Child abuse: an algorithm developed to detect battered children at Dijon University Hospital

Child abuse is partly under the radar of hospitals. In 2019, the University Hospital of Dijon (Burgundy, France) developed an algorithm to identify these physically abused children, aged 0 to 5 years.

It was still necessary to validate it. Now completed: CHU conducted a study, published in the Epidemiological Weekly Bulletin of Public Health in France, to test the suitability of this algorithm. Results: The positive predictive value of the algorithm is high; The younger the children, the better. For infants less than one year old, the algorithm can achieve an efficiency of 94.4%.


Let’s first see how the hospital developed this algorithm. The team used French hospital codes (what in France is called PMSI data, meaning the codes under which hospitals record diagnoses of hospitalizations), for children under the age of five who are hospitalized for physical abuse.

In the case of assault, the hospital stay can be coded either specifically (the code describes injuries and the fact that it is a voluntary physical attack), or inaccurately, without stating the source of the injury. This may understate the number of abuse cases, because being in either category does not change the hospital’s budget. Hence the importance of the algorithm, in order to more realistically identify children between the ages of 0 and 5 years who could be hospitalized due to injuries resulting from physical abuse.

We relied on two. Catherine Quantin, Principal, explains Biostatistics and Medical Informatics Unit at Dijon University Hospital A public health professor is behind the project. ” The first is to identify the icons that define clearly In the wording that the child was the victim of voluntary physical abuse – that’s what we call our group 1 – and the second is that we did a lot of work in the bibliography, searching the literature, looking for pests suspecious of abuse. This means, for example, that a child under a year old will have a serious fracture, this child does not walk, we think that he will not be able to fall on his own … A child will have a recurrence over time, there again, we still wonder why it recurs these fractures. There has been a lot of literature on this topic. Doctors reported suspicious cases of abuse. This type of potential situation is identified in Cluster 2: Suspected Abuse.

Fit, this algorithm?

Let us now get to the method of validating this algorithm, which is the topic of the current post.

The medical examiner reopened the medical records of the children in both groups several years ago. The pathologist completed a standardized network for the study, based on this analysis of medical records. In a second step, a group of three forensic pathologists studied the children’s injuries and the original description that brought them back around them. In each case, the trio determined whether what was in the medical file was consistent with the lesions presented by the child. What marks potentially abuse-related lesions as “confirmed, highly suspicious, or excluded”.

All that remains is to compare the data:

– Cases of medical records and analysis of forensic scientists.

– Taken from encoded PMSI data when the child leaves for hospital treatment.

Of the children admitted to hospital at the University Hospital of Dijon during the period in question, 170 were identified by the algorithm as probable or suspected cases of abuse.

Comparing the data showed that the younger the children (between one month and one year), the more reliable the tool. When dealing with infants (less than a year old) the algorithm showed an efficiency of more than 90%. “We have had very favorable results for babies from 1 month to 1 year old.‘, explains Catherine Kwanten.For example, when codes of abuse are entered into the hospital summary, for children between the ages of one month and one year, that abuse is on file in approximately 95% of cases.. Among children aged 0-5 years, the predictive value is about 80%, which remains sufficiently reliable from an epidemiological point of view.


How is it useful? “Already, to make a measure of this bad treatment‘, identifies Catherine Kwanten.It is impossible to list all cases of abuse. In the hospital, physical abuse is easy to identify. But international studies have shown that it is not possible to list all these abuses. So it’s interesting to have a barometer. The goal of this barometer is not to obtain completeness of states, but to be able to follow the evolution of this indicator in time and space.. “

Thus, after the algorithm was published in 2019, a first study showed that, after a period of relative stability of the index, there was a 50% increase in the frequency of hospitalizations due to physical abuse in France. This made it possible to alert public authorities and support the reopening of schools and nurseries with the French government.

Towards a machine learning?

The Dijon University Hospital team aspires to go further: extend this validation study to several French hospitals and consider an artificial intelligence tool that, when getting results, can improve the algorithm and index, and perhaps one day also share it with care services that can report to social services. At the end of the day, medical professionals will always be the ones following up on the reports. The algorithm can be used as a tool, as an aid for diagnosis and reporting.

Leave a Comment