INTRODUCTION

Hospital readmissions, defined as a subsequent admission of recently discharged patients are recognized as a significant contributor to health care costs and having a negative social impact [1,2,3]. Readmissions are more complex and almost twice as likely to result in a further readmission [4]. In Denmark, in 2017, 18% of discharges of people aged 67+ were followed by acute readmission within 30 days [5]. Although rates of readmission vary across countries and populations, the incidence of hospital readmissions in Denmark is comparable to readmission rates in large international cohort studies [1, 6, 7].

In Western countries, hospital readmission reduction programs have become a political priority. [8,9,10,11] As an outcome indicator, readmission intersects organizational boundaries within the health care system and identification of risk factors for readmission may be useful to distinguish between low- and high-risk groups and thereby facilitating allocation of clinical resources and tailored interventions across sectors [12, 13].

Comprehensive reviews of risk factors for readmission indicated that the underlying causes of readmission were multifaceted and found inconsistency and lack of clarity in the patterns of determinants pertaining to readmission in heterogeneous populations [14,15,16,17,18]. Similarly, prediction of risk of readmission seems to be a complex endeavor [19]. In a systematic review of 26 unique risk prediction models for hospital readmission, Kansagara et al. found that most models incorporated variables for medical comorbidity and use of prior medical services, while some studies have found different strengths of these determinants. Including determinants associated with overall health, illness severity, or social conditions improved the predictive ability [20].

The population of older people is largely heterogeneous in terms of health, illness, and care trajectories [12, 21,22,23]. Despite numerous studies on risk factors and generalized or disease-specific models for predicting hospital readmissions, there is a need to abstract risk factors and corroborate their applicability to a Danish population of older people. Hence, based on comprehensive information obtained from Danish population-based registers, the overall aim of this study was to identify high-risk patients, high-risk admissions, and high-risk circumstances within a heterogeneous, though selected, group of older people with general hospital admissions. The aim was to explore and identify patterns of generalized patient-level and admission-level risk factors of acute 30-day hospital readmission in a Danish nationwide cohort of patients aged 65+.

METHODS

Setting

Health care utilization strongly depends on the structures of the health care system and health policy [24, 25]. The Danish health care system is a universal, tax-financed health care service for the entire population. Every Danish citizen has a general practitioner (GP), and outside regular office hours, GPs on call serve the patients from central regional clinics providing telephone and face-to-face consultations. The GPs serve as gatekeepers for access to specialized care, and except in emergencies, the GPs make referrals to hospitals and specialists. Acute hospital admission is also available through a 24-h emergency call service [26].

Study Design

This study was conducted as a register-based cohort study analyzing individual-level linkable data obtained from ten population-based Danish nationwide registers.

Since 1968, all Danish citizens are assigned a unique identification number, which makes it possible to link information between nationwide registers and to follow care trajectories over time through different parts of the public sector [27]. Statistics Denmark (SD) offers remote access to the individual-level data necessary to perform research. Based on the availability of these data sources, we developed a database that comprised information for each individual on demographics, education, income, employment, housing, health, clinical, and administrative information on pre-hospital as well as in-hospital health care utilization and death. Linking these data sources enables researchers to track care trajectories and readmissions across settings in the Danish health care system. A recent publication has described the development of this database and cohort design in further details [28].

The study was registered under the North Denmark Region’s joint notification of health research (ID 2008-58-0028).

Study Population

In the Danish National Registry of Patients (DNRP), we identified all consecutive admissions of patients 65+, discharged from an inpatient hospital stay in a Danish hospital from 1 January 2007 to 30 September 2010 [29]. Inpatient stays in psychiatric, private hospitals and hospices were not included. Due to current policy incentives internationally to reduce the rates of hospital readmission among older people and for comparison with international research, we chose the chronological age 65 as cutoff.

The term index admission defined the initial inpatient stay in a series of admissions and determined the subsequent tracing of readmission in the follow-up period [28]. As patients could experience multiple admissions during the study period, each individual could contribute with several index admissions in the cohort.

Outcome

The outcome was a binary variable indicating whether the index admission was followed by a readmission and defined as the first acute, all-cause readmission within 30 days from index admission. The follow-up period was defined from date of discharge and extended for 30 days or until death, whichever came first.

Data Sources

The selection of candidate patient- and admission-level factors was inspired by previous research [6, 7, 16] and systematic reviews on risk factors and predictors of readmission [17, 18, 20] and afterwards discussed with experts according to clinical relevance. Potential risk factor variables comprised patient- and admission-level data, grouped into categories of socio-demographic, health status, and health care use as well as clinical and administrative determinants related to the pre-hospital and hospital setting. Using the personal identification number, clinical and administrative data on all inpatient admissions from the DNRP was linked with information from population-based registers and databases [25]. For each admission, we included variables to describe premorbid conditions and health care utilization 1 year prior to the index admission. In patients with multiple admissions, these variables were dynamic and varied for each admission.

Although the quality of information obtained from Danish nationwide registers has not been systematically validated, it is generally accepted that data quality as well as the completeness of the data in these registers is high [27].

For further details regarding the validation studies and quality of data sources used in this cohort study, see previous studies [27, 28, 30].

Baseline characteristics were described by mean and standard deviation (SD) for continuous variables and for categorical variables by numbers and percentages. As the majority of the non-negative variables were highly skewed, mean was preferred over median for reporting of continuous variables. We found high data completeness and few variables with missing observations. Following an examination of reasons for missing observations, missing data were included in the analyses as a category of “Unknown.”

Statistical modeling distinguishes between explanatory and predictive modeling [31, 32]. In this study, we combined the explanatory approach quantifying strength of associations and the predictive approach quantifying predictive ability.

Data was analyzed by univariable and multivariable logistic regression. Readmission was the dependent variable, explained by various groups of potential risk factor variables associated with hospital readmission. The unit of analysis was any patient’s index admission. For admission-level analysis, clustering was performed at individual level. Due to the need to examine non-linear associations, continuous variables entered the logistic regression model by restricted cubic splines with three knots [13]. Continuous variables with a high concentration in zero entered as categorical. For further details on how covariates entered the model, see Appendix (online).

The strength of the association would ideally be represented by the Wald test p value. However, due to the large sample size, many p values were calculated as zero. For each variable, we calculated a surrogate weight defined as the Wald test sum of squares divided by the degrees of freedom to quantify the associations [33]. As many of the variables were related and presumably correlated, we supplied the univariable weights with a multivariable weight, reflecting the importance of one variable when all others are considered known.

The quantification of the predictive ability measured by AUC using receiver operating characteristics (ROC) analysis was based on a multivariable and multilevel logistic regression including three models [34]. The overall model included all variables, the patient- and admission-level model included categories of patient-level variables or admission-level variables, and the final model included five subcategories of variables. Patient-level variables comprised various subcategories of demographic and socioeconomic variables, health characteristics, and previous health care utilization. Admission-level variables comprised subcategories of clinical and administrative variables related to the index admission. To adjust for bias due to overfitting and for internal validation, we chose a split-sample design for each category of variables when calculating AUCs [13, 35, 36]. For internal validation, the cohort was randomly divided into a two-third derivation cohort (n = 835,792) and a one-third validation cohort (n = 431,960) [13]. For external validation and trend analysis, the cohort was subsequently divided by making a non-random split according to the period of the index admission and divided into a two-third derivation (n = 845,165) and one-third validation cohort (n = 422,584). Bootstrap was used for producing confidence intervals (CIs).

All analyses were rerun for a population of acute-only index admissions. Statistical analyses were performed with STATA statistical software, version 14.0 (Stata Corporation, College Station, TX, USA).

Data Sharing Statement

Due to legal restrictions, no additional data is available.

RESULTS

We identified 1,267,752 index admissions in 479,854 patients aged 65+ in the entire cohort to be included in the analyses and in the acute-only cohort 908,696 index admissions in 395,398 patients. For socio-demographic characteristics, see Table 1. For health characteristics and health care utilization prior to the index admission and for clinical and administrative characteristics related to the index admission, see Table 2.

Table 1 Selected socio-demographic characteristics for the entire cohort (planned and acute index admissions) and the acute-only cohort (acute index admissions)
Full size table
Table 2 Selected health characteristics and health care use for the entire cohort (planned and acute index admissions) and the acute-only cohort (acute index admissions)
Full size table

In the following 30 days, 239,077 (18.9%) admissions were followed by acute hospital readmission, while 272,490 (21.5%) resulted in either acute readmission or death. For the acute-only cohort, the percentage of 30-day readmission as well as readmission or death were higher, with 199,466 (21.9%) and 230,303 (25.3%), respectively. For the entire cohort, the majority of index admissions were acute (n = 908,696, 71.7%), primarily for medical reasons (n = 755,489, 59.6%), with a mean length of stay of 6.1 days (SD 10.4). For the acute-only cohort, the percentage of admission due to medical reasons increased remarkably to 638,729 (70.3%) and so did the mean length of stay of 7.0 days (SD 11.1). Almost half of the index admissions comprised patients with at least one comorbid condition. The mean number of reimbursed prescriptions 6 months prior to admission was 8.3 (SD 5.1) and 8.7 (SD 5.2) for the acute-only cohort.

The results of the explanatory analyses of patient- and admission-level factors associated with readmission for the entire cohort as well as the acute-only cohort are summarized in Table 3. Based on the size of weights obtained from multivariable analyses, the ten factors most strongly associated with readmission for both cohorts were listed (see Table 4). For the entire cohort, acute index admission and a recent discharge were the two factors most strongly associated with increased risk of future acute readmission. Other factors comprised gender, employment status, number of reimbursed prescriptions, and the Charlson comorbidity index score, as well as the number of consultations with GPs or GPs on call. Aside from acute index admission, clinical and administrative factors related to the index admission were only weakly associated with risk of subsequent readmission. A similar pattern remained for the acute-only cohort. However, while variable “Number of previous 30-day readmissions” disappeared from the list, variables “Receiving home health care services” and “Number of ED visits” entered the list for the acute-only cohort.

Table 3 Associations between acute hospital readmission and patient- and admission-level factors calculated as univariable and multivariable weights, reported separately for the entire cohort and acute-only cohort
Full size table
Table 4 List of the ten most important patient- and admission-level factors associated with risk of acute readmission and acute readmission within 30 days
Full size table

The predictive ability of the three models was tested in the randomly derived validation cohort and results for the entire cohort are shown in Table 5 and for the acute-only cohort in Table 6.

Table 5 Predictive ability of all-cause 30-day readmission for the entire cohort—comparison of the logistic regression model between the overall model, patient- and admission-level model and the model based on subcategories of variables
Full size table
Table 6 Predictive ability of all-cause 30-day readmission for the acute-only cohort—comparison of the logistic regression model between the overall model, patient- and admission-level model and the model based on subcategories of variables
Full size table

In the randomly derived validation cohort, AUC was 0.709 (Table 5) and 0.691 for the acute-only cohort (Table 6). For the patient- and admission-level model as well as the models based on subcategories of variables, the predictive ability for both cohorts were slightly lower. Interestingly, the predictive ability for the subcategory “Health care use” was higher for the acute-only cohort than for the entire cohort, with AUCs of 0.650 and 0.647, respectively. With an AUC of 0.71 in the periodically derived validation cohort, we did not find any time trend.

DISCUSSION

In this study on patterns of risk factors and predictors for readmission in a cohort of Danish patients aged 65+, patient-level factors were the dominant contributors to the increased risk of acute readmission, indicating that patients at risk of readmission can be identified based on pre-hospital information obtainable and assessable at the time of the index admission.

We found that acute index admission and a recent discharge prior to the index admission were the factors most strongly associated with a subsequent 30-day readmission. Patients at increased risk of future readmissions suffered from chronic illnesses and consumed more drugs. In addition, they were frequent users of in- and outpatient health care services in the year prior to the index admission. These results correspond with previously performed studies demonstrating that complex patterns of care trajectories prior to admission and post-discharge indicate a higher risk of readmission [7, 12, 37].

Similar to previous studies, we found that male gender [17, 18] and socioeconomic factors [3, 4] are associated with a higher risk of readmission. Surprisingly, we did not identify any associations between readmission and previously identified risk factors such as age [4, 38], prolonged length of stay, and specific medical conditions related to the index admission [6, 37,38,39,40].

A number of studies found that risk of readmission increased with functional decline and poor overall health condition [37, 41]. Given the nature of data sources, our study lacked valid measures of functional and cognitive status [28]. Instead, we used information on home health care services for personal and practical care 1 month prior to the index episode as proxy for functional decline and dependency in daily living. Receiving home health care was only associated with readmission for the acute-only cohort. However, in both cohorts, the number of drugs used and frequent use of health care services prior to the index episode indicate that the population at increased risk of readmission suffered from multiple morbidities and experienced serious health challenges and complex care needs [41,42,43].

Corresponding to previous research, we found that underlying comorbidities rather than diagnosis at the index admission were associated with increased risk of readmission [44]. In contrast, a systematic review indicated that specific medical conditions or comorbidities influenced the risk of readmission; however, there was no consensus on which conditions and comorbidities [18]. These divergent findings might reflect differences in the study populations and clinical settings, or various definitions and lack of valid data to determine premorbid conditions and diagnoses [45,46,47]. Furthermore, we used a composite measure of the total weight of broad categories of relative heterogeneous diagnoses, which might have masked associations between readmission and specific comorbidities and medical conditions.

As risk of readmission was strongly associated with frequent pre-hospital use of in- and outpatient health care services this is most likely a strong indicator of more serious health conditions. As males have higher rates of hospitalization compared to women, some acute readmissions might be explained through gender differences in health and health behavior [48,49,50]. In this study, more than 70% of the index admissions were unplanned; this indicates that acute referral to hospital could be due to exacerbation and progression of illness and/or limited access to alternative acute, non-hospital services. The potential impact of patient preferences and gender differences and the impact of accessibility to alternatives to hospital readmission needs further investigation. Given the universal health insurance in Denmark, everyone has equal access to care. Thus, in countries and health care systems with non-universal insurances, the pre-hospital and access factors may be otherwise associated with risk of readmission than in the current study. Furthermore, risk for excess readmission and interventions to prevent 30-day readmission should be shared across sectors and support patient capacity for self-care [51].

The AUC of the overall model was 0.71 for acute 30-day readmission, which is considered acceptable [20] and comparable to similar studies [6, 7, 13, 52]. We found neither trend for season nor time. In models including patient- and admission-level variables, the predictive ability for the category of patient-level factors was higher than for admission-level factors. Finally, subcategories for health characteristics and health care use during the pre-index period accounted for the highest predictive ability among patient-level subcategories.

In this general population of older people, we would expect a huge amount of different pathways leading to readmission; each pathway represented with main effects and two- or maybe more-way interactions [12]. Including relevant interactions in the analysis might improve the predictive ability and would be feasible when analyzing specific subgroups and selected pathways, but omitted in this paper due to the general population at hand [31].

We opted for a general approach and included all patients aged 65+ admitted to a Danish hospital during a four-year period. Thus, the study population was heterogeneous and representative for an aging population. The large study size and the comprehensive database with the possibility to link demographic and social characteristics into patient pathways, including a complete medical history, comorbidity data, hospital data and follow-up data, and high data completeness, were major strengths of this study [28].

However, this study also exposed some weaknesses in register-based research. Due to the observational design, no firm causal explanations for readmission can be inferred. Furthermore, the interpretation, implications, and predictive ability within defined patient groups and specific clinical settings and health care systems might be limited [28, 30]. However, we did not intend to develop a predictive model centered on a specific patient population, clinical pathway, or clinical setting. Combining the explanatory and predictive approach, this analysis illustrates patterns of risk factors and predictors for readmission in a general Danish population of patients aged 65+. To improve the predictive ability, this general and complex model could be validated considering variable selection in different health care systems, clinical settings, and patient populations and improved by various subgroup analyses. For future research and longitudinal analyses, an update of the database would be preferable.

CONCLUSIONS

In a general population of older people, we found that pre-hospital factors rather than hospital factors account for the greatest risk of readmission and are the dominant contributors to predict acute all-cause 30-day readmission. Therefore, risk for excess readmission and approaches to prevent readmission should be shared across sectors.

PERSPECTIVES

The underlying causal relationships for readmission are multifaceted and simple explanatory and general predictive models based on broad categorically categories provide modest predictive value.