RELEVANCE OF MULTIVARIATE PATTERNS OF ROUTINE CARDIOPULMONARY MEASUREMENTS FOR EARLY PROGNOSIS IN THE ADULT RESPIRATORY-DISTRESS SYNDROME

The aim of this work was to investigate by multivariate methods early patterns of conventional cardiopulmonary measurements and their time courses and to relate them to final outcome in the adult respiratory distress syndrome (ARDS). Data of 27 patients treated by the same medical team between 1985 and 1988 were used. These included demographic data, risk factors, causes of death, cardiopulmonary variables, band forms in peripheral smear, APACHE II score and previously proposed indices for ARDS. Discriminant analysis was applied to data collected during the 2 days after fulfilment of diagnostic criteria indicating ARDS. After univariate analysis and stepwise procedures, mean pulmonary artery pressure (PAP), dynamic compliance (C(dyn)), respiratory rate (f) and positive end-expiratory pressure (PEEP) were selected as discriminant variables and used to compute a final discriminant function. The evaluation of this function was done using cross-validation and bootstrap estimators of prediction error. The cross-validation estimate of correct classifications was 94%, with a standard deviation of 6%, as computed with bootstrap techniques. Since two of the selected variables (PEEP and f) reflect medical therapy rather than patient response, a discriminant function was computed for each pair of variables, i.e. PAP and C(dyn) as well as f and PEEP, showing 90% and 85% of correct classifications, respectively. PaO2/FIO2, PAP and C(dyn) presented time trends related to outcome. Star plots using the relative importance of each measurement were employed to display the discriminant variables, and time trends were related to critical values. The selection of therapy-dependent measurements using objective methods suggests that while factors independent of staff manipulation should be used for patient classification, those dependent upon therapeutic strategies should be considered for a more exact prognosis. We conclude that the approach provided by multivariate patterns and time series analysis may improve the quality of medial decision making for ARDS patients.