EVALUATION OF COMPUTER-ECG PROGRAMS - STRANGE CASE OF THE GOLDEN STANDARD

A statistical model is formulated which permits analysis of expected magnitude of errors when the performance of computer-ECG programs is judged against human readers as a standard. It is demonstrated that the apparent performance of a program being tested is always worse than that of the standard even if the program were absolutely perfect. Furthermore, the program's apparent performance depends on the composition of the test population regarding true event rate. These observations explain why two groups of investigators can end up with different conclusions about a program's performance even if otherwise identical test procedures were employed. © 1979.