Background
Chronic obstructive pulmonary disease (COPD) is a major public health problem [
1,
2]. In the US, there are approximately 700,000 hospitalizations each year [
3] with one-fifth resulting in readmission within 30 days [
4]. To curb the healthcare burden, the Hospital Readmissions Reduction Program (HRRP) has started penalizing hospitals for higher than expected rate of 30-day readmission after COPD hospitalization [
5]. In addition to hospital-level quality improvement efforts, identification of patients at high risk for readmission and the development of interventions (e.g., care transition interventions) are of great interest to many stakeholders [
6‐
9].
As with claim-based models to predict 30-day readmission after hospitalization for other HRRP-targeted conditions (e.g., heart failure) [
10‐
13], several studies have identified predictors and developed prediction models for readmissions in patients hospitalized for COPD [
8,
9,
14‐
16]. These models incorporated the basic demographics (e.g., age, sex), comorbidities, and in-hospital management (e.g., medication use), with reporting C-statistics of 0.63 to 0.72 [
14,
16]. However, these prediction models have been criticized for their lack of detailed social factors (e.g., educational level, marital status) [
13,
17‐
19], and for the assumption that 30-day readmission is a homogeneous process [
17,
18,
20]. Indeed, the effects of inpatient management on the readmission risk diminishes rapidly after discharge, reaching a nadir at post-discharge day 7 [
18]. Despite the emerging evidence suggesting the involvement of non-clinical factors – such as social factors – in readmission processes [
21‐
23], little is known about whether these factors improve prediction ability and how their contribution varies by timing after COPD hospitalization. In addition, while several studies built prediction models using administrative datasets (e.g., Nationwide Readmission Database [NRD]), these datasets do not include the information on detailed social factors [
9,
13,
17‐
20].
To address this knowledge gap, we used nationally-representative sample of Medicare beneficiaries to test the hypothesis that the addition of social factors to prediction models quantitively improves the predictive ability for 30-day readmission risks in patients hospitalized for COPD. We also examined separately the predictive ability for early readmissions (within 7 days after discharge) and late readmissions (8–30 days after discharge).
Discussion
By using nationally-representative sample of US Medicare beneficiaries, we found a potential benefit of adding social factors to the CMS-based reference model to improve the predictive ability for readmission within 30 days after COPD hospitalization. When we examined early and late readmissions separately, the predictive ability of optimized models were also significantly higher than that of the corresponding reference model. The decision curve analysis indicates the greater net benefit of optimized model over the reference model for thresholds between 15 and 20% of probability of 30-day readmission. Additionally, the contribution of predictive factors (e.g., cardiac comorbidity, poverty status) to the readmission risk differed between early and late readmissions. To the best of our knowledge, this is the first study that has investigated the incremental benefit of social factors on predicting the risk of readmissions – including early and late readmissions – in patients hospitalized for COPD. Given that the current one-size-fits-all approach (i.e., HRRP) has not been successful at lowering numbers of 30-day COPD readmissions [
49,
50], our findings demonstrating the heterogeneity of the 30-day readmission should help identify patients at high risk for readmission and inform the development of more targeted preventive interventions.
Despite the evidence suggesting the associations between social factors and readmission risks in various disease conditions (e.g., heart failure) [
33‐
37], most studies in the COPD population have focused on patient and hospital factors as predictor for readmission [
8,
9,
14‐
16]. Of these, few studies have used other factors to develop models predicting readmissions [
7,
14]. For example, in an US-based study of patients with COPD (age 40–64 years) using a commercial insurance database, Sharif et al. reported that the C-statistic of prediction model for 30-day readmissions improved to 0.72 after adding provider-level (e.g., medication prescriptions) and system-level (e.g., hospital length-of-stay) factors to their reference model that had the C-statistic of 0.68 [
14]. Studies using large datasets (e.g., NRD, Medicare data) have shown that some proxy social factors (i.e., quartiles of household income that are estimated by ZIP code, insurance status) were related to COPD readmissions [
13,
51]. However, the NRD and Medicare data do not include detailed social factors (e.g., marital status, actual income, number of children). In another single-centre retrospective study of 109 Canadian patients, while the prediction performance was not examined, Wong et al. reported that marital status (single) was a significant predictive factor for readmission following hospitalization for COPD [
52]. In non-COPD populations, (e.g., acute myocardial infarction, heart failure, pneumonia), the emerging evidence has suggested the importance of social factors to improve the prediction ability for readmission risks [
22,
37,
53,
54]. In addition, an earlier study examined the association of the lower income with the risk of acute exacerbation of COPD in patients aged 40–65 years with COPD [
55]. While the previous study had the different design and outcomes (COPD exacerbation vs. 30-day readmission), the importance of social factors (e.g., poverty) in the association of and prediction for the COPD morbidity risk is consistent. Consequently, our study corroborates these prior studies, and extends them by demonstrating, in a nationally-representative sample of Medicare beneficiaries, the incremental benefit of social factors on the CMS models to predict readmissions in patients with COPD.
While we found the additional benefit of social factors on predicting both early and late readmission, our findings also support the heterogeneity of the “30-day readmission” outcome. For example, mechanical ventilation use was a predictor for early readmissions but not for overall 30-day readmissions or late readmissions. By contrast, cardiac comorbidity and poverty status were significant predictors for late readmissions. The relative decrease in the effect of acute clinical factors and in-hospital factors (such as the use of mechanical ventilation) over time is clinically plausible. Indeed, the effects of inpatient care on the risk for readmissions diminish rapidly within a week after discharge [
18], with a recovery from symptoms of COPD exacerbation [
28]. In contrast, as a patient returns to the community, the relative importance of social factors (e.g., poverty) and chronic conditions (e.g., comorbid cardiac diseases) increases over time. While poverty status was associated with a lower risk of early readmission in this study, it is possible that poverty functioned as a barrier to accessing health care due to the costs of seeking health care, which include not only out-of-pocket spending on care but also transportation costs [
56]. Consequently, patients with poverty may have avoided having ambulatory health care and/or presenting to hospitals until the later period, which would reduce the rate of early readmission but increase the rate of late readmission. The latter finding can also be explained by the observations from earlier studies. For example, the literature has suggested that poverty is associated with the lack of health literacy affecting adherence to post-discharge instructions [
54,
57]. Without social support, some patients would not be able to cope with the post-hospital syndrome, a transient condition of generalized risk after hospitalization [
58,
59]. By contrast, positive social support provided by family members has been associated with improved quality of life [
60] and better pulmonary rehabilitation adherence [
61], resulting in reduced late readmissions. While marital status and number of children may be indicators for social support, living status (e.g., living alone) has been reported as an important prognostic factor. Indeed, an observational study reported that, patients with COPD who lived alone had lower levels of physical activity and lower rates of pulmonary rehabilitation participation compared with patients who live with other family members [
62]. Taken together, hospital readmission is a complex construct involving multiple factors, such as patient factors, inpatient and outpatient care, and social factors. Accordingly, to reduce readmissions and improve outcomes in patients with COPD, it is imperative to developing multifaceted strategies targeting each of underlying constructs – e.g., provision of high quality of inpatient, transition, and outpatient care, improvement in access to ambulatory care after hospital discharge, and social support in the community [
63,
64].
Potential limitations
Our study has several potential limitations. First, as with other studies using claims data, misclassification of hospitalizations is possible. However, the definitions using the CMS publicly-reported readmission measure [
11] have a high specificity and positive predictive value (both ≥90%) [
65]. Furthermore, the Medicare data are widely used for clinical research because the data are rigorously tested and considered accurate. Second, we did not account for several clinical information including the chronic severity of COPD since the MCBS and Medicare data do not contain such clinical information. Third, while the predictive ability of the optimized model was not high, the study objective was
not to develop clinical prediction models but, rather, to examine the incremental benefit of social factors on predicting the readmission risk in the population. Fourth, because of the relatively limited sample size and unavailability of unique dataset that has social factors, we used bootstrap samples to develop the model (rather than the use of an external sample). In addition, the limited sample size precluded us from estimating the detailed predictive contribution (e.g., the predictive contribution of specific cancers for the risk of readmission). Yet, formal validation of our study in separate populations is necessary. In addition, the grouping of 38 comorbidities into 10 categories according to the organ system may not yield the same predictive ability with the original model. Fifth, as we excluded patients who left the hospital against medical advice according to the CMS readmission measurement. This might cause selection bias. Lastly, the study sample comprised Medicare beneficiaries, and, therefore, our inference may not be generalizable to non-Medicare individuals with COPD or other non-US settings where social factors and their effects may differ [
66]. Nonetheless, our study population accounts for 70% of hospitalizations for COPD in the US [
4].
Conclusions
Based on nationally-representative sample of Medicare beneficiaries hospitalized for COPD, we found that the addition of social factors to the prediction model quantitatively improves the predictive ability for early and late readmissions. We also found that inpatient care factor (e.g., the use of mechanical ventilation) was a predictor for early readmissions while comorbidity and social factors (e.g., poverty) were predictors for late readmissions, suggesting that the readmission is a complex and heterogeneous construct. Despite the modest predictive ability for the clinical use, the improvement of predictive ability has important implications for researchers and policy makers. For researchers, our observations should facilitate further investigations into better identification of patients with COPD at high risk for readmissions. For policy-makers, our findings underscore the importance of continued efforts to develop and implement preventive strategies (e.g., high quality inpatient, transition, and outpatient management, as well as optimization of post-discharge environment) to reduce readmissions in this high-risk population.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.