A network recognises heart attacks
On 1 December 1991 William Baxt published a prospective, blinded test of a neural network trained on 351 hospitalised patients: on 331 emergency patients with chest pain it recognised myocardial infarction with 97.2% sensitivity and 96.2% specificity, against 77.7% and 84.7% for the physicians.
Why it matters
A network trained on examples rather than written as rules beat the physicians treating the same patients in a prospective test, not on an archive. The author himself asked for confirmation on a larger sample.
The network was trained on clinical features of patients admitted with a high likelihood of infarction and tested on consecutive adult patients at a university hospital; physicians and network assessed the same people. The paper concludes cautiously that the network "may be" a valuable aid and that this must be confirmed on a larger sample. The abstract was read; the record gives neither the network's architecture nor its list of inputs.