GUS SLOTMAN, MD, FCCM
Director of Clinical Research
Department of Surgery, Inspira Health Network
VINELAND, New Jersey, United States
Disclosure information not submitted.
Title: Pre-Discharge Decision Trees Analysis Predicts 30-Day Pneumonia Re-Admission
Introduction: CMS penalizes hospitals for patients who are re-admitted within 30 days (R-A30) after hospital discharge for a primary diagnosis of pneumonia. Identifying pneumonia inpatients at highest risk for R-A30 is key to formulating interventions for reducing it. Pneumonia-specific R-A30 scoring is not standardized. Hypothesis: R-A30 can be predicted by decision trees analysis from clinical data captured during the sentinel pneumonia admission.
Methods: Data from 3350 consecutive patients hospitalized with a primary diagnosis of pneumonia were modeled for post-discharge R-A30 by decision trees analysis using Classification and Regression Trees (CART)/log likelihood algorithm, splitting the strongest categorical/continuous baseline covariates into optimal categories. Data included medical history, demographics, and hospital laboratory results in the first 48 hours of the sentinel pneumonia admission and at 96 and 48 hours pre-discharge (PD).
Results: Overall 508/3350 (15%) were R-A30. R-A30 was 33% for < 57 days since last hospitalization (node n=657), 45% for < 57 days + 96PD PMN >80%(node n=145), and 61% for < 57 days + 96PD PMN >80% + 96PM creatinine >1.85(node n=41). R-A30 was 18% for >57 days (node n=1186), 45% for >57 days + 48PD PMN >89%(node n=62). In smaller nodes, >57days plus low 96PD GFRcalc or high 96PD osmolality and/or low 48PD glucose predicted R-A30 at 36-67%.
Conclusions: CART decision tree analysis from sentinel pneumonia admission clinical data predicts R-A30 up to quadruple the overall rate of 15% in this study population. Days since last hospitalization and PMN count identify prime R-A30 at-risk patients. Under these parameters, a high creatine level further identifies at-risk patients, as do low GFR, high osmolality, and low glucose, effective in smaller node numbers. Although the node n for some R-A30 predictions is small, this information still may help direct focus of intensivists managing hospitalized patients with pneumonia toward individuals identified as high-risk for R-A30 by the decision-tree algorithms. Subsequently, this advance knowledge may lead to targeted interventions before and after discharge of at-risk pneumonia patients, possibly reducing R-A30 frequencies for them.