PTSD Symptoms Among a Sample of Homeless Former Foster Youth: A Predictive Approach Using Artificial Intelligence (AI)

Background: Reports estimate nearly 2 million unaccompanied homeless youth aged 13-24 in the United States each year. Nearly 40% of these youth were formerly a part of the foster care system. Among both homeless youth and former foster youth, rates of trauma experiences and PTSD symptomatology exist are significantly higher than those of the general public. In examination of issues such as homelessness, foster care, and PTSD, social sciences have historically utilized statistical modeling to examine the relationship between these variables. And while these models have produced valuable information for intervention and treatment programs aimed at treating PTSD symptoms, a number of shortcomings exist for these statistical models. Toward this end, modeling based on the use of artificial intelligence (AI) through machine learning may be effectively used to predict outcomes that typical regression models cannot. This analysis seeks to explore how effectively a number of AI modeling techniques can predict PTSD symptoms, compared to more traditional statistical models.

Methods: A series decision tree models were tested on two datasets. One from a sample 158 quantitative surveys were collected from a sample of homeless former foster youth form a drop-in center in Los Angeles. This data asks a series of foster care-specific questions in addition to a number of behavioral health and PTSD questions. The other originates from a sample of 352 homeless former foster youth from a larger study og HIV risk among homeless youth more generally. Decision trees for both datasets were run individually and combined, with both general variables common to both data sets, as well as those specific foster care variables from the smaller data set.

Results: Decision Trees rules were examined for the most general dataset which was a combination of the two datasets for the common subset of features and the most specific dataset in terms of foster care attributes which is FCHIV including all of its variables, in order to figure out the most important variables for predicting PTSD. For the general dataset, foster care entry age, number of foster care placements and gender were found to be the most important factors. For the specific dataset, comprised of the more specific foster care dataset, gender, being placed in foster care due to neglect, and youth opinion of their foster care experience were found to be the most important factors.

Discussion: Overall, specificity of foster care variables and datasets determined the predictive nature of PTSD symptoms, demonstrating the potential for predictive modeling of PTSD symptoms among foster care variables. These decision tree results could prove to be extremely important in the development of screening tools and interventions aimed at reducing the burden of PTSD symptoms, as it directs practitioners on specific characteristics and experiences that would likely impact the development of PTSD. Future work between social work constructs and AI methods should consider the importance of the type of data collected and needed, both to capture all necessary aspects of a population, but also to meet the specific data requirements needed for AI predictive models.