Detecting and Classifying Self-injurious Behavior in Autism Spectrum Disorder Using Machine Learning Techniques.

Researchers put small accelerometers on the kids with autism. The kids were 6-12 years old and had a history of head-hitting, hand-biting, or body-banging.

The team recorded every move for two hours. They fed the data into two machine-learning programs: k-NN and SVM. The goal was to spot true SIB moments and ignore normal play.

The wrist sensor caught 99 % of SIB episodes. The ankle sensor caught 94 %. Each alert popped up on a tablet within three seconds.

False alarms stayed below 3 %. The system worked for light and hard hits, with or without a jacket on.

Gilchrist et al. (2018) used the same wrist trick but looked at hand-flapping, not SIB. Their hit rate was 80-93 %. The new study pushes accuracy to 99 % by training the model only on injury-level hits.

Maharaj et al. (2020) used a Kinect camera and reached 92 % agreement with human eyes. Wearables beat the camera when the child turned away or wore loose clothes.

Takahashi et al. (2023) claim 100 % accuracy in a classroom, but they tracked 14 common behaviors like raising a hand. Their perfect score drops when the list shrinks to rare, fast SIB. The 94-99 % here is more honest for low-base-rate events.

You no longer need mom or dad to guess how many times Danny hit his head. Tape a $25 wristband on him, open the app, and you get a live count. Use the graph to see if your intervention is working session-by-session instead of waiting for tomorrow’s parent report.