The Effect of Sensor Feature Inputs on Joint Angle Prediction across Simple Movements

Abstract

The use of wearable sensors, such as inertial measurement units (IMUs), and machine learning for human intent recognition in health-related areas has grown considerably. However, there is limited research exploring how IMU quantity and placement affect human movement intent prediction (HMIP) at the joint level. The objective of this study was to analyze various combinations of IMU input signals to maximize the machine learning prediction accuracy for multiple simple movements. We trained a Random Forest algorithm to predict future joint angles across these movements using various sensor features. We hypothesized that joint angle prediction accuracy would increase with the addition of IMUs attached to adjacent body segments and that non-adjacent IMUs would not increase the prediction accuracy. The results indicated that the addition of adjacent IMUs to current joint angle inputs did not significantly increase the prediction accuracy (RMSE of 1.92° vs. 3.32° at the ankle, 8.78° vs. 12.54° at the knee, and 5.48° vs. 9.67° at the hip). Additionally, including non-adjacent IMUs did not increase the prediction accuracy (RMSE of 5.35° vs. 5.55° at the ankle, 20.29° vs. 20.71° at the knee, and 14.86° vs. 13.55° at the hip). These results demonstrated how future joint angle prediction during simple movements did not improve with the addition of IMUs alongside current joint angle inputs.

Keywords: accelerometers; gyroscopes; movement intent prediction; wearable sensors.

Figure 1

Experimental setup. (a) IMU placement of the lower limb and foot. (b) Frontal view of the placement of retroreflective markers and IMU sensors. (c) IMU sensor axes of flexion/extension, abduction/adduction, and internal/external rotation following the x-z-y coordinate system.

Figure 2

Lab-based data analysis and algorithm overview. Trials of multiple actions (action-agnostic) were trained with sensory information input for a continuous Random Forest predictor. Joint angle predictions were compared to angles obtained from motion capture using inverse kinematics.

Figure 3

Input signal combinations for predicting ankle, knee, and hip angles. The blue rectangles on the person represent the evaluated IMU sensor configuration. Two IMUs were evaluated to the neighboring joint as a reduced sensor set. The two IMU sensors were located at the shank and foot to predict ankle angles, the thigh and the shank to predict knee angles, and the torso and thigh to predict hip angles. An identical set of four IMU sensors was evaluated at the foot, shank, thigh, and torso to predict ankle, knee, and hip angles.

Figure 4

Prediction of RMSE of joint angles 100 ms into the future, showing significant results denoted as * p < 0.05, ** p < 0.01, and *** p < 0.001.

Figure 5

Bland–Altman plots corresponding to the difference between the predicted and measured ankle angles. Sensor combinations are displayed for ankle dorsiflexion (a–e) and ankle plantarflexion (f–j). The dotted red line represents the slope of the differences between the predicted and measured angles. The scatterplot colors are darkened for additional number sensor inputs and purple is shown for 4 IMUs. The 95% prediction limits and Pearson correlation coefficients assessing the proportional bias are displayed in the upper left corner of each plot. Error bars represent the standard error.

Figure 6

Bland–Altman plots corresponding to the difference between the predicted and measured knee angles. Sensor combinations are displayed for ankle dorsiflexion (a–e) and ankle plantarflexion (f–j). The dotted red line represents the slope of the differences between the predicted and measured angles. The scatterplot colors are darkened for additional number sensor inputs and purple is shown for 4 IMUs. The 95% prediction limits and Pearson correlation coefficients assessing the proportional bias are displayed in the upper left corner of each plot. Error bars represent the standard error.

Figure 7

Bland–Altman plots corresponding to the difference between the predicted and measured hip angles. Sensor combinations are displayed for ankle dorsiflexion (a–e) and ankle plantarflexion (f–j). The dotted red line represents the slope of the differences between the predicted and measured angles. The scatterplot colors are darkened for additional number sensor inputs and purple is shown for 4 IMUs. The 95% prediction limits and Pearson correlation coefficients assessing the proportional bias are displayed in the upper left corner of each plot. Error bars represent the standard error.