Background: Left arm and left leg lead-wire interchange (LA-LL) is difficult to detect because a normal ECG still appears normal with such interchange. Comparison of serial ECG simplifies the problem as long as there is no physiological change affecting limb lead morphology. The aim of this study is to design and evaluation a LA-LL interchange algorithm based on serial ECG differences. Methods: A large database of ECGs (n=146,000) was separated into serial ECG pairs for the patients with multiple ECGs. There were 89,600 patients and 25,700 had multiple ECGs. Many patients had more than one pair resulting in 56,800 pairs total. The average beats of serial ECG pairs, QT region only, were subtracted to construct a single 16 bit greyscale ECG image. The STT region was normalized for heart rate by a linear correction. The complete data set was twice as big, one copy for negative cases and a second copy with the LA-LL simulated interchange. We split the ECG images randomly into training (70%), test (20%) and validation (10%) subsets. The convolutional neural network (CNN) hyper parameters were selected based on trials on the training and validation sets. With the goal of high specificity, class weights were biased toward low false positives for training. CNN performance was based on sensitivity (SE), specificity (SP), positive predictive value (PPV) and F1 score. Results: The CNN SE and SP were 84.5% and 99.5% respectively. Assuming a prevalence of LA-LL reversal of 0.5% in clinical practice, the PPV and F1 were 46% and 0.60. Conclusions: Detecting LA-LL reversal is the most challenging in common lead-wire interchanges. Compared to the Kors algorithm with SE/SP of 17.5/99.5%, this algorithm shows a significant boost in sensitivity with the same specificity. The downside of the algorithm is the requirement for a previous ECG.