Fig. 4: Blink data acquisition, analysis and machine learning procedure.

A The Pupil Core eye tracker (Pupil Labs GmbH, Berlin, Germany. The upper left image is cited from the product webpage (https://pupil-labs.com/products/core) with permission), example of pupil image by infrared eye cameras, and recording setup of eye tracker wired to a smartphone. B Blink extraction from pupil data and blink feature extraction. Pupil Player extracts eye blinks based on pupil confidence. A differential filter was applied to extract blink onset and offset. Blink interval and duration were calculated from the onset and offset detection. Blink confidence was calculated as the proportion of the shaded area within the region enclosed by the dashed line calculated for each blink. C Data processing flow. The Base Blink Features #1 were transformed into three features through the Baseline Correction: Raw Value (retaining the original, unprocessed value), the Difference from Baseline (baseline: the mean prior to L-dopa administration within the data from the same individual), and the Absolute Value of the Difference from Baseline. These parameters were subsequently normalized in two ways: within the individual and across all patient data. The Base Blink Features #2 were initially converted into multiple statistical features using Statistical Feature Extraction, and then processed in the same manner as Base Features #1, yielding a total of 468 features. These features were then selected for each model in the training phase based on their contribution. Details of features can be found in Tables 2 and 3. D Machine learning procedure. Time-series data were segmented by a time window of 3 min and treated as independent data points. Clinical indicators such as dyskinesia, ON/OFF, MDS-UPDRS Part III total score and plasma L-dopa concentration were set as target variables, while blink features were set as feature variables. machine learning was conducted using DataRobot, the automated machine learning platform, and custom-Python programs. The DataRobot logo is cited from product webpage (https://www.datarobot.com/) with permission. See text for details.