JMIR Publications

ABSTRACT

Background:

Human movement activity is commonly recorded with inertial measurement unit (IMU) sensors in many science disciplines. The IMU data is used for an algorithmic detection of different postures and/or movements, which can be used further for detailed and quantified assessments of more complex behaviors, such as daily activities. Since the studies are increasingly targeting human behavior in real life environments, there is a growing need to strike balance between recording complexity and analytic yield. However, it is poorly understood how the sensor configurations affect the accuracy of classifier algorithms, and therefore, what are sufficient recording configurations for different analysis targets.

Objective:

This study aims to define the effects of IMU sensor configurations on automatic, high temporal resolution classification of posture and movement when the data represents naturalistic daily activity without excessively repetitive movements. The overall goal is to identify the minimal sensor configurations that enable sufficiently accurate motility assessment.

Methods:

We used a purposefully challenging IMU dataset from spontaneously behaving infants (N=41; age range 4–18 months) with a multisensor wearable suit. The classification accuracy was determined using synchronously recorded video, and the reference IMU recording configuration included four IMU sensors using with triaxial accelerometer and gyroscope modalities sampled at 52 Hz. Then, we reduced this configuration stepwise to test how the algorithmic classification performance of postures (N=7) movements (N=9) is affected by reducing IMU data sampling frequency, sensor modality, and sensor placement.

Results:

Our results show that reducing the number of sensors has a significant effect on classifier performance. Most strikingly, single sensor configurations cannot be held reliable for the studied tasks. Reducing sensor modalities to accelerometer only, i.e., dropping gyroscope data, leads to modest reduction in movement classification performance. However, sampling frequency could be reduced from 52 to 13 Hz with negligible effects on the classifications.

Conclusions:

The present findings highlight the significant effects of IMU sensor configurations on the analytic accuracy, and hence on the addressable study questions. Notably, the single sensor recordings employed in most of the literature and wearable solutions are of very limited use when assessing posture and movement during daily behaviours at a high temporal resolution. For the accurate assessment of movements and postures, the minimal configuration with an acceptable classifier performance includes: at least a combination of one upper and one lower extremity sensor, 13 Hz sampling frequency, and at least accelerometer but preferably also gyroscope. These findings have direct implications to the design of future studies and wearable solutions that aim to assess real-life daily behaviors with minimal cost and maximal detail.