• Meltdown detection in autistic children combining stress sensors and machine learning

      Singh, Sarah (2022-05)
      Children with autism spectrum disorder face many challenges on a daily basis, including their struggle to communicate their needs, especially in times of distress. This can lead to meltdowns, making it difficult for them to learn, make friends, or have a positive social or educational experience. Existing research detecting meltdowns, specifically using deep learning combined with either facial recognition [1] or a variety of sensors such as heart rate, electrodermal, and temperature sensors [2], have proved successful. However, optimization for practical application utilizing more affordable technology could improve upon the accessibility of these tools for the autistic community, especially working class families. This thesis provides a method to detect and prevent autistic meltdowns inspired by my son, aiming to make a wearable device that can be used whenever and wherever by combining heart rate monitors and electrodermal sensors as a more practical means of detection, as well as a more cost friendly option using low power equipment. The device was built on an STM32-F446RE nucleo board using the kernel based operating system FreeRTOS. A bluetooth android application was created using MIT APP Inventor 2, allowing easy access to sensor data. The device was tested on a child diagnosed with autism by wearing a finger glove with sensors attached during their every day homework routine. A simple logistic regression model was applied to calculate the slope of the sensor's data. The logistic regression model showed promising results with an accuracy score of 0.82 and a recall score of 0.83. This device can be easily modified into a wrist watch interface, making it more comfortable and practical for autistic children to wear. The low cost sensors and processor, combined with a lower cost method of machine learning gives families a better chance at owning a device that could help their child. Meltdown predictions will allow teachers and guardians an opportunity for early intervention and meltdown mitigation.