Risk Factors for Hospitalization Among Community-Dwelling Primary Care Older Patients: Development and Validation of a Predictive Model

Abstract

Background

Unplanned hospitalization often represents a costly and hazardous event for the older population.

Objectives

To develop and validate a predictive model for unplanned medical hospitalization from administrative data.

Research Design

Model development and validation.

Subjects

3919 patients aged ≥ 70 years who were followed for at least one year in primary care clinics of an academic medical center.

Measures

Risk factor data and the primary outcome of unplanned medical hospitalization were obtained from administrative data.

Results

Of 1932 patients in the development cohort, 299 (15%) were hospitalized during one year follow up. Five independent risk factors were identified in the preceding year: Deyo-Charlson comorbidity score ≥ 2 (adjusted relative risk [RR]=1.8, 95% confidence interval [CI] 1.4–2.2), any prior hospitalization (RR=1.8, 95% CI 1.5, 2.3), 6 or more primary care visits (RR=1.6, 95%, 95% CI 1.3–2.0), age ≥ 85 years (RR=1.4, 95% CI 1.1, 1.7), and unmarried status (RR=1.4, 95% CI 1.1, 1.7). A risk stratification system was created by adding 1 point for each factor present. Rates of hospitalization for the low- (0 factor), intermediate- (1–2 factors) and high-risk (≥ 3 factors) groups were 5%, 15%, and 34% (P<0.0001). The corresponding rates in the validation cohort, where 328/1987 (17%) were hospitalized, were 6%, 16%, and 36% (P<0.0001).

Conclusions

A predictive model based on administrative data has been successfully validated for prediction of unplanned hospitalization. This model will identify patients at high risk for hospitalization who may be candidates for preventive interventions.

INTRODUCTION

Hospitalization is common among older persons, and results in substantial morbidity and costs¹. The Medicare-age population experiences a rate of hospitalization of 17.5% or 6.4 million per year², comprising over 49% of acute hospital days³, and resulting in over $148.2 billion of annual Medicare expenditures⁴. Frequently, hospitalization can initiate the terminal downward spiral for an older person. Hazards of hospitalization for the elderly are myriad, and include delirium, falls, functional decline, institutionalization, and death^{1,5, 6}. The rate of serious iatrogenic hospital complications in older patients range from 29–38%^7–9, and are at least 3–5 fold higher in older compared with younger patients with similar diagnoses^{6, 10}. Moreover, these complications are often unrelated to admitting diagnoses, and more directly related to aspects of hospitalization itself, such as immobilization, dehydration, malnutrition, nosocomial infections, and psychoactive medications^1,11.

Previous studies examining risk factors for hospitalization^{12, 13} have focused on specific populations (e.g., nursing home, home care, frail elders)^13–18, specific diagnoses (e.g., anemia, depression)^19,20, and often use intensive data collection methods to identify risk factors, such as interview or functional assessment^21–27. The present study was intended to provide a predictive model that would estimate the risk for hospitalization using administrative data.

Specific objectives were to: (1) identify risk factors for acute, unplanned medical hospitalization from administrative data in a community-dwelling older population followed in primary care medical clinics; (2) develop a predictive model for hospitalization in an initial cohort; and (3) validate the model in a separate cohort. Our hypotheses were that factors such as older age and higher comorbidity would predict the risk of hospitalization. Our goal was to create a parsimonious model based on clinically important risk factors._ Ultimately, we hoped that we could design a predictive system that would be useful to proactively identify high-risk patients for preventive interventions or planning purposes.

METHODS

Study design

The study followed a prospective validation design. The predictive model was developed in an initial cohort, then tested in a validation cohort.

Study setting and sample

Beth Israel Deaconess Medical Center (BIDMC) is a 585-bed academic acute-care hospital affiliated with Harvard Medical School with over 40,000 admissions and 750,000 outpatient visits each year. Health Care Associates and Senior Health represent two primary care clinics at the BIDMC. Referrals to both clinics come primarily from the emergency department, hospital attendings and housestaff, subspecialist physicians, and from family or other patients. Patients followed in these clinics tend to receive the majority of their healthcare within the BIDMC system. The patients served are primarily fee-for-service, with <1% enrolled in Medicare managed care plans. For Health Care Associates, founded in the early 1970s, the staff includes 50% residents, and patients are followed for a median of 10 years or greater longitudinally. For Senior Health, founded in 2003, the staff includes 20% fellows, and patients have been followed for a median of 3–4 years longitudinally. These two clinics serve a diverse sample of elders in terms of educational and socioeconomic background, primary language and country of origin, and regional representation from throughout the greater Boston area.

The sample included all patients age ≥ 70 year who were seen at least once in primary care clinics (Health Care Associates or Senior Health) at the BIDMC during July 1, 2003 through June 30, 2004, and at least once during the subsequent year from July 1, 2004 through June 30, 2005. This criterion was selected to ensure that patients had been followed for an adequate time period to maximize the chances that relevant data would be available in the administrative database. All patients who died before June 30, 2005 were excluded, since they died prior to the period where the outcomes were ascertained (July 1, 2005 to June 30, 2006). A total sample of 3919 patients met these eligibility criteria. A split sample was created using a random selection process; half the sample (N=1932) was used as the initial cohort to develop the prediction model, and the remaining half (N= 1987) was used as the validation cohort. Once the sample was selected, all data were de-identified for all Health Insurance Portability and Accountability Act (HIPAA) identifiers by the BIDMC Information Technology department before analyses were performed. This study was conducted under a waiver of the Health Insurance Portability and Accountability Act from the BIDMC Institutional Review Board.

Risk factor variables

The data source was an administrative database called the BIDMC casemix_tsi data repository, which includes information on date and type of inpatient and outpatient visits, demographics, billing diagnoses, and charges. The diagnosis codes were those used for billing purposes on hospital discharges, clinic visits, and other claims (radiologic procedures, laboratory tests). These codes were compiled for each patient across the one year prediction period. Additional relational databases contain further details on each visit, such as laboratory results. Data from the electronic medical record were not obtained for this study or utilized for these analyses.

We identified potential risk factor variables based on the medical literature^{12,19,21–22, 25–26, 27–29} and expert opinion. Risk factor variables considered included age, gender, ethnicity, marital status, insurance type, Deyo-adapted Charlson comorbidity score²⁷, number of diagnoses, any previous hospitalization, any nursing home stay, number of primary care physician visits, receipt of any surgical procedure, receipt of any hemodialysis, and abnormal laboratory values. These variables were identified from all inpatient and outpatient visits during the prediction interval from 7/1/04 through 6/30/05.

Continuous variables were examined continuously and with cutpoints selected by previous studies, clinical judgment, and data distributions. Advanced age was defined as age ≥ 85 years representing the oldest-old and comprising the highest 20th percentile in the sample. Male gender, nonwhite race, Medicaid or uninsured status, and unmarried status have been used previously^21,28. The Deyo-adapted Charlson score ≥ 2 is a commonly used cutpoint to indicate high comorbidity^{30, 31}. An individual with a Deyo-Charlson score of ≥ 2^30,31 would have either 2 or more of these conditions [myocardial infarct, congestive heart failure, peripheral vascular disease, cerebrovascular disease, dementia, chronic pulmonary disease, connective tissue disease, ulcer disease, mild liver disease, or diabetes] or any one of these conditions [hemiplegia, moderate or severe renal disease, diabetes with end organ damage, any tumor, leukemia, lymphoma, moderate or severe liver disease, metastatic solid tumor, or AIDS]. Number of diagnoses with a cut-point of ≥ 10 was used, since it represented the sample median and was used previously³². However, this score was highly correlated with number of clinic visits (i.e., more diagnoses assigned at each visit), Pearson’s r = 0.70; thus, given the collinearity, we opted to use the Deyo-adapted Charlson score in the final model. Cut-points for laboratory values were selected based on previous studies^33–35. Missing laboratory values were assigned a “normal” value, as done previously^33,36, since most physicians do not order laboratories in the absence of a clinical indication.

Study outcome

The outcome was any unplanned medical hospitalization at the BIDMC during the one year period from 7/1/05 through 6/30/06. An unplanned hospitalization was coded as an urgent or emergent admission type in the administrative data.

Model development

Potential risk factor variables were first examined in bivariable analyses. The variables were selected by meeting all of the following criteria identified a priori: (1) prevalence of >15%; (2) relative risk (RR) > 1.5 for categorical outcomes; (3) P<0.05; and (4) clinical relevance. For closely related variables, one variable was selected according to best fulfillment of the a priori criteria. The panel of selected variables was further evaluated in multivariable analyses to yield the final independent predictor variables.

Statistical analysis and model validation

Baseline characteristics of the development and validation cohorts (Table 1) were compared using t-test statistics for continuous variables or chi-square statistics for categorical variables. For bivariable and multivariable analyses, logistic regression analysis was used. Model-based estimates were applied (i.e., log-binomial model) to calculate adjusted relative risks (RRs) and confidence intervals (CIs) from the parameter estimates and standard errors³⁷. To address missing data, we first verified that the missing data were not biased in their distribution between outcome groups. Then, we used SAS PROC MI to impute missing values for race and marital status. Sensitivity analyses were conducted to verify that imputation of missing values did not affect the results³⁸. Calibration of the model was evaluated with the Hosmer-Lemeshow test of goodness-of-fit statistic, and with analysis of residuals, including the inspection of residual plots (delta chi-square residual, delta deviance residual, and Cook’s distance).

Table 1.

Baseline Patient Characteristics in the Two Cohorts^*

Characteristics	Development Cohort†(N =1932)	Validation Cohort†(N = 1987)
Age, y (mean ± SD)	78.5 ± 6.0	78.7 ± 6.0
Male, No. (%)	737 (38.1)	774 (39.0)
Nonwhite, No. (%)	405 (23.1)	408 (22.2)
Married, No. (%)	887 (46.1)	918 (46.4)
Medicaid or uninsured, No. (%)	53 (2.7)	49 (2.5)
Enrolled from:
Health Care Associates, No. (%)	1567 (81.1)	1609 (81.0)
Senior Health, No. (%)	365 (18.9)	378 (19.0)
Deyo-Charlson score, mean ± SD	1.60 ± 1.99	1.59 ± 2.02
Score ≥ 2, No. (%)	759 (39.3)	745 (37.5)
Number of medical diagnoses, mean ± SD	13.6 ± 10.2	14.0 ± 11.1
Number ≥ 10, No. (%)	1099 (56.9)	1147 (57.7)
Any hospitalization in previous year, No. (%)	434 (22.5)	451 (22.7)
Any nursing home stay in previous year, No. (%)	89 (4.6)	119 (6.0)
Primary care visits in past year, mean ± SD	5.7 ± 4.1	5.8 ± 4.4
Primary care visits ≥ 6 in previous year, No. (%)	796 (41.2)	829 (41.7)

^*Missing values were present for some variables as follows. In the development cohort, race data (missing n=178), marital status (n=11); in the validation cohort, race data (n=151), marital status (n=9).

^†No statistically significant differences in baseline characteristics between the two cohorts.

The predictive model created in the initial cohort was subsequently tested in the validation cohort. Model performance in both cohorts was assessed using the C statistic, approximating the area under a receiver-operating characteristic curve³⁹, and ranging from 0.5 (no discrimination above chance) to 1.0 (perfect discrimination). A risk stratification system was developed by assigning one point to each of the final risk factors present. Based on distributional characteristics (i.e., groups with similar hospitalization rates were combined), we created 3 strata representing low-, intermediate-, and high-risk groups. Overall chi-square and Cochran-Armitage trend tests were used to compare hospitalization rates by risk strata in both cohorts. All analyses were conducted using SAS Version 9.1 (SAS Institute, Cary, NC).

RESULTS

Baseline characteristics of the initial cohort appear in Table 1. Of 1932 patients, 299 (15.4%) had at least 1 urgent or emergent hospitalization during the outcome interval. The mean length of hospital stay was 8.3 ± 10.4 days, median 4.0 days (range, 1.0–93 days).

Development of the predictive model

The 16 candidate risk factor variables considered for the predictive model appear in Table 2. Using the a priori selection criteria above, these variables were narrowed based on bivariable results. Five potential risk factors were selected and all were retained in the final predictive model (Table 3). These included Deyo-Charlson comorbidity score ≥ 2, any hospitalization in the prior year, 6 or more primary care visits in the previous year, age ≥ 85 years, and unmarried status. These factors were entered into a single predictive model to provide an overall estimate of the independent contribution of each variable to subsequent hospitalization. While any abnormal laboratory result met selection criteria for inclusion, this variable was found to be collinear with any hospitalization in the prior year (Pearson r=0.45, P<0.001), and may not be readily available in administrative databases at all hospitals. In addition, the inclusion of this variable did not substantively improve the performance of the model. For these reasons, this variable was not included in the final model.

Table 2.

Variables Considered as Risk Factors for Hospitalization in the Development Cohort^*

(N=1932)
		Hospitalization When Factor
Factor	Prev(%)	Present No. (%),	Absent, No. (%)	RR (95% CI)
Age ≥ 85 years	19.6	87/378 (23)	212/1554 (14)	1.7 (1.4, 2.1)
Male gender	38.2	116/737(16)	183/1195(15)	1.0 (0.8, 1.3)
Nonwhite race	23.1	67/405(17)	207/1349(15)	1.1 (0.8, 1.4)
Unmarried	53.9	195/1034(19)	103/887(12)	1.6 (1.3, 2.0)
Medicaid or uninsured	2.7	6/53 (11)	293/1879 (16)	0.7 (0.3, 1.6)
Deyo-Charlson score ≥ 2	39.3	184/759(24)	115/1173(10)	2.5 (2.0, 3.1)
Number of medical diagnoses ≥ 10	56.9	244/1099(22)	55/833(7)	3.4 (2.5, 4.5)
Any hospitalization in previous year	22.5	130/434(30)	169/1498(11)	2.7 (2.2, 3.3)
Any nursing home stay in previous year	4.6	33/89(37)	266/1843(14)	2.6 (1.9, 3.4)
Primary care visits ≥ 6 in previous year	41.2	180/796(23)	119/1136(10)	2.2 (1.7, 2.7)
Any surgery in previous year	4.4	18/85(21)	281/1847(15)	1.4 (0.9, 2.1)
Hemodialysis in previous year	0.4	5/7(71)	294/1925(15)	4.7 (2.9, 7.6)
Hematocrit < 30 ml/dL	9.8	72/190(38)	227/1742(13)	2.9 (2.3, 3.6)
Serum albumin < 3.0 mg/dL	2.5	18/49(37)	281/1883(15)	2.5 (1.7, 3.6)
Serum creatinine >1.5 mg/dL	12.5	72/241(30)	227/1691(13)	2.2 (1.8, 2.8)
Any abnormal laboratory result†	18.8	116/363(32)	183/1569(12)	2.7 (2.2, 3.4)

^*Prev=prevalence; RR=relative risk; CI=confidence interval. Missing values were present for some variables as follows: race data (missing n=178), marital status (n=11).

^†Abnormal laboratory results included hematocrit < 30 ml/dL, serum albumin < 3.0 mg/dL, or serum creatinine > 1.5 mg/dL in previous year. The most abnormal result during the previous year was used to score this variable.

Table 3.

Independent Risk Factors for Hospitalization^*

(N = 1932)
Risk Factor	Adjusted RR† (95% CI)
Deyo-Charlson score ≥ 2 (n = 759)	1.8 (1.4, 2.2)
Any hospitalization in previous year (n = 434)	1.8 (1.5, 2.3)
Primary care visits ≥ 6 in previous year (n = 796)	1.6 (1.3, 2.0)
Age ≥ 85 (n = 378)	1.4 (1.1, 1.7)
Unmarried (n = 1040 )‡	1.4 (1.1, 1.7)

^*RR=relative risk; CI=confidence interval. n=number of patients with the risk factor present.

^†Adjusted relative risk derived from multivariable models in PROC GENMOD.

^‡_Includes imputed values for 11 missing subjects.

Performance of the predictive model

Development cohort

The final model generated a C-statistic of 0.72, indicating good prediction above chance³⁹. The Hosmer-Lemeshow goodness-of-fit test chi-square= 4.7 (7 degrees of freedom, P=0.70) indicates that the model fits the data well. Inspection of residuals revealed 4 influential outlier data points; the results were not appreciably different with exclusion of these patients; thus, all patients were retained in the final model. A risk stratification system was created by assigning 1 point to each of the final risk factors present during the prediction interval. Three risk groups were created: low risk (0 factors), intermediate risk (1–2 factors), and high-risk (3–5 factors) groups. Rates of hospitalization increased from 5% to 15% to 34% across risk groups (χ² trend = 115.6, P <0.0001), with the corresponding risk gradient extending from 1.0 (referent) to 3.1 to 7.0 (Table 4).

Table 4.

Performance of the Predictive Model in the Two Cohorts^*

		Development Cohort (N=1932)		Validation Cohort (N=1987)
Risk Group	Number of risk factors	Hospitalization n/N (%)	RR (95% CI)	Hospitalization n/N (%)	RR (95% CI)
Low	0	22/461 (4.8)†	1.0 (referent)	30/473 (6.3)‡	1.0 (referent)
Intermediate	1 – 2	167/1143 (14.6)†	3.1 (2.0, 4.8)	184/1191 (15.5)‡	2.4 (1.7, 3.6)
High	≥ 3	110/328 (33.5)†	7.0 (4.5, 11.1)	115/323 (35.6)‡	5.6 (3.8, 8.4)

^*RR=relative risk; CI=confidence interval. Includes imputed values for marital status for 11 subjects in development cohort and 9 subjects in validation cohort.

^†χ²_trend = 115.6, P<0.0001

^‡χ²_trend = 1151.3, P<0.0001

Validation cohort

The validation cohort was similar to the development cohort in all baseline characteristics (Table 1). Of 1987 patients, 329 (16.6%) had an urgent or emergent hospitalization during the outcome interval. Mean length of hospital stay was 8.0 ± 10.1 days, median 5.0 days (range, 1.0–92 days).

The predictive model yielded a C-statistic of 0.73, demonstrating good prediction above chance. The Hosmer-Lemeshow goodness-of-fit test chi-square= 6.32 (6 degrees of freedom, P=0.39) indicates that the model fits the data well. Applying the risk stratification system (Table 4), rates of hospitalization increased from 6% to 16% to 36% across groups (χ² trend = 111.3, P <0.0001), representing a 6-fold increased risk of hospitalization between low- and high-risk groups.

DISCUSSION

We developed and successfully validated a predictive model for acute medical hospitalization in a primary care community-dwelling cohort. The overall rates of hospitalization in our cohorts (15–17%) are directly comparable to the 17% rate for Medicare patients nationally². Five factors that independently contribute to risk of hospitalization were identified in the preceding year: Deyo-Charlson comorbidity score ≥ 2, any prior hospitalization, ≥ 6 primary care visits, age ≥ 85 years, and unmarried status. Thus, we confirmed our hypothesis that both demographic and comorbidity factors would contribute to the risk of hospitalization. In patients identified as high risk, the rate of hospitalization was 34–36% in the following year.

This study serves to confirm risk factors identified in previous studies^{12, 19, 21, 25–28}, yet allows us to extend this work to facilitate risk stratification from administrative data, which may enhance identification of high risk patients. While laboratory results were predictive, these results may not be readily available in all administrative databases, and thus, were not included in the final model. Similar to previous studies, our work again confirms the importance of comorbidity or casemix in prediction of future resource utilization ^40–41. Similar to the Probability of Repeated Admission (Pra) questionnaire^22,25,26, our study identified prior hospitalization and more than 6 physician visits in the previous year as important risk factors. Moreover, the unmarried variable may also reflect the lack of an informal caregiver, similar to the Pra questionnaire. Our goal in this study was to create a parsimonious model, based on clinically important risk factors, and not primarily to maximize prediction. Thus, we did not choose to include all possible variables which may have improved prediction, variables with low prevalence in the sample, or variables which were highly collinear. Prediction may have been improved with inclusion of more variables, however, we wanted to develop a model based on a smaller number of clinically important risk factors.

Strengths of this study include the availability of a large sample of primary care patients, who are representative of the older population in one metropolitan area. The sample is diverse in terms of race, ethnicity, and socioeconomic status. The development and validation of the predictive model in two separate groups (randomized split sample) is another important strength, which lends support for the robustness and generalizability of the model⁴². However, future studies will be needed to verify transportability of the model to other populations.

Several caveats deserve comment. First, given the nature of administrative data, some data were missing for risk factors such as race and marital status; however, sensitivity analyses reveal that imputation of these missing values did not affect the results. Second, we were unable to examine many risk factors identified in previous studies (e.g., self-rated health, physical functioning). While including these factors may have led to better prediction, this would contradict our goal of identifying factors via administrative data. Third, we did not examine specific diagnoses, which will be an important area for future research. Moreover given the nature of our administrative data, we were unable to distinguish which diagnoses codes were for “rule-out” conditions, rather than existing diagnoses. Fourth, we required that patients be followed in their primary care clinics for at least one year to ensure that adequate information in our administrative data. Thus, while prediction would be useful for primary care patients followed longitudinally, applying this model would be more difficult for patients who are new to a healthcare system or followed elsewhere and may affect the generalizability of the model. Fifth, since this was a single-site study, we may have missed hospitalizations that occurred at other hospitals; however, it is important to note that the vast majority of patients receiving their primary care at the BIDMC also choose to be hospitalized there. Finally, our model would not apply to prediction of all types of hospitalization (e.g., surgical, gynecological), since we restricted ourselves to only medical hospitalizations. Generalizability may have been limited by inclusion of only medical hospitalizations at a single site, and future studies will be needed to verify effectiveness in other populations and for other types of hospitalizations.

While the identified risk factors are not necessarily directly amenable to intervention, they do serve to identify a highly vulnerable group that is old, frail, with considerable comorbidity, and high health care utilization. Thus, this predictive model may serve the important role of risk stratification, to identify a group at high risk for future hospitalization who might be candidates for preventive interventions. Such interventions might span a broad array of interventions, such as care coordination, intensive case management, preventive home visits, or optimization of transitional care^43–47. The high-cost area of unplanned hospitalization represents an area of particular importance for cost containment. Moreover, given the high impact and substantial rates of adverse outcomes associated with hospitalization, targeting the identified high risk group for preventive interventions makes sense from both clinical and policy perspectives.