Estimating neonatal mortality risk: an analysis of clinicians' judgments

Clinicians' estimates of mortality risk in the neonatal intensive care unit (NICU) have implications for patient triage, transfer, initiation and termination of life support, and allocation of medical resources. The accuracy of these judgments has not been studied, nor the differences between nurses and attending physicians.

1) evaluate the accuracy of subjective judgments of NICU unit mortality risk, 2) identify the key components of clinician judgments, 3) compare accuracy between attending physicians and nurses, and 4) examine the utility of combining an objectively computed risk and clinician judgments to improve predictions.

We obtained estimates of mortality risk on 544 admissions to two NICUs on the day of admission from the attending physician and primary nurse. These were compared with an objective computed mortality risk based on birth weight and the Score for Neonatal Acute Physiology (SNAP) using a linear judgment analysis model, as well as with actual outcomes.

Physicians and nurses had good discriminating power with actual mortality rates ranging from 0% among low risk patients to 67% among those with the highest mortality estimates. Physicians had a tendency to overestimate mortality risk. Clinicians base their estimates on the same factors and similar judgment weights as the empiric mortality risk model (22% birth weight, 62% illness severity (SNAP), 13% low Apgar, and 3% for intrauterine growth restriction). Clinicians place additional emphasis on therapeutic as well as physiologic factors. When the computed risk and physician judgment were combined, both made significant contributions in a logistic mortality risk model.

Clinician judgments of mortality risk are fairly accurate and similar to an objective illness severity index. This simple method provides insight into clinical decision making and has important applications in improving direct patient care, appropriate allocation of medical resources, and medical training.