A Deep Learning-Based Framework Oriented to Pathological Gait Recognition with Inertial Sensors

Abstract

Abnormal locomotor patterns may occur in case of either motor damages or neurological conditions, thus potentially jeopardizing an individual's safety. Pathological gait recognition (PGR) is a research field that aims to discriminate among different walking patterns. A PGR-oriented system may benefit from the simulation of gait disorders by healthy subjects, since the acquisition of actual pathological gaits would require either a higher experimental time or a larger sample size. Only a few works have exploited abnormal walking patterns, emulated by unimpaired individuals, to perform PGR with Deep Learning-based models. In this article, the authors present a workflow based on convolutional neural networks to recognize normal and pathological locomotor behaviors by means of inertial data related to nineteen healthy subjects. Although this is a preliminary feasibility study, its promising performance in terms of accuracy and computational time pave the way for a more realistic validation on actual pathological data. In light of this, classification outcomes could support clinicians in the early detection of gait disorders and the tracking of rehabilitation advances in real time.

Keywords: bioengineering; convolutional neural network; deep learning; gait disorders; gait recognition; inertial measurement units; rehabilitation.

Figure 1

The presented framework is based on the acquisition of inertial data by means of five IMU sensors, whose components are given as input to each of three DL-based models, which return the label associated with the walking pattern.

Figure 2

The classification models addressing PGR: (a) a multi-branch one-dimensional convolutional neural network; (b) a simplified multi-branch one-dimensional convolutional neural network; (c) a sequential one-dimensional convolutional neural network. Note that $N_w$ is the number of windows in the input dataset, which differs with both subjects and trials; $W_l$ is the window length, which is fixed; and $N_{ch}$ is the number of sensor channels.

Figure 3

Sensor combinations grouped by the number of sensors, which can be placed at the sternum (S), the left pelvis (LP), the right pelvis (RP), the left wrist (LW), and the right wrist (RW).

Figure 4

Radar plot comparing metrics computed on the test set for the multi-branch CNN and its simplified version, with *, ** and *** representing statistically significant comparisons with $p<0.05$, $p<0.01$, and $p<0.001$, respectively.

Figure 5

Radar plot comparing metrics computed on the test set for the simplified multi-branch and the sequential CNNs, with *, **, and *** representing statistically significant comparisons with $p<0.05$, $p<0.01$, and $p<0.001$, respectively.

Figure 6

Radar plot of the accuracy, recall, and inference time computed on the test set by feeding the sequential CNN with each IMU component separately.

Figure 7

Related works simulating abnormal walking patterns [2,6,7,11,22,23,24].