ADHD Screening Tool: Investigating the effectiveness of a tablet-based game with machine learning

Abstract

This study investigated the ability of a low-cost, mobile device-based game incorporating machine learning to screen participants between ages six and twelve years for attention-deficit/hyperactivity disorder inattentive subtype (ADHD_I). Relevant information from literature was incorporated into the game in light of the DSM-V diagnostic criteria. The game has seven back-to-back segments with unique layouts and a visual theme.The ADHD Screening Tool presents a novel patient-testing interface with a cloud-based machine learning classifier (MLC) integrated with a consensus algorithm. The game was tested with 39 clinically diagnosed participants (ADHD_Iand non-ADHD). Out of nine classifiers tested, the locally-deep support vector machine gave the best results: using leave-one-out cross-validation, this MLC classified data from five game segments for 38 participants with sensitivity of 92.9% and specificity of 82.9%. By making use of the consensus algorithm, the 39th participant was correctly classified according to the clinical diagnosis. The MLC and consensus algorithm were able to classify 39 participants with a sensitivity of 100% and specificity of 87.5%. To overcome participant class imbalance, the synthetic minority oversampling technique (SMOTE) was implemented on game segment data. The SMOTE two-class LDSVM yielded sensitivity of 90.7% and specificity of 94.4%. The study used an internet-connected, commercially available tablet.