The Bactra Review   An Introduction to Computational Learning Theory
I recall reading it --- in 1991 --- in Maureen Caudill and Charles Butler, Naturally Intelligent Systems (MIT Press, 1989). Since I can neither lay hands on that book, nor remember who did the original research, I apologize for any errors due to undergraduate memories in what follows.

Addendum, 6 Nov. 1999: The work was done by Donald Specht in 1963 (!) at Stanford. The digitized EKG readings formed a 46 dimensional vector. There were two versions of the device. The first used several linear elements (hyper-planes) to separate normal from abnormal cardiograms (each effectively testing for a different abnormality) and correctly identified 93 percent of normals and 76 percent of abnormals. The second used a single nonlinear element, and so a polynomial separating hyper-surface (as I described in the text), and got success rates of 97 percent and 90 percent, respectively. The corresponding rates for clinicians were only 95 and 54 percent, respectively. See pp. 81--84 for Caudill and Butler's book. (Thanks to Isa Bernardini for providing me with a copy.)