Diagnosis by machineBMJ 2008; 337 doi: https://doi.org/10.1136/bmj.a1703 (Published 17 September 2008) Cite this as: BMJ 2008;337:a1703
- Christopher Martyn, associate editor, BMJ
Two or three decades ago there was much excitement about how computers were going to make diagnosis faster, more accurate, and more reliable. In the United Kingdom, and doubtless many other countries too, some accident and emergency departments had an early generation personal computer running Bayesian algorithms for the diagnosis of acute abdominal disorders. If you typed in answers to the questions it posed you got a differential diagnosis and the associated probabilities for your particular case. Several studies were published showing that such algorithms increased diagnostic accuracy and reduced the numbers of negative laparotomies.
Artificial neural networks were more sophisticated than the Bayesian stuff. The buzz words were parallel distributed processing and connectionism, and part of the allure of this approach was the link with cognition and neuroscience. It was hoped that these networks would be capable of extracting patterns from the noisy, messy, and incomplete sort of information you end up with after taking a history from a patient. There was …