While artificial intelligence (AI) cannot replace the frontline staff working to end homelessness, it can be used to support them. AI and machine learning can be used to assist human decision-makers by providing a personalized picture of a client's risk factors. This can help a case worker in connecting the client with the specific mix of services that they need to successfully transition into a home.
Why Estimating Risk is Useful
A case manager working with a housing or shelter client is very similar to a family doctor working with a patient. Both will use a person’s history and their understanding of the person's current condition to make a careful decision. However, physicians also benefit from having tools to assess a patient's medical risk factors. For example, a blood pressure cuff can help evaluate the risk of cardiovascular disease or a lab test may indicate an increased risk of certain forms of cancer.
A vulnerable person is also at risk of a range of serious events. A case worker could make better, more customized care decisions based on their individual risk factors. Tools like the VI-SPDAT that lump all risk factors into a single score are being questioned and communities are now being encouraged to decide on their own definitions of vulnerability based on a range of factors.
Like a medical test, machine learning is very good at estimating the risk of a specific event. A risk profile can be created by developing several machine learning tests, each customized to predict the risk of a different event. For example, one machine learning test could estimate the risk of drug poisoning, while another could estimate the risk of a violent police interaction.
Practical Risk Estimation for Shelters
In collaboration with The Calgary Drop-In Centre, our team developed a tool to identify people at risk of becoming long-term or chronic shelter users.
Chronic shelter clients are a high-priority group for Housing First programming. Definitions of chronic homelessness exist and can be used to identify chronic shelter clients. The problem is that these definitions typically require a client to spend a long time in a shelter.
Connecting clients to supportive housing soon after entering the shelter is also a challenge. We know that most new shelter clients exit shelters quickly with very little help. Even if offering housing resources to every new person in the shelter was possible, it would be unnecessary.
Machine learning offers a middle ground. Using anonymized historical data records from the Drop-In Centre, we used a technique called rule set searching to develop a test that uses a client's shelter check-in data to identify those at risk of becoming chronic shelter users. In some cases, clients who are at risk can be identified as quickly as 30 days after they first enter the shelter.
However, machine learning needs to be compatible with not-for-profit information technology (IT) systems. Many high-performance machine learning algorithms require very high-end IT infrastructure and expertise. Our rule search approach generates threshold-style tests that can be run on simple databases or spreadsheets. For example, clients that satisfy the rule:
(number of sleeps) greater than 53 AND (counsellor consultations) less than 11
after 60 days in the shelter are at risk of becoming chronic shelter users.
Machine learning also needs to be accessible. Staff making important decisions for vulnerable people will be reluctant to use tools they cannot understand. Many machine learning algorithms are “black box” – even data experts cannot fully explain how they make decisions.
In contrast, our rule search approach draws from the research area of interpretable machine learning. Drop-In Centre staff were able to understand and validate our threshold tests. They pointed out that the 60-day test spots "under the radar" clients with a lot of stays but few counsellor interactions. This resonated with their personal experiences working with chronic clients.
Is Machine Learning Right for Your Organization?
A few questions to help you decide if machine learning is right for your organization:
- Will it tell you anything you don’t already know? Machine learning tests are great for high-volume shelters where staff have large caseloads. They likely won't provide new insight for smaller shelters where staff can spend a lot of time getting to know each client.
- Is your data ready? Rule search tests can work on very simple IT systems, but your data can't be fragmented across many files. Most machine learning tests need your data to be all in one spot.
- Do you have (occasional) access to data experts? The rule search approach we’ve discussed produces very simple tests that non-technical staff can use. However, the algorithm used to find the best threshold tests (the “search” part of rule search) needs to be run on your client data by a machine learning expert. This doesn't have to happen very often and could be done as part of a university collaboration or by using consultants.
- Are your staff prepared? The results of a machine learning test should be used to enhance but not hijack staff decision-making. Machine learning should never be the only factor used to make a front-line care decision.
Dr. Geoffrey Messier has been a professor in Electrical and Software Engineering at the University of Calgary since 2004. Since 2016, Geoffrey has been actively involved with emergency shelter and housing agencies as a volunteer project manager and data scientist. His volunteer work has included shelter client record linkage, the establishment of a facility where clients can access homecare services in the shelter and the coordination of personal protective equipment (PPE) during the COVID-19 pandemic. His research explores how data science and machine learning can be used to support front-line emergency shelter and housing program staff. You can find out more about his work and interests here.
This post is part of our #CAEH22 blog series which highlights research on preventing and ending homelessness that is being presented at the 2022 National Conference on Ending Homelessness, November 2-4 in Toronto, ON. Learn more about Geoffrey’s work through his presentation in the Role of Data and Planning in Homelessness Prevention session Thursday, November 3rd at 3:00 pm.