Automated Detection of Stereotypical Motor Movements in Children with Autism Spectrum Disorder Using Geometric Feature Fusion

Friday, May 11, 2018: 10:00 AM-1:30 PM
Hall Grote Zaal (de Doelen ICC Rotterdam)
C. Tralie1, M. S. Goodwin2 and G. Sapiro1, (1)Duke University, Durham, NC, (2)Northeastern University, Boston, MA

One of the two main diagnostic criteria for autism spectrum disorder (ASD) in the DSM-5 is restricted, repetitive patterns of behavior, interests, and/or activities, and one of the ways these behaviors manifest in ASD is stereotypical motor movements (SMM). Traditional measures of SMM primarily include rating scales, direct behavioral observation, and video-based methods, all of which can be subjective, inaccurate, time-intensive, and difficult to compare across different individuals with ASD. More reliably, accurately, and efficiently detecting and monitoring SMM over time could provide important insights for understanding and intervening upon a core symptom of ASD.


Leverage a novel set of features based on sliding windows and topological data analysis to computationally detect the onset and type of SMM using accelerometer data from children with ASD. Also demonstrate that novel features we developed enable a more parsimonious representation of periodic SMM when combined with previous methods that, collectively, boost automated classification performance reported in the published literature.


We used publicly available data from [Goodwin etal. 2014] to study automated classification of SMM in a subset of 6 subjects, with each session spanning approximately 20 minutes. Each subject had a 3-axis accelerometer on his/her left wrist, right wrist, and torso to measure stereotypical hand flapping and body-rocking. Each accelerometer time series was accompanied by annotated ground truth labels provided by human coders indicating time-stamped onset and offset for the following three operationally defined SMM: flap, rock, and flap+rock. We segmented all of the accelerometer data into 2-second windows that overlapped by 130 milliseconds. For each window, we extracted both recurrence quantification analysis (RQA) features (as in [Großekathöfer et al. 2017], 9 dimensions per accelerometer, 27 dimensions total) and applied novel topology-based “persistence” features that measure “roundness” of a sliding window reconstruction of the joint embedding of the 3-axis accelerometers using persistent homology (novel feature, 1 number per accelerometer, 3 dimensions total). After subsampling the windows corresponding to periods when no SMM was observed to balance the data, we analyzed the ability of our features to classify SMM into the three target classes by running simple cross-validation experiments.


After combining all labeled sessions, we tested using 10-fold cross validation with a decision tree using only RQA features, only persistence features, and both combined. For the task of identifying the class of a SMM in a 2-second window or the absence therein, RQA features produced 85.9% classification accuracy, and persistence features (which use an order of magnitude fewer dimensions) yielded a similar classification accuracy of 84.8%. Moreover, when these methods were combined, they together yielded automated classification performance of 90.6% accuracy, demonstrating that in addition to their parsimony, the newly developed feature enables increasingly more accurate automated detection of SMM.


Based on recent theory on periodicity analysis using geometry, this study demonstrates the feasibility and validity of achieving excellent performance for automatically detecting and classifying SMM computationally in a low dimensional space with only 3 degrees of freedom.