|
BACKGROUND/RATIONALE:
ICU patients represent 20 percent of all hospitalized patients consuming significant healthcare resources. An economical accurate risk adjustment method could permit analysis of outcomes related to practices, providers, and institutions and evidence based organization and care.
OBJECTIVE(S):
This study developed and validated the best model to predict mortality at hospital discharge, compared prediction of that system to established systems {APACHE, Atlas Outcomes, and National Surgical Quality Improvement Project (NSQIP)], and characterized the contribution of ICU site to patient outcome.
METHODS:
This is a cohort study of more than 29,000 first ICU admissions. With a randomly split 60/40 dataset, a logistic regression model was developed and validated. We benchmarked the VA risk adjustment model by comparing the mortality forecast of individual patients developed by fours systems (VA ICU, NSQIP, Atlas Outcomes, APACHE) using a 10 x 10 table, and comparing sensitivity, specificity, positive and negative predictive value. We used a two level hierarchical logistic regression model to ascertain the contribution of the clusters represented by each ICU to hospital mortality forming confidence intervals with Markov chain monte carlo simulation.
FINDINGS/RESULTS:
The 11, 612 cases in the validation set showed an AUROC of 0.885 with excellent discrimination across the range of risk. Examination of the contribution of comorbid disease to hospital mortality found that use of conditions from the index hospitalization was sufficient in the model. Comparison of this system to that of NSQIP and to Atlas Outcomes showed generally similar or slightly superior results (c statistic VA ICU 0.87 NSQIP 0.84, ASG 0.80; O/E ratio : VA ICU 0.98 against NSQIP 1.2, VA ICU 0.64 against 1.52 for ASG. positive predictive value VA ICU .72, NSQIP .57 ASG .7). There were significant differences among ICU with O/E ratios ranging from 0.56 to 1.39 and confidence intervals suggesting four ICU with better than expected outcomes and five with worse than expected outcomes. These differences were not related to the proportion of imputed normal laboratory values, or discharge practices since they persisted when the outcome was changed to mortality at 90 days.
IMPACT:
This risk adjustment model has broad organizational, research and quality improvement applications since it can be used to compare various important ICU outcomes (length of stay, survival) across the VA without manual chart review, permitting large samples and large numbers of sites. In addition, the large ICU databases are being accessed by intensivists in the VA to answer other questions.
PUBLICATIONS:
None at this time.
DRA:
Health Services and Systems
DRE:
Quality of Care, Technology Development and Assessment
Keywords:
Quality assessment, Risk adjustment
MeSH Terms:
none
|
