Researchers at the University of Tübingen tried to train an A.I. to learn to identify a specific fish called tench on pictures. When they looked into the A.I. to see how it tries to identify that fish, they found out that the #AI was looking for human fingers against a green background as most of the tench photos in the training data featured a fisherman holding the fish in their hands after catching it. Can you imagine how useful it would be in medicine to point out such things before an algorithm is developed? This way, we could avoid so many false negatives, false positives and bias built into the algorithms. For example, radiologists looked at the ChestXray14 dataset of X-ray images and discovered that many of the images of patients with pneumothorax had a chest drain on the image as treatment already started when doing the X-ray. A #MachineLearning algorithm tasked with identifying X-rays of pneumothorax would probably look for the drain in the images as a perfect sign of the condition (they have to perform as well as humans would do, not learn to diagnose the condition), this way, missing patients with the condition who have not been treated yet. Can you think of an example in your practice (medical or not) where such subtle observations exist? …see more