A Novel Framework for Quantifying Accuracy and Precision of Event Detection Algorithms in FES-Cycling.
Ronan Le GuillouMartin SchmollBenoît SijobertDavid Lobato BorgesEmerson Fachin-MartinsHenrique ResendeRoger Pissard-GibolletCharles FattalChristine Azevedo CostePublished in: Sensors (Basel, Switzerland) (2021)
Functional electrical stimulation (FES) is a technique used in rehabilitation, allowing the recreation or facilitation of a movement or function, by electrically inducing the activation of targeted muscles. FES during cycling often uses activation patterns which are based on the crank angle of the pedals. Dynamic changes in their underlying predefined geometrical models (e.g., change in seating position) can lead to desynchronised contractions. Adaptive algorithms with a real-time interpretation of anatomical segments can avoid this and open new possibilities for the automatic design of stimulation patterns. However, their ability to accurately and precisely detect stimulation triggering events has to be evaluated in order to ensure their adaptability to real-case applications in various conditions. In this study, three algorithms (Hilbert, BSgonio, and Gait Cycle Index (GCI) Observer) were evaluated on passive cycling inertial data of six participants with spinal cord injury (SCI). For standardised comparison, a linear phase reference baseline was used to define target events (i.e., 10%, 40%, 60%, and 90% of the cycle's progress). Limits of agreement (LoA) of ±10% of the cycle's duration and Lin's concordance correlation coefficient (CCC) were used to evaluate the accuracy and precision of the algorithm's event detections. The delays in the detection were determined for each algorithm over 780 events. Analysis showed that the Hilbert and BSgonio algorithms validated the selected criteria (LoA: +5.17/-6.34% and +2.25/-2.51%, respectively), while the GCI Observer did not (LoA: +8.59/-27.89%). When evaluating control algorithms, it is paramount to define appropriate criteria in the context of the targeted practical application. To this end, normalising delays in event detection to the cycle's duration enables the use of a criterion that stays invariable to changes in cadence. Lin's CCC, comparing both linear correlation and strength of agreement between methods, also provides a reliable way of confirming comparisons between new control methods and an existing reference.
Keyphrases
- machine learning
- deep learning
- big data
- artificial intelligence
- loop mediated isothermal amplification
- real time pcr
- high intensity
- spinal cord injury
- label free
- cancer therapy
- neural network
- electronic health record
- minimally invasive
- magnetic resonance
- high resolution
- magnetic resonance imaging
- computed tomography
- diffusion weighted imaging