Author + information
Background: Readmissions after heart failure (HF) are the focus of pay-for-performance initiatives. Early and accurate identification of patients at risk for readmissions may improve quality and reduce cost of care.
Methods: OM1 Linked Data contains linked claims and EMR data from 20 million US patients. Of ∼200,000 patients with HF, 24,615 met study criteria of at least 1 HF-related hospital admission and at least 6 months data preceding that admission. Logistic regression, random forests, classification and regression trees were used to identify pre-admission predictors of 30-day readmission or death. Models were tested by 10-fold cross-validation in CART and training-validation (67% versus 33%) in logistic regression. We computed the simple risk score by adding the assigned weights to each predictor based on the parameter estimates from the logistic regression.
Results: Of the 24,615 patients with index HF hospitalization, 3,109 (13%) were readmitted within 30 days and 365 (1.5%) died. Hospitalizations in the previous year (odds ratio [confidence interval]: 1.2 [1.1, 1.3]), procedures in the previous year (1.1 [1.0, 1.1]), 5-point higher Charlson comorbidity index (1.6 [1.3, 2.0]), and months since last hospitalization (0.9 [0.9, 1.0]) were the strongest predictors. The overall classification rate in the cross-validation cohort was 69%; 73% in 21,141 patients who did not have an outcome, and 46% in 3,474 patients who died or were readmitted. The median [Q1, Q3] risk score was 35 [18, 59] in the validation cohort (n=8208). The risk score correlated well with the readmission rate within each decile: R2=0.90 in both the weighted and unweighted regression analyses where the weight was the number of patients within each decile. Patients with risk scores of ≥5 were classified as high risk. In the validation cohort of 1,157 patients who were readmitted or died, 600 (51.9%) had a risk score of at least 5.
Conclusions: Our model correctly predicted 30-day readmission or death in 7 out of 10 patients, prior to index HF admission. We developed a simple score based on routinely available data to identify patients at high risk. We will use notes, and machine learning techniques to further improve predictive accuracy.
Poster Hall, Hall C
Friday, March 17, 2017, 3:45 p.m.-4:30 p.m.
Session Title: Acute Heart Failure: Evaluating Strategies to Prevent Readmissions
Abstract Category: 13. Heart Failure and Cardiomyopathies: Clinical
Presentation Number: 1163-264
- 2017 American College of Cardiology Foundation