Early Prediction of ADHD Using LSTM-RNN Model Based on Behavioral Activities

Abstract

Telling apart children who have Attention Deficit Hyperactivity Disorder (ADHD) from those who do not is the central problem this paper takes on, and the tool built for the job layers three pieces together: a recurrent LSTM network that reads behavioral sequences, a Multi-Agent Salp Swarm (MASS) routine that tunes an XGBoost feature-selection stage, and an oversampling method (SMOTE) applied specifically to correct the skew between the two diagnostic groups within the training set. Sequential input is where LSTM networks excel, so that piece is tasked with pulling time-based signal out of the raw behavioral records. What it extracts then passes through MASS-guided XGBoost, a stage whose job is to sharpen how cleanly the two groups separate in feature space. Since ADHD samples are the smaller of the two classes in most such datasets, SMOTE generates additional synthetic minority-class examples before any training happens, so the model isn’t skewed toward predicting the majority class by default. The resulting system achieves an F1-score of 0.99, precision of 0.9980, recall of 0.9952 and overall accuracy of 0.9978 on held-out data. These numbers suggest that the two groups are very cleanly separable under this approach. All in all, these results indicate that the stacking of a sequence-aware deep network on top of a traditional boosted-tree feature selection is a promising approach for early automated ADHD screening, one that could plausibly complement a clinician’s own judgment during diagnosis.