Modified Rankin Scale

Modified Rankin Scale in Stroke Trials and Economic Models

The modified Rankin Scale grades how far a stroke has left a person dependent on others, using a single global judgement instead of a list of items. Acute stroke trials commonly report the grade at about 90 days, and many stroke cost-effectiveness models use the same grades, alone or grouped, as the health states to which utilities, care costs and death rates are attached. The grade a patient receives therefore drives both sides of the incremental cost-effectiveness ratio, which is why rater agreement, the choice between dichotomised and ordinal analysis and the source of utilities for each grade all matter to the result. The lower-case abbreviation mRS refers to the scale, whereas MRS in the economics literature usually means the marginal rate of substitution, an unrelated idea.

Grades of the scale and how a rating is reached

The modern scale descends from a grading of disability that Rankin published in the Scottish Medical Journal in 1957. In 1988 van Swieten and colleagues reported agreement between observers on the modified version, which has six grades numbered 0 to 5. Trials add a seventh category, grade 6, for death, so outcome data usually run 0 to 6.

The rating is a clinician-reported outcome. A clinician or trained assessor interviews the patient and assigns one grade that summarises the person's overall dependence. There are no item scores to add. The 1988 table describes the grades as shown below, and the Stroke Therapy Academic Industry Roundtable (STAIR) labels grade 1 symptomatic but nondisabled.

Saver and colleagues reported in 2021 that recent trials had described mRS outcomes with many different and often conflicting labels, and the STAIR XI consensus proposed standard names for 23 distinct numeric outcomes. Its cut points include 0 to 5 against 6, which separates survival from death, and 0 to 4 against 5 to 6, which separates patients who do not need constant care from those who need it or have died. Under the 1988 wording, a split between grades 0 to 3 and 4 to 6 separates patients who can walk without assistance from those who cannot or have died. The NICE stroke guideline NG128 used the split between grades 0 to 2 and 3 to 6 as a critical outcome when it reviewed thrombectomy.

Agreement between raters and what misgrading does to a model

A single global grade depends on the rater's judgement, so agreement between raters has been studied since the scale was modified. In the 1988 study ten senior neurologists and 24 residents from two centres took part, and 100 patients were each interviewed separately by two physicians. The two agreed exactly in 65 patients, differed by one grade in 32 and by two grades in 3. Unweighted kappa across all pairs was 0.56, while the weighted kappa with quadratic weights was 0.91, because most disagreements were between neighbouring grades. The kappa statistic chosen therefore changes how reliable the scale appears.

Banks and Marotta reviewed the evidence in 2007, selecting 50 of 224 articles found by their search for detailed assessment. They concluded that inter-rater reliability is moderate and improves with structured interviews (kappa 0.56 against 0.78), that test-retest reliability of 0.81 to 0.95 had been reported, and that multiple types of evidence attest to the validity and reliability of the scale. Wilson and colleagues compared multiple raters in 2005 and found overall agreement of 43 per cent (kappa 0.25, weighted kappa 0.71) with the usual scale, against 81 per cent (kappa 0.74, weighted kappa 0.91) with a structured interview.

For an economic model the concern is misclassification between neighbouring grades. As an illustration of non-differential misclassification, if raters misgrade patients at random and to the same extent in both arms, the observed distributions across grades look more alike than the true ones. When grades carry different utilities and care costs, the effect a model carries forward would then tend to be smaller and less precise than the true effect, which is why the way the mRS was assessed in a trial matters to a model built on it.

Dichotomised and ordinal analysis in stroke trials

Many stroke trials once reported the proportion of patients in a good outcome group, such as grades 0 to 2, and compared that proportion between arms. Dichotomising throws away the movement of patients between grades within each group. An ordinal or shift analysis compares the whole distribution across grades, so an improvement of one grade among the most dependent patients counts as a benefit in the same way as an improvement of one grade among the least dependent.

The Optimising Analysis of Stroke Trials (OAST) Collaboration pooled individual patient data from 47 trials with 54,173 patients in 55 datasets and compared statistical tests for functional outcomes including the mRS and the Barthel Index. It found that methods using the ordered nature of the outcome, such as ordinal logistic regression, were statistically more efficient than methods that collapse the data into two groups. Greater efficiency means a trial can detect a given effect with fewer patients, which affects statistical power and research costs. The NICE stroke guideline NG128 treated mRS 0 to 2 or ordinal shift at 90 days and at 1 year as critical outcomes, together with mortality, in its evidence review on thrombectomy.

For economic evaluation the full distribution matters for a separate reason. A model with one state per grade needs the share of patients at each grade in each arm, and a trial report that gives only the proportion at grades 0 to 2 forces the analyst to assume how patients are spread within each group. Those assumptions move costs and QALYs, as the worked example below shows.

mRS grades as health states in cost-effectiveness models

A common structure for acute stroke treatments is a short-term decision tree that ends in a distribution of patients across mRS grades at about 90 days, followed by a long-term Markov model whose states are the grades or groups of them. Each state carries a utility, an annual cost of health and social care and a risk of death, and patients accumulate QALYs and costs over the time horizon. De Andrés-Nogales and colleagues, for example, used a lifetime Markov model with health states defined by the mRS and clinical inputs from the SWIFT PRIME trial to compare thrombectomy with a stent retriever after intravenous thrombolysis against thrombolysis alone from the perspective of the Spanish National Health System, and found the combined treatment dominant.

The expected QALYs in one cycle of such a model are a weighted sum over grades.

$$Q = \sum_{k=0}^{6} n_k \, u_k \, t$$

where $Q$ is the QALYs accrued by the cohort in the cycle, $n_k$ is the number of patients at grade $k$, $u_k$ is the utility attached to grade $k$ (with $u_6 = 0$ for death) and $t$ is the cycle length in years. Costs follow the same form, with an annual cost per grade in place of the utility.

The grade itself is not a utility. The mRS has no preference weights, and its numbers are ranks on an ordinal scaling, so the gap between grades 1 and 2 need not equal the gap between grades 4 and 5. Utilities for each grade come from a separate preference-based measure, usually the EQ-5D, completed by patients whose grade is known.

Utilities and value sets attached to each grade

Ali and colleagues used EQ-5D-3L responses from the Virtual International Stroke Trials Archive (VISTA) at 3 months, applied published value sets from 13 countries and estimated mean health state utility values for each mRS grade. They reported in 2017 that for mRS 0 the mean utility ranged from 0.88 to 0.95 across value sets, while for mRS 5 it ranged from −0.48 to 0.22, and that the mRS at 3 months accounted for 65 to 71 per cent of the variation in utilities. The value set chosen therefore matters most at the severe end, where some sets place grade 5 below death and others above it.

When a trial recorded the mRS but no EQ-5D, a mapping function can predict utilities from the grade. Rivero-Arias and colleagues developed such an algorithm in 2010 with data from stroke and transient ischaemic attack patients in the Oxford Vascular Study (OXVASC). They used ordinary least squares to predict UK EQ-5D tariffs from the mRS and a multinomial logistic model with Monte Carlo simulation to predict the response to each EQ-5D question, and validated both out of sample in patients with coronary or peripheral vascular disease. The two methods gave similar errors, and both understated the uncertainty around the actual mean tariff, producing confidence intervals that were too narrow. A cost-utility analysis that uses mapped utilities should carry that extra uncertainty into its sensitivity analysis.

Illustrative worked example with seven grades over one year

All numbers in this example are illustrative and chosen for easy arithmetic. Two cohorts of 100 patients with acute ischaemic stroke receive either standard care or a new treatment that costs GBP 6,100 more per patient. For simplicity each patient is assumed to stay at the 90-day grade for one year, so $t = 1$, and costs and QALYs are not discounted.

1. QALYs under standard care. The products are 10 × 0.95 = 9.5, 15 × 0.85 = 12.75, 15 × 0.70 = 10.5, 15 × 0.55 = 8.25 and 15 × 0.30 = 4.5, with zero for grades 5 and 6. The total is 9.5 + 12.75 + 10.5 + 8.25 + 4.5 = 45.5 QALYs.

2. QALYs under the new treatment. The products are 15 × 0.95 = 14.25, 20 × 0.85 = 17, 15 × 0.70 = 10.5, 15 × 0.55 = 8.25 and 10 × 0.30 = 3, with zero for grades 5 and 6. The total is 14.25 + 17 + 10.5 + 8.25 + 3 = 53 QALYs, a gain of 53 − 45.5 = 7.5 QALYs per 100 patients, or 0.075 per patient.

3. Care costs. Under standard care, 10 × 1,000 = 10,000, 15 × 2,000 = 30,000, 15 × 5,000 = 75,000, 15 × 12,000 = 180,000, 15 × 25,000 = 375,000 and 10 × 40,000 = 400,000, a total of GBP 1,070,000. Under the new treatment, 15 × 1,000 = 15,000, 20 × 2,000 = 40,000, 15 × 5,000 = 75,000, 15 × 12,000 = 180,000, 10 × 25,000 = 250,000 and 5 × 40,000 = 200,000, a total of GBP 760,000. Care costs fall by 1,070,000 − 760,000 = 310,000, or GBP 3,100 per patient.

4. Cost per QALY. The net extra cost per patient is 6,100 − 3,100 = 3,000.

$$\text{ICER} = \frac{\Delta C}{\Delta Q} = \frac{3000}{0.075} = 40000$$

where $\Delta C$ is the incremental cost per patient in GBP, $\Delta Q$ is the incremental QALYs per patient and the result is GBP 40,000 per QALY gained in the first year.

5. What a dichotomised result hides. In both arms the share at grades 0 to 2 is known, 40 patients under standard care and 50 under the new treatment. Suppose instead that the ten extra good outcomes had come from grade 3 rather than from grades 4 and 5. The five extra patients at grade 0 add 5 × 0.95 = 4.75 QALYs and the five at grade 1 add 5 × 0.85 = 4.25, while the ten fewer at grade 3 remove 10 × 0.55 = 5.5. The QALY change would be 4.75 + 4.25 − 5.5 = 3.5 QALYs per 100 patients. The ten fewer at grade 3 save 10 × 12,000 = 120,000 in care costs, so the cost change would be 5,000 + 10,000 − 120,000 = −105,000.

The same 10-point rise in the proportion at grades 0 to 2 is worth 7.5 QALYs and GBP 310,000 of care savings per 100 patients in the first pattern, but only 3.5 QALYs and GBP 105,000 in the second. Most of the saving in the first pattern comes from fewer patients at grade 5, the most expensive state, which a 0 to 2 split cannot see. The example also shows why the utility given to grade 5 is a sensitive input, since a value set that places that grade below zero would add to the QALY gain from moving patients out of it.

How the mRS differs from neighbouring measures

The mRS measures disability and functional status as one global grade, and several nearby measures cover parts of the same ground. The Barthel Index scores performance in specific activities of daily living and adds the items into a total, whereas the mRS asks for a single overall judgement that also reflects how far life has changed.

The EQ-5D is patient-reported and preference-based, so its index can enter a QALY directly, while the mRS is clinician-rated and has no preference weights. A grade becomes a model health state only once a utility and a cost are attached to it, so the grade and the health state are separate things. Chaisinanunkul and colleagues proposed a utility-weighted mRS as a patient-centred primary outcome for acute stroke trials, which brings the two ideas closer without making the grade itself a preference measure.

The grade also has two roles that are easy to mix up. The pre-stroke mRS describes the patient before the event and is used to select patients for treatment, as in recommendation 1.4.8 of NICE guideline NG128, which limits thrombectomy to people with a pre-stroke functional status of less than 3 on the scale, while the mRS at 90 days or later is the outcome of the stroke and its treatment.

Limitations and common misreadings

The mRS is quick to assess and familiar to stroke clinicians, and its grades map neatly onto model states. Several errors still recur when trial results on the scale feed economic models.

• Averaging grades. A mean mRS treats the grades as equally spaced, which an ordinal scale does not support. Models avoid this by attaching a utility to each grade instead of using the grade number.
• Relying on a single cut point. A dichotomised result loses the movement between grades within each group, which carries much of the cost and QALY difference, as the example shows.
• Treating mapped utilities as observed. Rivero-Arias and colleagues found that mapped mean tariffs came with confidence intervals that were too narrow, so a model built on them looks more certain than it is.
• Ignoring the value set. Ali and colleagues found that utilities for grade 5 ranged across zero in the 13 national value sets, so a model transferred between countries needs utilities from the right set.
• Confusing the scale with a global health measure. The mRS centres on dependence and activity after stroke. Two patients at the same grade can differ in pain or mood, which the grade does not capture and a preference-based measure would.

Sources

• Ali M, MacIsaac R, Quinn TJ, Bath PM, Veenstra DL, Xu Y, Brady MC, Patel A, Lees KR. Dependency and health utilities in stroke: data to inform cost-effectiveness analyses. European Stroke Journal. 2017;2(1):70-76. doi:10.1177/2396987316683780
• Banks JL, Marotta CA. Outcomes validity and reliability of the modified Rankin scale: implications for stroke clinical trials: a literature review and synthesis. Stroke. 2007;38:1091-1096. doi:10.1161/01.STR.0000258355.23810.c6
• Broderick JP, Adeoye O, Elm J. Evolution of the modified Rankin scale and its use in future stroke trials. Stroke. 2017;48(7):2007-2012. doi:10.1161/STROKEAHA.117.017866
• Chaisinanunkul N, Adeoye O, Lewis RJ, Grotta JC, Broderick J, Jovin TG, et al. Adopting a patient-centered approach to primary outcome analysis of acute stroke trials using a utility-weighted modified Rankin scale. Stroke. 2015;46:2238-2243. doi:10.1161/STROKEAHA.114.008547
• de Andrés-Nogales F, Álvarez M, de Miquel MÁ, Segura T, Gil A, Cardona P, Casado MÁ, Nogueira RG, Dávalos A. Cost-effectiveness of mechanical thrombectomy using stent retriever after intravenous tissue plasminogen activator compared with intravenous tissue plasminogen activator alone in the treatment of acute ischaemic stroke due to large vessel occlusion in Spain. European Stroke Journal. 2017;2(3):272-284. doi:10.1177/2396987317721865
• National Institute for Health and Care Excellence. Stroke and transient ischaemic attack in over 16s: diagnosis and initial management (NG128), recommendation 1.4.8 [2019] and evidence review D (thrombectomy), 2019, last updated 13 April 2022. https://www.nice.org.uk/guidance/ng128/chapter/recommendations and https://www.nice.org.uk/guidance/ng128/chapter/rationale-and-impact
• Optimising Analysis of Stroke Trials (OAST) Collaboration, Bath PM, Gray LJ, Collier T, Pocock S, Carpenter J. Can we improve the statistical analysis of stroke trials? Statistical reanalysis of functional outcomes in stroke trials. Stroke. 2007;38:1911-1915. PMID 17463316
• Rankin J. Cerebral vascular accidents in patients over the age of 60. II. Prognosis. Scottish Medical Journal. 1957;2(5):200-215. doi:10.1177/003693305700200504
• Rivero-Arias O, Ouellet M, Gray A, Wolstenholme J, Rothwell PM, Luengo-Fernandez R. Mapping the modified Rankin scale (mRS) measurement into the generic EuroQol (EQ-5D) health outcome. Medical Decision Making. 2010;30(3):341-354. doi:10.1177/0272989X09349961
• Saver JL, et al. Standardized nomenclature for modified Rankin scale global disability outcomes: consensus recommendations from Stroke Therapy Academic Industry Roundtable XI. Stroke. 2021;52:3054-3062. doi:10.1161/STROKEAHA.121.034480
• van Swieten JC, Koudstaal PJ, Visser MC, Schouten HJ, van Gijn J. Interobserver agreement for the assessment of handicap in stroke patients. Stroke. 1988;19(5):604-607. doi:10.1161/01.str.19.5.604
• Wilson JT, et al. Reliability of the modified Rankin scale across multiple raters: benefits of a structured interview. Stroke. 2005;36:777-781. PMID 15718510. doi:10.1161/01.STR.0000157596.13234.95