Accuracy of sensor-based classification of clinically relevant motor activities in daily life of children with mobility impairments

Abstract

Background: Wearable inertial sensors enable objective, long-term monitoring of motor activities in the children's habitual environment after rehabilitation. However, sophisticated algorithms are needed to derive clinically relevant outcome measures. Therefore, we developed three independent algorithms based on the needs of pediatric rehabilitation. The first algorithm estimates the duration of lying, sitting, and standing positions and the number of sit-to- stand transitions with data of a trunk and a thigh sensor. The second algorithm detects active wheeling periods and distinguishes it from passive wheeling with data of a wrist and a wheelchair sensor. The third algorithm detects walking periods, discriminates between free and assisted walking, and estimates the covered altitude change during stair climbing with data of a single ankle sensor and a sensor placed on walking aids. Research question: This study aimed to determine the accuracy of each algorithm in children undergoing rehabilitation. Methods: Thirty-one children and adolescents with various medical diagnoses and levels of mobility impairments performed a semi-structured activity circuit. They wore inertial sensors on both wrists, the sternum, and the thigh and shank of the less-affected side. Video recordings, which were labeled by two independent researchers, served as reference criteria to determine the algorithms' performance. Results: The activity classification accuracy was 97% for the posture detection algorithm, 96% for the wheeling detection algorithm, and 93% for the walking detection algorithm. Significance: This study presents three novel algorithms that provide a comprehensive and clinically relevant view of the children's motor activities. These algorithms are described reproducibly and can be applied to other inertial sensor technologies. Moreover, they were validated in children with mobility impairments and can be used in clinical practice and clinical trials to determine the children's motor performance in their habitual environment. To enable the evaluation of future algorithms, we published the labeled dataset.

Concurrent Validity of Different Sensor-Based Measures: Activity Counts Do Not Reflect Functional Hand Use in Children and Adolescents With Upper Limb Impairments

“…We used an existing dataset of 31 school-aged children and adolescents undergoing rehabilitation, which was established to validate sensor-based outcomes of upper and lower extremities. 15 The children's abilities to handle objects were determined with the Manual Ability Classification System (MACS). 16 They were able to walk or use a manual wheelchair for household distances, had cognitive abilities to follow instructions, had no medical conditions that prevented sensor placement, and provided informed consent to participate in the study.…”

“…18 The sensors were placed on both wrists, the sternum, and the thigh and ankle of the less-affected side with hook-and-loop straps. 15 Data of the wrist sensors were used to derive the hand use measures. In addition, the ankle sensor and an additional sensor on the spokes of the wheelchair were used to detect walking and active wheeling periods, respectively.…”

“…The participants were videotaped while performing a semi-structured activity circuit at the rehabilitation center. 15 They watched a movie on a tablet in their bedroom, played a selfselected game (e.g., card games, puzzles, etc.) in the living room, drank a glass of water in the restaurant, cycled in the gym hall, and played a self-selected outdoor game on the playground (e.g., catching and throwing balls, swinging, etc.).…”

“…They were detected with an activity recognition algorithm having an accuracy of >93%. 15 Then, the acceleration signal of these periods was set to zero before applying approach 1 to the adjusted signal.…”