- Email: [email protected]
A prognostic index for 30-day mortality after stroke
Journal of Clinical Epidemiology 54 (2001) 766–773
A prognostic index for 30-day mortality after stroke Yang Wanga*, Lynette L-Y Lima, Christopher Levib, Richard F Hellera, Janet Fischerc a
Centre for Clinical Epidemiology and Biostatistics, David Maddison Clinical Sciences Building, Royal Newcastle Hospital, Newcastle, New South Wales 2300, Australia b Department of Neurology, John Hunter Hospital, Locked Bag No 1, Hunter Regional Mail Center, 2310 NSW, Australia c Heart & Stroke Register, Centre for Clinical Epidemiology and Biostatistics, David Maddison Clinical Sciences Building, Royal Newcastle Hospital, Newcastle, New South Wales 2300, Australia Received 3 May 2000; received in revised form 27 November 2000; accepted 15 December 2000
Abstract The objective of this study was to develop a simplified scoring system to predict 30-day mortality in patients with acute ischemic stroke. A retrospective cohort study was performed in a tertiary referral hospital in the Hunter Region of Australia. A prognostic index was created by assigning points to the variables in a Cox model. The index included impaired consciousness (5 points), dysphagia (3 points), urinary incontinence (4 points), admission body temperature higher than 36.5C (2 points), and hyperglycemia without a clinical history of diabetes (2 points). A score of 11 or more defined a high-risk group. The index achieved a sensitivity, specificity, and positive predictive value of 68%, 98% and 75%, respectively, in the derivation sample and 57%, 97% and 68%, respectively, in the validation sample. The results provide a simple risk stratification instrument for clinical research and practice. Further evaluation of the model in a prospective cohort is warranted © 2001 Elsevier Science Inc. All rights reserved. Keywords: Ischemic stroke; Prognostic index; 30-day mortality
1. Introduction Stroke is the third most common cause of death in Western industrialised countries and a leading cause of chronic disability and suffering [1,2]. Every 1 in 10 deaths in Australia is due to stroke [3]. In caring for stroke patients, clinicians and investigators often need to estimate the risk of death. Accurate prediction of the mortality of acute stroke is important for several reasons. The role of clinicians in stroke management is not confined to treatment, and the confidence of the patient and family can be greatly enhanced by the ability to offer an accurate prognosis. A reliable prognosis allows better planing for supportive care, more accurate information to be given to relatives and resources to be allocated in a more efficient way. It may also allow patients to be stratified into different prognostic groups for clinical trials. Although several factors are known to influence the short-term prognosis in acute cerebrovascular disease (CVD), only a few attempts have been made to create prediction models for these patients [4–23] and none has come
* Corresponding author. Tel.: 61-2-4923 6302; fax: 61-2-4923 6148. E-mail address: [email protected] (Y. Wang)
into common use because of their complexity. As Barer has noted [24], the more complex these models become, the less likely they are to be widely applicable. To be useful to a clinician, a prognostic aid must be relevant to the general population of patients coming under his/her care, be as simple as possible, and address questions of practical importance. This study set out to establish the value of simple clinical measures determined early in the hospital stay that would predict 30-day mortality in acute ischemic stroke. Evaluation of these measures in a prospective study is on going. 2. Methods We retrospectively studied 440 patients admitted to a tertiary teaching hospital in the Hunter Region of Australia with a diagnosis of acute ischemic stroke (ICD code 433, 434, 436) [International Classification of Disease (ICD), 9th revision] [25]. The recruitment period was from July 1, 1995 to June 30, 1997, and the patients were recruited consecutively. The principal author reviewed all the medical records of these patients using a specially designed form. The patients were then randomly allocated to either a derivation or a validation group. The derivation sample was used to develop a prediction model. The validation sample was to test the validity of the prediction model. A research
0895-4356/01/$ – see front matter © 2001 Elsevier Science Inc. All rights reserved. PII: S0895-4356(01)00 3 3 8 - 9
Y. Wang et al. / Journal of Clinical Epidemiology 54 (2001) 766–773
nurse independently checked the coding variables in 11 randomly selected medical records in order to assess the interobserver agreement. Patients with the following conditions were excluded: 1) patients whose diagnoses were intracerebral hemorrhage (ICD code: 431) or subarachnoid hemorrhage (ICD code: 430); 2) patients whose final diagnoses were not clear, or just recorded as acute cerebrovascular accident. 2.1. Definitions and baseline measures Definition of admission consciousness level was classified into conscious, impaired consciousness, and unconscious, defined as follows: Conscious: alert with appropriate responses to verbal stimuli. Impaired consciousness: drowsy but responsive to verbal stimuli Unconscious: no eye opening to verbal stimuli. Definition of other clinical variables was: • Dysphagia was defined if “dysphagia” moderate or severe swallowing difficulties were noted in the medical records. • Urinary and fecal incontinence were noted positive if they were recorded as positive in the medical notes. • Stroke side was classified as single side affected if the affected side was only on the right or left side of the patient’s body or both sides affected if both sides of the patient’s body were recorded as being affected. • Admission systolic and diastolic blood pressures were recorded as continuous variables. Admission high blood pressure was defined as an admission systolic blood pressure higher than or equal to 160 mmHg or diastolic blood pressure higher than or equal to 95 mmHg. • Blood glucose and white blood cell (WBC) count were the earliest record result, and were recorded as continuous variables. Leucocytosis was defined positive if the WBC count exceeded 11 109/L. • The admission body temperature was the first tympanic temperature recorded in the medical notes. The admission body temperature was then divided into three groups (hypothermia, normothermia and hyperthermia) according to the classification previously used by Reith and Wang [26,27]. An admission body temperature higher than 37.5C was defined as hyperthermia; lower than or equal to 36.5C as hypothermia; between this range was defined as normothermia. • According to the admission glucose level and a clinical history of diabetes mellitus, the patients were then classified into three categories: 1. Normoglycemic without history of diabetes (NG). 2. Clinical history of diabetes (DM). 3. Hyperglycemic without a clinical history of diabetes (HG).
767
• The presence of the following comorbid conditions was assessed: history of hypertension, ischemic heart disease, previous stroke, transient ischemic attack (TIA), peripheral vascular disease, and atrial fibrillation. They were noted as positive if they were noted in the medical records. The above comorbid conditions were selected from the measures in the Charlson Index [28] and were available from the medical records. If the patients were unconscious or had significant impaired consciousness on admission, dysphagia, urinary incontinence and fecal incontinence and stroke side were evaluated when assessment was first recorded as being performed. If the patients were unconscious throughout the whole hospital stay, dysphagia, urinary incontinence and fecal incontinence were recorded as present, and the stroke side was recorded according to the information of a CT scan report. When the patients regained consciousness, the presence of a urinary catheter was recorded as urinary incontinence. Among the above variables, admission consciousness, body temperature, blood pressure and laboratory findings were taken within 24 h after admission. The other variables of physical signs were taken from the earliest record of physical examination. The comorbid conditions were deemed present if patients had a past history of these comorbid conditions or were found to have these comorbid conditions during the current admission. 2.2. Outcome measure The primary outcome of this study was 30-day mortality. The mortality status of patients was provided by the Hunter Area Health Services Heart & Stroke Register, which links patient’s records to mortality data provided by the Registry of Births, Death and Marriages of New South Wales, Australia. The records cover all deaths from residents of the Hunter Region, New South Wales Australia [29]. The causes of death were also coded in ICD-9. 2.3. Statistical analysis Chi-squared tests were performed in the univariate analyses for screening of variables for multivariate analysis. Approximately half the patients were randomly allocated to the derivation sample and the remainder to the validation sample. Cox proportional hazards models were fitted in the multivariate analysis. The output of Cox models gave Wald test P-values, which were to test the null hypothesis that the coefficients in the model are not significantly different from 0. Explanatory variables included in the model were those with P-values less than .1 in the univariate analysis. Variables with Wald P-values larger than .1 were removed from the model, and the likelihood ratio test was performed each time to assess the fit of the more parsimonious model. Those predictors that were substantially associated with the study outcome were assigned a point score by rounding their hazard ratios (HR) to the nearest integer. For each patient, a total score was calculated. Each level of the score
768
Y. Wang et al. / Journal of Clinical Epidemiology 54 (2001) 766–773
was tried as a cutoff point and the sensitivity, specificity, and positive predictive value (PPV) for that potential cutoff point was calculated. The PPV is of special clinical use in this context as it represents the proportion of patients with a particular score who will have the outcome of interest: 30day mortality. The optimum cutoff point was chosen on the basis of optimal sensitivity, specificity and PPV. Patients were then divided into high-risk and low-risk groups according to the cutoff point. After developing the initial prognostic index in the derivation sample, we tested the validity of the index in the validation sample using the sensitivity, specificity and PPV. Student’s t test was used to compare the difference of sensitivity, specificity and PPV between the derivation and validation samples. All analyses were performed using STATA data analysis system (Stata Corp. 1997 Statistical Software: Release 5.0 College Station, TX: Stata Corporation). 3. Results The initial screening identified 544 patients of whom 72 and 12 were excluded because the diagnosis was intracere-
bral hemorrhage or subarachnoid hemorrhage, respectively. Seven were excluded because the final discharge diagnosis was TIA. One patient died immediately after admission, and 12 were excluded because their diagnoses were not clearly classified as ischemic stroke or hemorrhagic stroke. The remaining 440 ischemic stroke patients were entered into the study. Of these patients, 223 patients (50.7%) were randomly allocated into the sample to derive the prognostic score. Within 30 days of hospital admission, 45 (10.2%) patient had died, among them 29 (64.4%) died due to strokerelated causes (hemorrhagic transformation, cerebral artery occlusion, etc.). After randomisation, 22 (48.9%) were allocated to the derivation sample and 23 (51.1%) from the validation sample. For the 11 randomly selected medical records with 242 variables checked independently by a research nurse, the proportion of agreement was 99% [30]. The baseline characteristics of this study and validation samples are shown in Table 1. There were no significant differences in baseline characteristics between the derivation sample and the remaining 217 patients who formed the validation sample.
Table 1 Baseline characteristics of the study and validation sample Variables Demographic variables Age, years (meanSD) Male sex, % Clinical findings Stroke severity N (%) Consciousness level Consciousness Impaired consciousness Unconsciousness Dysphagia Urinary incontinence Fecal incontinence Both sides of the brain affected Laboratory findings Glucose, mmol/L, (meanSD)a Temperature, C (meanSD)b White Blood Cell, 109/L (meanSD) Admission systolic blood pressure, mmHg Admission diastolic blood pressure, mmHg Hematocrit (L/L) Hemoglobin (g/L) Platelet ( 109/L) Comorbidities, n (%) Hypertension Ischemic heart disease Previous stroke TIA Diabetes mellitus Peripheral vascular disease Atrial fibrillation a
Derivation sample (N233)
Validation sample (N217)
P-value for difference
70 10 126 (57)
69 11 126 (58)
.185 .741
169 (76) 45 (20) 9 (4) 43 (19) 23 (10) 19 (9) 16 (7)
170 (78) 37 (17) 10 (5) 35 (16) 26 (12) 17 (8) 21 (10)
.696 .387 .578 .793 .344
7.5 2.4 36.7 0.7 91 2.7 163 28 91 20 0.42 0.07 144 26 255 82
7.5 2.5 36.7 0.6 9.2 3.0 164 27 92 17 0.42 0.06 145 24 257 84
.732 .215 .922 .563 .454 .913 .677 .740
124 (56) 80 (36) 64 (29) 53 (24) 41 (18) 42 (19) 37 (17)
131 (60) 74 (34) 68 (31) 55 (25) 44 (20) 32 (15) 31 (14)
.312 .697 .546 .701 .616 .252 .503
Limited to the analysis of 213 patients in the derivation sample and 203 in the validation sample. Limited to the analysis of 221 patients in the derivation sample and 216 in the validation sample.
b
Y. Wang et al. / Journal of Clinical Epidemiology 54 (2001) 766–773
3.1. Developing a prognostic index Table 2 shows the results of the univariate analysis. The following factors were significantly associated with mortality: impaired consciousness, unconsciousness, dysphagia, urinary and fecal incontinence, both sides of the body affected, admission body temperature, hyperglycemia without a clinical history of diabetes, history of previous stroke and atrial fibrillation. These variables were then entered into a multivariate Cox regression model. The first Cox regression model (Table 3) included all the variables selected from univariate analysis. Fecal incontinence, previous history of stroke, atrial fibrillation, leucocytosis and both sides affected became nonsignificant in this model. They were then removed to form the second model. To simplify the final model, all the variables in the second
769
model were dichotomised. The consciousness level was classified as impaired consciousness (combining the groups of impaired consciousness and unconsciousness) versus normal consciousness. Body temperature was classified as nonhypothermia (hyperthermia and normothermia) versus hypothermia; glycemic control was classified as non-HG (NG and DM) versus HG. The likelihood ratio test showed that the difference between the first and the second model was not significant (P-value: .151). The second Cox regression model is shown as Table 4. Impaired consciousness, dysphagia, urinary incontinence, hypothermia ( 36.5C) and hyperglycemia without a clinical history of diabetes comprised the five variables in the second model. Only hypothermia had a P-value slightly higher than .05 (P-value: .077). However, when this vari-
Table 2 Univariate analysis of predictors of 30-day mortality in the derivation sample Variables Age Sex Consciousness level
Dysphagia Urinary incontinence Fecal incontinence Stroke side Body temperature
WBC count HCT Hgb Admission high blood pressure: Diabetes and hyperglycemia: Normorglycemia with no diabetes (NG) Diabetes mellitus (DM) Hyperglycemia with diabetes (HG) Hypertension Ischemic Heart Disease Previous Stroke TIA Peripheral Vascular Disease Atrial fibrillation
Death/total
%
65 65 Male Female Consciousness Impaired consciousness Unconsciousness Yes No Yes No Yes No Single side Both sides 36.5 36.5 and 37.5 37.5 11.0 109/L 11.0 109/L 0.4 0.4 150 g/L 150 g/L Yes No
3/57 19/166 12/126 10/97 4/169 11/45 7/9 17/43 5/180 14/23 8/200 9/19 13/204 16/207 6/16 3/90 11/101 8/30 13/175 9/48 10/71 12/152 16/143 6/80 15/138 7/85
5 11 10 10 2 24 78 40 3 61 4 47 6 8 38 3 11 27 7 19 14 8 11 8 11 8
Yes No Yes No Yes No Yes No Yes No Yes No
10/140 4/38 8/35 9/124 13/99 8/80 14/143 10/64 12/159 4/53 18/170 5/42 17/181 9/37 13/186
7 11 23 7 13 10 10 16 8 8 11 12 9 24 7
P-value .177 .845 .001 .001 .001 .001 .001
.001 .020 .149 .376 .522
.024 .144 .960 .067 .517 .623 .001
770
Y. Wang et al. / Journal of Clinical Epidemiology 54 (2001) 766–773
Table 3 The first Cox regression model of 30-day mortality in the derivation sample Variable Consciousness status Consciousness Impaired consciousness Unconsciousness Dysphagia Urinary incontinence Fecal incontinence Both sides of the body affected Body temperature Hypothermia Normothermia Hyperthermia Hyperthermia and diabetes NG DM HG Leucocytosis Previous stroke Atrial fibrillation
Hazard ratio
S.E.
P-value
95% CI
1.0 4.2 9.7 2.7 2.9 1.1 2.3
— 1.7 5.3 0.9 1.0 0.5 2.4
— .001 .001 .001 .001 .818 .423
— 2.0–9.1 3.3–28.5 1.4–5.1 1.5–5.8 0.5–2.6 0.3–17.2
1.0 1.9 1.6
— 0.7 0.7
— .100 .299
— 0.09–4.1 0.7–3.6
1.0 1.7 2.6 1.1 1.0 0.6
— 0.6 0.7 0.2 0.2 0.2
— .121 0.0 .804 .949 .122
— 0.9–3.2 1.6–4.5 0.7–1.7 0.6–1.6 0.4–1.1
able was removed from the second model to generate a new model, the likelihood ratio test showed the difference between the two models was significant (P .002). Hypothermia was thus kept in the final model. To develop the prognostic system, we assigned “points” to each variable in the final model (Table 4). The points were derived by rounding the estimated hazard ratios from the Cox proportional hazards model to the nearest integer. Thus, the prediction model can be written as: prognostic indeximpaired consciousness (5 points) urinary incontinence (4 points) dysphagia (3 points) admission body temperature higher than 36.5C (2 points) hyperglycemia without a clinical history of diabetes (2 points). This prognostic index is a 0–16 scoring system. 3.2. Selecting of the optimum cutoff point of the prognostic index
ties, specificities and PPVs calculated for each point in the score are shown in Table 5. There are two levels of the score that could be selected as the optimum cutoff point: 10 or 11. With a score of 10 or higher, the sensitivity, specificity and PPV were 73%, 95% and 62%, respectively. With a score of 11 or higher, the sensitivity, specificity, and PPV were 68%, 98% and 75%, respectively. The risk of 30-day mortality for the patients with scores less than 11 was 3%, which compares with the risk of death of 75% for those with scores 11 or higher. To get the best PPV while at the same time maintaining satisfactory sensitivity and specificity, we chose 11 as the optimal cutoff point, defining the two groups: high risk (scores 11) and low risk (scores 11). Table 5 also shows the advantage of the prediction model over simply looking at single prediction variables. No single variable alone is as good as the model for optimising sensitivity, specificity, and PPV. Urinary incontinence, for example, if judged alone, has a sensitivity of 63%, specificity of 95% and PPV of 61%. Kaplan–Meier survival curves for the two risk groups in the derivation sample are drawn in Fig. 2. The log-rank test showed the difference between the two risk groups was highly significant (P .001). 3.3. Validation of the prognostic index When the score was applied to the remaining validation sample of 217 patients, the sensitivity, specificity, and PPV of a score of 11 or more were 57%, 97% and 68%, respectively, indicating stability of the model. There was no statistically significant difference between both estimates of sensitivity (P .648), specificity (P .715) or PPV (P .648). 4. Discussion We have developed a prognostic scoring system for patients with acute ischemic stroke during the early hospital stay. The system is built with five easily measured clinical predictors on which data were routinely available for all pa-
Fig. 1 shows that there is a clear trend for increasing 30day mortality with increasing index scores. The sensitivi-
Table 4 Cox proportional hazards model and allocation of points for the prognostic scoring system in the derivation sample Variables a
Impaired consciousness Dysphagia Urinary incontinence Body temperature (36.5C) HG
Hazard ratio S.E.
P-value 95% CI
Points
4.6 3.1 4.2 1.9 1.9
.001 .001 .001 .077 .018
5 3 4 2 2
1.7 0.9 1.2 0.7 0.4
2.3–9.6 1.8–5.6 2.4–7.3 0.9–4.1 1.2–2.9
a
Impaired consciousness included both previously defined as unconsciousness and with impairment of consciousness.
Fig. 1. Relationship between prognostic score and 30-day mortality.
Y. Wang et al. / Journal of Clinical Epidemiology 54 (2001) 766–773
771
Table 5 Validity of various score values and individual predictors of 30-day mortality in the derivation sample Cutoff points of the Index
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16 Variables Impaired consciousness Dysphagia Urinary incontinence Temperature 36.5C Hyperglycemia without diabetes
Sensitivity (%)
Specificity (%)
PPV (%)
96 95 95 95 82 82 82 77 73 68 68 59 45 14 14
27 66 70 76 85 87 91 93 95 98 98 98 99 100 100
13 24 26 31 38 40 51 57 62 75 79 81 91 100 100
82 77 63 86 36
82 87 95 44 86
33 39 61 14 22
Fig. 2. Kaplan–Meier survival curves for the two risk groups in the derivation sample. The log-rank test shows the difference between the two risk groups is highly significant (P .001).
PPV: positive predictive value.
tients. Impaired consciousness, dysphagia, urinary incontinence, admission body temperature higher than 36.5C, and hyperglycemia without a clinical history of diabetes constitute the five variables in the prediction model. A score of 11 or more suggests a more than 75% chance of death within 30 days, while scores of less than 11 suggests a less than 3% risk of death with the same time scale. In addition to acceptable sensitivity (68%) and specificity (98%), our prognostic index scores well for PPV (75%). The finding of a 75% likelihood of death within 30 days of stroke is of high potential clinical relevance. When our index was tested in the validation sample, there was no statistically significant difference of the sensitivity, specificity and PPV between the two samples, demonstrating the stability of the model. The variables used in this study were noted to be present if documented at any time during the acute hospital stay. A few studies have developed simplified scoring systems. However, there has, to our knowledge, been no published simple prediction model for short-term stroke mortality. The Guy’s score [31] and Fiorellis’ prediction model [14] were developed to predict 2- and 4-month mortality, respectively. Gompertz’s G-score (simplified Guy’s score) [32] and Fullerton’s prognostic index [10] were used to predict the outcome at 6 months. Wade et al. [12] develop his prognostic scoring system for prediction of mortality over a 2-year period. All the models are complicated and do not lend themselves to bedside use. Although the Cox model, which is inherently a multiplicative model, may perform better in predicting the outcome than our simple additive model, it is also complicated and
not ideal for use in routine clinical practice. The use of rounding hazard ratios from the Cox proportional hazards model into prediction scores is novel and is a very simple additive method that can be used by any clinician without a hand calculator. To our knowledge, only Kernan et al. [18] used a similar method to develop a prognostic system. However, that system was designed for patients with TIA or minor stroke, and contained few variables. Given this point, the method used in our study may be helpful in that it can be used for future prospective studies to develop simple prediction tools. It would, however, be preferable for our prognostic index to be tested in the other independent samples and in prospective studies. To develop the best model that is directly related to mortality, one must take account of relevant confounders and predictive factors such as age, stroke severity, body temperature, glucose, and history of diabetes. Such factors have not been completely accounted for in previous studies. This is the first study to include body temperature and dysphagia as important covariates. Tympanic body temperature measured in this study is slightly higher than the body temperature measured by other means (axillary or rectal), and it is closer to the core body temperature [33]. While tympanic body temperature may not be measured by all hospitals, the body temperature measured by other means can also be used in our prognostic index, since the temperature item in the index was broadly classified as hypothermia versus nonhypothermia. In our study, stroke patients who had impaired consciousness, urinary incontinence, or dysphagia during the acute hospital stay, had, respectively, 5, 4 or 3 times the chance of dying in the first 30 days after admission compared with those who were alert, continent or had no dysph-
772
Y. Wang et al. / Journal of Clinical Epidemiology 54 (2001) 766–773
agia. These three variables will reflect the severity of the neurological lesion. However, such findings may also be associated with confounding factors, such as urinary tract infection or sepsis of other types. Given the retrospective study design, it is not possible to exclude the effects or other complications in short-term stroke outcome. Level of consciousness was the main early clinical predictor of mortality in most of the previous studies [10,13,17,21,22,31,34,35]. The poorer outcome of stroke patients with urinary incontinence in the early phase of stroke is also in accordance with previous studies [36,37]. The reasons for this are likely to be multiple, including more extensive or strategic infarction, and also subtle effect on morale, which may influence response to rehabilitation [38]. Patients with dysphagia may also be at risk of dehydration and poor nutritional status [39–41]. Dysphagia after stroke could also be associated with a number of complications, such as aspiration-related chest infection [42]. Although urinary incontinence and dysphagia may not be routinely remarked upon the first 1–2 days of hospitalization, our results suggest that these factors are of prognostic relevance. Body temperature is a well-known predictor for stroke 30-day mortality [26,43]. In this study, we noted that the admission body temperature in excess of 36.5C contributed to 30-day death. In ischemic stroke, a central area of irreversible damaged tissue is surrounded by a zone of hypoperfusion, the “ischemic penumbra” [44], where neuronal function is impaired but potentially salvageable. It is probable that body temperature influences tissue survival in the penumbra. Ginsberg [45] suggested that body temperature be maintained in the range of 36.7C to 37.0C for at least the first few days after acute stroke. However, in our study, we found that in patients with admission body temperature higher than 36.5C, an increase in 30-day mortality was evident. Stress hyperglycemia (hyperglycemia without a clinical history of diabetes) is associated with an increased mortality in patients with myocardial infarction [46]. A similar association with stroke mortality has been noted in previous studies [47–49]. However, the mechanism underlying this process is still uncertain. Age has been found to be a significant predictor of mortality in most studies [12,14,20–23,31,50], but not all [10,13,22]. In our study, although there was a slight trend of increasing mortality with the increase of age, age was not a statistically significant predictor of 30-day mortality even in the univariate analysis. One possible explanation for this is that stroke severity variables contribute more than age to short-term mortality. A limitation of this study is its retrospective design. The variables were those recorded in the medical records rather than being prospectively and selectively collected. Some factors such as severity of motor paresis and smoking history were not taken into consideration, as without explicit prospective definitions they cannot be reliably drawn from the charts. While this means that standardisation of the study factors was not possible, it suggests that the results
may be more likely to reflect actual clinical practice. The retrospective nature avoids the danger of the data collection process influencing either the clinical care or introducing a selection bias to the study subjects. There is a possibility that some variables were incorrectly documented or were not recorded. In order to minimise this bias, we deliberately made all the variables simple (presence/absence), and selected the variables that clinicians would routinely examine (consciousness status, urinary incontinence, body temperature). Since the variables were simple and were easily collected and showed a high proportion of agreement between observers (99%), our prediction model should be generally applicable. It should be recognised, however, that the reviewer of the hospital records was not blind to the patient’s outcome. This may have introduced bias into the evaluation, and prospective data collection will be required to validate our findings. A further potential limitation in our study is the use of discharge diagnosis as opposed to admission diagnosis. Since this study is confined to the broad classification of ischemic stroke, the difference between admission and discharge diagnoses should be very small. In conclusion, using a retrospective but novel method, we have developed a prognostic index that uses simple and easily identified clinical features. Further prospective validation of this system will be needed before it can be incorporated into routine clinical practice. Such a system, however, may be useful in the design of clinical research protocols and interventional stroke trials.
Acknowledgments This study would not have been possible without the collaboration of the staff of the Heart and Stroke Register, the Hunter Area Health Service, who provided the primary data for the stroke patients. The authors wish to thank the staff of the Medical Information Department, John Hunter Hospital, who made available all the relevant medical records, and Patrick Fitzgerald, who provided statistical support for the data analysis.
References [1] Dyken M, Wolf P, Barnett H, Bergan J, Has W, Kannel W, Kuller L, Kurtzke J, Sundt T. Risk factors in stroke. A statement for physicians by the Subcommittee on Risk Factors. Stroke 1984;15:1105–11. [2] Toole J, Truscott B, Anderson W, Aronson P, Blaisdell F, Dustan M, Dyken M, Foley A, Hasley JJ, Langfitt T, Lawrence M, Mullan J. Report of the Joint Committee for Stroke Facilities: VII. Medical and surgical management of stroke. Stroke 1973;4:269–320. [3] Simons L, McCallum J, Friedlander Y, Simons J. Risk factors for ischemic stroke: Dubbo Study of the elderly. Stroke 1998;29(7):1341–6. [4] Marquardsen J. The natural history of acute cerebrovascular disease: a retrospective study of 769 patients. Acta Neurol Scand 1969; 45(Suppl 38):11 . [5] Sacco R, Wolf P, Kannel W, McNamara P. Survival and recurrence following stroke. Stroke 1982;13:290–5.
Y. Wang et al. / Journal of Clinical Epidemiology 54 (2001) 766–773 [6] Miah K, van Arbin M, Britton M, de Faire U, Helmers C, Maasing R. Prognosis in acute stroke with special reference to some cardiac factors. J Chron Dis 1983;36:279–88. [7] Frithz G, Werner I. Studies on cerebrovascular strokes. II. Clinical findings and short-term prognosis in a stroke material. Acta Med Scand 1976;199:133–40. [8] Sacco R, Foulkes M, Mohr J, Wolf P, Hier D, Price T. Determinants of early recurrence of cerebral infarction. The Stroke Data Bank. Stroke 1989;20:983–9. [9] Allen C, Harrison M, Wade D. The management of acute stroke. Tunbridge Wells: Cassell House, 1989. [10] Fullerton K, Mackenzie G, Stout R. Prognostic indices in stroke. Q J Med 1988;66(250):147–62. [11] Britton M, Faire Ud, Helmers C, Miah K. Prognostication in acute cerebrovascular disease. Acta Med Scand 1980;207:37–42. [12] Wade DT, Skilbeck CE, Wood VA, Hewer RL. Long-term survival after stroke. Age Ageing 1984;13:76–82. [13] Anderson C, Jamrozik K, Stewart-Wynne E. Predicting survival after stroke: experience from the Perth Community Stroke Study. Clin Exp Neurol 1992;29:117–28. [14] Fiorelli M, Alperovitch A, Argentino C, Sacchetti M, Toni D, Sette G, Cavalletti C, Gori M, Fieschi C. Prediction of long-term outcome in the early hours following acute ischemic stroke. Italian Acute Stroke Study Group. Arch Neurol 1995;52(3):250–5. [15] Slagado AV, Ferro JM, Gouveia-Oliveira A. Long-term prognosis of first-ever lacunar strokes. Stroke 1996;27:661–6. [16] Dighe MS, Aparasu RR, Rappaport HM. Factors predicting survival, changes in activity limitations, and disability in a geriatric post-stroke population. Gerontologist 1997;37(4):483–9. [17] Sheikh K, Brennan P, Meade T, Smith D, Goldenberg E. Predictors of mortality and disability in stroke. J Epidemiol Community Health 1983;37:70–4. [18] Kernan WN, Horwitz RI, Brass LM, Viscoli CM, Taylor KJW. A prognostic system for transient ischemia or minor stroke. Ann Intern Med 1991;114:552–7. [19] Rodrigues C, Joshi V. Predicting the immediate outcome of patients with cerebrovascular accident: a prognostic score. J Assoc Phys India 1991;39(2):175–80. [20] Lai SM, Alter M, Friday G, Sobel E. Prognosis for survival after initial stroke. Stroke 1995;26:2011–5. [21] Arboix A, Garcia-Eroles L, Massons J, Oliveres M. Predictive clinical factors of in-hospital mortality in 231 consecutive patients with cardioembolic cerebral infarction. Cerebrovasc Dis 1998;8(1):8–13. [22] Henon H, Godefroy O, Leys D, Mounier-Vehier F, Lucas C, Rondepierre P, Duhamel A, Pruvo J. Early predictors of death and disability after acute cerebral ischemic event. Stroke 1995;26(3):392–8. [23] Bonita R, Ford MA, Stewart AW. Predicting survival after stroke: a three-year follow-up. Stroke 1988;19:669–73. [24] Barer D, Mitchell J. Predicting the outcome of acute stroke: do multivariate models help? Q J Med 1988;70(261):27–39. [25] US Department of Health and Human Services. ICD-9-CM international classification of diseases 9th revision clinical modification, 9 ed., Vol. 1. Hobart Tasmania 7001, Australia: Information Services Division, Department of Community and Health Services, 1992. pp. 384–8. [26] Reith J, Jorgensen HS, Pedersen PM, Hirofumi N, Raaschou HO, Jeppesen LL, Olsen TS. Body temperature in acute stroke: relation to stroke severity, infarct size, mortality, and outcome. Lancet 1996; 347(8999):422–5. [27] Wang Y, Lim LL-Y, Levi C, Heller RF, Fisher J. The influence of admission body temperature on stroke mortality. Stroke 2000;31:404–9. [28] Charlson ME, Pompei P, Ales KL, MacKenzie CR. A new method of
[29] [30] [31] [32]
[33]
[34]
[35] [36]
[37] [38] [39] [40]
[41]
[42]
[43] [44]
[45] [46]
[47] [48]
[49] [50]
773
classifying prognostic comorbidity in longitudinal studies: development and validation. J Chron Dis 1987;40(5):373–83. Fisher J. Hunter Area heart and stroke register. Newcastle: Heart and Stroke Health Outcomes Council, 1997. pp. 1–3. Fleiss J. Statistical methods for rates and proportions, 2nd ed. New York: John Wiley & Sons, Inc, 1981. pp. 212–4. Allen C. Predicting the outcome of acute stroke: a prognostic score. J Neurol Neurosurg Psychiatry 1984;47:475–80. Gompertz P, Pound P, Ebrahim S. Predicting stroke outcome: Guy’s prognostic score in practice. J Neurol Neurosurg Psychiatry 1994;57: 932–5. Robinson J, Charlton J, Seal R, Spady D, Joffres MR. Oesophageal, rectal, axillary, tympanic and pulmonary artery temperatures during cardiac surgery. Can J Anaesth 1998;45(4):317–23. Czlonkowska A, Ryglewicz D, Lechowicz W. Basic analytical parameters as the predictive factors for 30-day case fatality rate in stroke. Acta Neurol Scand 1997;95(2):121–4. Chambers B, Norris J, Shurvell B, Hachinski V. Prognosis of acute stroke. Neurology 1987;37(2):211–5. Nakayama H, Jorgensen HS, Pedersen PM, Raaschou HO, Olsen TS. Prevalence and risk factors of incontinence after stroke. The Copenhagen Stroke Study. Stroke 1997;28(1):58–62. Barer DH. Continence after stroke: useful predictor or goal of therapy? Age Ageing 1989;18(3):183–91. Brittain KR, Peet SM, Castleden CM. Stroke and incontinence. Stroke 1998;29(2):524–8. Gordon C, Hewer RL, Wade DT. Dysphagia in acute stroke. Br Med J Clin Res 1987;295(6595):411–4. Barer DH. The natural history and functional consequences of dysphagia after hemispheric stroke. J Neurol Neurosurg Psychiatry 1989; 52(2):236–41. Axelsson K, Asplund K, Norberg A, Eriksson S. Eating problems and nutritional status during hospital stay of patients with severe stroke. J Am Dietetic Assoc 1989;89(8):1092–6. Smithard DG, O’Neill PA, Parks C, Morris J. Complications and outcome after acute stroke. Does dysphagia matter? Stroke 1996;27(7): 1200–4. [Published erratum appears in Stroke 1998 Jul;29(7):1480–1] Castillo J, Davalos A, Marrugat J, Noya M. Timing for fever-related brain damage in acute ischemic stroke. Stroke 1998;29(12):2455–60. Olsen T, Larsen B, Herning M, Skriver E, Lassen N. Blood flow and vascular reactivity in collaterally perfused brain tissue: evidence of an ischemic penumbra in patients with acute stroke. Stroke 1983;14: 332–42. Ginsberg M, Busto R. Combating hyperthermia in acute stroke: a significant clinical concern. Stroke 1998;29(2):529–34. Capes SE, Hunt D, Malmberg K, Gerstein HC. Stress hyperglycaemia and increased risk of death after myocardial infarction in patients with and without diabetes: a systematic overview. Lancet 2000; 355:2000. Jorgensen H, Nakayama H, Raaschou H, Olsen T. Stroke in patients with diabetes: the Copenhagen stroke study. Stroke 1994;25:1977–84. Weir CJ, Murray GD, Dyker AG, Lees KR. Is hyperglycemia an independent predictor of poor outcome after acute stroke? Results of a long term follow up study. Br Med J 1997;314(7090):1303–6. Candelise L, Landi G, Orazio E, Boccardi E. Prognostic significance of hyperglycemia in acute stroke. Arch Neurol 1985;45:661–3. Howard G, Walker MD, Becker CB, Bruce C, Feibel J, Mcleroy K, Toole JF, Yatsu F. Community hospital-based stroke programs: North Carolina, Oregon, and New York. III. Factors influencing survival after stroke: proportional hazards analysis of 4219 patients. Stroke 1986;17: 294–9.
A prognostic index for 30-day mortality after stroke Yang Wanga*, Lynette L-Y Lima, Christopher Levib, Richard F Hellera, Janet Fischerc a
Centre for Clinical Epidemiology and Biostatistics, David Maddison Clinical Sciences Building, Royal Newcastle Hospital, Newcastle, New South Wales 2300, Australia b Department of Neurology, John Hunter Hospital, Locked Bag No 1, Hunter Regional Mail Center, 2310 NSW, Australia c Heart & Stroke Register, Centre for Clinical Epidemiology and Biostatistics, David Maddison Clinical Sciences Building, Royal Newcastle Hospital, Newcastle, New South Wales 2300, Australia Received 3 May 2000; received in revised form 27 November 2000; accepted 15 December 2000
Abstract The objective of this study was to develop a simplified scoring system to predict 30-day mortality in patients with acute ischemic stroke. A retrospective cohort study was performed in a tertiary referral hospital in the Hunter Region of Australia. A prognostic index was created by assigning points to the variables in a Cox model. The index included impaired consciousness (5 points), dysphagia (3 points), urinary incontinence (4 points), admission body temperature higher than 36.5C (2 points), and hyperglycemia without a clinical history of diabetes (2 points). A score of 11 or more defined a high-risk group. The index achieved a sensitivity, specificity, and positive predictive value of 68%, 98% and 75%, respectively, in the derivation sample and 57%, 97% and 68%, respectively, in the validation sample. The results provide a simple risk stratification instrument for clinical research and practice. Further evaluation of the model in a prospective cohort is warranted © 2001 Elsevier Science Inc. All rights reserved. Keywords: Ischemic stroke; Prognostic index; 30-day mortality
1. Introduction Stroke is the third most common cause of death in Western industrialised countries and a leading cause of chronic disability and suffering [1,2]. Every 1 in 10 deaths in Australia is due to stroke [3]. In caring for stroke patients, clinicians and investigators often need to estimate the risk of death. Accurate prediction of the mortality of acute stroke is important for several reasons. The role of clinicians in stroke management is not confined to treatment, and the confidence of the patient and family can be greatly enhanced by the ability to offer an accurate prognosis. A reliable prognosis allows better planing for supportive care, more accurate information to be given to relatives and resources to be allocated in a more efficient way. It may also allow patients to be stratified into different prognostic groups for clinical trials. Although several factors are known to influence the short-term prognosis in acute cerebrovascular disease (CVD), only a few attempts have been made to create prediction models for these patients [4–23] and none has come
* Corresponding author. Tel.: 61-2-4923 6302; fax: 61-2-4923 6148. E-mail address: [email protected] (Y. Wang)
into common use because of their complexity. As Barer has noted [24], the more complex these models become, the less likely they are to be widely applicable. To be useful to a clinician, a prognostic aid must be relevant to the general population of patients coming under his/her care, be as simple as possible, and address questions of practical importance. This study set out to establish the value of simple clinical measures determined early in the hospital stay that would predict 30-day mortality in acute ischemic stroke. Evaluation of these measures in a prospective study is on going. 2. Methods We retrospectively studied 440 patients admitted to a tertiary teaching hospital in the Hunter Region of Australia with a diagnosis of acute ischemic stroke (ICD code 433, 434, 436) [International Classification of Disease (ICD), 9th revision] [25]. The recruitment period was from July 1, 1995 to June 30, 1997, and the patients were recruited consecutively. The principal author reviewed all the medical records of these patients using a specially designed form. The patients were then randomly allocated to either a derivation or a validation group. The derivation sample was used to develop a prediction model. The validation sample was to test the validity of the prediction model. A research
0895-4356/01/$ – see front matter © 2001 Elsevier Science Inc. All rights reserved. PII: S0895-4356(01)00 3 3 8 - 9
Y. Wang et al. / Journal of Clinical Epidemiology 54 (2001) 766–773
nurse independently checked the coding variables in 11 randomly selected medical records in order to assess the interobserver agreement. Patients with the following conditions were excluded: 1) patients whose diagnoses were intracerebral hemorrhage (ICD code: 431) or subarachnoid hemorrhage (ICD code: 430); 2) patients whose final diagnoses were not clear, or just recorded as acute cerebrovascular accident. 2.1. Definitions and baseline measures Definition of admission consciousness level was classified into conscious, impaired consciousness, and unconscious, defined as follows: Conscious: alert with appropriate responses to verbal stimuli. Impaired consciousness: drowsy but responsive to verbal stimuli Unconscious: no eye opening to verbal stimuli. Definition of other clinical variables was: • Dysphagia was defined if “dysphagia” moderate or severe swallowing difficulties were noted in the medical records. • Urinary and fecal incontinence were noted positive if they were recorded as positive in the medical notes. • Stroke side was classified as single side affected if the affected side was only on the right or left side of the patient’s body or both sides affected if both sides of the patient’s body were recorded as being affected. • Admission systolic and diastolic blood pressures were recorded as continuous variables. Admission high blood pressure was defined as an admission systolic blood pressure higher than or equal to 160 mmHg or diastolic blood pressure higher than or equal to 95 mmHg. • Blood glucose and white blood cell (WBC) count were the earliest record result, and were recorded as continuous variables. Leucocytosis was defined positive if the WBC count exceeded 11 109/L. • The admission body temperature was the first tympanic temperature recorded in the medical notes. The admission body temperature was then divided into three groups (hypothermia, normothermia and hyperthermia) according to the classification previously used by Reith and Wang [26,27]. An admission body temperature higher than 37.5C was defined as hyperthermia; lower than or equal to 36.5C as hypothermia; between this range was defined as normothermia. • According to the admission glucose level and a clinical history of diabetes mellitus, the patients were then classified into three categories: 1. Normoglycemic without history of diabetes (NG). 2. Clinical history of diabetes (DM). 3. Hyperglycemic without a clinical history of diabetes (HG).
767
• The presence of the following comorbid conditions was assessed: history of hypertension, ischemic heart disease, previous stroke, transient ischemic attack (TIA), peripheral vascular disease, and atrial fibrillation. They were noted as positive if they were noted in the medical records. The above comorbid conditions were selected from the measures in the Charlson Index [28] and were available from the medical records. If the patients were unconscious or had significant impaired consciousness on admission, dysphagia, urinary incontinence and fecal incontinence and stroke side were evaluated when assessment was first recorded as being performed. If the patients were unconscious throughout the whole hospital stay, dysphagia, urinary incontinence and fecal incontinence were recorded as present, and the stroke side was recorded according to the information of a CT scan report. When the patients regained consciousness, the presence of a urinary catheter was recorded as urinary incontinence. Among the above variables, admission consciousness, body temperature, blood pressure and laboratory findings were taken within 24 h after admission. The other variables of physical signs were taken from the earliest record of physical examination. The comorbid conditions were deemed present if patients had a past history of these comorbid conditions or were found to have these comorbid conditions during the current admission. 2.2. Outcome measure The primary outcome of this study was 30-day mortality. The mortality status of patients was provided by the Hunter Area Health Services Heart & Stroke Register, which links patient’s records to mortality data provided by the Registry of Births, Death and Marriages of New South Wales, Australia. The records cover all deaths from residents of the Hunter Region, New South Wales Australia [29]. The causes of death were also coded in ICD-9. 2.3. Statistical analysis Chi-squared tests were performed in the univariate analyses for screening of variables for multivariate analysis. Approximately half the patients were randomly allocated to the derivation sample and the remainder to the validation sample. Cox proportional hazards models were fitted in the multivariate analysis. The output of Cox models gave Wald test P-values, which were to test the null hypothesis that the coefficients in the model are not significantly different from 0. Explanatory variables included in the model were those with P-values less than .1 in the univariate analysis. Variables with Wald P-values larger than .1 were removed from the model, and the likelihood ratio test was performed each time to assess the fit of the more parsimonious model. Those predictors that were substantially associated with the study outcome were assigned a point score by rounding their hazard ratios (HR) to the nearest integer. For each patient, a total score was calculated. Each level of the score
768
Y. Wang et al. / Journal of Clinical Epidemiology 54 (2001) 766–773
was tried as a cutoff point and the sensitivity, specificity, and positive predictive value (PPV) for that potential cutoff point was calculated. The PPV is of special clinical use in this context as it represents the proportion of patients with a particular score who will have the outcome of interest: 30day mortality. The optimum cutoff point was chosen on the basis of optimal sensitivity, specificity and PPV. Patients were then divided into high-risk and low-risk groups according to the cutoff point. After developing the initial prognostic index in the derivation sample, we tested the validity of the index in the validation sample using the sensitivity, specificity and PPV. Student’s t test was used to compare the difference of sensitivity, specificity and PPV between the derivation and validation samples. All analyses were performed using STATA data analysis system (Stata Corp. 1997 Statistical Software: Release 5.0 College Station, TX: Stata Corporation). 3. Results The initial screening identified 544 patients of whom 72 and 12 were excluded because the diagnosis was intracere-
bral hemorrhage or subarachnoid hemorrhage, respectively. Seven were excluded because the final discharge diagnosis was TIA. One patient died immediately after admission, and 12 were excluded because their diagnoses were not clearly classified as ischemic stroke or hemorrhagic stroke. The remaining 440 ischemic stroke patients were entered into the study. Of these patients, 223 patients (50.7%) were randomly allocated into the sample to derive the prognostic score. Within 30 days of hospital admission, 45 (10.2%) patient had died, among them 29 (64.4%) died due to strokerelated causes (hemorrhagic transformation, cerebral artery occlusion, etc.). After randomisation, 22 (48.9%) were allocated to the derivation sample and 23 (51.1%) from the validation sample. For the 11 randomly selected medical records with 242 variables checked independently by a research nurse, the proportion of agreement was 99% [30]. The baseline characteristics of this study and validation samples are shown in Table 1. There were no significant differences in baseline characteristics between the derivation sample and the remaining 217 patients who formed the validation sample.
Table 1 Baseline characteristics of the study and validation sample Variables Demographic variables Age, years (meanSD) Male sex, % Clinical findings Stroke severity N (%) Consciousness level Consciousness Impaired consciousness Unconsciousness Dysphagia Urinary incontinence Fecal incontinence Both sides of the brain affected Laboratory findings Glucose, mmol/L, (meanSD)a Temperature, C (meanSD)b White Blood Cell, 109/L (meanSD) Admission systolic blood pressure, mmHg Admission diastolic blood pressure, mmHg Hematocrit (L/L) Hemoglobin (g/L) Platelet ( 109/L) Comorbidities, n (%) Hypertension Ischemic heart disease Previous stroke TIA Diabetes mellitus Peripheral vascular disease Atrial fibrillation a
Derivation sample (N233)
Validation sample (N217)
P-value for difference
70 10 126 (57)
69 11 126 (58)
.185 .741
169 (76) 45 (20) 9 (4) 43 (19) 23 (10) 19 (9) 16 (7)
170 (78) 37 (17) 10 (5) 35 (16) 26 (12) 17 (8) 21 (10)
.696 .387 .578 .793 .344
7.5 2.4 36.7 0.7 91 2.7 163 28 91 20 0.42 0.07 144 26 255 82
7.5 2.5 36.7 0.6 9.2 3.0 164 27 92 17 0.42 0.06 145 24 257 84
.732 .215 .922 .563 .454 .913 .677 .740
124 (56) 80 (36) 64 (29) 53 (24) 41 (18) 42 (19) 37 (17)
131 (60) 74 (34) 68 (31) 55 (25) 44 (20) 32 (15) 31 (14)
.312 .697 .546 .701 .616 .252 .503
Limited to the analysis of 213 patients in the derivation sample and 203 in the validation sample. Limited to the analysis of 221 patients in the derivation sample and 216 in the validation sample.
b
Y. Wang et al. / Journal of Clinical Epidemiology 54 (2001) 766–773
3.1. Developing a prognostic index Table 2 shows the results of the univariate analysis. The following factors were significantly associated with mortality: impaired consciousness, unconsciousness, dysphagia, urinary and fecal incontinence, both sides of the body affected, admission body temperature, hyperglycemia without a clinical history of diabetes, history of previous stroke and atrial fibrillation. These variables were then entered into a multivariate Cox regression model. The first Cox regression model (Table 3) included all the variables selected from univariate analysis. Fecal incontinence, previous history of stroke, atrial fibrillation, leucocytosis and both sides affected became nonsignificant in this model. They were then removed to form the second model. To simplify the final model, all the variables in the second
769
model were dichotomised. The consciousness level was classified as impaired consciousness (combining the groups of impaired consciousness and unconsciousness) versus normal consciousness. Body temperature was classified as nonhypothermia (hyperthermia and normothermia) versus hypothermia; glycemic control was classified as non-HG (NG and DM) versus HG. The likelihood ratio test showed that the difference between the first and the second model was not significant (P-value: .151). The second Cox regression model is shown as Table 4. Impaired consciousness, dysphagia, urinary incontinence, hypothermia ( 36.5C) and hyperglycemia without a clinical history of diabetes comprised the five variables in the second model. Only hypothermia had a P-value slightly higher than .05 (P-value: .077). However, when this vari-
Table 2 Univariate analysis of predictors of 30-day mortality in the derivation sample Variables Age Sex Consciousness level
Dysphagia Urinary incontinence Fecal incontinence Stroke side Body temperature
WBC count HCT Hgb Admission high blood pressure: Diabetes and hyperglycemia: Normorglycemia with no diabetes (NG) Diabetes mellitus (DM) Hyperglycemia with diabetes (HG) Hypertension Ischemic Heart Disease Previous Stroke TIA Peripheral Vascular Disease Atrial fibrillation
Death/total
%
65 65 Male Female Consciousness Impaired consciousness Unconsciousness Yes No Yes No Yes No Single side Both sides 36.5 36.5 and 37.5 37.5 11.0 109/L 11.0 109/L 0.4 0.4 150 g/L 150 g/L Yes No
3/57 19/166 12/126 10/97 4/169 11/45 7/9 17/43 5/180 14/23 8/200 9/19 13/204 16/207 6/16 3/90 11/101 8/30 13/175 9/48 10/71 12/152 16/143 6/80 15/138 7/85
5 11 10 10 2 24 78 40 3 61 4 47 6 8 38 3 11 27 7 19 14 8 11 8 11 8
Yes No Yes No Yes No Yes No Yes No Yes No
10/140 4/38 8/35 9/124 13/99 8/80 14/143 10/64 12/159 4/53 18/170 5/42 17/181 9/37 13/186
7 11 23 7 13 10 10 16 8 8 11 12 9 24 7
P-value .177 .845 .001 .001 .001 .001 .001
.001 .020 .149 .376 .522
.024 .144 .960 .067 .517 .623 .001
770
Y. Wang et al. / Journal of Clinical Epidemiology 54 (2001) 766–773
Table 3 The first Cox regression model of 30-day mortality in the derivation sample Variable Consciousness status Consciousness Impaired consciousness Unconsciousness Dysphagia Urinary incontinence Fecal incontinence Both sides of the body affected Body temperature Hypothermia Normothermia Hyperthermia Hyperthermia and diabetes NG DM HG Leucocytosis Previous stroke Atrial fibrillation
Hazard ratio
S.E.
P-value
95% CI
1.0 4.2 9.7 2.7 2.9 1.1 2.3
— 1.7 5.3 0.9 1.0 0.5 2.4
— .001 .001 .001 .001 .818 .423
— 2.0–9.1 3.3–28.5 1.4–5.1 1.5–5.8 0.5–2.6 0.3–17.2
1.0 1.9 1.6
— 0.7 0.7
— .100 .299
— 0.09–4.1 0.7–3.6
1.0 1.7 2.6 1.1 1.0 0.6
— 0.6 0.7 0.2 0.2 0.2
— .121 0.0 .804 .949 .122
— 0.9–3.2 1.6–4.5 0.7–1.7 0.6–1.6 0.4–1.1
able was removed from the second model to generate a new model, the likelihood ratio test showed the difference between the two models was significant (P .002). Hypothermia was thus kept in the final model. To develop the prognostic system, we assigned “points” to each variable in the final model (Table 4). The points were derived by rounding the estimated hazard ratios from the Cox proportional hazards model to the nearest integer. Thus, the prediction model can be written as: prognostic indeximpaired consciousness (5 points) urinary incontinence (4 points) dysphagia (3 points) admission body temperature higher than 36.5C (2 points) hyperglycemia without a clinical history of diabetes (2 points). This prognostic index is a 0–16 scoring system. 3.2. Selecting of the optimum cutoff point of the prognostic index
ties, specificities and PPVs calculated for each point in the score are shown in Table 5. There are two levels of the score that could be selected as the optimum cutoff point: 10 or 11. With a score of 10 or higher, the sensitivity, specificity and PPV were 73%, 95% and 62%, respectively. With a score of 11 or higher, the sensitivity, specificity, and PPV were 68%, 98% and 75%, respectively. The risk of 30-day mortality for the patients with scores less than 11 was 3%, which compares with the risk of death of 75% for those with scores 11 or higher. To get the best PPV while at the same time maintaining satisfactory sensitivity and specificity, we chose 11 as the optimal cutoff point, defining the two groups: high risk (scores 11) and low risk (scores 11). Table 5 also shows the advantage of the prediction model over simply looking at single prediction variables. No single variable alone is as good as the model for optimising sensitivity, specificity, and PPV. Urinary incontinence, for example, if judged alone, has a sensitivity of 63%, specificity of 95% and PPV of 61%. Kaplan–Meier survival curves for the two risk groups in the derivation sample are drawn in Fig. 2. The log-rank test showed the difference between the two risk groups was highly significant (P .001). 3.3. Validation of the prognostic index When the score was applied to the remaining validation sample of 217 patients, the sensitivity, specificity, and PPV of a score of 11 or more were 57%, 97% and 68%, respectively, indicating stability of the model. There was no statistically significant difference between both estimates of sensitivity (P .648), specificity (P .715) or PPV (P .648). 4. Discussion We have developed a prognostic scoring system for patients with acute ischemic stroke during the early hospital stay. The system is built with five easily measured clinical predictors on which data were routinely available for all pa-
Fig. 1 shows that there is a clear trend for increasing 30day mortality with increasing index scores. The sensitivi-
Table 4 Cox proportional hazards model and allocation of points for the prognostic scoring system in the derivation sample Variables a
Impaired consciousness Dysphagia Urinary incontinence Body temperature (36.5C) HG
Hazard ratio S.E.
P-value 95% CI
Points
4.6 3.1 4.2 1.9 1.9
.001 .001 .001 .077 .018
5 3 4 2 2
1.7 0.9 1.2 0.7 0.4
2.3–9.6 1.8–5.6 2.4–7.3 0.9–4.1 1.2–2.9
a
Impaired consciousness included both previously defined as unconsciousness and with impairment of consciousness.
Fig. 1. Relationship between prognostic score and 30-day mortality.
Y. Wang et al. / Journal of Clinical Epidemiology 54 (2001) 766–773
771
Table 5 Validity of various score values and individual predictors of 30-day mortality in the derivation sample Cutoff points of the Index
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16 Variables Impaired consciousness Dysphagia Urinary incontinence Temperature 36.5C Hyperglycemia without diabetes
Sensitivity (%)
Specificity (%)
PPV (%)
96 95 95 95 82 82 82 77 73 68 68 59 45 14 14
27 66 70 76 85 87 91 93 95 98 98 98 99 100 100
13 24 26 31 38 40 51 57 62 75 79 81 91 100 100
82 77 63 86 36
82 87 95 44 86
33 39 61 14 22
Fig. 2. Kaplan–Meier survival curves for the two risk groups in the derivation sample. The log-rank test shows the difference between the two risk groups is highly significant (P .001).
PPV: positive predictive value.
tients. Impaired consciousness, dysphagia, urinary incontinence, admission body temperature higher than 36.5C, and hyperglycemia without a clinical history of diabetes constitute the five variables in the prediction model. A score of 11 or more suggests a more than 75% chance of death within 30 days, while scores of less than 11 suggests a less than 3% risk of death with the same time scale. In addition to acceptable sensitivity (68%) and specificity (98%), our prognostic index scores well for PPV (75%). The finding of a 75% likelihood of death within 30 days of stroke is of high potential clinical relevance. When our index was tested in the validation sample, there was no statistically significant difference of the sensitivity, specificity and PPV between the two samples, demonstrating the stability of the model. The variables used in this study were noted to be present if documented at any time during the acute hospital stay. A few studies have developed simplified scoring systems. However, there has, to our knowledge, been no published simple prediction model for short-term stroke mortality. The Guy’s score [31] and Fiorellis’ prediction model [14] were developed to predict 2- and 4-month mortality, respectively. Gompertz’s G-score (simplified Guy’s score) [32] and Fullerton’s prognostic index [10] were used to predict the outcome at 6 months. Wade et al. [12] develop his prognostic scoring system for prediction of mortality over a 2-year period. All the models are complicated and do not lend themselves to bedside use. Although the Cox model, which is inherently a multiplicative model, may perform better in predicting the outcome than our simple additive model, it is also complicated and
not ideal for use in routine clinical practice. The use of rounding hazard ratios from the Cox proportional hazards model into prediction scores is novel and is a very simple additive method that can be used by any clinician without a hand calculator. To our knowledge, only Kernan et al. [18] used a similar method to develop a prognostic system. However, that system was designed for patients with TIA or minor stroke, and contained few variables. Given this point, the method used in our study may be helpful in that it can be used for future prospective studies to develop simple prediction tools. It would, however, be preferable for our prognostic index to be tested in the other independent samples and in prospective studies. To develop the best model that is directly related to mortality, one must take account of relevant confounders and predictive factors such as age, stroke severity, body temperature, glucose, and history of diabetes. Such factors have not been completely accounted for in previous studies. This is the first study to include body temperature and dysphagia as important covariates. Tympanic body temperature measured in this study is slightly higher than the body temperature measured by other means (axillary or rectal), and it is closer to the core body temperature [33]. While tympanic body temperature may not be measured by all hospitals, the body temperature measured by other means can also be used in our prognostic index, since the temperature item in the index was broadly classified as hypothermia versus nonhypothermia. In our study, stroke patients who had impaired consciousness, urinary incontinence, or dysphagia during the acute hospital stay, had, respectively, 5, 4 or 3 times the chance of dying in the first 30 days after admission compared with those who were alert, continent or had no dysph-
772
Y. Wang et al. / Journal of Clinical Epidemiology 54 (2001) 766–773
agia. These three variables will reflect the severity of the neurological lesion. However, such findings may also be associated with confounding factors, such as urinary tract infection or sepsis of other types. Given the retrospective study design, it is not possible to exclude the effects or other complications in short-term stroke outcome. Level of consciousness was the main early clinical predictor of mortality in most of the previous studies [10,13,17,21,22,31,34,35]. The poorer outcome of stroke patients with urinary incontinence in the early phase of stroke is also in accordance with previous studies [36,37]. The reasons for this are likely to be multiple, including more extensive or strategic infarction, and also subtle effect on morale, which may influence response to rehabilitation [38]. Patients with dysphagia may also be at risk of dehydration and poor nutritional status [39–41]. Dysphagia after stroke could also be associated with a number of complications, such as aspiration-related chest infection [42]. Although urinary incontinence and dysphagia may not be routinely remarked upon the first 1–2 days of hospitalization, our results suggest that these factors are of prognostic relevance. Body temperature is a well-known predictor for stroke 30-day mortality [26,43]. In this study, we noted that the admission body temperature in excess of 36.5C contributed to 30-day death. In ischemic stroke, a central area of irreversible damaged tissue is surrounded by a zone of hypoperfusion, the “ischemic penumbra” [44], where neuronal function is impaired but potentially salvageable. It is probable that body temperature influences tissue survival in the penumbra. Ginsberg [45] suggested that body temperature be maintained in the range of 36.7C to 37.0C for at least the first few days after acute stroke. However, in our study, we found that in patients with admission body temperature higher than 36.5C, an increase in 30-day mortality was evident. Stress hyperglycemia (hyperglycemia without a clinical history of diabetes) is associated with an increased mortality in patients with myocardial infarction [46]. A similar association with stroke mortality has been noted in previous studies [47–49]. However, the mechanism underlying this process is still uncertain. Age has been found to be a significant predictor of mortality in most studies [12,14,20–23,31,50], but not all [10,13,22]. In our study, although there was a slight trend of increasing mortality with the increase of age, age was not a statistically significant predictor of 30-day mortality even in the univariate analysis. One possible explanation for this is that stroke severity variables contribute more than age to short-term mortality. A limitation of this study is its retrospective design. The variables were those recorded in the medical records rather than being prospectively and selectively collected. Some factors such as severity of motor paresis and smoking history were not taken into consideration, as without explicit prospective definitions they cannot be reliably drawn from the charts. While this means that standardisation of the study factors was not possible, it suggests that the results
may be more likely to reflect actual clinical practice. The retrospective nature avoids the danger of the data collection process influencing either the clinical care or introducing a selection bias to the study subjects. There is a possibility that some variables were incorrectly documented or were not recorded. In order to minimise this bias, we deliberately made all the variables simple (presence/absence), and selected the variables that clinicians would routinely examine (consciousness status, urinary incontinence, body temperature). Since the variables were simple and were easily collected and showed a high proportion of agreement between observers (99%), our prediction model should be generally applicable. It should be recognised, however, that the reviewer of the hospital records was not blind to the patient’s outcome. This may have introduced bias into the evaluation, and prospective data collection will be required to validate our findings. A further potential limitation in our study is the use of discharge diagnosis as opposed to admission diagnosis. Since this study is confined to the broad classification of ischemic stroke, the difference between admission and discharge diagnoses should be very small. In conclusion, using a retrospective but novel method, we have developed a prognostic index that uses simple and easily identified clinical features. Further prospective validation of this system will be needed before it can be incorporated into routine clinical practice. Such a system, however, may be useful in the design of clinical research protocols and interventional stroke trials.
Acknowledgments This study would not have been possible without the collaboration of the staff of the Heart and Stroke Register, the Hunter Area Health Service, who provided the primary data for the stroke patients. The authors wish to thank the staff of the Medical Information Department, John Hunter Hospital, who made available all the relevant medical records, and Patrick Fitzgerald, who provided statistical support for the data analysis.
References [1] Dyken M, Wolf P, Barnett H, Bergan J, Has W, Kannel W, Kuller L, Kurtzke J, Sundt T. Risk factors in stroke. A statement for physicians by the Subcommittee on Risk Factors. Stroke 1984;15:1105–11. [2] Toole J, Truscott B, Anderson W, Aronson P, Blaisdell F, Dustan M, Dyken M, Foley A, Hasley JJ, Langfitt T, Lawrence M, Mullan J. Report of the Joint Committee for Stroke Facilities: VII. Medical and surgical management of stroke. Stroke 1973;4:269–320. [3] Simons L, McCallum J, Friedlander Y, Simons J. Risk factors for ischemic stroke: Dubbo Study of the elderly. Stroke 1998;29(7):1341–6. [4] Marquardsen J. The natural history of acute cerebrovascular disease: a retrospective study of 769 patients. Acta Neurol Scand 1969; 45(Suppl 38):11 . [5] Sacco R, Wolf P, Kannel W, McNamara P. Survival and recurrence following stroke. Stroke 1982;13:290–5.
Y. Wang et al. / Journal of Clinical Epidemiology 54 (2001) 766–773 [6] Miah K, van Arbin M, Britton M, de Faire U, Helmers C, Maasing R. Prognosis in acute stroke with special reference to some cardiac factors. J Chron Dis 1983;36:279–88. [7] Frithz G, Werner I. Studies on cerebrovascular strokes. II. Clinical findings and short-term prognosis in a stroke material. Acta Med Scand 1976;199:133–40. [8] Sacco R, Foulkes M, Mohr J, Wolf P, Hier D, Price T. Determinants of early recurrence of cerebral infarction. The Stroke Data Bank. Stroke 1989;20:983–9. [9] Allen C, Harrison M, Wade D. The management of acute stroke. Tunbridge Wells: Cassell House, 1989. [10] Fullerton K, Mackenzie G, Stout R. Prognostic indices in stroke. Q J Med 1988;66(250):147–62. [11] Britton M, Faire Ud, Helmers C, Miah K. Prognostication in acute cerebrovascular disease. Acta Med Scand 1980;207:37–42. [12] Wade DT, Skilbeck CE, Wood VA, Hewer RL. Long-term survival after stroke. Age Ageing 1984;13:76–82. [13] Anderson C, Jamrozik K, Stewart-Wynne E. Predicting survival after stroke: experience from the Perth Community Stroke Study. Clin Exp Neurol 1992;29:117–28. [14] Fiorelli M, Alperovitch A, Argentino C, Sacchetti M, Toni D, Sette G, Cavalletti C, Gori M, Fieschi C. Prediction of long-term outcome in the early hours following acute ischemic stroke. Italian Acute Stroke Study Group. Arch Neurol 1995;52(3):250–5. [15] Slagado AV, Ferro JM, Gouveia-Oliveira A. Long-term prognosis of first-ever lacunar strokes. Stroke 1996;27:661–6. [16] Dighe MS, Aparasu RR, Rappaport HM. Factors predicting survival, changes in activity limitations, and disability in a geriatric post-stroke population. Gerontologist 1997;37(4):483–9. [17] Sheikh K, Brennan P, Meade T, Smith D, Goldenberg E. Predictors of mortality and disability in stroke. J Epidemiol Community Health 1983;37:70–4. [18] Kernan WN, Horwitz RI, Brass LM, Viscoli CM, Taylor KJW. A prognostic system for transient ischemia or minor stroke. Ann Intern Med 1991;114:552–7. [19] Rodrigues C, Joshi V. Predicting the immediate outcome of patients with cerebrovascular accident: a prognostic score. J Assoc Phys India 1991;39(2):175–80. [20] Lai SM, Alter M, Friday G, Sobel E. Prognosis for survival after initial stroke. Stroke 1995;26:2011–5. [21] Arboix A, Garcia-Eroles L, Massons J, Oliveres M. Predictive clinical factors of in-hospital mortality in 231 consecutive patients with cardioembolic cerebral infarction. Cerebrovasc Dis 1998;8(1):8–13. [22] Henon H, Godefroy O, Leys D, Mounier-Vehier F, Lucas C, Rondepierre P, Duhamel A, Pruvo J. Early predictors of death and disability after acute cerebral ischemic event. Stroke 1995;26(3):392–8. [23] Bonita R, Ford MA, Stewart AW. Predicting survival after stroke: a three-year follow-up. Stroke 1988;19:669–73. [24] Barer D, Mitchell J. Predicting the outcome of acute stroke: do multivariate models help? Q J Med 1988;70(261):27–39. [25] US Department of Health and Human Services. ICD-9-CM international classification of diseases 9th revision clinical modification, 9 ed., Vol. 1. Hobart Tasmania 7001, Australia: Information Services Division, Department of Community and Health Services, 1992. pp. 384–8. [26] Reith J, Jorgensen HS, Pedersen PM, Hirofumi N, Raaschou HO, Jeppesen LL, Olsen TS. Body temperature in acute stroke: relation to stroke severity, infarct size, mortality, and outcome. Lancet 1996; 347(8999):422–5. [27] Wang Y, Lim LL-Y, Levi C, Heller RF, Fisher J. The influence of admission body temperature on stroke mortality. Stroke 2000;31:404–9. [28] Charlson ME, Pompei P, Ales KL, MacKenzie CR. A new method of
[29] [30] [31] [32]
[33]
[34]
[35] [36]
[37] [38] [39] [40]
[41]
[42]
[43] [44]
[45] [46]
[47] [48]
[49] [50]
773
classifying prognostic comorbidity in longitudinal studies: development and validation. J Chron Dis 1987;40(5):373–83. Fisher J. Hunter Area heart and stroke register. Newcastle: Heart and Stroke Health Outcomes Council, 1997. pp. 1–3. Fleiss J. Statistical methods for rates and proportions, 2nd ed. New York: John Wiley & Sons, Inc, 1981. pp. 212–4. Allen C. Predicting the outcome of acute stroke: a prognostic score. J Neurol Neurosurg Psychiatry 1984;47:475–80. Gompertz P, Pound P, Ebrahim S. Predicting stroke outcome: Guy’s prognostic score in practice. J Neurol Neurosurg Psychiatry 1994;57: 932–5. Robinson J, Charlton J, Seal R, Spady D, Joffres MR. Oesophageal, rectal, axillary, tympanic and pulmonary artery temperatures during cardiac surgery. Can J Anaesth 1998;45(4):317–23. Czlonkowska A, Ryglewicz D, Lechowicz W. Basic analytical parameters as the predictive factors for 30-day case fatality rate in stroke. Acta Neurol Scand 1997;95(2):121–4. Chambers B, Norris J, Shurvell B, Hachinski V. Prognosis of acute stroke. Neurology 1987;37(2):211–5. Nakayama H, Jorgensen HS, Pedersen PM, Raaschou HO, Olsen TS. Prevalence and risk factors of incontinence after stroke. The Copenhagen Stroke Study. Stroke 1997;28(1):58–62. Barer DH. Continence after stroke: useful predictor or goal of therapy? Age Ageing 1989;18(3):183–91. Brittain KR, Peet SM, Castleden CM. Stroke and incontinence. Stroke 1998;29(2):524–8. Gordon C, Hewer RL, Wade DT. Dysphagia in acute stroke. Br Med J Clin Res 1987;295(6595):411–4. Barer DH. The natural history and functional consequences of dysphagia after hemispheric stroke. J Neurol Neurosurg Psychiatry 1989; 52(2):236–41. Axelsson K, Asplund K, Norberg A, Eriksson S. Eating problems and nutritional status during hospital stay of patients with severe stroke. J Am Dietetic Assoc 1989;89(8):1092–6. Smithard DG, O’Neill PA, Parks C, Morris J. Complications and outcome after acute stroke. Does dysphagia matter? Stroke 1996;27(7): 1200–4. [Published erratum appears in Stroke 1998 Jul;29(7):1480–1] Castillo J, Davalos A, Marrugat J, Noya M. Timing for fever-related brain damage in acute ischemic stroke. Stroke 1998;29(12):2455–60. Olsen T, Larsen B, Herning M, Skriver E, Lassen N. Blood flow and vascular reactivity in collaterally perfused brain tissue: evidence of an ischemic penumbra in patients with acute stroke. Stroke 1983;14: 332–42. Ginsberg M, Busto R. Combating hyperthermia in acute stroke: a significant clinical concern. Stroke 1998;29(2):529–34. Capes SE, Hunt D, Malmberg K, Gerstein HC. Stress hyperglycaemia and increased risk of death after myocardial infarction in patients with and without diabetes: a systematic overview. Lancet 2000; 355:2000. Jorgensen H, Nakayama H, Raaschou H, Olsen T. Stroke in patients with diabetes: the Copenhagen stroke study. Stroke 1994;25:1977–84. Weir CJ, Murray GD, Dyker AG, Lees KR. Is hyperglycemia an independent predictor of poor outcome after acute stroke? Results of a long term follow up study. Br Med J 1997;314(7090):1303–6. Candelise L, Landi G, Orazio E, Boccardi E. Prognostic significance of hyperglycemia in acute stroke. Arch Neurol 1985;45:661–3. Howard G, Walker MD, Becker CB, Bruce C, Feibel J, Mcleroy K, Toole JF, Yatsu F. Community hospital-based stroke programs: North Carolina, Oregon, and New York. III. Factors influencing survival after stroke: proportional hazards analysis of 4219 patients. Stroke 1986;17: 294–9.