How Artificial Intelligence Detects Child Abuse (And Why It’s Hard To)

Author(s): Eera Bhatt

Originally published on Towards AI.

According to the CDC, at least one in every seven children was abused or neglected over a year, so this is a serious problem that needs attention.

With artificial intelligence, models can be developed to predict which children are at a risk of — or are already — being abused. But it’s hard to access data about child abuse, so not many studies have been done to resolve this yet. In a recent study, the authors review papers on PubMed to check on this area.

From the authors’ findings, the input data in most of these studies is pretty diverse. These studies mention convolutional and artificial neural networks along with natural language processing, which means that the input data includes images and text.

But the biggest problem here is that all studies reviewed have a high risk of bias, because of factors like sample size, validation, overfitting, and missing data.

Eventually, the authors narrow down their review to seven papers. They use a tool called PROBAST which associates all of the studies with a high risk of bias.

One of the biggest reasons for this is that the researchers simply do not use enough data.

In two of the studies, the authors use fewer than 200 cases in their datasets, which makes them pretty small. For instance, other medical studies — like ones to diagnose skin cancer — use datasets with over 120,000 cases, which is much more ideal than what these authors used. But since we have just a few hundred cases to analyze here, overfitting can happen. In other words, the algorithm created fits very well with its training data, but it doesn’t make accurate predictions.

To help our models predict accurately, a wide variety of data can also help us out. This is especially important in the area of child welfare because some types of data — like images — are limited or can’t be accessed due to privacy. Throughout these seven studies, the researchers use radiologic imaging, clinical text of medical records, demographic and clinical information, and self-figure drawings as input data.

Let’s dive into the methods actually used.

Convolutional neural networks. This method helps physicians analyze medical images of children to figure out whether they are victims of child abuse. For instance, CNNs help a radiologist determine whether certain bone fractures are accidental or caused by abuse.

Aside from pictures of injuries, Kissos et al. uses CNNs to distinguish between abused and non-abused kids based on how they draw themselves. A kid’s self-drawing is normally influenced by their own thoughts and emotions. So instead of interrogating a child to figure out whether they were abused — which feeds into their trauma — CNNs help us see the child’s emotional development in a way that’s healthy for them.

Natural language processing. Meanwhile, NLP draws meaning from the text that is displayed in medical records.

Let’s also consider why all seven of the studies discussed are at a high risk of bias according to PROBAST.

Validation. In machine learning, it’s important to test and validate a model on new data to make sure that it can really make accurate predictions. When a machine learning model starts memorizing its training data and performing incorrectly on new data, we call that overfitting.

Think about studying for a test. If you memorize the homework solutions, you won’t be able to solve new problems during the test. Especially in the medical field, and with an issue as serious as child abuse, we can’t send machine learning models out there that haven’t been thoroughly validated first.

In some of the studies from this paper, the authors use temporal validation which is generally regarded as a middle ground in between internal and external validation. They test their models on images and text that were sampled at a different time than their other data.

Missing data. In four of the seven studies discussed, there is no explanation for the data that is missing. The problem here is that missing data adds bias to the model and reduces its efficiency in general. (Remember overfitting?)

At the same time, though, missing data is so common. Especially with a topic as personal as child abuse, it’s rare to have equal image data available for every scenario. This issue is normally handled with multiple imputations, which means that we replace a missing data point with a value.

Future of artificial intelligence for child welfare. It’s pretty hard for a human to identify the line between accidents and child abuse, let alone a machine. There is no question about it — this is difficult work that these scientists work on.

As this work continues to be developed and the researchers refine their models, they often have to prioritize parts of the prediction that they want to be the most accurate. But when it comes to detecting child abuse, healthcare workers have to face similar pressures from this issue. If a false negative happens, injuries caused by abuse are classified as being normal.

This means that the child is returned to their abusive environment which puts their life at risk. But if a false positive happens, marks of injury on a child might be classified as abusive even though they are accidental.

Far beyond the machine learning model, this causes developmental and legal problems in the long run. If an allegedly abused child is separated from their innocent guardian, they don’t receive the care that they need to grow up. Not to mention that the child’s guardian is falsely convicted for acts of abuse that they never even committed.

Conclusion. Ultimately, it is completely up to the healthcare professionals, legal system, and protective services to decide what to do with the model’s results, depending on the child’s specific case.

Meanwhile, the authors of this paper offer great advice to any future researchers who want to build on these models. And with a subject as obscure as child abuse, it is essential to work with these guidelines in mind.

Further Reading:

[1] Hughes, K. (2023) ‘Research Summary: Use of Artificial Intelligence Tools in Social Work and Child Welfare Services’. San Diego: Child Welfare Development Services.

[2] Kawakami, A. et al. (2022) ‘Improving human-ai partnerships in child welfare: Understanding worker practices, challenges, and desires for algorithmic decision support’, CHI Conference on Human Factors in Computing Systems [Preprint]. doi:10.1145/3491102.3517439.

[3] Kissos, L. et al. (2020) ‘Can artificial intelligence achieve human-level performance? A pilot study of childhood sexual abuse detection in self-figure drawings’, Child Abuse & Neglect, 109, p. 104755. doi:10.1016/j.chiabu.2020.104755.

[4] Lupariello, F. et al. (2023) ‘Artificial Intelligence and child abuse and neglect: A systematic review’, Children, 10(10), p. 1659. doi:10.3390/children10101659.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI