Peer-reviewed article
The Development and Validation of Models of Risk for Behaviours That Challenge in Children With Developmental Disabilities: A Novel Machine Learning Approach.
Groves L, Davies G, Oliver C, Allen D, Bamford C et al. ·
Summary
Written by AI, not yet AICS's view · claude-haiku-4-5-20251001This study developed machine learning models to predict which children with developmental disabilities are at risk of behaviours that challenge, such as self-injury, aggression, and property destruction. The researchers trained algorithms on data from 778 caregivers and tested them on an independent group of 121 caregivers followed over 12 months. The best-performing models correctly identified 83.5% of children at risk and could predict whether challenging behaviours would persist or emerge over a year. The authors suggest these prediction tools could help clinicians target early interventions to children most in need.
Key points
- Random forest and Gaussian mixture model algorithms achieved 83.5% accuracy in identifying children at risk of behaviours that challenge
- Models predicted both persistence and incidence of behaviours over 12 months with 83.3-83.5% accuracy
- The study used two independent datasets: one for training (n=778) and one for external validation (n=121)
- Behaviours assessed included self-injurious behaviour, aggression, property destruction, and any combination of challenging behaviours
- Prediction models have potential applicability to clinical practice for targeting early preventative interventions
Limitations
The document does not specify the countries where the study was conducted, limiting understanding of generalisability to different health systems and populations. The study describes model development and validation but does not evaluate whether using these predictions to guide interventions actually improves child outcomes.
Abstract
Background Children with developmental disabilities show a high prevalence of behaviours that challenge (BtC). Thus, harnessing known risk markers to target early intervention to children at the greatest risk of BtC is essential. In this study, machine learning techniques were used to develop prediction models of risk (no, low and high severity behaviour) for different BtC (self-injurious behaviour, aggression, property destruction, 'any BtC'). A secondary aim was to assess the external validation of these models to predict future behaviour. Method Caregivers of individuals with developmental disabilities completed the Self-injury, Aggression and Destruction Screening Questionnaire. One dataset (n = 778) was used to train and test models to establish internal validation. Algorithms were created using random forest classifiers, K-nearest neighbours, multiple logistic regressions and Gaussian mixture models (GMM) for each type of behaviour. External validation utilising a second dataset of caregivers (n = 121) completing the SAD-SQ at baseline and 12 months later was then conducted. Outcomes Across internal and external validation, the random forest classifiers and GMM algorithms for any BtC showed the highest number of correct classifications with fair to good recall and precision, with 83.5% of people at risk of BtC correctly predicted. Predictions of persistence and incidence of behaviour over 12 months was also good (83.5% and 83.3%, respectively). Interpretation The novel prediction models showed the ability to predict BtC for children with developmental disabilities. Such models have applicability to clinical practice to inform provision of early preventative interventions for BtC.