TL;DR
The authors introduce ChildCI, a new open-source framework and dataset (ChildCIdb v1) containing touch and stylus interaction data from 438 children aged 18 months to 8 years. To demonstrate its potential for e-Health and e-Learning applications, they show that machine learning models can predict a child's developmental age group with over 90% accuracy based purely on how they color a digital tree.
What Problem Does This Paper Address?
Children are interacting with mobile devices earlier than ever, yet the field of Child-Computer Interaction (CCI) is still in its infancy. A major bottleneck is the severe lack of large-scale, publicly available datasets.
Most existing studies suffer from two main flaws: they capture data from a single session (missing long-term developmental changes) and they focus on narrow age ranges (typically 2–11 years). This leaves the critical early sensorimotor and preoperational stages—defined by Piaget's theory of cognitive development—largely unmapped. Furthermore, existing datasets rarely combine both finger-touch and stylus inputs or include rich metadata like academic performance or clinical conditions (e.g., ADHD).
Methods
The authors developed a custom Android application designed to capture both fine and gross motor skills through a series of gamified tests.
The app breaks interaction down into two main blocks:
- Touch Analysis (Finger): Tasks include tap/reaction time tests, drag-and-drop exercises, and multi-touch zoom-in/zoom-out gestures.
- Stylus Analysis (Pen): Tasks include tracing a spiral and a free-form drawing/coloring test.
To demonstrate the dataset's utility, the authors built an ML pipeline to classify children into three age groups (1–3, 3–6, and 6–8 years old) using data exclusively from the coloring test.
The ML Pipeline:
- Feature Engineering: They extracted 148 global features. While 114 were standard Human-Computer Interaction (HCI) metrics (kinematics, time, pressure), 34 were novel drawing-specific features introduced in this paper (e.g., stroke length, time spent drawing inside vs. outside the margins).
- Feature Selection: They compared Fisher Discriminant Ratio (FDR), Sequential Forward Floating Search (SFFS), and Genetic Algorithms (GA) to find the most informative features.
- Classification: They benchmarked seven different classifiers, ranging from simple Naive Bayes to Support Vector Machines (SVM) and Multi-Layer Perceptrons (MLP).
Datasets
The primary contribution is ChildCIdb v1, a new benchmark dataset containing data from 438 children ranging from 18 months to 8 years old.
Unlike previous datasets, it includes comprehensive metadata:
- Exact age, gender, and handedness.
- Clinical indicators (premature birth status, ADHD diagnosis).
- School grades.
- A pre-test emotional self-assessment where children select a face representing their current mood.
The dataset is explicitly designed for longitudinal tracking; the authors intend to follow this same cohort over a 5-year period.
Key Insights and Ideas
- Drawing Style as a Developmental Fingerprint: The way a child interacts with a screen changes predictably as their neuromotor skills mature. Younger children rely almost exclusively on simple swipe gestures, whereas older children exhibit highly organized, precise strokes.
- Context Matters More Than Kinematics: The newly proposed drawing-specific features (like the ability to stay within margins) were selected more frequently by optimization algorithms (24.2% of the time) than traditional kinematic or pressure features. This proves that task-specific context is vital for accurate CCI modeling.
- Unsupervised Alignment with Psychology: When the optimized feature vectors were projected into a 2D space using UMAP, they naturally clustered into three distinct groups that perfectly mirrored the children's actual biological age brackets. This provides strong quantitative backing for traditional stage-based psychological development theories (like Piaget's).
Key Results
- High-Accuracy Age Detection: The combination of SFFS (Feature Selection) + Support Vector Machine (SVM) achieved the highest age-group classification accuracy of 90.45%.
- Classifier Performance: Advanced classifiers significantly outperformed simpler ones. Random Forest and MLP achieved strong accuracies of 88.69% and 85.98% respectively, while Naive Bayes lagged behind at 69.63%.
- Feature Dominance: Geometry (20.3%) and Kinematics (18.5%) proved to be highly discriminative categories, whereas screen pressure (11.1%) and direction changes (11.2%) were less helpful in distinguishing age groups.
Limitations and Future Work
While ChildCIdb v1 is a massive step forward, the authors note several areas for improvement:
- Single-Task Limitation: The age detection experiment in this paper only utilized data from the coloring test (Test 6). Future work will incorporate data from the other five touch and stylus tasks.
- Lack of Clinical Grounding: The framework's insights have not yet been directly correlated with gold-standard clinical benchmarks like the NIH Toolbox or the Bayley Scales of Infant and Toddler Development.
- Longitudinal Execution: The current release represents "Year 1" data. The true value of the framework will unlock as multi-session data is gathered over the planned 5-year window, allowing researchers to track individual developmental trajectories over time.
Summary updated