Readmissions are common and costly, and can lead to further clinical decline of the patient. Many readmissions may be potentially preventable, the risk being modifiable by the quality and type of care provided. Identifying this subset of patients would help direct quality improvement (QI) efforts more effectively in reducing readmission rates. Focusing on specific subgroups in the VA who are at relatively high risk of readmission [patients with heart failure (HF), acute myocardial infarction (AMI), and pneumonia (PNA)] may yield a higher reward in terms of improved patient care and successful QI initiatives. These conditions are also publicly reported on the VA and CMS Hospital Compare websites, and are being used to assess the performance of VA hospitals. By furthering our understanding of potentially preventable readmissions, we will be better positioned to target those areas that need QI as well as understand those processes of care that indicate high quality of care and that may be less likely to lead to readmission.
1) Estimate risk-adjusted models to predict 30-day readmissions for patients discharged with HF, AMI, or PNA from an acute-care VA hospital; 2) Investigate rates of potentially preventable readmissions for HF, AMI, and PNA using readmission classification software [3M Potentially Preventable Readmissions (PPRs)] designed to identify potentially preventable readmissions based on administrative data; 3) Develop chart abstraction tools to identify potentially preventable readmissions for patients discharged with HF, AMI, or PNA; 4) Apply chart abstraction tools to VA electronic medical records (EMR) to classify HF, AMI, and PNA all-cause readmissions; and 5) Re-estimate hospital-specific risk-adjusted rates of potentially preventable readmissions in VA using supplemental automated data.
We conducted a 3-year retrospective observational study using FY2006-2010 VA inpatient and outpatient administrative and EMR data, supplemented by CMS Medicare files. Assessment of potential preventability of 30-day readmissions among 100 cases per condition-cohort was compared between EMR-abstracted data and the 3M PPR software. We developed and pilot-tested EMR abstraction tools to assess processes of care (i.e., quality) during the index and post-discharge periods. Two trained nurse-abstractors reviewed EMRs, with inter-rater agreement >90 in all conditions. We compared PPR-flagged and non-flagged cases on total and section-specific mean quality scores using t-tests. We also reclassified the PPR algorithm by incorporating lab values, vital signs, prior utilization, and medication data into the risk adjustment; we then estimated regression models to predict PPR readmissions (yes/no) and examined whether the re-classified algorithms improved model performance or led to changes in hospital ranks and performance using hospitals' original PPR observed-to-expected ratios versus enhanced PPR observed-to-expected ratios (with supplemental data).
1) Hospital risk-adjusted 30-day readmission rates ranged from 14.9% to 23.6% for AMI; 16.5% to 31.0% for HF, and 12.8% to 20.7% for PNA.
2) The overall all-condition PPR rate was 10.8%; hospital risk-adjusted rates ranged from 5.9%-18.6%. The range in hospital risk-adjusted PPR rates was 8.0%-60.0% for AMI, 9.0%-31.3% for HF, and 3.9%-29.2% for PNA.
3) Explicit criteria to assess processes of care for each condition were selected and refined by expert clinical panels following the RAND/UCLA Appropriateness Method. These were incorporated into 4 sections of chart abstraction tools: admission work-up, evaluation and treatment during stay, discharge readiness, and post-discharge period (a maximum obtainable quality score was 100 overall and 25 for each section).
4) For all three cohorts, we did not find any significant associations between PPR outcomes and the overall or section quality scores, whether we used equal, section, or Delphi-based weights for each chart-based data element. For example, among 100 HF cases, the overall mean quality score was 61.5+10.3. Section scores were highest for discharge readiness (18.8+2.4) and lowest for post-discharge care (7.3+8.1). Mean overall and section quality scores did not differ by PPR status; respective PPR-Yes vs. PPR-No overall scores were 61.2 and 63.4 (p=0.47).
5) We did not find improved adjusted R-square estimates in the reclassified condition-specific model when we examined the association between the quality score and the original versus reclassified PPR algorithms. However, specifically among PNA readmissions, we found that adding prior utilization data and vital signs to predict PPRs increased the c-statistic from 0.577 to 0.626. Prior utilization also significantly increased the odds of PPRs; in particular, the number of prior admissions had an odds ratio (OR) of 1.20, confidence interval (CI)=1.18-1.22. Finally, 9.2% of hospitals changed performance quartiles when supplemental clinical data were incorporated into the models.
Given increasing reliance on readmission as a measure of hospital performance, it is critical for hospitals to be able to identify those readmissions that are more likely to be preventable and therefore better targets for quality improvement. Although the PPRs represent an attractive alternative to the CMS all-cause readmission measure, we did not find a significant difference in the quality of care (as measured by inpatient and post-discharge processes of care) between cases flagged as PPRs and non-PPRs. This may be due to: problems in using administrative data-based readmission measures; limitations in the data present in the EMR (e.g., patient-provider communication) which limit our ability to determine preventability; and the multitude of factors associated with readmission we did not examine (e.g., socioeconomic factors, psychosocial factors, access issues). Future studies at the hospital level are needed to further explore the utility of the PPRs in identifying high-risk groups of patients for QI/intervention. In addition, future studies should also explore other ways of identifying potentially preventable readmissions (i.e., through provider and/or patient perceptions or direct observation).
External Links for this Project
Grant Number: I01HX000367-01A1
None at this time.