RF Sensing for Automated Classification of Stimming Behaviors in ASD: A First Step Toward Contactless Screening

The automated analysis and classification of physical behaviors associated with Autism Spectrum Disorder (ASD) offers a promising application of machine learning techniques, with the potential to support and streamline conventional assessment methods typically conducted by expert clinicians. However, established methods rely heavily on wearable sensors that can cause discomfort to ASD patients with sensory processing issues as well as video data that can compromise the privacy of vulnerable individuals. As such, recent developments in contactless sensing technologies such as Radio Frequency (RF) Sensing have generated interest as a viable alternative. This study explores the potential of RF Sensing for classification of ASD related physical behaviours. Four stimming actions are derived from a public 'stimming' dataset and recorded via radar. The raw data are then preprocessed using established RF Sensing techniques and the resulting doppler spectrograms are used to fine-tune a number of pre-trained image recognition models. All pre-trained models achieve cross-validation test accuracies of 90% and above, with resnet50 achieving 97.5%. These results serve to prove the efficacy of established RF Sensing methods for stimming classification, suggesting further validation through the use of a comprehensive dataset that includes more behaviors and participants with varying biometrics.