Attention-Deficit/Hyperactivity Disorder (ADHD) is acknowledged as a neurodevelopmental disorder that is becoming increasingly prevalent among children in Vietnam. This condition poses significant challenges, potentially hindering their academic and social accomplishments, thereby underscoring the importance of early diagnosis. This research seeks to design and conduct an experiment using eye-tracking technology, which provides a swift and accurate analysis of eye movement patterns to identify and characterize ADHD symptoms in children. A cohort of 34 participants from the Ben Tre Mental Hospital in Vietnam was enlisted for the study, comprising 21 individuals diagnosed with ADHD and 13 non-ADHD individuals serving as controls. All subjects underwent a 40 s assessment involving tasks designed to reflect their attention, all monitored through eye-tracking systems. A total of 216 features of gaze shifts and pupil diameter were meticulously extracted in specific tasks and across the entire assessment period. Subsequent analysis employed the ANOVA F-test score to pinpoint the characteristics significantly associated with ADHD. The F-test score results revealed that features based on gaze velocity and pupil dilation exhibit significant differences between control and ADHD-afflicted children, providing insights into vision control and attention characteristics across groups. To identify the best feature set, two ensemble learning models were employed to analyze the top 16, 32, 108, and 162 features exhibiting the highest F-scores. To facilitate ADHD recognition, three classical machine learning models, along with two ensemble learning models, were utilized on the best feature set. The Random Forest model demonstrated superior discriminative ability with a mean AUROC score of 0.94, while the XGBoost model presented the highest classification performance with a mean accuracy of 88% during the repeated 4-fold cross-validation on the best-16-feature set. These results underscore the model’s robustness in distinguishing between ADHD and non-ADHD participants, offering a promising tool for early diagnosis, particularly in rural areas where access to specialized medical resources may be limited.