A simple neural network was described on paper and then modeled using electrical circuits in 1943 by the neurophysiologist Warren McCulloch and the mathematician Walter Pitts. This idea was advanced through the ’50s, ’60s, and 70’s mainly as an academic pursuit until the 1980’s when the “instant physician” showed promising clinical application. This neural network stored a massive number of medical records, learned from those medical records, and could determine the best diagnosis and treatment when presented with a set of symptoms. Just like a human doctor; but more accurate, unbiased, and a heck of a lot quicker.
The “instant physician” is just one of many Artificial Neural Networks, or ANNs, which are mathematical algorithms generated by computers that capture knowledge contained within data and analyze it. Essentially, they are artificial human brains that learn things, remember those things, and then use those things to figure out other things that are often not grossly apparent in large swaths of data. They receive inputs (information we give them), think about it (process the inputs), and then generate an output (an answer). More often than not these answers are a yes or no or true or false, like, is this disease present? Is this person at risk of developing a specific disease? Will this person die in the next 5 years given the disease they have?
Nowadays, however, ANNs are much more sophisticated and integrated with many parts of human medicine involving clinical diagnosis, prediction of cancer and length of stay in hospital postoperatively, speech recognition, predicting risk of heart disease and osteoarthritis, radiograph and ECG analysis, and drug development. They are also used in more nebulous endeavors like clinical modeling and detecting nonlinear relationships in research that the human brain often cannot comprehend. And, even now in veterinary medicine, these networks can read radiographs and predict chronic kidney disease in cats. But ANNs in veterinary medicine is nowhere near as complex as those in human medicine.
But they will get there. As we know, veterinary medicine always trails behind human medicine, and much of what we do in our field is often modeled off the successes of our anthropocentric colleagues.
The utility of ANNs will surely become more apparent as veterinarians strive for efficiency, accuracy, automation, and getting as close as possible to gold standard care. We all want that as veterinarians, right? Wouldn’t it be awesome to be able to predict the risk of osteoarthritis in a 6-month-old German Shepherd so that preventive measures can be taken sooner? Or determine the presence of a fragmented medial coronoid process radiographically, without having to perform a costly CT scan? Or determine the risk of surgical/post-operative complications in a brachycephalic dog prior to surgery? These would all be great, and owners would love it.
When questions are asked, boundaries are often lifted, and advancements are made. I don’t know anything substantial about computer programming, writing code, artificial intelligence, or designing ANNs but I believe those who do fashion these networks to process information and think like us. They are pushing the threshold of machine learning and ANNs will, without doubt and in time, become so sophisticated that they will begin to think on their own, without our input. They may even replace us. What will we, as veterinarians, do then?