ABSTRACT
Background: The severity of antipsychotic-induced cervical dystonia has traditionally been evaluated visually. However, recent advances in information technology made quantification possible in this field through the introduction of engineering methodologies like machine learning.Methods: This study was conducted from June 2021 to March 2023. Psychiatrists rated the severity of cervical dystonia into 4 levels (0: none, 1: minimal, 2: mild, and 3: moderate) for 101 videoclips, recorded from 87 psychiatric patients receiving antipsychotics. The Face Mesh function of the open-source framework MediaPipe was employed to calculate the tilt angles of anterocollis or retrocollis, laterocollis, and torticollis. These were calculated to examine the range of tilt angles for the 4 levels of severity of the different types of cervical dystonia.Results: The tilt angles calculated using Face Mesh for each level of dystonia were 0° ≤ θ < 6° for none, 6° ≤ θ < 11° for minimal, 11° ≤ θ < 25° for mild, and 25° ≤ θ for moderate laterocollis; 0° ≤ θ < 11° for none, 11° ≤ θ < 18° for minimal, 18° ≤ θ <25° for mild, and 25° ≤ θ for moderate anterocollis or retrocollis; and 0° ≤ θ < 9° for none, 9° ≤ θ < 17° for minimal, 17° ≤ θ < 32° for mild, and 32° ≤ θ for moderate torticollis.Conclusion: While further validation with new cases is needed, the range of tilt angles in this study could provide a standard for future artificial intelligence devices for cervical dystonia.