INTRODUCTION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Annual epidemics of respiratory syncytial virus (RSV) infection in the United States are responsible for approximately 100,000 hospital admissions, primarily in infants. These epidemics result in significant morbidity and major health care expenditure. Since the major epidemiologic impact of RSV-associated illness was first described more than three decades ago, control of yearly epidemics has been attempted using several biologically different approaches. Initially, in the 1960s, control of epidemics with formalin-inactivated vaccines was examined; however, this initiative was unsuccessful, resulting in more severe illness and mortality in vaccinated patients than in control patients (1). In 1985, the first antiviral agent, ribavirin, was FDA-approved for use in hospitalized high risk pediatric patients with RSV infection. During the next decade improvement in important outcomes such as duration of hospitalization, need for mechanical ventilation, and mortality were not consistently demonstrated (2) and resulted in an American Academy of Pediatrics expert committee statement reducing the strength of its prior ribavirin recommendation (6).

Recent FDA-approved therapies to ameliorate the impact of RSV epidemics have been by way of a pooled adult immune globulin product (RespiGam) and a monoclonal immune globulin product (Synagis) (7, 8). The pivotal trials conducted in infants with bronchopulmonary dysplasia (BPD) and/or prematurity demonstrated reduced hospital admission rate and hospital days among treated patients (7, 8). No positive impact on severity of RSV disease was observed based on decreased mortality or reduced need for or duration of mechanical ventilation (7, 8). Further study will be required to determine the place of immune globulin therapies in these and in other pediatric high risk conditions as well as cost effectiveness in actual clinical practice (9, 10).

A major problem encountered in past RSV antiviral and immunoglobulin therapeutic studies has been inadequate sample size to allow adequate control of multiple possible confounding variables such as comorbid conditions (11). It is currently known that patients with comorbid conditions such as congenital heart disease, chronic lung conditions, immune compromised states, nosocomial infection, prematurity, and failure to thrive status are at increased risk for more severe illness than are previously well patients in terms of length of hospital stay (4, 21). In past studies patients were commonly entered into clinical trials based on meeting specific diagnostic conditions such as the presence of chronic heart disease (CHD), BPD, or prematurity and being within a specified age range (i.e., younger than 2 yr of age) (7, 15, 26). However, recent population-based cohort studies that have utilized multivariate statistical methods have demonstrated that the independent variable age (or weight) is associated with hospital duration on a continuous scale and that the presence of more than one risk variable increases the duration of hospitalization above that of any single variable (4, 27). Assigning a severity of illness assessment to individual patients with RSV in treatment groups who are represented by many combinations of risk factors and patient age has not been reported but could be potentially useful to further establish that compared groups had similar overall illness severity.

Severity of illness models, which are analogous to clinical prediction rules, have been used extensively in both adult and pediatric critical care settings to accurately predict mortality outcome in groups of patients; the methodology for development of severity of outcome models has been well described (28). We hypothesized that multivariate severity of illness models based on the dependent variable "prolonged hospital duration" could be created from predictor independent variables measurable early after hospital admission. We examined a previously established nine epidemic year database of RSV hospitalization at our children's hospital to test this hypothesis.

	METHODS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Study Population and Data Collection

The study population comprised all pediatric patients hospitalized with laboratory-proven RSV-associated illness at the University of Michigan's C. S. Mott Children's Hospital from July 1, 1983 through June 30, 1992. All study participants were retrospectively identified from hospital virology laboratory records of children with RSV-positive nasopharyngeal smears by immunofluorescence and/or RSV- positive nasopharyngeal cultures. Data were collected by review of individual medical records. All patients younger than 2 yr of age hospitalized with laboratory-confirmed RSV at the University of Michigan Hospitals were eligible. Patients with nosocomial illness were excluded from model development since the duration of hospitalization attributed to RSV could not be determined in such cases.

Demographic data abstracted included sex, chronological and gestational ages, weight in kilograms, and all underlying diagnoses. Patient age in weeks was calculated in days from the date of hospital admission minus date of birth divided by seven. The durations of hospitalization, intensive care, and mechanical ventilation were calculated as the date of completion minus the date of initiation of these therapies in days.

Definitions of diagnostic conditions have previously been described (4). Briefly, well patients were defined as having no preexisting medical diagnoses and born at a gestational age of more than 36 wk. Premature infants were defined as having a gestational age of 36 wk or less. Patients considered to have failure to thrive (FTT) status were verified by correcting for gestational age and plotting on standard growth charts. Patients with preexisting congenital heart conditions (with the exception of patients with isolated patent ductus arteriosus that had been previously repaired or spontaneously resolved and those with trivial peripheral pulmonic stenosis), myocarditis, or cardiomyopathy were considered to have CHD. Patients with pre-existing diagnoses of BPD or congenital diaphragmatic hernia were classified as having pulmonary conditions type A. Patients with all other pulmonary diagnoses such as history of aspiration pneumonia, airway obstruction secondary to bronchomalacia, trachemalacia, or laryngomalacia, history of reactive airway disease, history of apnea of prematurity, pulmonary muscle weakness, and other miscellaneous pulmonary conditions were classified as having pulmonary conditions type B. Patients with known congenital immune disorders, those neutropenic or receiving chemotherapeutic agents, or those receiving chronic corticosteriods were classified as immunocompromised. Patients with major disability of an organ system (excluding heart and/or lung conditions), disease of the endocrine system, metabolic disease, neurologic disease (except febrile seizures), and genetic syndromes such as trisomy 21 or Hurler's syndrome were classified as having other miscellaneous conditions. Because of the limited number of patients in the immunocompromised group, this group was also included under the miscellaneous conditions category. RSV infection was considered hospital-acquired (nosocomial) if the RSV-positive culture or nasopharyngeal smear was collected at least 4 d after admission, and admission was not associated with a suspected respiratory or febrile illness. Children treated with ribavirin (Virazole; ICN Pharmaceuticals, Irvine, CA) antiviral therapy were identified along with dates of use and the duration of treatment determined. The presence of a diagnosis of apnea (related to illness severity) and survival outcome status were also recorded. Mechanical ventilation was considered to have occurred on the first hospital day (coded "early mechanical ventilation") if the date of intubation minus the date of admission was one or zero. Similarly, ribavirin was considered to have been used on the first day (coded "early ribavirin") if the date of ribavirin administration minus the date of hospital admission was one or zero. Survival was considered to occur if a patient was successfully discharged from the hospital alive. "Prolonged length of hospital stay" was defined as present if the patient discharge date minus admission date was greater than or equal to seven days. This length of stay cut point approximated the upper quartile of RSV hospitalization.

Data Acquisition and Analysis

Descriptive analyses were conducted to characterize the study population and assess outcome of hospitalization. These analyses included frequency distributions for categorical variables, and medians with ranges or means with standard deviations for continuous variables. Patient chronological age, weight in kilograms, and length of stay were log-transformed to improve normality prior to statistical testing. Univariate comparisons between groups with and without prolonged hospital duration were performed with SYSTAT version 5.2.1 (SYSTAT Inc., Evanston, IL) for the Macintosh computer. A Mann-Whitney U test was used for all continuous univariate variable comparisons between groups, and Pearson's chi-square was used for dichotomous variable comparisons between groups. Multivariate logistic regression analyses were performed using SAS, version 6.10 (SAS Institute Inc., Cary, NC) for the Macintosh computer (33). These analyses were conducted to evaluate the relationship between predictor variables (independent variables) and the dependent variable "prolonged duration of hospitalization." Estimated coefficients for independent variables (beta), standard errors (SE), chi-square of the maximum likelihood Wald test with p value, and odds ratios with 95% confidence intervals are described. A receiver operator characteristic curve was constructed from a classification table created with probabilities from 0.00 to 1.00 at intervals of 0.02, where from the logistic model p = 1/(1 + e^r) and r =  [k +  b_i × _i]. Sensitivity was defined as the proportion of the number of correctly identified patients with prolonged length of stay divided by the total number of patients with prolonged length of stay. Specificity was defined as the proportion of number of correctly identified patients without prolonged length of stay divided by the total number of patients without prolonged length of stay. The false positive ratio (FPR) was equal to 1 minus the specificity. Sensitivity, specificity, and FPR were converted to percents by multiplying the corresponding ratio times 100. Receiver operator characteristic curve plots, area under the curve estimated by the c statistic, and the Hosmer Lemeshow goodness-of-fit test were calculated using SAS version 6.10 for the Macintosh computer (33). Bootstrap and split-sample methods of model validation were performed using a published SAS macro (34).

Informed Consent/Institutional Review

The University of Michigan Institutional Review Board waived the requirement of informed consent for this retrospective cohort study.

	RESULTS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

A total of 962 data records of hospitalized patients with RSV infection documented by antigen detection or culture were identified in this study of nine epidemics from 1983 to 1992. A total of 802 hospitalized patients were younger than 2 yr of age and had community-acquired illness secondary to RSV; this represents the population studied in this report. Mean duration of hospitalization was 5.5 ± 7.2 d (median, 4.0 d). Prolonged hospitalization greater than 6 d was observed in 23% (182 of 802) and approximated the upper quartile of RSV hospital stay. Of the population studied, 9% (73 of 802) were mechanically ventilated and one patient died. The frequency of various diagnoses in this population is described in Table 1. Diagnoses were not mutually exclusive, with some patients having more than one diagnostic condition. Four hundred seventy-two (59%) patients were male, 197 (26%) had a history of prematurity, 56 (7%) had a history of CHD, 160 (20%) had a history of pulmonary disease type A or B, 59 (7.4%) had FTT, and 83 (10.3%) were categorized as having miscellaneous conditions. The percent of prolonged hospitalization observed by demographic information, diagnostic conditions, and treatment status is depicted in Table 1. The categorical variable observed most strongly associated with long hospital stay was need for mechanical ventilation (88% on the first day and 90% on any day). However, less than 10% of the total population and only 66 of 182 (36%) patients with prolonged hospital stays would be identified by this variable alone. Other diagnostic conditions associated with prolonged hospital stay varied from 39% for pulmonary conditions type B to 67% for failure to thrive. Variables associated with prolonged hospital stay by way of univariate analysis were the following: weight, prematurity, CHD, pulmonary condition A, pulmonary condition B, FTT, miscellaneous conditions, apnea, ICU admission, need for mechanical ventilation, and ribavirin administration.

A logistic regression model of the dependent variable "prolonged length of hospital stay" using independent variables available during the first hospital day of admission is summarized in Table 2. Nine independent variables measurable early in the hospital course were associated with the dependent variable "prolonged duration of hospitalization" at the p < 0.05 level. These were decreased patient weight (log-transformed), the presence of pulmonary conditions A or B, failure to thrive status, the diagnosis of other miscellaneous conditions, gestational age less than 37 wk, CHD status, early ribavirin, and early mechanical ventilation. Two variables, apnea and sex, were not associated with prolonged length of stay in the multivariate analysis that controlled for the nine other variables. The same nine variables model resulted from either backward or forward stepwise regression.

The actual observed versus predicted hospital duration outcome in the entire population divided into 10 deciles of risk is shown in Table 3. The Hosmer Lemeshow Goodness-of-Fit Test for the nine variables model with 8 degrees of freedom was 10.76 (p = 0.216). This indicated the patient outcomes observed versus predicted by the model from this data set were not different for any deciles of risk.

A receiver operator characteristic curve of the multivariate logistic regression model, as described in Table 2, is shown in Figure 1. The area under the curve for the logistic regression model described in Table 2 was 0.894, indicating excellent model discrimination. High sensitivity and low false positive ratios are seen in the following examples: sensitivity of 96% and FPR of 58% resulted from a p = 0.06 cut point, sensitivity of 89% and FPR of 34% resulted from a p = 0.100 cut point, sensitivity of 86% and FPR of 23% resulted from a p = 0.14 cut point, sensitivity of 81% and FPR of 16% resulted from a p = 0.20 cut point, sensitivity of 71% and FPR of 11% resulted from a p = 0.28 cut point, and sensitivity of 67% and FPR of 9% resulted from a p = 0.34 cut point. The later two cut points correctly classified 85% of the population's hospital outcomes.

View larger version (15K): [in this window] [in a new window]   Figure 1.   Receiver operator characteristic curve from nine variables logistic regression model of the dependent variable "prolonged hospital duration" for 802 patients hospitalized for RSV infection. Area under the ROC curve = 0.0894.

Additional models were also examined and are depicted in Table 4. A logistic model that included the additional variables of mechanical ventilation, intensive care unit admission, and ribavirin use at any time during hospitalization resulted in only a small improvement in the AUC (0.906) compared with the simpler model composed of variables measured earlier in the hospital stay (0.894) described in Table 2. Also described in Table 4 is a simple six variables model composed of variables available on hospital admission that has an observed AUC of 0.86. These examples would suggest that variables measured on admission or during the first 24 h of hospitalization predicted prolonged hospitalization associated with RSV infection nearly as well as variables measured late in hospitalization.

In the univariate analyses from Table 1, there was a trend for age being associated with prolonged hospitalization, which did not reach statistical significance (p = 0.065) compared with patient weight (p < 0.001). In each of the multivariate models described above, log age was substituted for log weight in order to determine whether age was significantly associated with outcome after controlling for other variables such as patient diagnosis. In each model, log age was strongly associated with prolonged hospitalization (p < 0.001). ROC curves were also created and the resultant AUC were only slightly lower than those models where log weight was used.

Bootstrap validation resulted in the following seven variables being selected in at least 95% of models: log weight, prematurity, CHD, pulmonary A, pulmonary B, failure to thrive status, and mechanical ventilation on the first hospital day. Split-sample testing resulted in similar area under the ROC curves of 0.894 ± 0.015 for model-learning sample and of 0.874 ± 0.020 for model-testing sample.

	DISCUSSION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

RSV is a common infectious disease of infants and young children that is responsible for a large number of hospital admissions during winter and spring epidemics. For many years, a high national research priority has existed to develop effective therapies to ameliorate the impact of annual RSV epidemics in infants and children. This has included the development of vaccines, an antiviral agent, passive immunity through immune globulin administration, and combinations of therapies. Efficacy trials have for the most part been limited to study of patients with selected preexisting conditions such as prematurity, congenital heart disease or bronchopulmonary dysplasia. Randomization has not always ensured adequate balance between compared treatment groups in terms of important risk factors because of a lack of statistical power in small ribavirin studies and because of chance occurrence in more moderate-sized recent RSV-IGIV investigations (19, 26).

In past clinical trials evaluating RSV therapies, combining selected high risk diagnostic conditions within a specified age range followed by randomization to different treatment groups has been used (7, 15). However, because multiple variables, including age, on a continuous scale have recently been described to be independently associated with hospital length of stay, severity of illness may not be the same in treatment groups even after randomization unless very large sample sizes exist (4, 27). Presently, severity of illness assessment of individual patients is not possible when several factors exist simultaneously. For example, the expected probability of long length of hospital stay for a 6-mo-old premature infant weighing 4 kg with a history of BPD could not be compared with a 12-mo-old infant weighing 5 kg with a history of prematurity, BPD, FTT, and CHD.

The model reported here used a dependent outcome variable that was objectively defined and measured on the basis of admission and discharge dates. Although mortality risk has been used as the outcome variable for severity of illness models in intensive care settings, this is not a suitable or practical outcome for RSV hospitalization. Mortality is a very uncommon observed outcome in the RSV population and has not been reported in any therapeutic trial to be associated with treatment status. Hospital duration on the other hand has been a commonly examined outcome in both efficacy and effectiveness comparison studies (2, 7, 16, 20, 26). Hospital duration has the additional potential advantage of being directly associated with hospitalization cost which is responsible for nearly two-thirds of the total population economic impact of RSV in Canada (35). Recent clinical trials of RSV IGIV efficacy and cohort outcomes research studies evaluating ribavirin effectiveness in actual clinical practice have used hospital duration as a primary outcome measure (3, 7, 20, 26). Independent variables (predictor variables) used in the present model were factors which could easily and objectively be obtained from the medical record and admission information. The derived model did not show evidence of deviation of observed versus predicted outcomes as measured by the Homer-Lemeshow test and a ROC curve analysis showed excellent area under the curve characteristics; by utilizing various probability cut points, one could maximize the sensitivity and specificity according to intended uses.

An important potential use for severity of illness models as described in this study may be to demonstrate in clinical trials that comparison groups are similar with respect not only to narrow study entry criteria, but also a more global severity of illness measure that represents an assimilation of several variables derived from a large multiyear patient data sample. New drug therapies are often FDA approved based on narrowly defined subsets of patients in which efficacy has been demonstrated under ideal experimental conditions of clinical trials rather than average less than ideal conditions of actual clinical practice. However, following approval, such agents are often used in patients in which the therapy was not specifically tested and under ordinary clinical settings rather than ideal settings of experimental studies. Outcomes research often utilizes an observational analytic cohort study design to assess the effectiveness of therapies once they have been FDA approved and in actual clinical practice use. The development of severity of illness measures developed from large patient data bases could be utilized to verify that treatment groups are similar with respect to known variables associated with outcome. Should groups be different with respect to severity of illness measures, statistical methods could be utilized to adjust for differences in illness severity.

Other uses of severity of illness models may include selection of patients with high probability of long hospital stay for therapeutic studies. As seen in Table 1, no risk factor by itself was highly predictive of prolonged hospital length of stay except for the requirement of mechanical ventilation; this variable was present in only a small fraction of the population of RSV-hospitalized patients. By using a multivariate severity of illness model to predict prolonged length of stay, one would have the ability to prognosticate from the entire RSV population of hospitalized patients by accounting for multiple risk factors. Such models could conceivably reduce the number of patients required to show a therapeutic effect since fewer patients with short lengths of stay could be enrolled; this could also result in reducing overall study cost if fewer total patients were enrolled. Alternatively, models could also be used to stratify results of therapeutic trials in RSV-hospitalized patients. A new therapy may not be useful in patients with short predicted length of stay but may be efficacious in patients with longer predicted hospital durations. If most patients had short lengths of stay, the efficacy of a therapy may be missed when all patients were combined.

Additional uses of RSV severity of illness length of stay models may be to assist in the assessment of hospital length of stay for RSV infection in different types of settings and cost containment. If unlike settings (for example, community hospital versus academic center) had dissimilar length of stays for RSV-hospitalized populations but the severity of illness could be demonstrated to be different, then this may be used to explain the observed difference. If settings had the same illness severity, but length of stay was significantly longer in one setting, then other factors would need to be examined to account for differences. By using such tools, institutions and third party providers may be better able to assess outcomes observed in different settings. Yet another potential use may be to assist in selection of patients to receive expensive therapies. Patient groups with short hospital predicted lengths of stay would be unlikely to benefit from such therapies, whereas patients with long length of stays would be expected to have a greater likelihood of cost benefit.

Potential shortcomings of the present study need to be commented upon. First, this was a single center study which may not be generalizable to other settings. Prospective validation in different populations is optimally required prior to using these models in actual clinical practice for patient care decisions. Second, the study was conducted over a 9-yr period where length of stay may change over time and epidemic period. We have previously reported, however, that hospital duration associated with RSV infection has not varied over multiple epidemics at this institution (4). Third, this study did not investigate some variables (hypoxia and chest radiographic findings) described in a recent cohort study to be associated with length of stay (27). We believe these variables would be problematic and of limited potential value in a model such as described here because of problems with interobserver reliability, complexity in measurement, and possible cost. For example, the administration of oxygen route to deliver a specified FI_O2 in infants was not described in a prior report (27). The common methods by which oxygen may be administered to infants are by mechanical ventilator, hood, face mask, and nasal cannula. With the later two methods, accurate measurement is not routinely possible when compared with FI_O2 delivered by mechanical ventilator and hood. Measurement of oxyhemoglobin saturation reported poor interobserver reliability with a kappa statistic of 0.61 (27, 32). The use of the hypoxia variable in patients with cyanotic heart disease or chronic pulmonary disease with oxygen requirements was not described. In the current study, the requirement of mechanical ventilation would have represented a subset of the patients coded as hypoxic in the prior report. This variable (requirement of mechanical ventilation) has the major advantage of being objectively defined and would not be subject to the issues just discussed. Finally, findings on chest radiography have also been reported to be associated with hospital length of stay (27). However, the precise definition of findings on chest radiographs was not described and interobserver reliability between centers was not reported but may be expected to be poor. Additionally, not all patients may require and receive routine chest radiography during their hospitalization for RSV. Obtaining such studies for all patients may result in unnecessary increased hospital cost.

We conclude that severity of illness models of RSV-associated hospital length of stay are feasible and that variables that may be objectively measured early in hospitalization accurately predict prolonged hospital stay in this RSV-hospitalized population, as seen with a sensitivity of 89% and FPR of 34% at a p = 0.10 cut point and sensitivity 81% and FPR 16% at a p = 0.20 cut point. Such models may have multiple potential uses related to future RSV-outcomes research studies such as selection of patients for efficacy and effectiveness studies, validation of similar treatment groups severity of illness status, length of stay adjustment from diverse hospital settings, and as a tool that could assist in selection of patients to receive expensive or limited available therapies. Validation, in terms of generalizability across a range of institution types, with a prospective cohort data sample is now needed.