Article Text
PDF
Abstract
Background We developed and validated a risk score to predict delirium after stroke which was derived from our prospective cohort study where several risk factors were identified.
Methods Using the β coefficients from the logistic regression model, we allocated a score to values of the risk factors. In the first model, stroke severity, stroke subtype, infection, stroke localisation, pre-existent cognitive decline and age were included. The second model included age, stroke severity, stroke subtype and infection. A third model only included age and stroke severity. The risk score was validated in an independent dataset.
Results The area under the curve (AUC) of the first model was 0.85 (sensitivity 86%, specificity 74%). In the second model, the AUC was 0.84 (sensitivity 80%, specificity 75%). The third model had an AUC of 0.80 (sensitivity 79%, specificity 73%). In the validation set, model 1 had an AUC of 0.83 (sensitivity 78%, specificity 77%). The second had an AUC of 0.83 (sensitivity 76%, specificity 81%). The third model gave an AUC of 0.82 (sensitivity of 73%, specificity 75%). We conclude that model 2 is easy to use in clinical practice and slightly better than model 3 and, therefore, was used to create risk tables to use as a tool in clinical practice.
Conclusions A model including age, stroke severity, stroke subtype and infection can be used to identify patients who have a high risk to develop delirium in the early phase of stroke.
- STROKE
- COGNITION
- NEUROPSYCHOLOGY
Statistics from Altmetric.com
Introduction
Delirium is a common psychiatric syndrome in the care of elderly patients.1 The incidence of delirium after stroke in recent studies varies between 10% and 13%.2–4 Delirium after stroke is associated with a higher mortality, a worse functional outcome and an increased risk of institutionalisation.3 ,5 Moreover, it is an independent predictor for severe cognitive impairment after 2 years.6 Age, pre-existent cognitive decline, severe neurological deficits and medical complications are the most important risk factors.4 ,7 ,8 We recently identified pre-existing cognitive decline, infection, right-sided hemispheric stroke, anterior circulation large-vessel stroke, stroke severity and brain atrophy as independent risk factors for delirium in the acute phase after stroke.3
Since delirium after stroke is associated with a worse prognosis, we wanted to facilitate early identification of stroke patients at risk for delirium. In the present study, we developed a simple risk score from the risk factors found in an earlier study, and we validated this risk score in an independent cohort.
Methods
Recruitment of the first cohort
We derived the risk prediction score from our previously prospective cohort study of 527 acutely admitted stroke patients.3 Criteria for stroke were neurologic deficit of sudden onset, including language and speech problems, lasting longer than 24 h for which no other cause than stroke could be found. Patients with subarachnoid haemorrhage and TIA were excluded. Patients had to be older than 18 years.
The study was approved by the medical ethical committee of the St Elisabeth Hospital, Tilburg. Informed consent was given by the patient or a caregiver.3
Development of the risk models
Pre-existing cognitive decline, infection, right-sided hemispheric stroke, anterior circulation large-vessel stroke, stroke severity and brain atrophy were independent predictors of delirium.3 Age was an independent risk factor if brain atrophy was left out of the model. For the current study, our aim was to develop a risk score that is available on the day of admission and that can be easily obtained. Therefore, we used age instead of brain atrophy. By means of the β coefficients from the logistic regression model, we allocated a score to each risk factor (see table 1). At first we calculated the area under the curve (AUC) derived from the risks that were calculated with the logistic regression model. We compared this AUC with the AUC derived from the risk scores using the β coefficients. In total, we tested three models. In the first model, stroke severity (measured with the National Institute of Health Stroke Scale9 (NIHSS)), stroke subtype with Oxfordshire Community Stroke Project criteria,10 infection (infection was scored at both screening dates using data from the medical records from the day of hospitalisation until the day of screening. We used the following data: pyrexia, high leukocytosis and/or raised ESR with a positive blood, sputum or urine culture and/or infiltrate on chest x-ray, or for which antibiotics were prescribed, stroke localisation (left or right hemisphere), Informant Questionnaire on Cognitive Decline in the Elderly (IQCODE)11 ,12 above 50 years, and age were included. In the second model, we included only variables that are easily available for the clinician, namely age, NIHSS, stroke subtype and infection. The third model was further simplified and only included age and NIHSS.
Characteristics of the 527 consecutive stroke patients (first cohort)
Recruitment of the validation cohort
The risk score was validated in an independent dataset.
For this validation set, 332 consecutive patients with stroke who were admitted to the stroke unit of the St Elisabeth hospital in Tilburg, The Netherlands, were investigated for the presence and risk factors of delirium. Criteria for stroke were neurological deficit of sudden onset lasting longer than 24 h for which no other cause than stroke could be found. Patients with ischaemic and haemorrhagic stroke were included. Patients with subarachnoid haemorrhage and transient ischaemic attack (TIA) were excluded. Patients had to be older than 18 years.
Of the 332 consecutive stroke patients, 59 were excluded. Twenty-two patients died before screening, or were excluded because death appeared imminent. Twenty-four patients were already discharged home before screening. Five patients refused informed consent, four had a severe language barrier, one had severe mental retardation, in 1 it was not possible to obtain an IQCODE, and two patients were transferred to another hospital. Hence, 273 patients were included in the analysis. Every patient was screened for delirium between days 2 and 4 after admission and a second time between days 5 and 7. Delirium was assessed with the Confusion Assessment Method (CAM).13 In the risk model, all patients with delirium on the first screening or on the second screening were included.
Testing the validity of the models
The risk models were analysed by using a receiver operating characteristics curve (ROC) by plotting the sensitivity versus the 1-specficity. The AUC of the ROC was calculated to measure the ability of the model to correctly classify subjects who will develop delirium. The maximum of the sum of sensitivity and 1-specificity is presented as the optimal cut-off point. Using the logistic regression model, individual risks according to presence or absence of the risk factors were calculated and included in a risk table to assess a clinical tool for assisting in identifying patients at risk for delirium. The analyses were performed with SPSS statistics 10 (IBM, Somers, New York, USA).
The study was approved by the medical ethical committee of the St Elisabeth Hospital, Tilburg. Informed consent was given by the patient or a caregiver.
Results
The Characteristics of the first cohort are given in table 1. The incidence of delirium was 11.8%.3 table 2 shows the baseline characteristics of the validation dataset. The prevalence of delirium was 15% which is slightly higher than in our first study, but not statistically significant (p=0.2).
Characteristics of the 273 consecutive stroke patients in the validation study
Results of the risk models in the first cohort
The AUC of the model with all significant risk factors was 0.85 (95% CI 0.80 to 0.90). Second, atrophy was replaced by age which resulted in an identical AUC of 0.85 (95% CI 0.80 to 0.90). Therefore, we used age instead of atrophy in all subsequent analyses. A risk score was calculated for each patient using the scores from table 3. When the concrete values were transformed to these risk scores, the AUC of the full model was 0.86 (95% CI 0.81 to 0.90). A cut-off value of 18 resulted in a sensitivity of 86% and a specificity of 74%. In the second model, we simplified the model by leaving out stroke localisation and IQCODE. The AUC of this model was 0.84 (95% CI 0.80 to 0.89). A cut-off value of 13 resulted in a sensitivity of 80% and a specificity of 75%. In the third model with age and NIHSS, the AUC was 0.80 (95% CI 0.75 to 0.85). In this model, the cut-off value of 10 resulted in a sensitivity of 79% and a specificity of 73%.
Risk score for delirium after stroke for model 1, model 2 and model 3
Results of the validity of the models
The same cut-off values of table 3 were used in this dataset. If model 1 was used, the AUC was 0.83 (95% CI 0.76 to 0.90). Cut-off values of 18 resulted in a sensitivity of 78% and a specificity of 77%. In model 2, the AUC was 0.83 (95% CI 0.77 to 0.90) with a sensitivity of 76% and a specificity of 81%, with the cut-of value of 13. In model 3, the AUC was 0.82 (95% CI 0.75 to 0.89) with a sensitivity of 73% and a specificity of 75%, with a cut-off value of 10.
Model 2 is more easy to use in clinical practice than model 1 and is slightly better than model 3 in terms of sensitivity and specificity. We therefore used this model in creating a risk table for all values of the included risk factors. Besides the individual calculated risks, we used the colour green to express a low risk for delirium (5%), orange for an intermediate risk (>5% <20%), and red for a high risk of delirium (>20%) (see table 4).
Risks calculated for values of the risk factors included in model 2
Discussion
We derived and validated a simple score to predict delirium after stroke in the first week of admission based on age, stroke severity, stroke subtype and infection. To the best of our knowledge, this is the first study with the specific aim to predict delirium in stroke patients.
We used data from our first prospective cohort study3 conducted in two large stroke units of two general hospitals. For validation of the risk score, we used an independent dataset from patients admitted on a large stroke unit of a general hospital. An advantage of our study is that we used information that is available in routine clinical practice. We believe that our findings are generalisable to most stroke units working according to international standards.
In the present study, we only used variables that are available at time of admission on the stroke unit which make high-risk patients easy to identify immediately, and subsequently preventive measures can be initiated. Non-pharamacological preventive measures reduce incidence and duration of delirium in patients aged 70 years or older admitted to a general medicine service.14 Additionally, it also facilitates pharmacological intervention in high-risk patients, prevention with haloperidol reduced duration and severity of delirium and shortened hospital length of stay in elderly hip-surgery patients at risk for delirium.15 In stroke patients, intervention studies on delirium are not available. Our model may help in identifying subjects at risk for delirium. The ultimate goal of risk assessment is prevention and early start of therapy which may result in lower incidence of delirium, a shorter duration and ultimately a better outcome. These goals have to be studied in future prospective research.
For critically ill patients a PRE-DELIRIC (PREdiction of DELIRium in ICu patients) model, based on variables important for critical ill patients was developed to predict the risk of delirium.16 We could not find such a model for stroke patients. As a consequence, there is no information available from previous studies that could guide us to choose values for the identification of a low, intermediate or a high risk of delirium. Therefore, the cut-off values that were chosen may be considered subjective. Since delirium after stroke has a worse prognosis,3 ,6 it is important to detect these patients as early as possible. Hence, for the risk table (table 4), we have chosen a relatively low cut-off value of 20%, as a high risk of delirium after stroke. This table can function as an easy instrument in the clinical setting to assess risk for delirium. Validation of this table in other populations of stroke patients is necessary.
A limitation of our study is that we used a static model. The health status of a stroke patient can change during their admission, but our model does not provide for this. On the other hand, in our prospective study,3 we found that most patients develop delirium in the first 2–4 days after admission, and few patients had delirium at time of the second screening that was not present at first screening. Furthermore, patients who were discharged before the first screening were not included in our study. The risk model, therefore, cannot be applied to this group but, probably, delirium is very rare in this group.
Another limitation could be the assessment of delirium with the CAM. The presence of aphasia, in which changes in cognition and behaviour are more difficult to assess, could result in an underestimation of delirium. In the first cohort study, we also used the Delirium Rating Scale (DRS),17 which quantifies multiple parameters affected by delirium, to minimise this bias. In the validation study, we found an almost similar incidence of delirium (15% vs 12%) indicating this effect is probable minimal.
In conclusion, the risk of delirium in the acute phase after stroke can be predicted with a simple model which is easy to use in routine clinical practice. It will facilitate early identification of high-risk patients admitted to a stroke unit. Further research is needed to confirm the diagnostic quality of our model, and to study whether early identification of patients at risk for delirium would result in a better outcome.
References
Footnotes
-
Contributors AWO was involved in acquisition of data, statistical analysis and interpretation of data and drafted the manuscript. PLMdeK participated in the design of the study and revision of the manuscript. JFvEckvdS was involved in acquisition of the data and revision of the manuscript. LJK participated in the design of the study and revision of the manuscript. GR was involved in acquisition of the data, statistical analysis, interpretation of the data and revision of the manuscript. All authors read and approved the final manuscript.
-
Competing interests None.
-
Ethics approval METC St Elisabeth Hospital.
-
Provenance and peer review Not commissioned; externally peer reviewed.
Linked Articles
- Editorial commentary