AI Supports Residential Care Placement Decision Process for Individuals with Autism Spectrum Disorder and Complex Medical Issues

Introduction:

The process of deciding which residential program is appropriate for an individual with profound autism spectrum disorder (ASD) is a challenging one, that is affected by the presentation of the disorder and co-occurring conditions, as well as the specific focus and type of care provided at the residential program. Improper placements can lead to additional stress, a delay in receiving the appropriate care, and the strain associated with being later moved to a new placement. While there are no guidelines on what to look for when making placement decisions, experienced staff members can make use of their knowledge of what characteristics have and have not worked well in the past. However, this knowledge can be lost when a staff member retires or moves on. This work demonstrates how AI models are able to capture the knowledge of which characteristics affect a successful placement for one particular program. 

Methods:

To improve the ability of residential care programs to screen potential residents, this work creates and evaluates a non-linear ADALINE classifier to predict from a small cohort of forty-five current and past residents whether an individual will respond well to the care at one specific residential care program, The Center for Discovery (TCFD). Whether a resident responded well or not was decided holistically by staff at TCFD and is supported by behavior data from a behavior intervention plan. Furthermore, a sensitivity analysis was performed to determine the influence of different attributes on the decision making process both for the cohort as a whole as well as for individual subgroups.

Results:

The classifier was able to achieve a good placement decision with over 80% balanced accuracy using a combination of the first day of sleep data available for a resident and their medical diagnoses history. The most influential individual attribute was the number ofsleep interruptions, which have been linked to high-risk behaviors, such as self-injurious and aggressive behaviors, in this population. Additionally, sleep quality has been strongly linked to overall mental and physical health in the general population. Subsequent subgroup analysis indicates that the model predicts better for individuals who stayed less than 2000 days at TCFD and had their first day of data be eight to thirty days post admission. 

Conclusions:

An AI model can be used to support the placement decision making process, which is particularly effective if a large amount of data on prior placements is available. While this work used data from one particular residential program, the same approach could be used at other programs. This has the potential to reduce inappropriate placement decisions, which improves the delivery of proper care for children with disabilities and reduces the burden on a strained care support system.