# Eye-tracking during free visual exploration of familiar dramatic character faces facilitates rapid and accurate stroke recognition

## TL;DR

Eye-tracking during visual exploration of familiar faces can help detect strokes quickly and accurately.

## Contribution

A novel ecological paradigm using eye-tracking and machine learning for rapid stroke recognition.

## Key findings

- Stroke patients showed prolonged fixation duration, restricted saccadic movements, and reduced scanpath length.
- A machine learning model achieved 87.18% accuracy and an AUROC of 0.92 in identifying stroke patients.
- Multi-type eye movement features improve stroke recognition accuracy and logistical efficiency.

## Abstract

Stroke patients often experience significant impairments, making rapid and accurate detection crucial for timely intervention and early warning. However, existing diagnostic methods such as advanced neuroimaging are often time-consuming, highly dependent on operator expertise, or costly and complex to deploy, limiting their scalability in resource-restricted settings. Eye movement patterns in stroke patients present a promising opportunity for efficient detection, given their close ties to underlying neurocognitive mechanisms and potential diagnostic sensitivity. Nevertheless, the lack of a feasible task paradigm and robust detection strategy has hindered the practical application of eye movement-based stroke identification. This study aimed to capture eye movement dysfunction associated with stroke through an ecological paradigm and develop a machine learning model with improved diagnostic accuracy.

We recorded eye movement signals in stroke patients (N = 16) and healthy controls (N = 23) during free visual exploration of familiar dramatic character faces. A diverse set of eye movement features, encompassing saccadic, fixation, and scanpath features, was extracted and analyzed. These features were subsequently employed to construct machine learning models for the recognition of stroke patients.

We identified distinctive eye movement patterns in the stroke group, including prolonged fixation duration, restricted saccadic movements, and reduced scanpath length, which reflect underlying visual processing impairments. Furthermore, by integrating these multi-dimensional eye movement features, our machine learning model achieved a high accuracy of 87.18% and an excellent area under the receiver operating characteristic curve (AUROC) of 0.92 in distinguishing stroke patients.

This study demonstrates that ecologically valid eye-tracking, combined with multi-type feature analytics, serves as a practical screening tool with the potential to significantly improve identification accuracy and alleviate logistical burdens in community and primary care settings.

Eye movement patterns offer a promising opportunity for efficient stroke detection due to their link to neurocognitive mechanisms. Based on this correlation, we develop an ecological paradigm, free visual exploration of familiar dramatic character faces, to capture eye movement dysfunctions in stroke patients. By leveraging multi type eye movement features, we can facilitate rapid and accurate stroke recognition. Diagram illustrating a process for rapid and accurate stroke recognition. The top section shows a man with a diagram of the brain, linking eye movement dysfunctions to cerebral lesions in stroke. Another section shows a flowchart for processing eye movement features, using standardization and LASSO selection, applied to machine learning algorithms to classify individuals as stroke or healthy. The bottom section depicts an eye-tracking experiment with stroke and healthy participants, highlighting multi-type eye movement features such as saccadic, fixation, and scanpath features.