| HOME | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|---|
| ||||||||||||||||||||||||||||||||
Correspondence: Address correspondence to Teresa M. Damush, PhD, Regenstrief Institute for Health Care (RG6), 1050 Wishard Blvd., Indianapolis, IN 46202. E-mail: tdamush{at}regenstrief.org
 |
Abstract |
|---|
 TOP Abstract MethodsResultsDiscussionReferences |
|---|
75 years of age or were 50 years of age with severe disease. As a secondary analysis, we assessed patient characteristics at baseline involving five domains of health, including utilization and satisfaction. We followed participants for 12 months and recorded the occurrence of nonelective hospitalization within the study period. We developed a multivariate model using logistic regression to predict this outcome. Results: The following patient characteristics independently predicted an increased risk for nonelective hospitalization: having the diagnosis of congestive heart failure, diabetes mellitus, or anemia; and having more medications prescribed, having a lower body mass index, and having more emergency department visits during the previous year. Better physical functioning reduced the risk of hospitalization. Implications: Moderate accuracy of a prediction model (0.73) was observed. In addition to focusing on patients with chronic disease, helping them maintain physical functioning may help reduce nonelective hospitalization.
Key Words: Ambulatory care • Hospitalization • Risk factors
Although the prediction models have significantly contributed to our understanding of patients at high risk (Bazargan, Bazargan, & Baker, 1998; Boult et al., 1993; Coleman et al., 1998; Culler, Parchman, & Przbylski, 1998; Freedman, Beck, Robertson, Calonge, & Gade, 1996; Komaromy et al., 1996; Miller et al., 1998; Parkerson, Broadhead, & Tse, 1995; Roos, 1989; Shelton, Sager, & Schrader, 2000; Smith et al., 1983; Weinberger et al., 1986; Wolinsky, Culler, Callahan, & Johnson, 1994), the accuracy of the prediction is moderate. That is, when reported, the area under the receiver operating characteristic (ROC) curve ranges from 0.61 to 0.74 (Boult et al., 1993; Coleman et al., 1998; Freedman et al., 1996; Parkerson et al., 1995; Shelton et al., 2000; Wolinsky et al., 1994). We reviewed these models to determine if there might be ways to improve the accuracy of prediction. The number of patient risk factors considered in previous studies ranged from 14 to 40 (Bazargan et al., 1998; Boult et al., 1993; Coleman et al., 1998; Culler et al., 1998; Freedman et al., 1996; Komaromy et al., 1996; Miller et al., 1998; Parkerson et al., 1995; Roos, 1989; Shelton et al., 2000; Smith et al., 1983; Weinberger et al., 1986; Wolinsky et al., 1994). The risk factors included five domains: demographics or social support (Bazargan et al., 1998; Boult et al., 1993; Coleman et al., 1998; Culler et al., 1998; Freedman et al., 1996; Komaromy et al., 1996; Miller et al., 1998; Parkerson et al., 1995; Roos, 1989; Shelton et al., 2000; Smith et al., 1983; Weinberger et al., 1986; Wolinsky et al., 1994), disease or disease severity (Bazargan et al., 1998; Boult et al., 1993; Coleman et al., 1998; Culler et al., 1998; Freedman et al., 1996; Miller et al., 1998; Parkerson et al., 1995; Roos, 1989; Shelton et al., 2000; Wolinsky et al., 1994), physical examination and laboratory tests (Miller et al., 1998; Smith et al., 1983), health-related quality of life or physical functioning (Bazargan et al., 1998; Boult et al., 1993; Coleman et al., 1998; Culler et al., 1998; Parkerson et al., 1995; Roos, 1989; Shelton et al., 2000; Weinberger et al., 1986; Wolinsky et al., 1994), and previous resource utilization (Boult et al., 1993; Coleman et al., 1998; Freedman et al., 1996; Roos, 1989; Shelton et al., 2000; Smith et al., 1983; Wolinsky et al., 1994). None of the previous models included all domains, and this may have reduced the potential for achieving higher accuracy. Another potential limitation of previous models is that the outcome for most models was all hospitalizations rather than nonelective hospitalizations (Culler et al., 1998; Smith et al., 1983). Because elective hospitalization for procedures is less likely to be related to risk factors for poor health outcomes, this could contribute to lowering the accuracy of prediction.
In this study, we attempted to improve the accuracy of a predictive model by considering (a) a large number of potential risk factors; (b) patient characteristics in all five domains; (c) variables not previously considered that could affect outcomes, such as patient satisfaction, which has been associated with compliance (Sherbourne, Hays, Ordway, DiMatteo, & Kravitz, 1992); and (d) nonelective hospitalization as the outcome variable. We also wished to identify patients living in the inner city who could be targeted for an intervention to reduce nonelective hospitalization. A recent randomized controlled trial has made available a unique database of frail and older adult patients, of whom 56% were African American, containing multiple patient biomedical, functional, and psychosocial characteristics in primary care (Dexter et al., 1998; Gramelspacher, Zhou, Hanna, & Tierney, 1997).
|
Methods |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
Patients who were at risk for acute deterioration were included if they were 75 years old or older, or if they were 50 years old or older with one of the following chronic conditions: (a) coronary heart disease, (b) congestive heart failure, (c) chronic obstructive pulmonary disease, (d) cancer other than nonmelanomatous skin cancer, (e) cerebrovascular disease, (f) renal insufficiency, or (g) severe liver disease. A computer program identified eligible patients among those with scheduled appointments at the General Medicine Clinic (GMC) by using problem lists created by physicians and test results stored in the Regenstrief Medical Record System (RMRS; McDonald, Tierney, Overhage, Martin, & Wilson, 1992).
In the clinic waiting room, research assistants approached eligible patients who kept appointments at the GMC and interviewed those interested in participation. Patients were excluded if they were residents in a nursing home or prison, if they were hearing impaired, or if they did not speak English as noted in their medical record or revealed at baseline (eight patients were excluded because they did not speak English). They were also excluded if their score on the Pfeiffer Mental Status Questionnaire (Pfeiffer, 1975) indicated cognitive dysfunction.
As part of the original randomized trial, 30 half-day practice sessions associated with four teams in the GMC were each randomized to one of the following four study groups: (a) control (usual care), (b) received computer reminders to discuss advance directives, (c) received computer reminders to discuss health care representatives, or (d) received computer reminders to discuss advance directives and health care representatives with the patient at the time of the visit. Physicians associated with these half-day sessions were assigned to the corresponding study group (Dexter et al., 1998; Gramelspacher et al., 1997). Patients were closed out by interview between 11 and 16 months after enrollment, depending on their scheduled clinic visit. Of 1,049 patients enrolled during a primary care visit, we excluded 8 who had only 1 day or less of follow-up, leaving an inception cohort of 1,041. Further details regarding patient recruitment are published elsewhere (Gramelspacher et al., 1997).
Independent Variables
We obtained patient information by interview and by query of the RMRS. We assessed independent variables from five domains of patient risk factors that have previously been studied: demographics or social support; disease or disease severity; physical examination and laboratory tests; health-related quality of life (HRQOL); and previous resource utilization or compliance. In addition, we added another domain of patient satisfaction that may be related to nonelective hospitalization.
Demographic variables included age and education in years; gender; and race or ethnicity. Perceived poverty status was measured in three categories: financial resources are sufficient to make one comfortable, coded as 3; have just enough to make ends meet, coded as 2; and do not have enough to make ends meet, coded as 1. We also measured social variables and included dichotomous categories of Medicare or Medicaid coverage, marital status, living arrangements, and caregiver support.
Disease status was assessed from the seven conditions needed for eligibility, the total number of medications, and the presence of diabetes mellitus. In addition, each patient was assigned a measure of disease severity from 1 to 51 Ambulatory Care Groups (ACGs; Weiner, Starfield, Steinwachs, & Mumford, 1991) The ACG measure is a rigorous measure of outpatient comorbidity that has been validated in a number of settings, including inner-city Medicaid populations (Tierney, Harris, Bates, Culler, & Wolinsky, 1995). Because our study sample consisted of older and sicker adults, patients fell into a limited number of ACGs. Therefore, the groups were collapsed into 6 groups with similar severities of illness (Dexter et al., 1998). For physical examination and laboratory tests, we obtained measures from the RMRS for computing the body mass index (BMI) and the most recent values for hemoglobin, blood glucose, serum creatinine, serum albumin, and serum potassium.
We assessed health-related quality of life during the 4 weeks prior to enrollment by using the RAND Medical Outcomes Study (MOS) short-form health survey (SF-36), including all eight subscales: general health perceptions, physical function, mental health, social function, vitality, bodily pain, emotional role, and physical role (Ware & Sherbourne, 1992). We transformed all scale scores to a 0–100 scale in which a higher score indicates better health.
Previous resource utilization, including the frequency of hospitalizations, emergency department visits, and total GMC visits (scheduled plus walk-in visits) during the previous year, was obtained through the RMRS (McDonald et al., 1992).
Patient compliance was also assessed and reported by four categories. First, the total number of GMC visits including walk-in visits during the prior year was reported. Second, compliance was reported by the percentage of total GMC visits that were walk in (without an appointment) during the prior year. Third, no-show appointments during the prior year were expressed as a percentage of total scheduled GMC visits. Fourth, we also computed the percentage of GMC visits that were made to the enrollment physician in the past year (i.e., continuity of care).
Patient satisfaction was measured by using two questionnaires. The first included 10 items from the American Board of Internal Medicine (ABIM) instrument (Webster, 1988), which evaluates the relationship between patients and physicians in internal medicine residencies. The second questionnaire consisted of 5 items from the RAND MOS-VSQ (Rubin et al., 1993) and evaluated satisfaction with the current medical visit. Response options for both the ABIM and MOS questionnaires are (a) excellent, (b) very good, (c) good, (d) fair, and (e) poor. The two instruments measure different but complementary aspects of outpatient care (Stump, Dexter, Tierney, & Wolinsky, 1995).
Dependent Variable
To predict health care utilization, we followed clinical outcomes over the next 12 months. This time frame allowed us to measure the use of health services that were rendered relatively close to the enrollment visit. The primary outcome variable was the presence of a nonelective admission to the hospital within 12 months after enrollment. We identified admissions by using the RMRS (McDonald et al., 1992). Repeated prior internal audits and studies have shown that more than 90% of primary care patients in this setting receive all of their care at this facility (Smith et al., 1983; Wallihan, Stump & Callahan, 1999). In addition, we reviewed the records of all participants and verified their continued contact with the primary care practice. We reviewed discharge summaries of hospital admissions during the 12 months and classified admissions as elective or nonelective. We defined elective admissions as nonurgent, scheduled events for which delayed admission would present negligible risk to the patients, such as elective orthopedic surgery, hernia repair, and elective cardiac catheterization. All other procedures requiring hospitalization were considered nonelective (Smith et al., 1996). We focused on nonelective admissions instead of all admissions in order to understand variables that placed patients at risk for nonelective hospitalization.
Statistical Analysis
To determine if continuous and dichotomous independent variables were associated with the presence of an event (i.e., nonelective hospitalization), we performed t tests and chi-square tests, respectively. All independent variables within each domain with a p value less than.20 from univariate comparisons were considered eligible for the multivariate model selection process. We selected a final multivariate model by using the stepwise procedure at a significance level of.15 for entry into the model and a significance level of.05 to remain in the model. We used logistic regression to model a nonelective admission occurring within 12 months. In addition, we conducted bootstrap validation by using 1,000 samples to estimate average regression parameters and standard deviations of the final model (Mick & Ratain, 1994). We also performed all analyses by using survival analysis techniques to model time to first nonelective admission as the outcome. Results were similar; therefore, we present only results from the logistical regression analysis.
|
Results |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
Univariate Analyses of Nonelective Hospital Admission
Of the 1,041 patients, 216 (20.7%) had one or more nonelective hospital admissions during the 12 months of follow-up. The total number of nonelective admissions was 390, with a mean of 1.81 per patient per 12 months.
Tables 1 and 2 present univariate comparisons of dichotomous and continuous independent predictors across nonelective admission status within 12 months. In the domain of demographics and social support (Table 1), only having Medicare had a significance level of.05 or smaller. In the domain of disease or disease severity, the presence of coronary artery disease, stroke or transient ischemic attacks, congestive heart failure, chronic obstructive pulmonary disease, chronic renal insufficiency, and diabetes mellitus were all more prevalent among those hospitalized. In addition, those nonelectively hospitalized had a greater number of medications at baseline. Finally, there were fewer patients considered to be in the chronic medical, stable group (ACG 9; Weiner et al., 1991) in the hospitalized group than other ACGs.
|
|
Multivariable Model for Nonelective Hospital Admission
The model for predicting a nonelective admission during the subsequent 12 months is shown in Table 3. Notably, no variables from the domain of demographics and social support or from the domain of patient satisfaction entered the model. Independent and significant predictors of a nonelective admission during the subsequent 12 months included the following: presence of congestive heart failure, diabetes mellitus, and anemia; the number of medications at baseline; having a lower BMI; and emergency department visits in the prior year. The independent variable inversely associated with a nonelective admission during the ensuing 12 months was better physical functioning.
|
|
Discussion |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
Among the risk factors identified in this study, most are supported by previous studies in the literature. Specifically, the presence of the diagnosis of congestive heart failure (Bazargan et al., 1998; Coleman et al., 1998; Freedman et al., 1996; Miller et al., 1998) or diabetes mellitus (Boult et al., 1993; Coleman et al., 1998; Culler et al., 1998; Freedman et al., 1996; Miller et al., 1998), higher number of medications (Shelton et al., 2000), and a history of previous utilization (Boult et al., 1993; Coleman et al., 1998; Shelton et al., 2000; Wolinsky et al., 1994) have been reported as predictive of subsequent hospitalizations. Moreover, older adult physical functioning has been reported as a predictor of subsequent disability (Guralnik, Ferrucci, Simonsick, Saliver, & Wallace, 1995). Thus, interventions that target physical functioning among frail and older adults may reduce the likelihood of undergoing nonelective hospitalization. In addition, we found that having a lower BMI was predictive of nonelective hospital admissions. Previous studies have reported lower BMI as a significant predictor of mortality among both community-dwelling older adults (Landi et al., 1999) and seriously ill, hospitalized patients (Galanos et al., 1997). Loss of body mass appears to be an indicator of future health status. Thus, interventions focusing on nutrition may be beneficial for the ill and elderly populations.
Our identifying these risk factors lends further support to their importance as predictor variables for nonelective hospitalization among frail and older adult patients. In addition, the rate of prior hospitalizations of 0.38 per year is also consistent with a rate for frail and elderly persons; that is, it is a higher rate than that for the civilian noninstitutionalized population (Adams & Marano, 1995).
Our other objective was to identify patients for whom interventions should target to reduce these outcomes. Of the final prediction model, six characteristics (presence of congestive heart failure, diabetes mellitus, and anemia; number of medications; BMI; and emergency department visits in the prior year) are usually available in the medical record. Only physical functioning from the SF-36 is not routinely assessed. Thus, health care personnel should be able to identify patients who could be targeted for an intervention designed to reduce poor health outcomes and the likelihood for nonelective hospitalization among frail and older adult patients in ambulatory care.
The results of this study are limited. During the follow-up period, other events could have influenced patient outcomes that may not have been included in the model. This is a limitation of other models as well. Second, the generalizability of this study may be limited to older or middle-aged adults with severe disease who are receiving health care at an urban hospital affiliated with a university. Nonelective hospital admission rates may differ in other health care settings.
It is possible that we are reaching the limits in accuracy in attempting to predict a diverse outcome such as nonelective hospitalization in a diverse population of frail and older adults. The results of this study and of the previous models suggest that further efforts in research in predictive models might be better directed to subpopulations, for example, selecting cohorts of frail and older adults with the single diagnosis of congestive heart failure, or diabetes mellitus. This would allow examination of more specific clinical variables such as measures of left ventricular ejection fractions, glycosylated hemoglobin levels, and the quality of patient care. The difficulty of using these more specific clinical variables is that they are frequently not available in medical records because the tests were not ordered or not recorded, or the quality was not observed. However, in subpopulations of frail and older adults with these conditions, these tests and observations are more likely to be found, and their absence could reflect lower quality of care (Aronow, 1994). Developing predictive models for subpopulations is not a new idea (Wolinsky, Overhage, Stump, Lubitz, & Smith, 1997), but increased research efforts in this direction could contribute to both improving the accuracy of prediction and providing guidance for potential interventions.
The results of the current study provide a tool to identify patients at risk for nonelective hospitalizations among the frail and older adult population. Increased research efforts in subpopulations of persons with specific diseases may improve accuracy and provide more specific directions for interventions.
|
Footnotes |
|---|
1 Regenstrief Institute for Health Care, Indianapolis, IN.
2 Indiana University Center for Aging Research, Indianapolis.
3 Department of Medicine, Indiana University School of Medicine, Indianapolis.
4 Richard L. Roudebush Veterans Affairs Medical Center, Indianapolis, IN.
Decision Editor: Laurence G. Branch, PhD
Received for publication July 8, 2002. Accepted for publication October 17, 2002.
|
References |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
This article has been cited by other articles:
 |
|
 |
|
  D. Lyon, G. A Lancaster, S. Taylor, C. Dowrick, and H. Chellaswamy Predicting the likelihood of emergency admission to hospital of older people: development and validation of the Emergency Admission Risk Likelihood Index (EARLI) Fam. Pract., April 1, 2007; 24(2): 158 - 167. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||
| HOME | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|---|
