COMPUTATION MODES OF MULTIVARIATE POSITIVE PREDICTIVE CHARACTERISTICS

Multivariate receiver operator characteristics (ROCs) and positive pre dictive characteristics (PPCs), based on a combination of two or more clinical variables, are usually computed by varying dichotomy limits f or each variable independently. This approach has a similar number of degrees of freedom (i.e., uses the same number of programmable paramet ers) as the approach which defines test positive cases based on a line ar combination of all the clinical variables involved. Either approach can be implemented without any assumption about the underlying probab ility distributions by using an exhaustive computer search. Both appro aches were compared in a demonstration study of predicting 2-year all- cause mortality after acute myocardial infarction, based on applying v arious time- and spectral-domain indices of signal-averaged ECGs from a research survey. The results showed that the optimum mode for the co mputation of ROCs and PPCs depends on the character of the particular data used. Therefore, in order to increase the precision of the retros pective multifactoral studies, both approaches to ROC and PPC computat ion should be used and compared in each individual investigation.