Answers are grounded in this paper and cite their sources.
Objective monitoring of repetitive motor behaviors, or stimming, could support quantitative assessment in autism research and assistive technology, but wearable systems reported to date are typically validated on very few participants and benchmarked on desktop hardware, leaving their actual deployability unresolved. Here, a compact wrist‐worn nine‐degree‐of‐freedom (9DoF) inertial sensing node is presented together with a deployment‐oriented rule for selecting a wearable classifier. Recordings from participants performing four stimming behaviors and two nonstimming control activities are used to compare three feature‐based classifiers and four deep‐learning architectures across four sensing configurations under fully subject‐independent validation. Accuracy is found not to be the limiting factor: across the feature‐based classifiers, macro‐ F 1 exceeds 0.97 for every model and sensing configuration, and stimming is separated from nonstimming activity with an F 1 above 0.998 using the accelerometer alone. Deployment cost, by contrast, varies by more than an order of magnitude, and porting the networks to the target microcontroller reveals on‐device inference spanning 3.7–631 ms, modifying the ranking obtained on a workstation. Model selection for wearable stimming monitoring is therefore governed by memory and latency rather than by accuracy, and on‐device measurement is shown to be a prerequisite rather than a refinement.
No takes yet. Share an insight, caveat, or question.
Angel et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: