AI system could predict safety problems in social housing before they happen

The project, called PRISM (Predictive Risk Intelligence for Social housing Maintenance), is a collaboration between the University of Cambridge and two local councils: Cambridge City Council and South Cambridgeshire District Council. It is supported by the Local Government AI Accelerator, a new initiative from ai@cam, the University’s flagship mission on artificial intelligence.

Instead of waiting for a tenant to report a leaking roof or a damp bedroom wall, a computer model would scan data from thousands of properties and flag the ones most likely to deteriorate, and the residents most likely to be harmed if they do.

“At the moment we’re very much waiting for things to break before we act,” said Peter Campbell, Head of Housing at South Cambridgeshire District Council, which manages around 5,500 social housing properties. “Quite often when things break, it’s not only the item itself that gets damaged, but also the damage caused by the break. For example, it's not just the roof that needs replacing; it's where the water has gotten in and damaged the rest of the property.”

The two councils together manage thousands of tenancies across an unusually wide geography – the urban density of Cambridge city, and the more suburban and rural sprawl of South Cambridgeshire, where it can take over an hour to travel between two addresses.

Campbell says that better data could make his teams far more efficient. In a previous role, he introduced route-planning software for repair staff and watched their daily visit count jump from six to eight. 

Combining satellites with repair logs

The system being developed by Professor Ronita Bardhan and Dr Ramit Debnath from Cambridge’s Department of Architecture and the Centre for Human-Inspired AI (CHIA) combines three sources of data into a single risk score for each property.

The first source is satellite data. Bardhan's team has spent years developing AI algorithms to detect heat loss from buildings using thermal imagery captured by low Earth orbit satellites. That research produced a building-level dataset covering all of England and Wales, mapping energy efficiency property by property.

The second source is conventional housing data: construction type, Energy Performance Certificate ratings, records of damp and mould and repair histories.

The third source is what the researchers call ‘soft’ data: fuel poverty indicators, rent arrears, and accumulated logs of tenant contacts that councils already hold but rarely use at scale.

“The housing officers have a much more grounded idea of how they see vulnerability,” said Debnath, who is also Executive Director of Cambridge’s Centre for Human-Inspired AI (CHIA). “They have information about things like fuel poverty, repair logs, tenancy history and health calls. The interesting bit, which is unique to this project, is that we’re predicting not just on observation data, but also on data from lived experience.”

The result, the researchers say, will be a dashboard displaying a map of risk hotspots – not just flagging buildings in poor condition, but highlighting where a vulnerable person lives in one.

Campbell gives a concrete example of what the dashboard might do. Imagine two identical properties, both with a crack in an outside wall. “One is occupied by a family who are out at work all day, so the heat loss caused by the crack has a minimal effect on them,” he said. “The identical property next door is occupied by a single person who is housebound with disabilities, and the heat loss could have a much bigger impact. The tool would allow us to target the person most in need – it’s not just about the properties, it’s about the people who live in them.”

A shifting approach to social housing

The project reflects a broader change in how social housing is regulated in England. Following a series of high-profile cases involving damp, mould and disrepair – including the death of two-year-old Awaab Ishak in 2020 – the government has tightened expectations on councils and housing associations to use data more proactively in managing their social housing stock. 

“There’s been a changing approach to the way social housing is managed through the housing regulator,” said Campbell. “There's an expectation from government to make better use of data in order to plan our services.”

One area of particular concern is reaching tenants who, for whatever reason, have little contact with their council — people with mental health problems, elderly residents who rarely seek help, or those who mask problems rather than report them.

“What we’re doing now is identifying people with whom we’ve had absolutely no contact and prioritising them for a home visit,” said Campbell. “But we don’t have the resources to do that for everybody, all the time.”

Keeping humans in the loop

The researchers emphasise that PRISM is not designed to make automated decisions about people’s homes or welfare. All alerts generated by the model would be reviewed by housing officers, not acted on directly by machines. 

“AI helps, but welfare decisions stay with trained officers,” said Bardhan. The team has also been careful about privacy: the model is built to work with anonymised data and is designed so that nothing in the outputs can be traced back to a named individual.

“Modelling a building with a machine learning model is relatively straightforward,” said Debnath. “Removing all the risks around personal data and making it tight within the context of its ethics: that’s the complex part.”

The project is designed as a proof of concept over 12 months. If it works, both councils say they hope it could serve as a template for social housing authorities elsewhere in the UK. The researchers have already built a roadmap to help other councils replicate it.

“This is just a starting point,” said Bardhan. “But we hope it can be replicated across different councils across the country.”

Ronia Bardhan is a Fellow of Selwyn College, Cambridge. Ramit Debnath is a Fellow of Churchill College, Cambridge. 

Sarah Collins

Stats

Post a comment

You need to be signed in to post comments. You can sign in here.

Comments

There are no comments yet.