Time-Series Prediction of Proximal Aggression Onset in Minimally-Verbal Youth with Autism Spectrum Disorder Using Physiological Biosignals

Abstract

It has been suggested that changes in physiological arousal precede potentially dangerous aggressive behavior in youth with autism spectrum disorder (ASD) who are minimally verbal (MV-ASD). The current work tests this hypothesis through time-series analyses on biosignals acquired prior to proximal aggression onset. We implement ridge-regularized logistic regression models on physiological biosensor data wirelessly recorded from 15 MV-ASD youth over 64 independent naturalistic observations in a hospital inpatient unit. Our results demonstrate proof-of-concept, feasibility, and incipient validity predicting aggression onset 1 minute before it occurs using global, person-dependent, and hybrid classifier models.

Multi-level modeling with nonlinear movement metrics to classify self-injurious behaviors in autism spectrum disorder

“…SIBs here lasted for about two seconds at minimum, though more subtle movements, such as picking, lasted longer, which lasted ~ 10 to 90 s. SIB and non-SIB data were balanced as in other work to address skewness 28,39,[79][80][81] . Balanced data were used for training, and tenfold cross-validation was used 18,26,30 . This validation method was implemented to reflect the likely use cases in SIB interventions, including training and validating a model using data from each unique individual.…”

“…Classification models in earlier studies were typically specific to each participant, with training and testing completed on each individual 18,29 . When group-level models were employed, accuracy levels tended to decrease, for example from 80% for individual models to 69% for group-level models in Ozdenizci et al 30 . Additionally, machine-learning based classification methods used in earlier studies (e.g., SVMs or neural networks) can have low interpretability, and other more accessible models should also be explored, such as regression 30 .…”

“…When group-level models were employed, accuracy levels tended to decrease, for example from 80% for individual models to 69% for group-level models in Ozdenizci et al 30 . Additionally, machine-learning based classification methods used in earlier studies (e.g., SVMs or neural networks) can have low interpretability, and other more accessible models should also be explored, such as regression 30 . Interpretable models could provide information about predictors of SIB onset, which would be particularly relevant for clinicians and caregivers seeking to manage this behavior (see Cantin-Garside et al 23 , Dunlap et al 12 and Williams et al 9 for further discussion on the need to capture triggers of SIB).…”

“…Classifiers may have had stronger performance if trained on natural data, (versus simulated SIB), and their generalizability is unknown when used on more than one participant with more than one behavior. One study also examined aggression towards others among youth with ASD 30 . Naturally-collected episodes of aggression were classified with high accuracy using physiological and movement sensors (area under the curve: AUC = 71-80 for individuals; AUC = 69 for group performance), though SIB was not included in the activities of interest 30 .…”

“…One study also examined aggression towards others among youth with ASD 30 . Naturally-collected episodes of aggression were classified with high accuracy using physiological and movement sensors (area under the curve: AUC = 71-80 for individuals; AUC = 69 for group performance), though SIB was not included in the activities of interest 30 . Importantly, sensory aversions prevalent in SIB 31 may preclude the physiological sensors that require skin contact, which were used in Ozdenizci et al 30 , so other sensor and classification methods may be preferable for our application.…”