Currently, clinical screening and diagnosis for Attention Deficit Hyperactivity Disorder (ADHD) relies on complex, time-consuming, expert-dependent thorough clinical assessment. This study presents an algorithm for automatic ADHD detection using electrocardiogram (ECG) signals. A one-dimensional Convolutional Neural Network (CNN) was trained at first, then a score-weighted visual explanation method for CNN method called Score-CAM was used to generate heatmaps. The derive temporal and spectral features of the heatmaps then utilized in classifiers for ADHD detection. The best classifier achieves an average classification accuracy of 95.20%, precision of 95.21%, recall of 95.15% and F1-score of 0.9515. This pilot study highlights the potential of using deep learning features in bio-signal processing. This research could be translated into an efficient, low-cost screening technique for ADHD.