Evaluating Sparse Inertial Measurement Unit Configurations for Inferring Treadmill Running Motion

Abstract

Inertial measurement units (IMUs) are used to analyze running performance. While leveraging one sensor to estimate kinematic and kinetic variables is common, sparsity limits the number of digital biomarkers that can be evaluated. Shallow recurrent decoder networks (SHRED) can reconstruct a dense set of time-series signals from a single input sensor and have been successful in human mobility applications, highlighting the potential for this algorithm to monitor running. We trained and tested subject-specific SHRED models of nine subjects running on a treadmill to map from one input sensor to the remaining three IMUs. We varied the type of input to reflect experimental parameters that are important in running studies-sensor location, sensor type, sampling rate, and running speed-and compared the error of inferred signals from each input type. Sensor location and type did not impact SHRED inference accuracy, while decreasing the sampling rate affected the accuracy of ankle measurements. All ankle acceleration inferences from these models remained below the minimal detectable change threshold of 12.0 m/s². SHRED models trained and tested at multiple speeds did not accurately infer IMU measurements below this threshold. SHRED may broaden the scope of motion analysis by expanding datasets with fewer sensors.

Keywords: IMU; accelerometer; machine learning; running; sampling rate; sparse sensing.

Figure 1

Architecture of a shallow recurrent decoder (SHRED) network. Sparse time-series measurements from an inertial measurement unit (IMU) feed into a recurrent neural network. The encoded signals then feed into a shallow decoder to learn their spatial relationship with the other sensors. The output contains inferred time-series measurements of IMU signals at multiple sensor locations. Modified from [22] with authors’ permission using BioRender.

Figure 2

Left ankle signals predicted from individualized SHRED models with input sensor locations at the chest (pink), hip (green), or right ankle (blue). Errors in (a) acceleration and (c) angular velocity signal inference are measured as root mean squared error (RMSE) and aggregated across all subjects’ models with average ± 1 standard deviation. Signals are further separated by the directions of vertical, anteroposterior (AP), and mediolateral (ML) acceleration and transverse, frontal, and sagittal plane angular velocity. A sample time interval of the inferred left ankle (b) vertical acceleration and (d) sagittal plane angular velocity is shown for a representative subject’s three SHRED models, along with the true signal measurement (dashed line).

Figure 3

Left ankle acceleration from individualized SHRED models with input sensor types of uniaxial accelerometer (light blue), triaxial accelerometer (blue), or triaxial accelerometer and triaxial gyroscope (dark blue). (a) Error of signal inference is measured as root mean squared error (RMSE) and aggregated across all subjects’ models to show average ± 1 standard deviation. Signals are further separated by the vertical, anteroposterior (AP), and mediolateral (ML) directions. (b) A sample time interval of the inferred left ankle vertical acceleration is shown for a representative subject’s three SHRED models, along with the true signal measurement (dashed line).

Figure 4

Linear acceleration from individualized SHRED models with sampling rates of 16, 32, 64, and 128 Hz. All models were trained with a right ankle triaxial accelerometer input. (a) Error of signal inference is measured as root mean squared error (RMSE) and aggregated across all subjects’ models, shown as average ± 1 standard deviation. Signals are further separated by the output locations of chest, hip, and left ankle. (b) A sample time interval of the inferred left ankle vertical acceleration is shown for a representative subject’s four SHRED models with varying sampling rates, along with the true signal measurement (dashed line).

Figure 5

Left ankle acceleration from individualized SHRED models with different training data running speeds. Interpolation models were trained with two speeds and tested with an unseen speed of 2.2 m/s, while extrapolation models were trained using two running speeds and tested with an unseen speed of 2.7 m/s. (a) Error of signal inference is measured as root mean squared error (RMSE) and aggregated across all subjects’ models as average ±1 standard deviation. (b) A sample time interval of the inferred left ankle vertical acceleration is shown for a representative subject’s interpolation SHRED results, along with the true signal measurement (dashed line).