A frailty scale built for people experiencing homelessness: useful AI measures those the models don't see

By Western News (Western University) · October 30, 2026.

A team from the London Health Sciences Centre Research Institute (LHSCRI) and ICES Western, in Canada, will develop an AI-powered frailty scale designed for people who have experienced homelessness. The project is led by Stephanie Dixon, professor of medicine and of epidemiology and biostatistics at the Schulich School of Medicine & Dentistry. According to the university, it will use anonymized data from more than 35,000 Ontario residents with a history of homelessness.

A frailty scale is a tool that estimates a person's overall health and their risk of complications, hospitalization or serious illness. The problem, according to the researchers, is that existing scales were designed mainly for older adults with stable housing. They give heavy weight to age and little to trauma, chronic stress, substance use, mental health problems or barriers to accessing care.

The project builds on a finding the university itself cites: research suggests that people experiencing homelessness may suffer the effects of frailty years earlier than the general population. If so, an age-calibrated scale leaves out precisely those who need it most.

The long-term goal, according to Western, is for physicians and shelter health staff to identify earlier the people at greatest risk of poor outcomes and to act proactively. The work will also document gaps in health data collection, that is, where these populations are less visible to the system.

The design is "co-designed": community partners, healthcare professionals and people with lived experience of homelessness are taking part. Dixon, who has spent nearly 15 years at ICES Western, points to factors that are often overlooked, such as the cumulative effects of adverse childhood experiences, substance use, history of infectious disease and mental health. The study is supported by the AMS Healthcare Fellowship in Compassion and Artificial Intelligence. The release provides no model performance metrics, since the project is at an early stage.

**Our take.** The thesis: the most common bias in healthcare AI is not a flaw in the algorithm but in who is included in the training data. A model well calibrated for a stable, aging population can be systematically blind to another. Here, AI is not meant to replace the clinician but to correct a blind spot that already exists in traditional tools.

There are reasons for caution, and they are worth stating. Today there are no results to evaluate, only a plan. A risk scale is only worthwhile if it is validated with external data and if it changes real decisions. If there are no beds, primary care or support behind the alert, earlier detection improves nothing. Moreover, administrative data only capture those who have already been in contact with the system, and the release itself acknowledges those gaps. It is to the project's credit that it makes them an object of study.

More broadly, the medicine of the future we promise at Zendoric, with eradicated diseases and extended health, will only be abundance if it also reaches those who are on the margins today. That horizon is not reached through more powerful models alone. It requires more representative data and design with the people affected. This project, modest and at an early stage, points in that direction: before distributing the fruits of medical AI, we must make sure the system knows these people exist.