Machine learning in children’s services: does it work?

• Vicky Clayton, Michael Sanders, Eva Schoenwald, Lee Surkis, Daniel Gibbons, September 2020

Natural language processing techniques were used to turn reports and assessments into information that can be used as input to a model. Machine learning techniques were then used to learn patterns in historical data associated with risks and protective factors, and examine whether those factors were present in unseen cases. The study aimed to understand whether machine learning models, applied in this way, correctly identify the cases at risk of the outcome and those that are not. Four different ways of designing the models were assessed.

About the project:

Practitioners in the four local authorities identified two outcomes to predict for their own authority and in total we predicted eight outcomes:

• PREDICTION 1: Does a child/young person’s case come in as a “re-contact” within 12 months of their case being NFA-ed (“no further action”-ed), and does the case then escalate to the child being on a Child Protection Plan (CPP) or being Child Looked After (CLA)?
• PREDICTION 2: Does the child/young person’s case progress to the child being subject to a Child Protection Plan (CPP) or being a Child Looked After (CLA) within 6-12 months of a contact?
• PREDICTION 3: Is the child/young person’s case open to children’s social care – but the child/young person not subject to a Child Protection Plan (CPP) or being Child Looked After (CLA) – within 12 months of their case being designated “No Further Action”?
• PREDICTION 4: Is the child or young person’s case which is already open to children’s social care being escalated (to the child being subject to a Child Protection Plan, being Looked After, being adopted, being subject to a Residence Order or being subject to a Special Guardianship Order) between three months and two years of the referral start date?
• PREDICTION 5: Does the child/young person’s case progress to the child being subject to a Child Protection Plan (CPP) or the child being Child Looked After (CLA) within 6-12 months of a contact?
• PREDICTION 6: After successfully finishing early help, is the child/young person referred to statutory children’s services within 12 months?
• PREDICTION 7: Does the child/young person’s case progress to the child being subject to a Child Protection Plan (CPP) within 1-12 months of the assessment authorisation date?
• PREDICTION 8: Does the child/young person progress to the child being a Child Looked After (CLA) within 1-12 months of the assessment authorisation date?

Findings

The study tested whether the patterns learned enabled the models to predict well on cases the model hadn’t yet “seen” but whose outcome was already known. The number of misclassifcations the model makes of cases as “at risk” or “not at risk” of the outcome give an indication of the numbers and types of misclassifcations the models would make if they were used to assist social workers in practice.

Metrics which summarise the overall model performance count the correct and incorrect classifications of the model. Two different ways of summarising the performance of the model were devised: average precision and “area under the curve” (AUC). Both metrics are measured on a scale of 0 to 1 with 0 being the worst possible model and 1 being the best possible model.

WWCSC deemed the average precision metric is more appropriate for use because the proportion of cases at risk of the outcome is quite small, ranging from 2-17 per cent with seven of the outcomes being 2-7 per cent.

Before beginning the analysis, researchers deemed the model would be a “success” if it scored above 0.65 average precision, although this is lower than the threshold WWCSC would recommend for putting a model into practice but provides a useful low benchmark.

Overall, none of the 24 models had average precision scores which exceeded the threshold for success. Ten of the 24 models have AUCs greater than 0.65 but this “success” reflects that the model correctly identifies most of the cases not at risk as not at risk, an outcome the authors say is an “easy win”.

Conclusion

The authors did not find evidence that the models created using machine learning techniques work well in children’s social care. In particular, the models missed a large proportion of children at risk which – were the models to be used in practice – could risk discouraging social workers from investigating valid concerns further, potentially putting children and young people at risk.

For just over half of the models, adding more cases may improve the model performance; however, using data further back in time is unlikely to help.

The authors did not seek to definitively answer the question of whether machine learning will ever work in children’s social care but said it illustrated some of the challenges faced when using these approaches in children’s social care. For local authorities already piloting machine learning, WWCSC encouraged them to be transparent about the challenges they experience.

Implications for practice

None of the models that we tested cleared the threshold we established to say that they did an adequate job of predicting outcomes. This means there are good reasons to be skeptical about the value of any predictive analytics tool at the moment. If you’re using one to make decisions, you should consider whether or not the predictions are really any better than guessing before you put your professional judgment to one side.

If you’re in a position to decide whether or not to use a predictive analytics engine in your service or practice, you should ask for a transparent measure of the model’s quality before doing so – this would include false positives, false negatives, and level of bias in the model against particular groups.

• What Works for Children’s Social Care: These studies have been identified by What Works for Children’s Social Care as being influential studies in our understanding of digital interventions and applications in children’s social care services. WWCSC has produced the implications for practice for each study too.

Read more in CYP Now's Digital Innovation in Social Care Special Report