Seven prospective studies have systematically examined risk factors for delirium in elderly, hospitalized patients [9, 12-14, 16, 30, 31]. The risk factors generally identified included age, dementia, severe illness, metabolic and electrolyte imbalance, the use of psychoactive medications, and infections. However, most studies failed to distinguish prevalent cases of delirium at admission from incident cases occurring during hospitalization, a necessary step to define risk factors that truly precede the onset of delirium. In addition, only one of the previous studies [16] used a standardized, validated instrument to identify delirium. Finally, no study proposed a predictive model for delirium that was validated in an independent sample.

Previous studies focused on identifying precipitating factors for delirium (hospitalization-related factors such as medications, procedures, and intercurrent infections) or have simultaneously examined precipitating and predisposing conditions [9, 12-14, 16, 30, 31]. Although identifying these precipitating factors is undoubtedly important, an essential, prerequisite step is elucidation of baseline (or host) characteristics that indicate a particular vulnerability to delirium.

The focus of our study was to better define host or baseline vulnerability factors that would assist clinicians in identifying, at admission, those patients with a high risk for developing delirium during hospitalization. Our objectives were to estimate the incidence of delirium in an elderly hospitalized cohort using a standardized, validated instrument; to identify admission characteristics associated with the new occurrence of delirium; and to develop and validate a predictive model for the occurrence of delirium in the hospital based on the characteristics present at admission.

Methods

Top Methods Results Discussion Author & Article Info References

Patients

The potential study patients were consecutive weekday admissions from the emergency service to six general medicine floors at Yale-New Haven Hospital from June 1988 through March 1989. Eligibility criteria, which were met by 135 patients, included age of 70 years or more; no evidence of delirium at admission by clinical evaluation; and no history of severe dementia reported by the admitting attending physician or nurse or documented in the medical record. The latter criterion was selected to exclude patients with severe underlying dementia; patients with mild-to-moderate cognitive impairment were still eligible for the study. Exclusion criteria included an inability to undergo interviewing for such reasons as terminal illness, violent behavior, or intubation (n = 14); brief hospitalization (< 48 hours) (n = 5); language barrier (n = 2); refusal by the patient, family, or physician (n = 3); and other reasons (n = 4). Of the 135 eligible patients, 107 (79%) were enrolled in the study.

Clinical Evaluation

Trained clinician-researchers carried out daily structured interviews with the patients and their primary nurses from entry until hospital discharge. The baseline patient interview, completed within 48 hours of admission, was designed to gather demographic information; to discover whether the patient had a history of confusion; and to determine the patient's pre-illness social activity level, activities of daily living [32], instrumental activities of daily living [33], and social network and supports [34]. The interview also included the Mini-Mental State Examination [35], the Geriatric Depression Scale [36], standard vision (Jaeger- and Snellen-type tests) and hearing tests (Welch-Allyn audioscope), and questions designed to screen for hearing loss [37, 38]. Baseline information obtained from the nurses included an overall rating of illness severity [39]. Early in the hospital admission, family members or caregivers underwent a structured interview that included a modified Blessed Dementia Rating Scale [40, 41] and an estimation of the duration of any cognitive impairment.

Thereafter, the clinician-researchers interviewed the patients and their nurses daily and reviewed medical records to detect any new cases of delirium. The discharge assessment included determination of discharge location. Medical records were reviewed for admission medical diagnoses (inclusive list based on review of all admission notes), medications, and laboratory data. Informed consent was obtained from the patient or, for those with significant cognitive impairment, from the closest relative. The study was approved by the institutional review board of Yale University School of Medicine.

Identification of Delirium

Suspected cases of delirium were identified during the daily patient and nurse interviews. The Confusion Assessment Method [42] questionnaire, a previously validated instrument with a sensitivity rate ranging from 94% to 100% and a specificity rate ranging from 90% to 95%, was completed daily by two observers (a clinician-researcher and a nurse). The principal investigator (SKI) saw each patient with suspected delirium within 24 hours and established all final diagnoses of delirium based on fulfillment of the Confusion Assessment Method diagnostic criteria, which included the following: acute onset and fluctuating course; inattention; and either disorganized thinking or altered level of consciousness [42]. A case of delirium developing at any time during hospitalization was included in the analysis; however, a given patient could only develop delirium once (recurrent episodes of delirium were not counted when determining the relative-risk estimates).

Definition of Variables

We used clinically meaningful cutpoints to categorize variables. Patients were considered to have vision impairment if their corrected vision was worse than 20/70 on both near and distant binocular tests. Patients were considered to have hearing loss if they heard fewer than three of eight tones on the audioscope test (tones at a hearing level of 40 dB and at frequencies of 500, 1000, 2000, and 4000 Hz), had a score of 4 of 8 or less on the questions screening for hearing loss [37, 38], or wore a hearing aid. Patients were considered to have chronic cognitive impairment if they scored less than 24 of 30 on the Mini-Mental State Examination at admission [35]. The chronicity of the cognitive impairment in each patient with a score of less than 24 was confirmed by a modified Blessed Dementia Rating Scale score of 4 or more or by a duration of cognitive symptoms of at least 6 months. Illness severity was rated at admission using an APACHE II score [43] and a subjective overall rating by the nurse (adapted from Charlson and colleagues' observer-rated ordinal scale [39]) trichotomized as mild, moderate, or severe. Severe illness was considered present if the nurse gave a rating of severe or the APACHE score exceeded 16. An abnormal blood urea nitrogen/creatinine ratio, used as an index of dehydration, was defined as 18 or more [4]. Prominent depressive symptoms were considered present if the Geriatric Depression Scale score was equal to or exceeded the median score of 8 of 30. The size of a patient's social support network was estimated by the sum of the number of children, close relatives, and friends he or she saw at least once a month. A "low number" of social supports was indicated by six or fewer such contacts per month. Three social support types—instrumental support, emotional support, and presence of a confidante—were rated as present or absent. Having a "few" support types was indicated by the presence of one or none of these support types.

Validation Study

To validate the predictive model in an independent sample, a second cohort of patients was assembled.

Patients

Potential participants in the validation study included 202 patients who were enrolled from November 1989 through June 1990 in another ongoing study examining the frequency of functional decline in hospitalized elderly persons at Yale-New Haven Hospital. The inclusion and exclusion criteria for this study were identical to those for the study in the development cohort, except that patients with preexisting dementia were not excluded. For comparability with the initial development cohort, therefore, 25 patients (12%) with a documented history of dementia, Alzheimer disease, or chronic organic brain syndrome in their medical record were not considered eligible for the validation sample. Of the 177 eligible patients, 174 (98%) were included in the validation sample; 3 patients were excluded because missing data did not allow assessment of the outcome of delirium.

Procedure

The clinical evaluation and identification of patients with delirium were similar in the two cohorts. In all but three instances, the same assessment instruments were used. Only near binocular vision (Jaeger-type test), not distant vision, was tested in the validation cohort. For assessment of hearing, the Whisper test [44] was used. Hearing impairment was defined as present if the patient heard correctly fewer than 7 of 12 numbers on the Whisper test or wore a hearing aid. A shortened 15-item version of the Geriatric Depression Scale [45] was used. Significant depressive symptoms were considered present if the patient had a score that was equal to or exceeded the median of 4 of 15.

Statistical Analyses

In bivariate analyses, rates of delirium were calculated when each risk factor was present or absent. Crude relative risks were determined as the number of events (delirium) per person-days of observation in the group with the risk factor present relative to the number of such events per person-days of observation in the group without the risk factor. Ninety-five percent CIs were calculated when appropriate [46].

Variables with relative risk estimates of 1.5 or greater and clinical relevance were selected for evaluation in a proportional hazards model [47]. To avoid redundancy, clusters of related variables, such as measures of illness severity, were tested individually and in combinations; the factor most strongly associated with the outcome in the bivariate analysis was chosen for further evaluation. A proportional hazards model was used to identify the independent contributions of the risk factors to the outcome of delirium. To avoid excluding possibly important factors, the final risk factors were selected using a forward stepwise algorithm (the limit to enter a variable was set at P < 0.10 and to remove a variable at P > 0.15 for the log-likelihood ratio). Trends in outcome rates were assessed by the chi-square test for order in proportions.

The predictive model for delirium was created in the development cohort, then tested in the validation cohort. Significant differences in baseline characteristics between the development and validation cohorts were determined using Student t-tests for continuous variables and chi-square tests for dichotomous variables. Rates of delirium in the two cohorts were compared using Kaplan-Meier estimates and the log-rank test at 14 days [48]. Performance of the risk stratification system was measured using receiver-operating characteristic (ROC) analysis [49]. The values for area under the ROC curve range from 0 to 1, with 1 corresponding to perfect prediction, 0.5 to random performance (equivalent to chance alone), and 0 to completely incorrect prediction. All analyses were carried out using the SAS [50] and BMDP [51] programs.

Results

Top Methods Results Discussion Author & Article Info References

Characteristics of the development cohort are shown in Table 1. Delirium developed in 27 of 107 patients (25%) in the development cohort; the median onset of delirium was hospital day 4 (range, hospital day 2 to hospital day 18). Of the 27 patients diagnosed as having delirium, 24 met Confusion Assessment Method criteria for delirium on the basis of the daily patient interviews and 3 met the Confusion Assessment Method criteria on the basis of the nurse interview and medical record review. In the latter three cases of delirium, patients had transient nocturnal episodes with inattention, disorganized speech, altered level of consciousness, visual hallucinations, agitation, and disorientation.

Development of the Predictive Model

Potential risk factors for delirium in the development cohort are summarized in Table 2. The relative risks for seven of these variables achieved statistical significance in the bivariate analysis, with 95% CIs excluding 1.0; these variables included vision impairment, hearing impairment, cognitive impairment, three measures of illness severity (nurse rating of severe, APACHE score > 16, and a composite measure), and blood urea nitrogen/creatinine ratio of 18 or more. We created a composite measure of illness severity, defined by either a nurse rating of severe or an APACHE score of more than 16. This composite rating, which incorporated aspects of both subjective illness severity rating [39] and objective criteria [43], was the illness severity measure chosen for evaluation in the multivariable model.

Other variables that were examined in bivariate analyses but were not included in the final model because of an exposure rate of less than 2% or a relative risk estimate of less than 1.5 included race, education, marital status, living alone, religiousness, body mass index (weight/height²), number and type of admission medical diagnoses, number and type of admission medications, electrolyte abnormalities, leukocytosis, anemia, acid-base imbalance, and abnormalities in glucose, calcium, or liver function tests. Because of the small numbers of patients with the same medical diagnosis at admission, we grouped diagnoses by related cause or severity (for example, acute myocardial infarction, sepsis, respiratory failure, diabetic ketoacidosis), as well as by organ system (for example, cardiac, pulmonary, renal, gastrointestinal). Medical diagnosis as a variable did not achieve quantitative or statistical significance [48] with any of these methods. Number and type of admission medications—analyzed individually and by clinical groupings (for example, sedative-hypnotics, narcotics, anticholinergics)—were not quantitatively or statistically significant [48] predictors of delirium.

For development of the final predictive model, we evaluated all 13 variables with a relative risk of 1.5 or more Table 2 in a stepwise proportional hazards model. The four variables selected in the final model were vision impairment, severe illness (composite measure), cognitive impairment, and high blood urea nitrogen/creatinine ratio (Table 3). The adjusted relative risks, derived from the coefficients in the proportional hazards model, are estimates of the independent contribution of each variable to the risk for developing delirium that were calculated while controlling for the other variables in the model.

Performance of the Predictive Model

Development Cohort

We determined the effect of the number of risk factors on the risk for developing delirium during hospitalization. The proportion of patients with newly developed delirium increased linearly with the number of risk factors present at baseline.

A risk stratification system was developed by assigning 1 point to each of the final risk factors present. Based on the distributional characteristics in our population (groups with similar delirium rates were combined), we created a low-risk group with no risk factors present, an intermediate-risk group with 1 to 2 risk factors present, and a high-risk group with 3 to 4 risk factors present. The performance of this risk stratification system is shown in Table 4. Rates of delirium for the low-, intermediate-, and high-risk groups were 9%, 23%, and 83%, respectively (chi-square trend, P < 0.0001). The area under the ROC curve was 0.74 (CI, 0.63 to 0.85) for the risk stratification system.

Validation Cohort

In the validation cohort, delirium developed in 29 (17%) patients; the median day of onset was hospital day 6 (range, hospital day 3 to hospital day 39). Kaplan-Meier estimates of rates of onset of delirium in the development and validation cohorts during the first 14 days differed statistically (33% compared with 19%, respectively; P = 0.04). As shown in Table 1, the validation cohort had statistically higher levels of education, a greater number of patients living alone, and higher rates of vision impairment and history of confusion at baseline than did the development cohort. Applying the risk stratification system to the validation cohort, we found that the rates of delirium for the low-, intermediate-, and high-risk groups were 3%, 16%, and 32%, respectively (chi-square trend, P < 0.002) (Table 4). Although the validation cohort had lower rates of delirium at each risk level when compared with the development cohort, the trend remained statistically significant. The area under the ROC curve was 0.66 (CI, 0.55 to 0.77) for the risk stratification system in the validation cohort. The predictive performance was somewhat less strong than that observed in the development cohort but statistically exceeded chance alone.

Clinical Outcomes Related to Delirium

We next examined the performance of the risk stratification system for predicting the most important clinical outcomes associated with delirium, namely death during hospitalization or nursing home placement (from the hospital), as a measure of the predictive validity of the risk stratification system [52]. Previous studies showed that delirium is associated with an increased risk for nursing home placement [4, 20, 28, 53, 54], perhaps because of its associated increased morbidity and functional decline. Moreover, one recent study [55] showed that delirium may be substantially more persistent than was previously believed. The assessment of death in combination with nursing home placement avoids potential inferential errors that may arise because patients who die can no longer be placed in a nursing home [56]. For the development cohort, the rate of death or nursing home placement increased from 9% to 16% to 42% in the low-, intermediate-, and high-risk groups, respectively, yielding a fourfold increase overall (Table 5). In the validation cohort, the rate of death or nursing home placement increased from 3% in the low-risk group to 14% in the intermediate-risk group to 26% in the high-risk group, for an eightfold increase overall. In both cohorts, the increasing trends were statistically significant. There were one, four, and zero deaths in the development cohort and one, six, and four deaths in the validation cohort for the low-, intermediate-, and high-risk groups, respectively.

Most but not all of the deaths and nursing home placements occurred in patients who developed delirium. In the development cohort, these clinical outcomes occurred four times more often among patients with delirium than among patients without delirium; in the validation cohort, the rate was six times higher in patients with delirium.

Discussion

Top Methods Results Discussion Author & Article Info References

In our prospective study, delirium developed during hospitalization in 25% and 17% of two successive cohorts of elderly general medical patients. These rates are similar to the incidence rates described in previous studies that have examined similar populations [11-16]. The median onset of delirium occurred early after hospitalization in these cohorts (hospital days 4 and 6), which is also in accordance with previous studies [12-14].

Our final predictive model is based on four well-documented risk factors for delirium—vision impairment [2, 16, 57, 58], severe illness [12-14], preexisting cognitive impairment [9, 13, 14, 30], and dehydration [4, 12, 14]—and offers quantitative confirmation of risk factors suggested by previous studies. The strengths of our study include the use of a standardized, validated diagnostic instrument for delirium, the Confusion Assessment Method [42], as well as the routine and regular assessment for delirium throughout hospitalization. In addition, the predictive model was prospectively validated in a clinically distinct population. Finally, the predictive model created gradients for clinical outcomes potentially related to delirium in both the development and validation cohorts. This finding confirms that our risk stratification system has predictive validity.

The validation cohort and the development cohort had some statistically significant differences. Although the inclusion and exclusion criteria were similar in both cohorts, patients with dementia were not excluded initially from the validation cohort and were excluded subsequently only if evidence of cognitive impairment was documented in the medical record. Thus, the validation cohort had a greater number of cognitively impaired patients at baseline. In addition, differences existed in educational level, number of patients living alone, and vision impairment (possibly due to differences in the measurement of vision). Finally, the validation cohort had a smaller proportion of patients with delirium when compared with the development cohort, resulting in relatively fewer outcome events and a greater potential for unstable predictive estimates. This factor may explain the lower relative risks and wider CIs for the individual risk factors in the validation cohort. Although the within-strata rates of delirium decline in the validation cohort, the across-strata relative risks are just as high as in the development cohort (see Table 4). The predictive model did create a distinct and statistically significant risk gradient in the validation cohort. Thus, the overall predictive capacity of the risk stratification system is retained across two distinct populations—an important strength of this study.

Some risk factors, such as vision impairment and severe illness, had a low prevalence in the development cohort but were associated with a relatively high risk for delirium. However, extremely low prevalence rates may result in unstable predictive estimates (unstable point estimates and widened confidence intervals). The prevalence of these risk factors may vary among study populations, and their predictive ability will need to be tested in other samples.

In contrast to results from previous studies, older age was not a statistically significant independent predictor of delirium. This finding may be attributable to the age cutoff ( 70 years) among our patients and the relatively narrow age range that was represented or to better adjustment for other age-related factors such as illness severity. The number of admission medical diagnoses was not a statistically significant risk factor for delirium, suggesting that the number of medical diagnoses is a weak indicator of illness severity and comorbidity. Our findings suggest that among elderly persons, the underlying illness severity, as indicated by the nurse's subjective rating or the APACHE score, is a more important predictor of delirium than the specific medical diagnoses involved. In addition, the number and type of admission medications were not statistically significant predictors of delirium in this study. Finally, abnormal laboratory findings (for example, electrolyte abnormalities, leukocytosis, anemia, acid-base imbalance, and abnormal glucose, calcium, or liver function tests) were not predictive of patients at risk for developing delirium. Because many of these risk factors were present in small numbers of patients, our study may have lacked sufficient power to assess adequately their predictive ability.

Because our purpose was to characterize host vulnerability and to estimate the risk for development of delirium at hospital admission, precipitating factors occurring during hospitalization were not examined. Adverse drug effects, invasive procedures, intercurrent illness, sleep deprivation, and forced immobilization may precipitate delirium during hospitalization in elderly patients and must be examined in future studies, with particular attention given to the interaction of these factors with baseline patient vulnerability. In previous studies, investigators assumed that a given precipitating factor had the same effect in any patient. In fact, the development of delirium probably involves a complex and interacting web of precipitating factors that act on hosts with varying degrees of vulnerability. For instance, in a vulnerable patient (for example, one who is severely ill or cognitively impaired), delirium may be precipitated by a slight insult (such as one dose of sleeping medication), whereas in an intact patient, delirium may result from the cumulative effect of many noxious insults (for example, multiple invasive procedures, multiple medications with psychoactive side effects, and intensive care unit stay). The predictive model for baseline risk developed in our study provides an important initial step in elucidating the complex mechanisms underlying delirium.

An important limitation of our study was the relatively small sample size and number of outcome events (delirium) in both cohorts. The sample sizes were restricted because of the detailed clinical follow-up required for each patient. In addition, our selection criteria were designed to exclude severely demented patients because the diagnostic criteria for delirium do not apply to severely demented patients [42]. This decision meant loss of generalizability because severely demented patients are at high risk for development of delirium; however, the critical issue of validity of diagnosis was considered more important. Excluding severely demented patients may have reduced the strength of preexisting cognitive impairment as a predictive variable, although a 3.5-fold increased risk was still found. Notably, patients in both cohorts did represent a broad range of cognitive performance, with sizable proportions of cognitively impaired patients (see Table 1). Examining delirium in larger samples including severely demented patients must be done in future studies.

This model is the first for the prediction of delirium in hospitalized elderly persons in which validity has been tested in a separate group of patients. Identifying patients at high risk for developing delirium during hospitalization will enable physicians to monitor patients' cognitive status closely and to minimize interventions that may contribute to delirium during hospitalization. In addition, targeting high-risk patients may be important for early discharge planning. Conversely, the risk system also works well to identify patients at low risk for delirium. Clinical investigators may find this risk system useful to identify and stratify patients for intervention trials, particularly much-needed studies to identify, prevent, and treat delirium and its complications in the hospitalized elderly.

Despite the performance of our final predictive model in one independent sample, future studies are needed to verify its usefulness in other populations and institutions. In addition, study of other groups at high risk for delirium, such as surgical and intensive care unit patients, is needed. Finally, future studies are needed to determine if interventions aimed at the identified risk factors, such as volume repletion and vision aids, can substantially reduce the risk for delirium in hospitalized elderly patients.