Abstract 242 Poster Session III, Monday, 5/3 (poster 89)

Introduction: None of the currently available physiology based mortality risk prediction models incorporate subjective judgements of health care professionals, a source of additional information that could improve predictor performance and make such systems more acceptable to health care professionals. This study compared the performance of subjective mortality estimates by nurse's and physician's with a physiology-based method, PRISM III. Then, using Bayesian statistics, we combined the PRISM III estimate with the health care provider estimate. The performance of the Bayesian model was compared to the initial two prediction models, PRISM III and subjective predictions.

Methods: For each of the 642 consecutive eligible patients, an exact mortality estimate and the degree of certainty (continuous scale from 1 to 5) associated with the estimate was collected from the attending, fellow, resident and nurse responsible for the patient's care. Bayesian statistics were used to combine the PRISM III and subjective predictions to create a third Bayesian estimate of mortality. Model performance comparisons utilized the area under the curve (AUC) and the goodness-of-fit Chi-squares statistics.

Results: The table shows the AUCs and the goodness-of-fit Chi-squares for the mortality estimates of the health care providers, PRISM III, and the Bayesian predictor. While the AUCs of the health providers were not significantly different from the AUCs of PRISM III, the Bayesian AUCs were higher than both the health providers AUCs (p<0.09, for all) and PRISM III AUCs. Similarly, the calibration statistics for the Bayesian estimates were superior to the calibration statistics for both the health care providers and PRISM III models.

Table 1 No caption available
Full size table

Discussion: The results from this study demonstrate that physician mortality predictions and PRISM III predictions performed equally well. The Bayesian model that combines physician and PRISM III mortality predictions is more accurate than either alone and may be more acceptable to physicians. A methodology using subjective outcome predictions maybe more relevant to individual patient decision support.