Leveraging Artificial Intelligence to Prevent Homelessness in Los Angeles County

Homelessness and health are deeply intertwined, each shaping and reinforcing the other. Poor health can push individuals into housing instability, while homelessness introduces new challenges and worsens existing conditions. The result is a cycle that not only affects individuals' well-being but also places growing strain on the health care system. For hospitals, this cycle translates into higher utilization, more complex care needs, and an urgent need to prevent homelessness before it occurs.

The Los Angeles County Department of Health Services, an association member, leverages artificial intelligence to respond to that need through the Los Angeles County Homeless Prevention Unit (HPU). In 2017, the University of California, Los Angeles' California Policy Lab (UCLA CPL) studied whether data on an individual's interactions with county institutions and programs could predict their risk of becoming unhoused. In partnership with county leadership, CPL leveraged millions of records linked across nine county administrative data sources. The lab determined that data could predict a person's risk of homelessness as far as 12 months in advance, providing a significant opportunity for the county to provide stabilizing interventions.

Inspired by the UCLA CPL's research, the LA County Department of Health Services established and staffed the HPU with funding from the American Rescue Plan Act. The HPU uses data from the county's Health Services, Mental Health, Public Social Services, and Child and Family Services departments, as well as the sheriff's office. The County Chief Information Office (CIO) manages the "InfoHub" database, which enables real-time, cross-departmental data sharing to support predictive modeling.

The CPL and CIO jointly monitor the model for accuracy and false negative rates across demographic groups. They adjust the model to mitigate bias, even when doing so reduces predictive performance. This reflects a necessary trade-off, as reducing bias often requires limiting the model's reliance on historically skewed patterns that may boost overall accuracy but disadvantage certain groups. For example, if a model learns from historical data that certain groups were less likely to receive housing assistance, it might underestimate their risk of homelessness. Correcting that bias requires adjusting the model in ways that can reduce its overall accuracy. Program demographics show high representation of clients of color and an over-enrollment of women, while an ongoing concern is under-enrollment of Latino men.

How the HPU Operates

The InfoHub links data from 11 county agencies-including health, mental health, social service benefits, arrests, probation, and homelessness service records-for millions of individuals into a secure and regularly updated data pool. The HPU begins by reviewing a risk list that the CPL and CIO generate quarterly using InfoHub data. The risk list is compiled by an algorithm, which analyzes de-identified data and assigns each individual a point-in-time probability score, estimating their likelihood of entering homelessness within the next 12 months. The probability score is determined by individuals' interactions with county institutions, and it captures events such as previous housing instability, recent hospitalizations, mental health crisis holds, and criminal legal involvement. The program targets individuals with risk scores between approximately 30-33%, who have a high risk of homelessness and are still eligible for prevention services. About 25% of identified individuals are removed because they are currently homeless, incarcerated, or in conservatorship. Of those who remain eligible, roughly half are successfully contacted through staff outreach, primarily via cold calling.

Once enrolled in the program, clients receive a two-pronged intervention focused on stabilization. The first component is financial assistance, typically ranging from $4,000 to $6,000 for single adults. The second component includes referrals to external resources such as food banks, legal services, and other community-based support. Each client is assigned to a case manager who works on budgeting, stabilization, and connection to services.

Results

The program assists approximately 700 clients annually, and 89% exit the program with permanent housing. Given the program's early success, a randomized controlled trial is underway to evaluate the program's effectiveness in preventing homelessness. In the study, only the treatment group is contacted and offered services, while the control group does not receive outreach. Results are expected in early 2027.

While the HPU was originally housed within the LA County Department of Public Health, it now operates under the newly established Department of Homeless Services and Housing. With this shift in governance, HPU leadership hopes to further scale and embed proactive homelessness prevention across county systems.

The HPU offers a compelling example of how responsible use of artificial intelligence (AI) and strong governance can be used to address upstream drivers of health and social need. By pairing predictive analytics with robust data governance, cross-sector collaboration, and human-centered service delivery, the program demonstrates how hospitals can deploy AI to improve outcomes for high-risk populations. This work closely aligns with Essential Hospitals Institute's recent brief on how essential hospitals are mitigating housing instability through partnerships, data sharing, and innovative care models. For essential hospitals, the HPU underscores the importance of aligning AI use with mission-driven goals, investing in governance and data infrastructure, and ensuring that technology enhances the effort and builds trusted relationships with patients and communities.

Read more about the HPU in the California Policy Lab's brief, The Homelessness Prevention Unit: A Proactive Approach to Preventing Homelessness in Los Angeles County. To learn more about how essential hospitals are adopting AI, join the association's AI interest group.