Evaluation of pattern recognition rules for the apex of the heart

The observer variability and accuracy of rules for the automatic recognition of the left ventricular (LV) apex are reported. To form a data base for examining this problem 18 co-workers were asked to independently identify the apex of 79 left ventriculograms (LVgrams) in the 30-degree right anterior oblique projection at end-disastole (ED) and end-systole (ES), and to assign a confidence score from 0% to 100% for each apex. The group consisted of five clinical cardiologists (G1), six trained analysts (G2), and seven other experienced observers (G3). The LVgrams were selected to include those from 31 normal individuals and those from 48 patients with mixed akinesis/dykinesis as coded by conventional scoring techniques. There was no significant difference in the confidence for the ED vs ES apices. However, confidence of G1 was significantly lower than that of G2. The standard deviation in the observer-detected apex correlated (r = -0.92) with the mean observer confidence, which suggests a basis for quantifying "confidence" in the automated apex detection. The apex determined with respect to the superior aspect of the aortic valve proved superior to all other tested methods. the concept of "peakedness index" is introduced and used to estimate observer variability.