Early and explainable identification of handwriting difficulties in children via game-based machine learning

Handwriting is a fundamental, but complex skill. Many children find it hard to master due to motor, cognitive, or attentional factors. Structured and timely training can reduce struggles, but some children are training-resistant, signaling clinical risks. Technological tools, including serious games and machine learning, can enrich training, but it is crucial to verify in an online manner and with explainable methods its efficacy with attention to temporal evolution, for early intervention in cases of resistance. This study addresses such challenge by evaluating game-based features for clinical risk detection in first-graders. Sixty-six children (age 6.45 ± 0.31 years) underwent three-month tablet-based handwriting training at school, with clinical assessments and teacher observations collected before and after training. We extracted kinematic, smoothness, temporal, frequency, and performance features, selecting those showing strong learning trajectories (|r| ≥ 0.5 or R 2 ≥ 0.5). Models were trained on aggregated features using nested 5-fold cross-validation with hyperparameter tuning, oversampling and progressive evaluation using cumulative data from the first month, first two months, and full period. Clinical risk detection outperformed teacher-identified weakness detection (Average Precision: 0.86 vs. 0.67), achieving test ROC-AUC = 0.850 ± 0.063 and AP = 0.772 ± 0.057, with discriminative signal emerging by day 36 (11 gameplay days), and smoothness and pressure features guiding classification. Misclassification patterns revealed attention-related difficulties in false negatives and compensated motor inefficiencies in false positives. These findings demonstrate that tracking learning dynamics over time via game-based features enables early identification of training-resistant children, supporting educators in timely intervention beyond static assessments.