How AI can detect patterns of children at risk


Strängnäs Municipality

AI method

Text classification


Public sector
- Social services

The challenge

One of the tasks social workers struggle with is to pick up on signs of domestic violence or abuse in a home where children are present. Detecting these signs and patterns involves interviews and going through different reports, and sometimes time is simply not enough. Also the time spent on administrative work is necessary for actually meeting children, or executing the necessary measures required in time.

The opportunity

By applying a classification language model that could handle concern reports the municipality could speed up the process of going through a lot of data and mapping it into different categories that could help the social worker see patterns. Saving more time for actually talking to the children, for instance, chances increase to contribute to a safer environment.

02/ Background

There is always room for improving processes, however, with rising average age of the population in the municipality causing lower tax income as they retire, Strängnäs needed to look for more cost-efficient ways of improvement. When Strängnäs Municipality was reviewing the process of how they handle concern reports for children at risk, they realized that there was room for improvement that previous technologies couldn’t fulfill. 

Strängnäs Municipality was at the same time looking into implementing AI as a pilot since they saw great potential with the technology in terms of making processes more efficient, and the need to gain experience in this new field. For the case of concern reports the purpose, in addition to efficiency, was to detect possible signs of violence or conflict at an early stage. The outcome could then be a quicker intervention and more precise course of action from the social workers with the help of an AI model.

Contributing to system effects at a national level will help Strängnäs municipality too - we are gaining maturity on our way to using data as a strategic resource.

Frédéric Rambaud
Head of digital strategy, Strängnäs Municipality

03/ How the AI model works

The model is built on text classification which essentially makes it possible to categorize the text content in the concern reports in some categories. The pilot project focused on the following groups: "Trauma", "Behavior", "Separation conflict”, "Upbringing", "Injuries to the body", "Social network" and "Sexual abuse". The categories were selected by Strängnäs themselves and reflect some of the areas to be looked upon in the process. Being able to identify those deficiencies in children’s care is part of a wider national way of working. 

The model goes through the text from a journal by sections and gives each section a value indicating if one or several of the categories are met. In the future, this could help a social worker determine how stressed the matter might be and what measures it will require. In extension, a language model could be trained in a similar way to identify signs of other categories important to the child's well being. Eventually, this could lead to analyzing texts with a more holistic perspective: is the child given good care. 

Even at a trial stage, the model shows clear potential of how it could provide operational support to social workers. Social workers could receive concerning flags more efficiently and could therefore intervene quicker, positively affecting the case outcomes for the involved children and youth.

The image shows the kind of results that the model could provide when a text from a journal has been entered for classification.

04/ Data requirements

Good quality data is often the most important component of a successful AI project. For this model to work properly, Strängnäs needed to label at least 300 examples that reflect each category. 

When the first round of annotation is completed, you should assess how the model works and what categories seem to perform better or worse to steer the work and make sure that the data is in balance to avoid any possible bias.

Since the data needs to be anonymized, make sure that you’re consistent in the way different attributes are replaced, such as names, locations, and so on. However, it’s important to still distinguish parents as ‘mom’ and ‘dad’ instead of just calling both ‘parent’.

05/ Results and next steps

There are still some challenges with data anonymization which is particularly important in the public sector due to the personal nature of the information that is being processed. A next step could therefore be to increase data protection, or improve anonymization and encryption processes. In addition, more data needs to be labeled to further improve the performance of the model. Once done the solution is scalable and could be rolled out in municipalities all over Sweden, since 95% of them are using the same way of working. The benefits of this would be more time to focus on taking action instead of admin work for the social workers across the country, as well as patterns being detected earlier and essentially preventing more children and youths from being harmed in any way.

Strängnäs Municipality is now also taking steps at national level, participating in the newly started project “Data readiness lab”, contributing to developing tools and guidelines for data readiness, annotation, anonymization, and evaluation. Contributing to system effects will help Strängnäs municipality too, gaining maturity on its way to using data as a strategic resource.