Automated Detection of Stereotypical Motor Movements in Autism Spectrum Disorder Using Recurrence Quantification Analysis

Abstract

A number of recent studies using accelerometer features as input to machine learning classifiers show promising results for automatically detecting stereotypical motor movements (SMM) in individuals with Autism Spectrum Disorder (ASD). However, replicating these results across different types of accelerometers and their position on the body still remains a challenge. We introduce a new set of features in this domain based on recurrence plot and quantification analyses that are orientation invariant and able to capture non-linear dynamics of SMM. Applying these features to an existing published data set containing acceleration data, we achieve up to 9% average increase in accuracy compared to current state-of-the-art published results. Furthermore, we provide evidence that a single torso sensor can automatically detect multiple types of SMM in ASD, and that our approach allows recognition of SMM with high accuracy in individuals when using a person-independent classifier.

Keywords: ASD; autism; recurrence analysis; recurrence plots; repetitive behavior; stereotypical motor movement.

Figure 1

Accelerometer readings of one second in length from the class flapping. The accelerometer was mounted to the right wrist. Each line corresponds to one of the three acceleration axes.

Figure 2

Recurrence plots computed from the data displayed in Figure 1 and for three different values of the threshold parameter ε.

Figure 3

This figure illustrates the results from Experiment 2. Each line characterizes the dependency between the training set size (y-axis) and the reached accuracy (x-axis) for one participant by means of RF classifiers. The accuracy values were derived in a k-fold cross-validation where the folds correspond to the recording sessions of Study 1. The training examples where randomly selected from all available training data.

Figure 4

This plot shows average precision and recall values based on training set size, corresponding to Figure 3 and RF classifiers.

Figure 5

This plot shows the classification accuracy based on sensor position with RF classifiers. Each bar corresponds to one sensor from one participant. Sensors are grouped by participants, where the right-most group summarizes the average accuracy per sensor. We estimated the accuracy by means of k-fold cross-validation where folds correspond to recording sessions.

Figure 6

This figure illustrates the five most important features yielded by the RF classifier for data collected from participant 2. Each bar corresponds to a single feature. RQA features were extracted from a sensor position (right, left, torso) and the optimal ε value found in cross-validation was used.

Figure 7

This figure illustrates the five most important features yielded by the RF classifier for data collected from participant 4. Each bar corresponds to a single feature. RQA features were extracted from a sensor position (right, left, torso) using the optimal ε value found in cross-validation.