An algorithm to detect violence against children

The starting hypothesis of the research team from University Hospital Who is Dijon? Although it is an important public health problem, epidemiological documentation is still inaccurate with regard to child maltreatment, and the few existing data suggest an underestimation of this phenomenon. The National Child Protection Observatory estimates that in 2019, at least 94 minors died within the family due to violence and/or murder.

However, digitization has made available a great deal of epidemiological data, which has been used in scientific research for more than 20 years. For example, in France, the Medicines Information Systems (PMSI) program collects data on all public and private hospitals. However, the CHU team wanted to create a tool to overcome the problem of under-reporting of abuse cases.

Create an algorithm

As described in the weekly Epidemiology (BEH) of Public Health in France, which published the study on May 17, the team worked to create an algorithm to identify children aged 0 to 5 years, who had benefited from hospitalization for infections that may have be caused by it. of physical abuse. The objective was through the study to validate the case selection algorithm.

The tool classifies children into two groups. In group 1 ‘highly potential abuse’ the child, at the time of hospital admission, displays traumatic lesions described as intentional. In the second group “suspected abuse”, he shows traumatic injuries that appear to be suspicious of physical abuse, by their characteristics or number, without sufficiently suspicious to be included in the first group.

corrected at 80%

Then CHU conducted a study to validate the utility of the tool. Among hospitalized children aged 0–5 years, 170 were identified by the algorithm as having presented hospital admission in relation to potential or suspected maltreatment: 54 for group 1 (high probability of maltreatment); 102 for group 2 (suspected abuse). An additional 14 children were identified in the two groups, and placed in group 1, after a group opinion, because of the painful lesions that appeared.

According to the team, the algorithm’s positive predictive value for identifying highly probable cases of child maltreatment appears to be over 80%, regardless of the age studied.

The study of medical files by a group of forensic physicians who benefited from specific knowledge in the field of child abuse, enabled at the same time to match the pests with the facts reported by those around and thus identify the cases of abuse and the homogeneity of the case study.

restricted tool

However, some limitations were noted, starting with the use of PMSI, which limits the study to cases of children who are hospitalized for physical abuse. Therefore, the tool does not take into account cases of child victims of physical abuse who did not receive hospital treatment.

It also does not allow identifying other types of abuse (psychological, sexual, etc.). The team therefore encourages extending this validation study to other centres, in order to allow for broader validation of the algorithm.

research method

All files of children aged 0-5 who were identified in the local PMSI database by the algorithm were included as ‘highly probable maltreatment’ (group 1) between 2008 and 2019 and as ‘suspected maltreatment’ (group 2). Between 2013 and 2019. A study of children’s medical records was then conducted by a medical examiner between November 2020 and April 2021. The medical file should be made up of electronic or paper documents that re-trace the hospital path.

In a second step, a group of three forensic pathologists studied the child’s injuries and the mechanism of injury reported by the attendants as the source of the injuries. In each case it was determined whether the alleged mechanism in the medical file was compatible with the lesions presented by the child. This approach allowed the team to determine whether traumatic injuries from abuse were confirmed, highly suspicious, suspicious, or excluded from the medical file as a traumatic aetiology.

Thus a comparison with the results of the algorithm was possible.

Leave a Comment