How the public sector employs AI in social welfare provision

Summary:

An AI-powered family screening tool has improved the effectiveness and efficiency of the call screening process for child abuse reports. The system has achieved 77 percent accuracy in identifying children at risk of being placed into protective care.

Client: 

Allegheny County is a medium size county in Pennsylvania, USA, and comprises the greater metropolitan area of the city of Pittsburgh.

Problem Statement: 

All 50 US states have laws requiring suspected child abuse to be reported to child protection services (CPS), usually through a call center. Call screeners must quickly decide whether to investigate based on the information received. This process required staff to manually review extensive data to determine if a case should be investigated. Without standardized protocols for using the data or for systematically weighting the information, an analysis found that 27% of highest risk cases were being screened out and 48% of the lowest risk cases were being screened in.

In 2014, the Allegheny County Department of Human Services (DHS) aimed to improve child welfare screening and sought innovative solutions to enhance decision-making and child safety.

Results: 

• Achieved 77 percent accuracy in identifying children at risk of placement, based on case data from 2010-2015.
• Improved call screening process by ensuring more high-risk children were screened in and low-risk children were screened out.
• The proportion of low-risk cases being screened in for investigation dropped from half to one third, prioritizing high-risk cases for investigation.

AI Solution Overview:

The Allegheny Family Screening Tool (AFST) was developed using AI techniques to identify patterns in historical data. Its primary aim is to assist DHS call screeners in making informed decisions when responding to calls alleging child maltreatment. The AFST serves as a predictive risk modeling AI tool that quickly integrates and analyzes hundreds of data elements for each person involved in a maltreatment allegation. This data, stored in the DHS Data Warehouse, is rapidly processed and visualized, resulting in a synthesized ‘Family Screening Score.’ This score predicts the long-term likelihood of future involvement in child welfare.

By combining insights from the Family Screening Score with traditional information, the AI tool enhances the ability to predict the long-term likelihood of a child needing to be removed from their home. The algorithm indicates that a higher score correlates with a greater chance of future out-of-home placement. When the score reaches the highest levels, triggering a ‘mandatory screen in’ threshold, the allegations must be investigated. In other cases, the score complements clinical judgment, providing additional information to aid in the call screening decision-making process. The Family Screening Score is strictly for call screening purposes and is not used for making investigative or other child welfare decisions, nor is it shared beyond the call screening process.

References: 

1. The Allegheny County Family Screening Tool A case study on the use of AI in government. https://www.centreforpublicimpact.org/assets/documents/ai-case-study-child-protection.pdf
2. Child welfare algorithm faces Justice Department scrutiny. https://apnews.com/article/justice-scrutinizes-pittsburgh-child-welfare-ai-tool-4f61f45bfc3245fd2556e886c2da988b
3.  Allegheny Family Screening Tool. https://www.alleghenycounty.us/Services/Human-Services-DHS/DHS-News-and-Events/Accomplishments-and-Innovations/Allegheny-Family-Screening-Tool
4. A case study of algorithm-assisted decision making in child maltreatment hotline screening decisions. https://proceedings.mlr.press/v81/chouldechova18a/chouldechova18a.pdf

Industry: Public Services

Vendor:  A consortium of researchers from Auckland University of Technology (New Zealand) the University of Southern California, the University of California at Berkeley, and the University of Auckland (New Zealand).

Client/Clients: Allegheny County

Publication Date: 2016