpretable machine learning models for mental workload prediction using physiological signals, achieving competitive accuracy compared to black-box approaches. Logistic regression and decision tree models demonstrated high performance with minimal features, emphasizing the importance of preprocessing strategies. Error analysis revealed systematic scenario-, time-, and individual-dependent misclassification patterns, highlighting the complexity of mental workload prediction. These findings have practical implications for adaptive systems in safety-critical environments and pave the way for real-time workload monitoring. Future research should focus on expanding datasets, refining ground truth definitions, and exploring person-specific modeling approaches.