The Development and Validation of Models of Risk for Behaviours That Challenge in Children With Developmental Disabilities: A Novel Machine Learning Approach.
Laura Groves, Guy Davies, Chris Oliver and 15 others
PMID 42011951WHAT IT FOUND
Models built from a short questionnaire identified which children with developmental disabilities would show challenging behaviour a year later, including most of those with the most severe behaviour.
Telling low severity from high was unreliable, and the authors say clinical use needs more testing.
Key findings
01In the internal data the best performing model, a random forest for any challenging behaviour, correctly classified 74.49% of children with no behaviour, 54.27% of those with low severity behaviour and 79.49% of those with high severity behaviour.
02On the separate group of children followed up, one model would flag 83.50% of those needing an intervention, including 95.24% of those at highest severity, and miss 16.50%, most of them low severity.
03The models identified children who showed some level of behaviour (recall 71.96% to 81.08%) far better than they separated low from high severity (recall 19.35% to 69.28%), and misclassifying the low severity group was the most common error across models.
STILL TO COME
How it was doneWhat they found
Read the rest of this summary
You get three full summaries a month, free, and we do not ask for a card. Search, the TL;DRs and your library stay unlimited either way.
What it does not show
The two datasets differed significantly on nearly every variable measured, including age, health problems, repetitive behaviour, obsessive behaviour, impulsivity and overactivity, so the external test was a hard one. The external group was small (121 children) and only 45.5% of those who completed the first questionnaire were followed up a year later. Few children changed behaviour over the year, which the authors say limits what can be said about predicting new cases or recovery. The models were least accurate for the low severity group, the group most likely to be seen before behaviour becomes severe. Predictors were mostly characteristics of the child; the authors note that caregiver and environmental factors were not included, which may limit accuracy for more specific behaviours. The models have not been tested as part of clinical care, and the authors say future studies should evaluate them further and that any implementation should be cautious and combined with clinical judgement.
Declared interests
The work was funded by the British Academy of Childhood Disability-Castang Foundation, by the European Research Council under the Horizon 2020 Framework Programme (CartographY GA. 804752) and by Cerebra. The authors declare no conflicts of interest.
The easy way to misread this
Do not read this as a screening tool you can use tomorrow. The models were tested on data that had already been collected rather than in a clinic, they were poor at separating low from high severity behaviour, and children who showed no behaviour were often flagged as being at risk (25.68% in the internal check and 55.56% in the external one). The authors themselves say any use in practice should be cautious and should sit alongside clinical judgement.
Summarised by AI from the full paper, without a clinician reviewing it. Check it against the source before it changes what you do. Read it on PubMed →