← Back to Digital Poster Abstracts
DP60: SIMULATION-BASED VALIDATION OF MOTION EXTRACTION FOR QUANTITATIVE TRAIN OF FOUR ASSESSMENT
Adam Wolach, MD; Nikolaus Gravenstein, MD
University of Florida Department of Anesthesiology
Introduction: The accessibility of quantitative Train of Four (qTOF) for monitoring recovery from neuromuscular blockade (NMB), a strong recommendation of current society guidelines,1 is limited by expensive deployment and cumbersome hardware. In our previous work, we demonstrated proof-of-concept for qTOF via an AI machine vision hand-tracking system utilizing OpenCV and MediaPipe.2 In this work, we present an alternative video processing technique, motion extraction (ME), which we theorize can provide motion analysis agnostic to procedural factors that might limit our previous technique, such as lighting, camera position, and rolling shutter camera effects. The technique utilizes time-shifting and color inversion to transform motion into brightness data3, which is then analyzed to calculate qTOF.
Methods: The ME technique involed frame stabilization and subsequent integration of brightness data representing motion (Figure 1). Pattern matching isolated 1, 2, 3 or 4 twitches at 2hz and a twitch number or qTOF ratio were returned. The technique was tested against a corpus of short videos of a 3D simulated hand built in the Unity® game engine with known qTOF ratios, varied camera angles and simulated physiologic tremor. Detection accuracy was calculated with a mean and standard deviation (SD) of the difference between observed and ground truth, and a coefficient of determination (R2). We used Bland-Altman analysis to identify agreement and systematic error.
Figure 1
Results and Discussion: 660 trials were conducted across each twitch strength and camera position. The mean and SD of differences of observed and ground truth ratios were 0.00168 and 0.030449, respectively (Figure 2). R2 was 0.9918. Motion extraction produced more consistent results with a higher R2. Bland-Altman analysis showed very low error in the clinically sensitive domain near qTOF 0.9. The system demonstrated systematic error with in the domain of high qTOF values (>120%) (Figure 3). Systematic error could be due to thresholds used to account for noise, the parameters used for detection and integration of each twitch, and our reliance on simulated hand motion in testing.
Figure 2
Figure 3
Given the strong agreement with ground truth, particularly near qTOF of 0.9, this work demonstrates that ME presents a promising avenue for tracking patient movement and providing access to qTOF. Future work will involve threshold optimization, combined techniques with hand tracking, and in vivo testing of motion extraction compared to gold-standard qTOF methods.
References:
1. Thilen SR, Weigel WA, Todd MM, et al. 2023 American Society of Anesthesiologists Practice Guidelines for Monitoring and Antagonism of Neuromuscular Blockade: A Report by the American Society of Anesthesiologists Task Force on Neuromuscular Blockade. Anesthesiology. 2023;138(1):13. doi:10.1097/ALN.0000000000004379
2. Wolach A. Simulation-Based Validation of AI Computer Vision for Quantitative Train of Four. Presented at: 2025 STA Annual Meeting; January 10, 2025. https://www.stahq.org/account/annual-meeting-archive/2025
3. Motion Extraction.; 2023. Accessed January 11, 2026. https://www.youtube.com/watch?v=NSS6yAMZF78
4. Carignan B, Daneault JF, Duval C. Quantifying the importance of high frequency components on the amplitude of physiological tremor. Exp Brain Res. 2010;202(2):299-306. doi:10.1007/s00221-009-2132-7