MDS-Comorbidity Index using register data has prognostic impact in Swedish MDS patients

Comorbidities influence the mortality in patients with myelodysplastic syndromes, and a growing body of evidence suggest that comorbidity history should be used in addition to established prognostic indices. A comorbidity index specific for MDS, the MDS-CI, was introduced a decade ago. In this study we aim to construct an MDS-CI version based on diagnoses from register data only, to expand its use beyond the clinical setting to retrospective and register based studies. We further test this version on a Swedish population-based MDS cohort of 2947 patients, and compare its prognostic accuracy to that of Charlson Comorbidity Index. Our register based MDS-CI divided patients into three risk groups of similar proportions as have been published for the original MDS-CI. Compared to low risk patients, intermediate and high risk patients had 50 % and 70 % higher mortality, respectively. The prognostic value of MDS-CI was equal to that of Charlson comorbidity index. Adding MDS-CI to the established prognostic factors IPSS-R and age increased the prognostic accuracy. In summary, we demonstrate that MDS-CI can be adequately estimated from diagnoses recorded in registers only, and that it is a useful tool in any future study on myelodysplastic syndromes with a need to adjust for comorbidities.


Introduction
Patients with myelodysplastic syndromes (MDS) are often elderly at diagnosis, with a reported median age of 75 years at diagnosis [1,2].Many have one or more concomitant medical conditions.Although MDS is a disease with high mortality, a significant portion of patients live more than five years and comorbidities may thus have time to influence the expected survival [3].
For patients with life-threatening diseases, the impact of comorbidities on overall survival, and as effect modifiers for treatment alternatives, have been acknowledged by clinicians and epidemiologists for decades.The pioneering work by Feinstein [4] and Charlson [5] led to the Charlson Comorbidity Index (CCI), which has proven useful in several groups of patients [6].More recently, an index specifically aimed at MDS patients, the myelodysplastic syndrome comorbidity index (MDS-CI), has been proposed by Della Porta et al. [7].This index is a further development from the Sorror score or HCT-CI [8], which is used in patients under consideration for hematopoietic stem cell transplantations.MDS-CI includes fewer comorbidity categories than CCI and HCT-CI, and is thus easier to calculate in a clinical setting.Several groups have demonstrated a correlation between MDS-CI or CCI and mortality in MDS patients [7,[9][10][11], and this correlation seems to go beyond what is already known from established prognostic markers such as IPSS-R and age.Consequently, it has been argued that comorbidities should be included in decision making algorithms for MDS also outside the transplant setting [12].
Although the proposed and established indices all predict survival in large patient cohorts, their usefulness in retrospective studies is hampered by their clinical origin.The indices may include results from laboratory tests such as creatinine levels, from physiological assessments such as the left ventricular ejection fraction, or details on co-medication such as insulin treatment.Such information is often not readily available outside longitudinal studies or clinical trials, at least not if we expect them to be uniformly recorded.Consequently, numerous attempts have been made to transform the CCI into versions strictly based on previous ICD (International Classification of Diseases)-coded diagnoses [13,14].Some of these have also modernised the CCI, to exclude comorbidities that have become less clinically relevant in recent years due to medical progress, and adjusted the original weighting of the variables such as lowering the weight on AIDS and increasing weights on dementia and chronic heart failure [15].
In this study, we aim to construct a version of the MDS-CI based on ICD-coded diagnoses, for use in register-based studies.We validate MDS-CI in the so far largest cohort of MDS patients.We further compare it to the registry based CCI, with regard to correlation to overall survival in MDS patients.We also compare the prevalence and impact on survival of diagnosis based CCI and MDS-CI categories in patients with MDS to that in age-and sex-matched controls from the general population.

Study population
All patients with MDS (excluding chronic myelomonocytic leukaemia and other forms of myelodysplastic/myeloproliferative neoplasms) and registered in the Swedish MDS register (SMDSR) between January 2009, when the register was initiated, and December 2018 were included.In all, the cohort included 2947 patients.
The SMDSR aims at including all newly diagnosed adult Swedish patients with MDS, with diagnostic, prognostic, and clinical data from time of diagnosis.From 2012 and onwards, follow-up data at one year and every third year after diagnosis is recorded.Compared to the Swedish Cancer Register, to which reporting is mandatory for clinicians and pathologists, the coverage was 97 % for the years 2009-2015 [16].
To compare the prevalence of comorbidity and the impact on survival of the comorbidity indices between MDS patients and the general population, 5 population comparators for each MDS-case were chosen from the Swedish Total Population Register.The comparators were matched for sex, birth year and county of residence; they had to be free of MDS and MDS/ MPN and alive at the time of diagnosis of their corresponding MDS patient.
Patients were followed for death or permanent emigration in the Swedish Total Population Register until November 20th 2019.

Construction of register based comorbidity indices
To calculate comorbidity indices, we used data from the National Patient Register, which holds records of ICD-10 coded hospital discharge diagnoses.The National Patient Register covers all Swedish in-patient care from 1987 and onwards.Since 2001, diagnoses from specialised out-patient care are also recorded [17].As comorbidities, we considered all diagnoses recorded in health care contacts during ten years preceding MDS diagnosis.For integrity reasons, we had no access to the psychiatric diagnoses in the F chapter of ICD-10, and our coverage of dementia diagnoses was hence somewhat limited.However, several dementia diagnoses, including Alzheimer's disease, are coded both in the F (psychiatry) and G (neurology) chapters and were consequently retrievable.For categories including previous malignancies, we instead used the Swedish Cancer Register, to which reporting has been mandatory since 1958.In an earlier assessment of completeness, the register was estimated to cover more than 96% of incident cancer cases in Sweden [18].To construct the CCI, the unweighted Royal College of Surgeons' version was used [19], which has in a critical review been pointed out as especially valid and useful [20].CCI was categorised into low risk (score 0),
The proposed diagnosis based MDS-CI is presented in Table 1 with comparison to the diagnosis based CCI.To separate them from the original versions, we will use the abbreviations D-MDS-CI and D-CCI for the diagnosis based MDS-CI and CCI in this article.For severe pulmonary disease, which in MDS-CI is defined by physiological and clinical measures, only diagnoses of chronic obstructive pulmonary disease and pulmonary fibrosis were used.For hepatic disease, diagnoses of cirrhosis, fibrosis, hepatic failure, and chronic non-viral hepatitis were used.For renal disease, diagnoses of chronic kidney disease, dialysis and renal transplants were used.

Statistical analysis
Differences in mortality were compared by Cox proportional hazards regression [21] and by Kaplan-Meier curves [22] with Hall-Wellner confidence bands [23].To determine the different indices' ability to predict mortality, a concordance statistic (c-statistic) was estimated for the proportional hazards models, according to the methods proposed by Uno et al. [24].Uno's c-statistic, or c-index, is a summary estimate of the predictive potential of one or more co-variates to a certain endpoint, where 0.5 means no predictive value and 1.0 means full concordance between exposure and outcome.Using the same methods, the c statistics for different explanatory models were further compared, using 50 perturbation samples to estimate standard error.

Ethical considerations
The study was approved by the Ethics Committee of Uppsala University (2014/176).Record linkages were performed at Statistics Sweden and the National Board of Health and Welfare.All data were deidentified before analysis.

Results
Baseline characteristics are presented in Table 2. Median age at diagnosis was 76 years and median follow-up was 1.9 years (range 0-10.9 years).A majority (59 %) were male.Among all 2947 MDS patients, 59 % had a D-MDS-CI score of 0, while 33 % had a score of 1-2 % and 8 % had a score of 3 or more (Table 3).These proportions are similar to those reported by Della Porta et al. in the original MDS-CI, where the proportions were 65 %, 29 %, and 6 %, respectively.
Both D-CCI and D-MDS-CI predicted mortality in a meaningful way.Compared to low risk (patients with no comorbidities), the mortality was 50 % higher in intermediate risk patients, and 70 % higher in high risk patients according to D-MDS-CI (p < 0.0001) (Table 3, Fig. 1).Patients with D-CCI intermediate risk had 30 % higher mortality, and those with high risk had 70 % higher mortality, compared to those with low risk.If D-MDS-CI was dichotomised into any comorbidity versus no comorbidity (Fig. 2), those with comorbidity had 50 % higher mortality than those without comorbidity.As expected, D-MDS-CI and D-CCI did not only predict mortality, but also the chance of being treated with an allogeneic stem cell transplant (SCT).In patients with low D-MDS-CI index, 12 % underwent SCT, whereas only 4% among those in the intermediate category and only two patients (<1 %) of those with high comorbidity underwent SCT.However, differences in transplant rates did not explain the differences in mortality between patients with more or less comorbidities, and adjustment for transplant in multivariable analyses did not alter the results (data available on request).
Compared to matched controls from the general population, most comorbidity categories were more prevalent in MDS patients (Table 3).The highest prevalence ratios were found for connective tissue disease, liver disease, and AIDS, but all of those were based on small number of patients.The correlation between co-morbidity and mortality was higher in comparators than in MDS patients.
Both D-CCI and D-MDS-CI had c-index significantly higher than 0.5 for overall mortality; D-CCI had a c-index of 0.58 (95 % CI 0.56-0.59)and D-MDS-CI had a c-index of 0.57 (95 % CI 0.56-0.58).Neither of these were better than the other (p = 0.28).A dichotomised D-MDS-CI also resulted in a c-index of 0.57 (95 % CI 0.56-0.58).IPSS-R, for comparison, had a c-index of 0.68 and provided significantly better predictive information (p < 0.0001).On the other hand, adding either D- CCI to IPSS-R (c-index 0.71) or D-MDS-CI to IPSS-R (c-index 0.70) increased the predictive potential over IPSS-R alone (p < 0.0001 for both comparisons).Age (over or under age 75) gave similar predictive information (c-index 0.57) as D-CCI or D-MDS-CI (p = 0.2 and p = 0.6, respectively).
Comorbidities were not randomly distributed among MDS patients (Table 4).Patients with high risk D-MDS-CI at diagnosis were more often men, above median age, transfusion dependent, or patients with a therapy-related MDS.Meanwhile, comorbidity risk classes did not correlate with IPSS or IPSS-R risk classes, underlining why the combination of IPSS-R with a comorbidity index provides better information than either of them alone.

Discussion
From a large, population-based cohort, we can now confirm the usefulness of the MDS-CI proposed by the study groups in Pavia and Düsseldorf.We can further demonstrate that similar results can be achieved from records of diagnoses only, which will expand its use to register based research, and to retrospective studies on medical records.
The present study does not compare our register based MDS-CI to the original MDS-CI based on results from physiological examinations, as we did not have that information.Hence, we cannot conclude that information on diagnoses alone is equally informative as the clinical data included in the original MDS-CI.However, we demonstrate that diagnoses alone can divide patients in groups of similar size as in the original publication, and that these comorbidities has an impact on prognosis.
Comorbidity and age are patient characteristics, whereas IPSS-R is based on disease characteristics.While comorbidity history provides less prognostic information than IPSS-R, we demonstrate that the combination of IPSS-R with a comorbidity index outperforms IPSS-R alone.Hence, comorbidities should be taken into account when estimating prognosis in MDS, just as we do with the age of the patient.
In our data, CCI and MDS-CI gave similar prognostic information and can both be recommended in future MDS studies.An advantage with MDS-CI is the low number of comorbidity categories, and the fact that Table 3 Frequency of each D-CCI and D-MDS-CI comorbidity in MDS patients and in age and sex matched comparators from the general population, and their impact on overall survival in cases and comparators.Hazard ratios of death are adjusted for sex and age.HR=hazard ratio, CI=confidence interval.some of the rarer CCI categories seemed to contribute very little information.Furthermore the fewer number of categories makes the MDS-CI easier to use in a clinical setting.On the other hand, unweighted CCI (categorised into 0, 1, or ≥2 comorbidities) divided the cohort into three risk groups of similar size, whereas most patients were classified as MDS-CI low risk, and only 8% were classified as high risk.This could be an important statistical strength for CCI in studies on smaller patient cohorts.
Judging from our data, weighting of MDS-CI is not particularly useful.Cardiac comorbidities, which in MDS-CI gets double weight [7], were not more closely linked to mortality than other comorbidity categories.As the high risk group was also very small, and did not have significantly lower survival than the intermediate risk group, a simplified approach would be to divide patients into two groups; those with comorbidities and those without.As demonstrated in Fig. 2, those two groups have distinctly different survival.The prognostic information gained from a two group model was also equal to that from a three group model, as measured by c-index.
Apart from the original Pavia/Düsseldorf publication [7], MDS-CI has also been linked to mortality in MDS patients by groups in Genova [9], Rome [25], Osaka [26], Toronto [10], Sevilla [27], Tianjin [28], and Amsterdam [29].Several of these groups have also demonstrated that MDS-CI provides additional information over IPSS-R [9,27,29,30], just as we confirm in our study.The original version of CCI has also been linked to mortality in MDS patients, independently of IPSS-R [31].This is the first published MDS-CI based on diagnoses from registers, but register based CCI has previously been used in a study of MDS prognosis [11].Researchers from the Mayo clinic have also demonstrated a correlation between number of recorded comorbidity diagnoses and risk of death in MDS patients, but without the use of a specific comorbidity index [32].
In our data, we also had the possibility of comparing comorbidities in MDS patients with those in an age and sex matched sample of the general population.From these analyses, we can draw two conclusions.Firstly, MDS patients seem to have more comorbidities than the general population.We did not find a remarkable correlation to one specific disease, which could be evaluated in future studies as a possible risk factor for MDS.Instead, we saw a pattern where most disease categories were more frequent in MDS patients, resulting in an intermediate/high risk MDS-CI in 41% of patients compared to 30% of comparators.This pattern can in part be due to surveillance bias, if MDS patients have been more closely followed by health care before their diagnosis.Secondly, comorbidities have a relatively lower correlation to mortality in MDS patients than in comparators, as MDS patients have a much higher baseline mortality.Pulmonary diseases, on the other hand, had almost as high correlation to mortality in MDS patients as in comparators (approximately doubled risk of death), suggesting that the pulmonary diseases included in the index are a matter of concern also in MDS patients.
A limitation of our study is the possible under-reporting of comorbidities.The National Patient Register does not cover all primary health care, and less severe health issues may not have been reported if they never resulted in specialised or in-patient care.For the minority of patients diagnosed before 2011 (18 %) the Patient Register did not cover out-patient care more than eight or nine years before diagnosis.Furthermore, patients with multiple comorbidities and short life expectancy may have declined further investigation and thus never been diagnosed with MDS.Such non-differential misclassification would bias our results towards the null hypothesis, and the true correlation between comorbidities and mortality in MDS might in reality be somewhat larger.
In summary, we present the largest to date investigation of comorbidities in MDS.We demonstrate that comorbidities can be reliably estimated retrospectively from register data, and that the classification of patients and correlation to mortality is similar to the results obtained in prospectively gathered data.We further strengthen the case for inclusion of comorbidities in estimates of MDS prognosis, both in clinical and research settings.

Table 2
Baseline clinical parameters in 2947 MDS patients.
a SLD=single lineage dysplasia, MLD=multi lineage dysplasia, RS=ring sideroblasts, EB=excess blasts, U=unclassified b Missing data on 46 cases for subtype, 545 cases for IPSS, 741 cases for IPSS-R, and 9 cases for transfusion dependency G. Larfors et al.