One size does not fit all: a support vector machine exploration of multiclass cognitive state classifications using physiological measures.

Introduction: This study aims to develop and evaluate support vector machines (SVMs) learning models for predicting cognitive workload (CWL) based on physiological data. The objectives include creating robust binary classifiers, expanding these to multiclass models for nuanced CWL prediction, and exploring the benefits of individualized models for enhanced accuracy. Cognitive workload assessment is critical for operator performance and safety in high-demand domains like aviation. Traditional CWL assessment methods rely on subjective reports or isolated metrics, which lack real-time applicability. Machine learning offers a promising solution for integrating physiological data to monitor and predict CWL dynamically. SVMs provide transparent and auditable decision-making pipelines, making them particularly suitable for safety-critical environments.

Methods: Physiological data, including electrocardiogram (ECG) and pupillometry metrics, were collected from three participants performing tasks with varying demand levels in a low-fidelity aviation simulator. Binary and multiclass SVMs were trained to classify task demand and subjective CWL ratings, with models tailored to individual and combined subject datasets. Feature selection approaches evaluated the impact of streamlined input variables on model performance.

Results: Binary SVMs achieved accuracies of 70.5% and 80.4% for task demand and subjective workload predictions, respectively, using all features. Multiclass models demonstrated comparable discrimination (AUC-ROC: 0.75-0.79), providing finer resolution across CWL levels. Individualized models outperformed combined-subject models, showing a 13% average improvement in accuracy. SVMs effectively predict CWL from physiological data, with individualized multiclass models offering superior granularity and accuracy.

Discussion: These findings emphasize the potential of tailored machine learning approaches for real-time workload monitoring in fields that can justify the added time and expense required for personalization. The results support the development of adaptive automation systems in aviation and other high-stakes domains, enabling dynamic interventions to mitigate cognitive overload and enhance operator performance and safety.