Serial Assessment of Mortality in the Neonatal Intensive Care Unit by Algorithm and Intuition: Certainty, Uncertainty, and Informed Consent

Abstract

Objectives. Does predictive power for outcomes of neonatal intensive care unit (NICU) patients get better with time? Or does it get worse? We determined the predictive power of Score for Neonatal Acute Physiology (SNAP) scores and clinical intuitions as a function of day of life (DOL) for newborn infants admitted to our NICU. Methods. We identified 369 infants admitted to our NICU during 1996–1997 who required mechanical ventilation. We calculated SNAP scores on DOL 1, 3, 4, 5, 7, 10, 14, 21, 28, and weekly thereafter until either death or extubation. We also asked nurses, residents, fellows, and attendings on each day of mechanical ventilation: “Do you think this child is going to live to go home to their family, or die before hospital discharge?” Results. Two thousand twenty-eight SNAP scores were calculated for 285 infants. On DOL 1, SNAP for nonsurvivors (24 ± 8.7 [standard deviation]) was significantly higher than SNAP for survivors (13 ± 6.1). However, this difference diminished steadily and by DOL 10 was no longer statistically significant (12.7 ± 4.9 vs 10.0 ± 4.8). On each NICU day, at all ranges of SNAP scores, there were at least as many infants who would ultimately survive as would die. Consequently, the positive predictive value of any SNAP value for subsequent mortality was <0.5 on all NICU days. Prediction profiles were obtained for 230 ventilated infants reflecting over 11 000 intuitions obtained on 2867 patient days. One hundred fifty-seven (81%) of 192 survivor profiles displayed consistent accurate prediction profiles—at least 90% of their NICU ventilation days were characterized by 100% prediction of survival. Twenty-five (13%) of 192 surviving infants survived somewhat unexpectedly; that is, after at least 1 day characterized by at least 1 estimate of “death.” Thirty-three (60%) of the 55 nonsurvivors died before DOL 10. Eighty-two percent of the prediction profiles for these early dying infants were homogeneous, dismal, and accurate. Twenty-two (40%) of the 55 nonsurvivors died after DOL 10. Seventeen (78%) of these 22 late-dying infants were predicted to live by many observers on many hospital days. Sixty-one (30%) of 230 profiled patients had at least 1 NICU day characterized by at least 1 prediction of death; 26/61 (43%) of these patients were incorrectly predicted; that is, they survived. Seventeen infants who were predicted to die during but survived nonetheless were assessed neurologically at 1 year. Fourteen (82%) of these 17 were not neurologically normal—8 were clearly abnormal, 1 suspicious, and 5 had died. Conclusions. If absolute certainty about mortality is the only criterion that can justify a decision to withhold or withdraw life-sustaining treatment in the NICU, these data would make such decisions difficult on the first day of life, and increasingly problematic thereafter. However, if we acknowledge that medicine is inevitably an inexact science and that clinical predictions can never be perfect, we can ask the more interesting question of whether good but less-than-perfect predictions of imprecise but ethically relevant clinical outcomes can still be useful. We think that they can—and that they must.