Advanced modelling of residuum to improve bionic limbs

Introduction: Current methods to assess human activity recognition (HAR) and fatigue do not account for environmental factors and noise which occur down range. Wearable sensors and machine learning (ML) models offer the promise of determining fatigue in the field however the number of sensors and comparison to gold standards are problematic when translating to the field. Fatigue can be quantified as the level of performance to perform a task relying on a cascade of overlapping redundant systems, influenced by physical, central and psychological factors1. The complexity of fatigue makes ML a promising computational approach2. This work investigates using an ML model with single wearable sensor to predict cognitive and physical fatigue in an outdoor environment. Methods: There are challenges moving into the field including logistics, validation, battery power, equipment size, assessment distractions and validation protocols. The research used a novel protocol to collect data for clinical comparison and computational model training in an unstructured environment with varying fatigue levels with self-pacing. The trail was segmented by terrain surface, slope or obstacle using observation, satellite imagery and GPS. Sensor data was collected from a Zephyr BioHarness (Medtronic, Minneapolis, USA). Activity was voluntary and self-paced so transitions could not be determined a priori. Terrain included slope, crossing obstacles (cross river or climb fence) and surfaces including road, gravel, clay, mud, long grass. A goal was set to run 100 km with 10,000 feet of ascent within 25 hours to trigger goal assessment with possible action crisis, which required each one-hour cycle to complete 4 km trail in 35 minutes. Mental load used a software tool implementing the NASA Multi Attribute Test. Clinical comparison was performed against, Stroop, Jump Height, Finger Tap Test (FTT), Trail Making A (PVSAT), spatial memory. Separate convolutional neural networks were developed and trained to predict HAR, physical fatigue and cognitive fatigue. Results: HAR results for 5 features accuracy 0.98 and minimum class precision for one-off activity 0.802. Fatigue prediction accuracy was highest for FTT right hand (R2 0.71) and Jump Height (R2 0.78). Discussion/conclusions: This experiment showed a single sensor and ML model can achieve accurate results for HAR, cognitive fatigue and physical fatigue equivalent to gold standard tests performed in clinic. Additionally, a field environment can be calibrated to enable translational real world dataset noise and equipment acceptable to an operator. Further research is required for more individuals, terrain and equipment carriage, sleep deprivation, dehydration, and recovery cycles. References 1Russell B, McDaid A, Toscano W, et al. Moving the lab into the mountains: a pilot study of human activity recognition in unstructured environments. Sensors 2021; 21(2):654. https://doi.org/10.3390/s21020654 2Russell B, McDaid A, Toscano W, et al. Predicting fatigue in long duration mountain events with a single sensor and deep learning model. Sensors 2021; 21(16):5442. https://doi.org/10.3390/s21165442