Machine learning models for stroke detection by observing the eye-movement features under five-color visual stimuli in traditional Chinese medicine

Abstract

To develop a novel diagnostic modality to identify and diagnose stroke in daily life scenarios for improving the therapeutic effects and prognoses of patients. In this study, 16 stroke patients and 24 age-matched healthy participants as controls were recruited for comparative analysis. Leveraging a portable eye-tracking device and integrating traditional Chinese medicine theory with modern color psychology principles, we recorded the eye movement signals and calculated eye movement features. Meanwhile, the stroke recognition models based on eye movement features were further trained by using random forest (RF), k-nearest neighbors (KNN), decision tree (DT), gradient boosting classifier (GBC), XGBoost, and CatBoost. The stroke group and the healthy group showed significant differences in some eye movement features (P< .05). The models trained based on eye movement characteristics had good performances in recognizing stroke individuals, with accuracies ranging from 77.40% to 88.45%. Under the red stimulus, the eye movement model trained by RF became the best machine learning model with a recall of 84.65%, a precision of 86.48%, a F1 score of 85.47%. Among the six algorithms, RF and CatBoost performed better in classification. This study pioneers the application of traditional Chinese medicine’s five-color stimuli to visual observation tasks. On the basis of the combined design, the eye-movement models can accurately identify stroke, and the developed high-performance models may be used in daily life scenarios.