EF Predict — Predicting the risk of EF development difficulties
A machine learning tool that helps child development centres and kindergartens screen for risks involving EF (executive function) skills from early childhood, so children can be referred for care before developmental difficulties worsen.
Translated from Thai · reviewed by the CF team
Brain skills that shape the future
EF (executive function) refers to brain skills involving planning, focus, working memory and managing several things at once. Research indicates that developing EF from early childhood helps reduce the risk of future behavioural problems, including aggression, substance use and gambling. This system is designed for use only by child development centres and kindergartens. Its results may be inaccurate and must be used alongside ongoing developmental observation and assessment; they are not a standalone diagnosis.
From field data to a predictive model
The team worked with teachers, doctors, public health academics and early childhood specialists to design a data collection set. They collected data from around 2,000 young children at child development centres in Lopburi area using an online form completed by teachers. The data was then analysed to examine the relationship between 27 factors and 5 areas of EF, before selecting the 13 factors most strongly associated with EF to build the model.
Real-world data linked to standard EF scores
Links data on children's environmental factors with actual EF assessment scores (MU.EF-101/102) from the Neuroscience Research Centre, Mahidol University.
A two-layer XGBoost model with joint voting
After finding that basic models (Logistic Regression, Decision Tree, SVM) were only 30-40% accurate, the team switched to XGBoost, with two sub-models (EF101 for development and EF102 for behaviour) voting together.
Screening across 5 areas of EF
Inhibitory control, Shift (cognitive flexibility), emotional regulation, working memory, and planning/organisation, using 13 data points such as birth order, diet, free-play time and screen time.
No user data is stored
The system immediately deletes the information entered after processing and does not store it, to protect the privacy of children and families.
The model's Recall score (the accuracy metric selected by the team because they prioritise not missing at-risk cases) ranges from 49–87%, depending on the EF area and set of labels. The team also openly states that the relationship for each individual factor is still relatively weak (the highest Pearson r is only 0.05–0.14), so no single factor can predict outcomes accurately on its own.
Because the system does not record usage data at all, there are no figures for the number of children actually screened using the tool or the number of centres/schools currently using it. This is a structural limitation of the system itself, not missing data.
Developed by
ChangeFusion
Data collection and testing area
The High-Scope Project at child development centres in Thawung District, Lopburi Province, implemented by Thawung Hospital.
Supported by
Thai Health Promotion Foundation (ThaiHealth)
Academic partner
Neuroscience Research Centre, Institute of Molecular Biosciences, Mahidol University — the source of the EF assessment scores used as the reference.
The team openly states that the model reflects patterns only in the Lopburi area. Collecting data nationally to make the model more widely usable "is not feasible at this time", but the team is willing to advise those interested in starting national-level data collection.
0 entries
Everything the field team has recorded, newest first
No updates for this project yet
How can we
help you?
Choose the question that fits your role. Answers are summarised from our own records of the work, and you can talk to the CF team from any card.
What change does this work create, and how do you know it is effective?
We have developed an EF risk-screening model with Recall ranging from 49–87%, depending on the EF domain and label set, but we do not yet have figures on outcomes for children after screening.
EF Predict is a tool for screening the risk of EF (executive function) skills difficulties in early childhood, designed to help child development centres and nursery schools refer children for care before developmental problems worsen. This Project We used data from around 2,000 young children in Lop Buri to train and validate the model, linking information on environmental factors with standard EF assessment scores. This Project
The model's Recall ranges from 49–87%, depending on the EF domain and label set. We prioritised avoiding missed at-risk cases. This Project However, the relationships with individual factors are relatively weak, with the highest Pearson r values at 0.05–0.14, so no single factor can reliably predict outcomes on its own. This Project
We do not yet have figures for the number of children actually screened or information on how screening affects outcomes for children, because the system does not record usage data. This Project
How was this work carried out, and what lessons can be applied going forward?
We collected data from teachers, linked it to standard EF scores, and used 13 factors to build an XGBoost model with two sub-models that vote together. The lesson is to use screening results alongside ongoing observation, not as a standalone diagnosis.
We began by designing a data-collection set with teachers, doctors, public health academics and early childhood experts. Teachers completed an online form to collect data on around 2,000 young children at child development centres in the Lopburi area. This Project We then analysed 27 factors against 5 EF domains and selected the 13 factors with the strongest relationships to build the model, using standard EF assessment scores as the reference. This Project
Basic models such as Logistic Regression, Decision Tree and SVM achieved an accuracy of only 30-40%, so we switched to two XGBoost sub-models, which vote together on development and behaviour. This Project The screening tool covers Inhibitory control, Shift, emotion regulation, working memory, and planning or organisation. This Project
The key lesson is that model results may be inaccurate, and the relationships between individual factors are still weak. The results should therefore be used alongside ongoing observation and developmental assessment, not as a standalone diagnosis. This Project The system also forgets the data entered as soon as processing is complete and does not store user information, to protect the privacy of children and their families. This Project
What kinds of areas is this approach suitable for, and what is needed to get started in our area?
This model was built using data from Lopburi and designed for use by child development centres and kindergartens, but we do not yet have documented requirements for starting implementation in other areas.
This model was developed using data on young children from child development centres in Lopburi, so it reflects only the pattern in that area. This Project Its intended users are child development centres and kindergartens, and screening must be used alongside observation and ongoing developmental assessment. This Project
The development approach involved working with teachers, doctors, public health academics and early childhood specialists, designing a data collection set, and having teachers complete an online form. This Project The data used to train the model linked children’s environmental factors with standard EF assessment scores, and the system does not save the data entered after processing. This Project
The CF team said that collecting data nationwide to expand its use is not currently feasible. This Project We do not yet have documented requirements regarding resources, procedures or conditions for starting to use or build the model in other areas.
What methods or tools can we start using right away?
The tool is an EF risk-screening system for child development centres and kindergartens, to be used alongside ongoing observation. However, we have no record of how to access it or get started.
The tool developed is EF Predict, a Machine Learning system that helps child development centres and kindergartens screen young children for risks related to EF (executive function) skills. This Project The system assesses 5 areas: Inhibitory control, Shift, emotional regulation, working memory, and planning or organisation, using 13 data points such as birth order, diet, free-play time, and screen time. This Project
The screening result is not a diagnosis on its own, as the results may be inaccurate. It should therefore be used alongside ongoing observation and developmental assessment. This Project During development, we used data from around 2,000 young children at child development centres in the Lopburi area. The model had a Recall value of 49–87%, depending on the EF area and labelled dataset. This Project
We have no record of how to access the tool, the steps for getting started, or whether it is currently available for organisations to use.
How does this work connect to systemic issues, and what kind of collaboration is needed?
This work addresses the need to screen for developmental risks related to EF skills from early childhood and was developed with several groups of experts. However, we do not yet have a record of the partnerships we will need in the next phase.
EF Predict works at the source of developmental problems by helping screen for risks related to EF (executive function) skills in young children, so that child development centres and kindergartens can refer them for care before problems escalate. This Project This work aligns with the child and youth well-being framework, which takes a multidimensional view of development rather than focusing only on learning or physical health. IssueChildren, Youth and Well-being
Development involved collaboration with teachers, doctors, public health academics and early childhood experts in designing the data collection. This Project The data collection and pilot site was the High-Scope project at Tha Wung District Child Development Centre, Lop Buri Province, run by Tha Wung Hospital, and the EF scores were based on data from the Research Centre for Neuroscience, Mahidol University. This Project
One learning is that the model reflects the circumstances of the Lop Buri area only, and nationwide data collection to make it more widely usable is not feasible at present. This Project We do not yet have a record of what types of partners we need to collaborate with, or what forms of collaboration and resources will be needed in the next phase.