Validation of the Charlson Comorbidity Index for Predicting Functional Outcome of Stroke

Article Outline

• Abstract
• Methods
  • Participants
  • Measures
    • Comorbidity
  • Function
    • Functional recovery measure
    • Measure of functioning at 3 months
  • Statistical Analyses
    • Creation of algorithms: estimating the impact of comorbidities on function
    • Creation of algorithms: weights
    • Comparison of predictive capacity
    • Determination of whether the fit statistics were model-dependent
• Results
• Discussion
  • Study Limitations
• Conclusions
• Appendix
• References
• Copyright

Abstract 

Tessier A, Finch L, Daskalopoulou SS, Mayo NE. Validation of the Charlson Comorbidity Index for predicting functional outcome of stroke.

Objective

To determine whether a separate comorbidity index is needed to predict functional outcome after stroke, we compared the predictability of the Charlson Comorbidity Index (CMI) and the Functional Comorbidity Index (FCI) to that of a stroke-specific comorbidity index with function quantified with a measure developed with a Rasch model as outcome.

Design

Two prospective inception cohort studies, in 1996 through 1998 and in 2002 through 2005, with up to 9 months of follow-up.

Setting

Participants enrolled in 2 studies were recruited from acute care hospitals in the Montreal area.

Participants

For study one, 1027 persons with a first stroke discharged into the community were eligible; the 437 who were interviewed a second time at 6 months were included in the analysis. In study two, 235 of 262 patients with stroke were enrolled.

Interventions

Not applicable.

Main Outcome Measures

To predict recovery, we developed 3 stroke-specific comorbidity algorithms based on the estimated strength of association between comorbidities and stroke function. The various indices were compared on the basis of their predictive ability with a c statistic.

Results

In study 1, the c statistics were .758, .763, .766, and .763 for the stroke-specific algorithms 1, 2, and 3 and the CMI, respectively. In study 2, the c statistics were .680, .700, .704, .714, and .714 for the algorithms 1, 2, and 3, the CMI, and the FCI, respectively.

Conclusions

For purposes of case-mix adjustment, the CMI seems to be more than adequate.

Key Words: Comorbidity, Rehabilitation, Stroke

List of Abbreviations: CI, confidence interval, CMI, Charlson Comorbidity Index, FCI, Functional Comorbidity Index, IADLs, Instrumental Activities of Daily Living, ICD-9, International Classification of Diseases, 9th Revision, ICF, International Classification of Functioning, Disability and Health, OR, odds ratio, OARS-IADL, Older American Resources and Services–Instrumental Activity of Daily Living, RNLI, Reintegration to Normal Living Index, ROC, receiver operating characteristic, SF-36, Medical Outcomes Study 36-Item Short-Form Health Survey, WHO, World Health Organization

STROKE IS A LEADING CAUSE of serious, long-term disability in the United States.¹ In Canada, stroke affects over 50,000 persons a year, and the costs of treating stroke are substantial, initially because of acute care hospital costs and subsequently because of continued care for the 70% of persons who have suboptimal outcome.² Understanding the factors³ present at baseline that affect outcome and health care utilization in the short-term and long-term poststroke is crucial. First, interventions targeting these high-priority factors could be implemented. Second, in research, accounting for comorbidity as part of case-mix adjustment explains an important portion of the variation in the outcome.⁴, ⁵ Some of the main variables that can predict functional outcome are stroke severity, age, comorbidity, education, and depression.⁶, ⁷ Comorbidity is defined as a clinical condition that exists at time of the onset of the event and is likely to influence the outcome under study.³

There are various approaches to comorbidity assessment. The impact of comorbidity is often assessed by simply summing the number of conditions from a specific list.⁸ Using this approach, each condition is weighted equally without considering its relative impact. Other approaches incorporate disease impact, such as that developed by Charlson et al⁹ to predict mortality. Each condition is weighted on the basis of its association with 1-year mortality. The CMI,⁹ the Cumulative Illness Rating Scale,¹⁰ the Index of Co-Existent Disease,¹¹ and the Kaplan-Feinstein Index⁹ are examples of valid and reliable methods to measure comorbidity¹²; however, most of these indices focus on predicting mortality. Furthermore, it is unknown whether the influence of comorbidities on functional outcome among stroke survivors is similar to that on mortality. Therefore, it is justified to consider whether an index specific to functional outcome is needed.

There are also comorbidity indices developed specifically to predict function. The FCI was developed by Groll et al¹³ using a North American population affected mostly by orthopedic problems. Two other indices were developed with stroke survivors from Japan and Italy,¹⁴, ¹⁵ and 1 with Canadian institutionalized elderly.¹⁶ All these indices quantified function with a single measure such as the SF-36 physical function subscale, SF-36 physical component summary, FIM instrument, or Functional Autonomy Measurement System. There are numerous tests and indices that attempt to measure 1 or more components of function, and the outcome chosen for prediction may influence the weight placed on comorbidities. Modern psychometric theory based on item response theory and Rasch modeling has shown that it is possible to combine items from different tests and indices into a single measure that has interval-like properties and that covers the whole range of the construct.¹⁷ The mathematical properties of measure produced by these models are optimal for true measurement and adequate statistical analysis.¹⁸

At the moment, there is no agreed-on comorbidity index for predicting functional outcomes after stroke, nor is it clear that comorbidity indices predicting mortality will not predict function. The purpose of this study is to compare the predictability of stroke-specific comorbidity indices to the CMI⁹ and FCI,¹³ and identify whether a separate comorbidity index is needed to predict function after stroke.

Back to Article Outline

Methods 

First, we developed predictive models linking comorbidity to function; comorbidities were defined through a list of stroke-specific conditions and were weighted according to their impact on function. The predictive ability of the stroke-specific functional outcome comorbidity algorithms was then compared with the predictive ability of the CMI.⁹ The second step consisted of confirming the comparative predictability of stroke-specific algorithms to the CMI⁹ and FCI¹³ in a separate sample.

Participants 

The data for study 1 came from the Montreal Stroke Cohort, a prospective inception cohort aimed at estimating the extent of activity and participation at 6 months poststroke.² Participants were recruited from 10 acute care hospitals in the Montreal area from 1996 to 1998. Only patients with confirmed initial stroke were eligible; 98% of participants had a computed tomography scan. They were contacted in person during their hospital stay or by mail after discharge and asked to participate in the study (mean length of stay, 15.8±14.8d). In all, 1027 persons with a first stroke discharged into the community were eligible. The 612 patients who agreed were interviewed at baseline, and the 437 participants who were interviewed a second time by telephone at 6 months (between 3 and 9 months; mean, 190.3±57.5d poststroke) were included in this analysis. The persons interviewed on time did not differ in initial stroke severity nor on any of the measures of outcome from those interviewed later.

For replication purposes, data from another sample of patients hospitalized for stroke were analyzed. Persons were excluded if a diagnosis of stroke was not confirmed by imaging or clinical examination within 24 to 72 hours. In addition, persons with the following conditions were excluded: transient ischemic attacks, hemiplegia resulting from nonvascular causes, subdural hematoma or subarachnoid hemorrhage, severe diseases such as end-stage cancer, pulmonary, cardiac, or renal disease, and severe cognitive impairments. Of the 262 persons eligible, 235 (89%) elected to participate. Their functional level was assessed within 3 days poststroke and at 3 months at the hospital by trained health professionals. Both studies had ethics approval from McGill University Institutional Review Board and from the research ethics committees of all participating hospitals.², ¹⁹

Measures 

In study 1, comorbidities were ascertained from the medical chart; a list of 16 comorbidities commonly encountered among persons with acute stroke were examined and coded according to the ICD-9. The list of conditions included only those prevalent before the stroke so as not to include health conditions that arose from the stroke. This list had previously been compiled to meet the needs of comorbidity adjustment for research purposes in our research unit.¹⁹ These conditions included cardiovascular, respiratory, neurologic, musculoskeletal, metabolic, ophthalmic, and other diseases. Cerebrovascular diseases and hemiplegia were excluded because we considered only people with a first stroke. Comorbidity was also ascertained with the CMI⁹ and FCI.¹³ The FCI has been developed with physical function as the outcome. The CMI and FCI correlate fairly strongly (Spearman ρ=.62; P<.001). The FCI is significantly associated with the physical function subscale and physical component subscale scores of the SF-36.²⁰

Function 

We defined function according to the WHO's ICF model.²¹ Function is an umbrella term composed of all positive aspects of body functions and structures, activities, and participation, whereas disability is composed of the negative aspects. For both studies, the outcome was developed through a partial credit Rasch modeling using RUMM 2020 software.^a

In study 1, we analyzed 39 items from 5 tests and indices. After item reduction, the resulting functional recovery measure¹⁹ consisted of 12 items from 4 tests and indices (SF-36 physical function scale, RNLI, OARS-IADL, Barthel Index) encompassing IADLs, physical functioning, and participation (appendix 1). Rasch analysis ordered the items by difficulty from the easiest item, “Can you get to the toilet independently?” to the hardest, “Does your health limit you in performing vigorous activities?” The items score was 0, 1, or 2, and the total score ranged from 0 to 18 and has excellent internal reliability (person hierarchy reliability, .97). The item and person fit criteria were the standardized fit residuals ±2. The global model fit criteria was a nonsignificant chi-square or F statistic.

For study 2, another measure of functioning was developed through Rasch analysis by combining 204 items from 14 tests and indices evaluating capacity through direct observation and performance using a person's rating of difficulty. It has been found that the combination of self-reported and performance-based measures refined prognostic information.²² The resulting measure of functioning at 3 months was composed of 52 items (including the 12 from the functional recovery measure¹⁹) from 7 tests and indices (SF-36, Chedoke-McMaster Stroke Assessment, Stroke Rehabilitation Assessment of Movement, Berg Balance Scale, Stroke Impact Scale, Preference-Based Stroke Index, 5-m walking speed) and encompassed limb movements, balance, mobility, activities of daily living, and life role participation (see appendix 2 for list of the items and indices). The item difficulty ranged from the easiest, “facilitate finger flexion,” to the hardest, “walking speed >1.3m/s,” with the item of average difficulty being “walk 3 steps sideways.” The item score was 0 or 1. The total score ranged from 0 to 51 and had excellent internal reliability (person hierarchy separation, .99). Both measures met the requirements of Rasch model, confirming that the items adequately define the continuum of function. Unidimensionality was ascertained by principal components analysis of the person item residuals with the criterion that the variance of the first principle component be less than 10%.

Statistical Analyses 

Stroke-specific algorithms were created by estimating the strength of the association between the comorbidity index and functional outcome in the first study. Despite the fact that function here could be considered on an interval-like scale because of the Rasch model, linear regression was not used to identify the predictive ability of comorbidity because the CMI⁹ was based on the strength of the prediction of mortality and morbidity (dichotomous outcomes) after hospitalization, and the weights assigned to comorbidities were based on the OR. To use OR-based weights, function was hierarchically categorized according to its distribution and analyzed with an ordinal regression model, which yields OR. The continuation odds model was used, which creates a summary OR across all possible consecutive cut-points of functional outcome using maximum likelihood methods. The OR is interpreted as the impact of comorbidity on the probability of a worse functional outcome regardless of how worse was defined (constant across the cut-points).

Three stroke-specific algorithms were created using different weights for the comorbidities. The weights for the algorithms were based on the magnitude of the point estimate of the summary OR. The weights for algorithm 1 were attributed using the criteria applied by Charlson et al⁹; conditions with an OR less than 1.2 were dropped (weighted 0), between 1.2 and 1.49 were weighted 1, between 1.5 and 2.49 were weighted 2, between 2.5 and 3.49 were weighted 3, and greater than 3.5 were weighted 6. To include more conditions, 2 other algorithms were created with different criteria to allocate weights. The weights for algorithm 2 were as follows: OR <.99, weighted 0; OR between 1 and 1.19, weighted 1; and OR >1.2, weighted 2. The weights for algorithm 3 were as follows: OR <.99, weighted 1; OR between 1 and 1.19, weighted 2; and OR >1.2, weighted 3. A comorbidity score was calculated using each algorithm. A CMI⁹ score was also computed for each sample. Because of incomplete information, a score on the FCI¹³ could be calculated only for study 2.

The c statistic was used to compare the predictive ability of the different methods. The c statistic quantifies the area-under-the-ROC curve—a measure of the discrimination ability of various predictors of an event or outcome. The c statistic represents the proportion of all possible pairs of observations where the member with the outcome has a higher predicted value than the member without it. Values range from 0 to 1, with 1 indicating perfect prediction and 0.5 being chance prediction; values greater than 0.7 indicate acceptable prediction; greater than 0.8, excellent prediction.⁴ The c statistics were derived from the ordinal regression models and calculated for each scoring method. Spearman correlations were also calculated for each index against the ordinal outcomes.

We allowed the outcomes to be first dichotomous, and used logistic regression to correspond to how the CMI⁹ was developed; and second continuous, and used linear regression because this was the model used to develop the FCI.¹³ The R² statistic was calculated from a logistic regression. It compares 2 averages: the average predicted probability for those with and those without the outcome. As long as the average predicted probability is greater for those with the outcome than those without, this is evidence of predictability. The magnitude has no real meaning because with dichotomous variables, the R² statistic is always very small. However, it is useful for comparing across models. We used the maximum rescaled R² statistic from a logistic regression, where 0 reflects no prediction and 1 reflects perfect prediction. Maximum rescaled R² statistic is similar to an R² from linear regression, which we also calculated. The models were not adjusted for confounders; comorbidity is most often used as a confounder, not as a variable under study. The statistical analyses were performed using the macro published by Scott et al²³ with SAS software.^b

Back to Article Outline

Results 

Selected characteristics of both samples are presented in table 1. The mean age of the participants was 70 and 72 years for studies 1 and 2, respectively. They were predominantly men, with ischemic stroke. About 80% had a moderate or severe stroke, and the main discharge destination was home at the time of follow-up after 14.7 and 16.9 days in the hospital for studies 1 and 2, respectively. Hypertension and diabetes were among the most common conditions. People in study 1 had on average 4 comorbidities (maximum, 12); for study 2, the mean was 2 (maximum, 5). The CMI score was 1.1 and 1.5 for studies 1 and 2, respectively. Their SF-36 physical function scale was not available for study 1 and was 74.6 for study 2.

Table 1. Characteristics of the Participants in the 2 Studies

Characteristics	Study 1 (n=437)	Study 2 (n=235)
Mean age ± SD (y)	69.7±12.5	71.6±20.6
Men/women, n (%)	246(56)/191(44)	146(62)/89(38)
Type of stroke, n (%)
Ischemic	379(87)	202(86)
Hemorrhagic	56(13)	33(14)
Stroke severity, n (%)
Very mild	0(0)	42(18)
Mild	61(14)	52(22)
Moderate	188(43)	96(41)
Severe	188(43)	45(19)
Comorbidities, n (%)
Hypertension	242(57)	170(72)
Arthritis	153(36)	20(9)
Hearing impairment	109(25)	11(5)
Cataract	101(24)	18(8)
Bronchitis	101(24)	17(7)
Myocardial infarct	89(21)	34(14)
Angina	84(20)	16(7)
Diabetes	81(19)	60(26)
Cancer	40(9)	51(22)
Mean no. of comorbidities ± SD	4.1±2.3	2.0±1.0
Mean CMI score ± SD	1.1±1.2	1.5±1.5
Mean physical function before stroke ± SD	NA	74.6±29.4
Mean length of stay in acute care ± SD	14.7±12.7	16.9±20.9
Discharge destination, n (%)
Rehabilitation	271(44)	21(9)
Home	326(53)	181(77)
Long-term care	0(0)	31(13)

NOTE. The comorbidities are listed according to their prevalence in study 1.

Abbreviation: NA, not available.

Table 2 presents the level of functioning at the time of discharge of the participants as indicated by the Rasch-modeled functional outcomes and each of the constituents of these outcomes. For study 1, the median score on the functional recovery measure was 72.2 out of 100; for study 2, the median score on the measure of functioning at 3 months was 36 out of 100. Considering the constituents of the measures, most participants function reasonably well on basic activities, with scores in the higher range on most scales. The average score on physical function was relatively low for both samples (study 1 median, 70; study 2 median, 50), but 50% had a score of 70 or higher. The norm for this age group is 75.7.²⁴ The correlations between the functional recovery measure, the measure of functioning at 3 months, and the individual measures used to derive the outcome were all within the moderate to high range.

Table 2. Level of Functioning of the Participants and Correlations Between the Functional Outcomes and Their Constituents

Study	Median (Interquartile)	r
Study 1 (n=437)
Functional recovery measure	72.2(50.0)
Constituents
Barthel Index (0–100)	100.0(10.0)	0.59
OARS-IADL (0–14)	13.0(4.0)	0.76
RNLI (22–0)	3.0(8.0)	−0.69
Physical function (0–100)	70.0(45.0)	0.68
Study 2 (n=235)
Measure of functioning at 3mo	36.0(19.0)
Constituents
Physical function (0–100)	50.0(55.0)	0.82
CMSA (1–42)	37.0(8.0)	0.93
STREAM (0–100)	93.9(18.8)	0.87
BBS (0–56)	50.0(13.0)	0.85
SIS-16 (0–100)	75.0(38.4)	0.89
PBSI (0–100)	71.9(31.0)	0.80
Gait speed	5.0(2.0)	0.85

NOTE. All correlations are significant at P<.001.

Abbreviations: BBS, Berg Balance Scale; CMSA, Chedoke-McMaster Stroke Assessment; PBSI, Preference-Based Stroke Index; STREAM, Stroke Rehabilitation Assessment; SIS-16, Stroke Impact Scale−16-question version.

Table 3 presents the comorbidities with their estimated ORs from study 1 along with the different weights applied by each index. Weights for comorbidities not included in the indices were assigned 0 by default (blank on the table). For example, Parkinson's disease increases the risk of having a poorer functional outcome by 1.43, which corresponds to a weight of 1 in algorithm 1, 2 in algorithm 2, and 3 in algorithm 3. In the CMI and FCI, Parkinson's disease is not included and hence, by default, was assigned a weight of 0. In table 3, the statistical properties of each of the indices are also shown. The number of comorbidities in each index varied from 7 to 19. The theoretical score range of the CMI was the widest, whereas algorithm 1 had the narrowest range. The average value of the indices varied from 0.2 to 5.2. Thirty-seven percent of scores on the CMI are either the maximum or the minimum score.

Table 3. Comparison of Stroke-Specific Algorithms, CMI, and FCI

Comorbidity	OR	Stroke-Specific Algorithms			CMI	FCI
		1	2	3
Parkinson's disease	1.43	1	2	3	NE	NE
Chronic obstructive pulmonary disease/emphysema	1.36	1	2	3	1	1
Thyroid disease	1.34	1	2	3	NE	NE
Glaucoma	1.33	1	2	3	NE	NE
Liver disease	1.30	1	2	3	1	NE
Asthma	1.26	1	2	3	NE	1
Gastrointestinal ulcer	1.24	1	2	3	1	NE
Cancer	1.15	0	1	2	2	NE
Diabetes mellitus	1.15	0	1	2	1	1
Hypertension	1.13	0	1	2	NE	NE
Myocardial infarction	1.05	0	1	2	1	1
Arthritis or connective tissue disease	1.03	0	1	2	1	1
Angina	1.01	0	1	2	NE	1
Bronchitis	0.97	0	0	1	NE	NE
Impaired hearing	0.94	0	0	1	NE	1
Cataract	0.82	0	0	1	NE	NE
Congestive heart disease	NE	NE	NE	NE	1	1
Peripheral vascular disease	NE	NE	NE	NE	1	1
Dementia	NE	NE	NE	NE	1	NE
Cerebrovascular disease	NE	NE	NE	NE	1	1
Hemiplegia	NE	NE	NE	NE	NE	NE
Moderate or severe renal disease	NE	NE	NE	NE	2	NE
Diabetes with end-organ damage	NE	NE	NE	NE	2	NE
Leukemia	NE	NE	NE	NE	2	NE
Lymphoma	NE	NE	NE	NE	2	NE
Moderate or severe liver disease	NE	NE	NE	NE	3	NE
Metastatic solid tumor	NE	NE	NE	NE	6	NE
Acquired immune deficiency virus	NE	NE	NE	NE	6	NE
Osteoporosis	NE	NE	NE	NE	NE	1
Neurologic disease	NE	NE	NE	NE	NE	1
Upper gastrointestinal disease	NE	NE	NE	NE	NE	1
Anxiety disorder or depression	NE	NE	NE	NE	NE	1
Visual impairment	NE	NE	NE	NE	NE	1
No. of conditions		7	13	16	18	15
Theoretical range		0–7	0–20	0–36	0–37	0–15
Study 1
Mean ± SD		0.5±0.8	2.5±2.2	5.2±4.1	1.1±1.2
Median		0	2.0	5.0	1.0
Observed range		5	12	23	6
% score max/min		0/67	0/18	0/12	0/37
Study 2
Mean ± SD		0.2±0.4	1.8±1.3	3.6±2.4	1.5±1.5	2.5±2.2
Median		0.0	2	3	1.0	2
Observed range		2	7	12	6	12
% score max/min		1/86	0/11	0/11	1/30	0/18

NOTE. Algorithm 1: OR <1.2 weighted as 0; OR 1.2 to 1.49 weighted as 1; OR 1.5 to 2.49 weighted as 2; OR 2.5 to 3.49 weighted as 3; OR >3.5 weighted as 6. Algorithm 2: OR <.99 weighted as 0; OR 1 to 1.19 weighted as 1; OR >1.2 weighted as 2. Algorithm 3: OR <.99 weighted as 1; OR 1 to 1.19 weighted as 2; OR >1.2 weighted as 3.

Abbreviations: NE, not estimated; % score max/min, percentage of scores that are either the maximum or minimum score.

Table 4 presents model fit statistics for different indices and for each of the statistical models (ordinal, logistic, linear). The ability of each model to predict functional outcome was compared with the c statistic. All indices had comparable c statistics. All indices except algorithm 1 were significantly correlated with function.

Table 4. Model Fit Statistics of the Indices

Study	c statistics	Spearman ρ	R2	Maximum Rescale R2	R2
			From Logistic Regression	From Logistic Regression	From Multiple Linear Regression
Study 1
Algorithm 1	.758	−.03	.002	.003	.014
Algorithm 2	.763	−.08	.003	.004	.010
Algorithm 3	.766	−.13⁎	.007	.010	.002
CMI	.763	−.11⁎	.004	.006	.003
Study 2
Algorithm 1	.680	.00	.002	.002	.000
Algorithm 2	.700	−.15⁎	.019	.025	.017
Algorithm 3	.704	−.17⁎	.027	.036	.026
CMI	.714	−.20⁎	.046	.061	.041
FCI	.714	−.27⁎	.043	.058	.078

The R² and the maximum rescaled R² obtained from logistic regression showed a minimal difference between the indices and indicated that only a small proportion of variability in the outcomes was explained by comorbidity alone, which was expected, because a huge amount of variability in outcome is more often explained by the extent of the damage or disease process, whereas comorbidity tends to account for what is unexplained by known factors. When the outcome was treated as a continuous variable in a linear regression, the index with the highest proportion of variance explained was the FCI¹³ (7.75%), only slightly higher than the CMI (4.14%). The results of the different fit statistics used to compare the indices are consistent (see table 4).

Back to Article Outline

Discussion 

Comorbidities can compromise recovery from a stroke. The CMI⁹ is widely used in studies predicting function,²⁵, ²⁶, ²⁷ although its use can be questioned because it was developed to predict mortality and morbidity. To account better for the effect of comorbidities when assessing rehabilitation programs, several comorbidity indices have been recently developed to predict functional outcomes.¹³, ¹⁴, ¹⁵, ¹⁶, ²⁰ To verify the need for a stroke-specific comorbidity index, we developed 3 stroke-specific comorbidity algorithms.

In this study, we defined function specifically for stroke. How function is defined may have important implications for determining the contribution of a specific comorbidity index. We used a conceptually driven measure of function that reflects not only capacity but also the patient's perspective on performance and goes beyond physical functioning. The interval-like measures were developed integrating the WHO's ICF framework, and the items were quantified and ordered using Rasch analysis (functional recovery measure and measure of functioning at 3 months).

In addition to the diversity of outcome being predicted, the statistical methods used in the literature to estimate the predictive values of the existing indices also varied. For example, Liu et al,¹⁴ using simple correlation, found a slightly higher value for their index (ρ=−.51) than the CMI (ρ=−.23). Bravo et al,¹⁶ using linear regression and R², showed that their index explained 10% more variance in function than the CMI (Bravo index: R²=.26; 95% CI, .18–.35; CMI: R²=.16; 95% CI, not reported). The Bravo index also had a 7% better predictive power (ROC=.86; 95% CI, .80–.90) than the CMI (ROC=.79; 95% CI, not reported). Groll et al,¹³ using linear models, found significant correlations between the FCI and SF-36 subscales, with ρ ranging from −.296 to −.548, whereas the correlations with the CMI were not significant and ranged from −.112 to −.249. Also reported was a significant relation between the FCI and physical function (and physical health) in a multiple linear regression controlling for other factors. The CMI was not found to be related to function. The proportion of variance in physical function explained by the FCI alone was not reported; the proportion explained with 9 other variables in the model ranged from .14 to .39, depending on outcome and period.

Our results suggest that the CMI⁹ provides as good a prediction of long-term functional outcome as other indices developed specifically to predict function or developed specifically for the stroke population. High comorbidity as measured by the CMI has also been associated with poorer outcome in the short term after acute stroke (OR=1.36; P=.038).²⁷ In our samples, the CMI showed acceptable to excellent prediction for function in the community (c=.763), and slightly higher prediction than the stroke-specific algorithms and the FCI (c=.714). Our results support that an index developed to predict mortality can be used to predict function; similarly, Gabriel et al²⁸ found that a comorbidity index developed to predict function could also predict mortality. This suggests that function and mortality are related and that there might be the possibility for using interchangeably these 2 types of comorbidity indices.

The small proportion of variance explained by the indices needs to be considered in the context of the data used—those arising from research studies that had inclusion and exclusion criteria. Variance explained depends on the range of both the outcome and the exposure (comorbidity). The outcome covers the full range of the construct (10 logits where 1 logit is similar to an SD); the exposure, comorbidity, has a theoretical range of 0 to 37 in the case of the CMI, but in this sample, the observed range was 6. Comorbidity adjustment is used to compare across settings, and in reality, the range of comorbidity may greatly exceed that observed here. As a result, in real-life clinical situations, the proportion of variance in outcome explained by comorbidity may be larger, with a larger range of comorbidity.

Another source of variation in quantifying the predictive ability of different comorbidity indices is the statistical model used to link comorbidity to outcome. We compared predictive ability using logistic, ordinal, and linear models to correspond to treating the outcome as binary, ordered categories, and continuous. The CMI performed as well as, if not better than, the other indices, independent of the model used.

The algorithms were developed with data derived from 2 samples of stroke survivors at 3 to 9 months. This is an optimal period because (1) function tends to stabilize after 3 months,²⁹ (2) many aspects of function cannot be assessed until the person fully experiences community challenges, and (3) there is a great deal of variability in the recovery process early on that would affect the ability of the comorbidity index to predict a moving target. Therefore, targeting this period is appropriate. The samples are similar to stroke populations in the literature.²⁵, ³⁰, ³¹ The sample for study 1 was restricted to community dwellers, but this sample is relevant because a large proportion of the overall stroke population (69%–87% of stroke survivors, depending on the study) are discharged back to the community,³² and this proportion is rising with time.³³ The need to predict function is most important for this group. The information is used to identify need for services and to allocate adequate resources not only to improve function but also to manage comorbidities.

Study Limitations 

This study has a number of strengths and limitations. Both studies were inception cohorts, and there was substantial documentation of the study participants.², ¹⁹ The sample sizes were relatively large for clinical studies of stroke. The outcomes were comprehensive, with good mathematical properties. However, there is no standard list of conditions to assess comorbidity. The list that we used is not inclusive, and the exclusion of conditions such as obesity, depression, or dementia before the occurrence of stroke may have influenced the predictive power of the indices; however, the CMI⁹ includes dementia. Furthermore, these comorbidities are relatively infrequent in the stroke population (11%–18% have obesity; 8%–9% have depression; 1%–2% have dementia)¹⁴, ²⁵, ³⁴, ³⁵ and inconsistently recorded in the medical record. Although there were 2 different formats of the outcome, because of the conceptualization and the mathematical approach (Rasch modeling), the measure of functioning at 3 months, used in study 2, is compatible with the functional recovery measure used in study 1. These measures are comparable because similar items were included.

Back to Article Outline

Conclusions 

The CMI, although originally developed for predicting mortality and subsequently validated for some indicators of morbidity, predicted function as well as, if not better than, stroke-specific comorbidity indices. For purposes of case-mix adjustment, the CMI seems to be more than adequate for explaining variability in functional outcome poststroke; it is widely known and easily obtained from either clinical or administrative data. Use of standard assessments for comorbidity will facilitate comparisons across studies and settings.

Suppliers

Back to Article Outline

Appendix 

Appendix 1. THE ITEMS OF THE FUNCTIONAL RECOVERY MEASURE USED IN STUDY 1 AND THEIR LEVEL OF DIFFICULTY

NOTE. Items are from the following indices: SF-36 physical function scale, RNLI, OARS-IADL, and Barthel Index.

Abbreviations: N, no; Y, yes.

Appendix 2. THE ITEMS OF THE MEASURE OF FUNCTIONING AT 3 MONTHS USED WITH STUDY 2

NOTE. The items are ordered by difficulty. Boldface represents those for which persons rate their difficulty in performing the activity; nonshaded items are those items for which performance is observed and rated. Items are from the following indices: SF-36, Chedoke-McMaster Stroke Assessment, Stroke Rehabilitation Assessment of Movement, Berg Balance Scale, Stroke Impact Scale, Preference-Based Stroke Index, and 5-m walking speed.

Back to Article Outline

References 

1. Rosamond W, Flegal K, Friday G, et al. Heart disease and stroke statistics—2007 update: a report from the American Heart Association Statistics Committee and Stroke Statistics Subcommittee. Circulation. 2007;115:e69–e171
   • View In Article
   • CrossRef
2. Mayo NE, Wood-Dauphinee S, Cote R, Durcan L, Carlton J. Activity, participation, and quality of life 6 months poststroke. Arch Phys Med Rehabil. 2002;83:1035–1042
   • View In Article
   • Abstract
   • Full Text
   • Full-Text PDF (84 KB)
   • CrossRef
3. Powell H, Lim LL, Heller RF. Accuracy of administrative data to assess comorbidity in patients with heart disease: an Australian perspective. J Clin Epidemiol. 2001;54:687–693
   • View In Article
   • Abstract
   • Full Text
   • Full-Text PDF (61 KB)
   • CrossRef
4. Schneeweiss S, Seeger JD, Maclure M, Wang PS, Avorn J, Glynn RJ. Performance of comorbidity scores to control for confounding in epidemiologic studies using claims data. Am J Epidemiol. 2001;154:854–864
   • View In Article
   • MEDLINE
   • CrossRef
5. Mayo NE, Nadeau L, Levesque L, Miller S, Poissant L, Tamblyn R. Does the addition of functional status indicators to case-mix adjustment indices improve prediction of hospitalization, institutionalization, and death in the elderly?. Med Care. 2005;43:1194–1202
   • View In Article
   • MEDLINE
   • CrossRef
6. Stineman MG, Maislin G, Williams SV. Applying quantitative methods to the prediction of full functional recovery in adult rehabilitation patients. Arch Phys Med Rehabil. 1993;74:787–795
   • View In Article
   • MEDLINE
   • CrossRef
7. Aprile I, Piazzini D, Bertolini C, et al. Predictive variables on disability and quality of life in stroke outpatients undergoing rehabilitation. Neurol Sci. 2006;27:40–46
   • View In Article
   • MEDLINE
8. Guralnik JM. Assessing the impact of comorbidity in the older population. Ann Epidemiol. 1996;6:376–380
   • View In Article
   • Abstract
   • Full-Text PDF (634 KB)
   • CrossRef
9. Charlson ME, Pompei P, Ales KL, MacKenzie CR. A new method of classifying prognostic comorbidity in longitudinal studies: development and validation. J Chronic Dis. 1987;40:373–383
   • View In Article
   • MEDLINE
   • CrossRef
10. Giaquinto S. Comorbidity in post-stroke rehabilitation. Eur J Neurol. 2003;10:235–238
    • View In Article
    • MEDLINE
    • CrossRef
11. Imamura K, McKinnon M, Middleton R, Black N. Reliability of a comorbidity measure: the Index of Co-Existent Disease (ICED). J Clin Epidemiol. 1997;50:1011–1016
    • View In Article
    • Abstract
    • Full-Text PDF (698 KB)
    • CrossRef
12. de Groot V, Beckerman H, Lankhorst GJ, Bouter LM. How to measure comorbidity: a critical review of available methods. J Clin Epidemiol. 2003;56:221–229
    • View In Article
    • Abstract
    • Full Text
    • Full-Text PDF (91 KB)
    • CrossRef
13. Groll D, To T, Bombardier C, Wright J. The development of a comorbidity index with physical function as the outcome. J Clin Epidemiol. 2005;58:595–602
    • View In Article
    • Abstract
    • Full Text
    • Full-Text PDF (196 KB)
    • CrossRef
14. Liu M, Domen K, Chino N. Comorbidity measures for stroke outcome research: a preliminary study. Arch Phys Med Rehabil. 1997;78:166–172
    • View In Article
    • Abstract
    • Full-Text PDF (994 KB)
    • CrossRef
15. Ferriero G, Franchignoni F, Benevolo E, Ottonello M, Scocchi M, Xanthi M. The influence of comorbidities and complications on discharge function in stroke rehabilitation inpatients. Eura Medicophys. 2006;42:91–96
    • View In Article
    • MEDLINE
16. Bravo G, Dubois MF, Hebert R, De Wals P, Messier L. A prospective evaluation of the Charlson Comorbidity Index for use in long-term care patients. J Am Geriatr Soc. 2002;50:740–745
    • View In Article
    • MEDLINE
    • CrossRef
17. Wright BD, Mok M. An overview of the family of Rasch measurement models. In:  Smith RM,  Smith EV editor. Introduction to Rasch measurement. Maple Grove: JAM Pr; 2004;p. 1–24
    • View In Article
18. Bezruczko N. Rasch model essentials: Rasch measurement in health sciences. In: Maple Grove: JAM Pr; 2005;p. 35–71
    • View In Article
19. Finch L, Higgins J, Wood-Dauphinee S, Mayo NE. Development of a measure of functioning for stroke recovery (The functional recovery measure). Disabil Rehabil. 2007;15:1–16
    • View In Article
    • MEDLINE
    • CrossRef
20. Groll D, Heyland D, Caeser M, Wright J. Assessment of long-term physical function in acute respiratory distress syndrome (ARDS) patients: comparison of the Charlson Comorbidity Index and the Functional Comorbidity Index. Am J Phys Med Rehabil. 2006;85:574–581
    • View In Article
    • MEDLINE
    • CrossRef
21. World Health Organization. International classification of functioning, disability and health: ICF. Geneva: WHO; 2001;
    • View In Article
22. Reuben DB, Seeman TE, Keeler E, et al. Refining the categorization of physical functional status: the added value of combining self-reported and performance-based measures. J Gerontol A Biol Sci Med Sci. 2004;59:M1056–M1061
    • View In Article
    • CrossRef
23. Scott SC, Goldberg MS, Mayo NE. Statistical assessment of ordinal outcomes in comparative studies. J Clin Epidemiol. 1997;50:45–55
    • View In Article
    • Abstract
    • Full-Text PDF (1286 KB)
    • CrossRef
24. Hopman WM, Towheed T, Anastassiades T, et al. Canadian normative data for the SF-36 health survey. CMAJ. 2000;163:265–271
    • View In Article
    • MEDLINE
25. Fischer U, Arnold M, Nedeltchev K, et al. Impact of comorbidity on ischemic stroke outcome. Acta Neurol Scand. 2006;113:108–113
    • View In Article
    • MEDLINE
    • CrossRef
26. Kelly PJ, Stein J, Shafqat S, et al. Functional recovery after rehabilitation for cerebellar stroke. Stroke. 2001;32:530–534
    • View In Article
    • CrossRef
27. Goldstein LB, Samsa G, Matchar D, Horner R. Charlson index comorbidity adjustment for ischemic stroke outcome studies. Stroke. 2004;35:1941–1945
    • View In Article
    • CrossRef
28. Gabriel SE, Crowson CS, O'Fallon WM. A comparison of two comorbidity instruments in arthritis. J Clin Epidemiol. 1999;52:1137–1142
    • View In Article
    • Abstract
    • Full Text
    • Full-Text PDF (164 KB)
    • CrossRef
29. Jorgensen HS, Nakayama H, Raaschou HO, Vive-Larsen J, Stoier M, Olsen TS. Outcome and time course of recovery in stroke, part II: time course of recovery (The Copenhagen Stroke Study). Arch Phys Med Rehabil. 1995;76:406–412
    • View In Article
    • Abstract
    • Full-Text PDF (666 KB)
    • CrossRef
30. Massucci M, Perdon L, Agosti M, et al. Prognostic factors of activity limitation and discharge destination after stroke rehabilitation. Am J Phys Med Rehabil. 2006;85:963–970
    • View In Article
    • MEDLINE
    • CrossRef
31. Lai SM, Duncan PW, Keighley J. Prediction of functional outcome after stroke: comparison of the Orpington Prognostic Scale and the NIH Stroke Scale. Stroke. 1998;29:1838–1842
    • View In Article
    • MEDLINE
    • CrossRef
32. Ottenbacher KJ, Smith PM, Illig SB, Linn RT, Ostir GV, Granger CV. Trends in length of stay, living setting, functional outcome, and mortality following medical rehabilitation. JAMA. 2004;292:1687–1695
    • View In Article
    • CrossRef
33. Mayo NE, Nadeau L, Daskalopoulou SS, Cote R. The evolution of stroke in Quebec: a 15-year perspective. Neurology. 2007;68:1122–1127
    • View In Article
    • CrossRef
34. Razinia T, Saver JL, Liebeskind DS, Ali LK, Buck B, Ovbiagele B. Body mass index and hospital discharge outcomes after ischemic stroke. Arch Neurol. 2007;64:388–391
    • View In Article
    • MEDLINE
    • CrossRef
35. Salaycik KJ, Kelly-Hayes M, Beiser A, et al. Depressive symptoms and risk of stroke: the Framingham Study. Stroke. 2007;38:16–21
    • View In Article
    • CrossRef

• a RUMM Laboratory Ltd. Available at: http://www.rummlab.com.au. Accessed March 26, 2008.
• b Version 9.1; SAS Institute, 100 SAS Campus Dr, Cary, NC 27513-2414.