Spontaneous Blink Features Extracted from Video for Multi-Class Classification of Parkinson’s Disease

Parkinson’s disease (PD) presents both motor and non-motor impairments, including alterations in spontaneous blinking behavior. Oculomotor biomarkers, often preceding major motor symptoms, offer a non-invasive and quantitative approach to characterize oculomotor impairments. In this paper, we propose an automated framework for PD classification based on spontaneous blink recording. We introduce a novel dataset comprising healthy controls, PD patients without medication, PD patients in both medication states (PD_OFF and PD_ON). We extracted blink-related features— blink rate, mean inter-blink interval (IBI), and IBI variability—to assess their discriminative power in multi-class classification. Among several standard machine learning algorithms, gradient boosting achieved the highest accuracy (84.44%), successfully distinguishing all four groups. Other models (RF, LGBM, KNN) also demonstrated robust performance. These findings indicate that blink-based metrics can reliably capture both disease presence and medication effects, supporting their potential as non-invasive biomarkers for monitoring PD progression and treatment response.