A recent validation analysis at our center among allogeneic hematopoietic cell transplant (HCT) recipients did not find the HCT-specific comorbidity index (HCT-CI) to clearly segregate patient's transplant-related risk. We hypothesized that the discriminating and predictive power of the HCT-CI for mortality could be improved by eliminating the assignment of categorical weights to comorbidities and instead replacing them with hazard ratios (HR) from a Fine and Gray adjusted regression model. This approach allowed us to look carefully at each component of the comorbidity index. We developed the modified comorbidity index (MCI) using a cohort of 444 adult allogeneic HCT recipients using a pure multiplicative model. Compared with low-risk patients, the HR for non-relapse mortality (NRM) using the HCT-CI was 1.3 (95% confidence intervals, 0.7–2.4) for intermediate risk and 1.6 (0.9–2.8) for high-risk patients, and with the MCI was 1.6 (0.9–2.8) and 2.7 (1.5–5.0), respectively. In conclusion, we are introducing the MCI which may have higher discriminating and predictive power for overall survival and NRM. Validation of the HCT-CI and the MCI in larger and separate cohorts of HCT recipients is still needed.
An increasing number of hematopoietic cell transplant (HCT) recipients have pre-existing comorbidities at the time of transplant. This has led to the evaluation of comorbidity indices as independent predictors of outcomes such as overall survival (OS) and non-relapse mortality (NRM).1, 2, 3, 4 Common scales used include the Charlson Comorbidity Index,1 the Kaplan–Feinstein Scale2 and the Adult Comorbidity Inventory-27.3 More recently, the HCT-specific comorbidity index (HCT-CI) was proposed to better predict outcomes of HCT recipients based on their pre-existing comorbidities.4 This index was proposed and subsequently evaluated within two randomly divided datasets. Further retrospective studies have shown the HCT-CI to be useful in predicting NRM in allogeneic HCT recipients.5, 6 Recently, we investigated the utility of this tool in a retrospective analysis of 373 adult allogeneic HCT recipients.7 We found that the HCT-CI has the potential for widespread applicability but its generalizability and sensitivity may be limited and should be further investigated at other centers. We now hypothesize that the discriminating and predictive power of the HCT-CI for the outcomes of OS and NRM could be improved by eliminating the assignment of categorical weights (0–3) to the comorbidities, and replacing them directly with the hazard ratios (HR) from the adjusted regression model to calculate the comorbidity index. We introduce a modified comorbidity index (MCI) that may be more efficient at predicting comorbidity risk for OS and NRM in allogeneic HCT recipients.
Patients and methods
This analysis consisted of adult allogeneic HCT) recipients who received a transplant at the University Of Minnesota between 2000 and 2008. The study cohort of 444 patients included 192 matched-related donor (MRD), 19 matched-unrelated marrow donor (MUD) and 233 unrelated umbilical cord blood (UCB) transplant recipients who received either a myeloablative (MA, N=169) or non-myeloablative (NMA, N=275) conditioning regimen. No significant differences were noted in the distribution of HCT-CI scores between the MRD, MUD and UCB groups or between the MA and NMA groups. Our institutional MA and NMA regimens and graft-verus-host disease prophylaxis and treatment regimens have been described earlier.8, 9, 10, 11 All patients were transplanted on protocols approved by our institutional review board and signed an informed consent based on the principles of the Declaration of Helsinki.
Baseline demographic data, treatment characteristics and outcome data were prospectively collected using standardized methods by the Clinical Data Management Core for the Blood and Marrow Transplant Program at the University of Minnesota. Data regarding pre-transplant comorbidities were abstracted retrospectively for patients transplanted before 2008 by a detailed review of all medical records including those from referring physicians and treating institutions as available. A second investigator independently reviewed medical charts of 110 randomly selected patients. There was excellent inter-rater agreement between the two investigators (κ coefficient 0.87 (95% confidence interval (CI), 0.81–0.95)) and no specific domain was identified where there was inconsistent scoring of the comorbidities. Comorbidities were collected and assessed prospectively for patients transplanted during 2008.
Assignment of weights for the HCT-CI
Comorbidities for the HCT-CI were assigned integer weights based on the derived scores from Sorror et al.4 These weights were based on adjusted HR in which an HR of 1.3–2.0 for a specific comorbidity was given a weight of 1, comorbidities with adjusted HR of 2.1–3.0 were assigned a weight of 2 and HR of 3.1 or more were assigned 3. The HCT-CI score was the sum of these integer weights. The patients were then assigned to one of three risk groups: 0 (low risk), 1–2 (intermediate risk) and 3 or more (high risk).
Assignment of weights for the MCI
The hypothesis underlying the MCI was that by removing assignment to a limited number of integer weights for each comorbidity, and instead, using a multiplicative model that uses the actual HR, we would have higher discriminating and predictive power for the outcomes of OS and NRM. As in the HCT-CI, parameter estimates were obtained by performing a regression analysis on NRM after adjusting for all other comorbidities as well as age (<50 years versus ⩾50 years), disease risk (standard versus high), donor type (MRD versus MUD versus UCB) and conditioning intensity (MA versus NMA). Instead of converting the adjusted HR to categorical weights, we used a pure multiplicative model in which we used the exact parameter estimates (β) from the HR, calculated as eβxi and obtained the final score directly by exponentiating the sum of all parameter estimates. The equation for the final score is
Π (HR for each comorbidity)=Π eβ for each parameter*xi=exponent [(β for arrhythmia)*xi +(β for cardiac disorders)*xi+(β for heart valve disease)*xi+(β for inflammatory bowel disease)*xi+(β for peptic ulcer)*xi+(β for diabetes)*xi+(β for obesity)*xi+(β for cerebrovascular disease) *xi+(β for psychiatric disturbance)*xi+(β for mild altered hepatic function)*xi+(β for moderate/severe altered hepatic function)*xi +(β for infection)*xi+(β for rheumatologic disorders)*xi+(β for renal insufficiency)*xi+(β for moderate pulmonary abnormalities)*xi+(β for severe pulmonary abnormalities)*xi+(β for prior solid tumor)*xi]
xi is a binary indicator for the presence of each comorbidity for patient i.
After obtaining the final multiplicative-derived CI score, patients were assigned to three risk groups: 1 (low risk), 1.01–2.0 (intermediate risk) and >2.0 (high risk). Cut points were determined as follows: the low-risk category was assigned equivalently to the HCT-CI in that it primarily included patients without any comorbidities, the cut point between the intermediate and high-risk categories was generated with the goal of creating equivalent sample sizes in both groups as well as using a natural integer cut point which in this case was 2.
Correlations between the HCT-CI and the MCI were evaluated both by the χ2 test for the three risk groups and the Spearman correlation coefficient for the continuously measured final score. Kaplan–Meier curves were used to estimate the probabilities of OS and cumulative incidence was used to estimate the probability of NRM through 2 years post transplant, treating relapse as a competing risk.12, 13 Parameter estimates for the revised score were obtained by using multiple regression analysis (Cox regression for survival and the Fine and Gray competing hazards method for NRM).14, 15 Cox regression was used to develop the HCT-CI but any future development or modification of the MCI would preferably use Fine and Gray regression so we decided to use this in creating the MCI. Likelihood ratios were calculated for each index score after adjusting for age (<50 years versus ⩾50 years), disease risk (standard versus high), donor type (MRD versus MUD versus UCB), conditioning (MA versus NMA) and baseline Karnofsky score (⩾90 versus <90). In addition to confounding, interactions were investigated between the comorbidity scores and age and baseline Karnofsky score. A formal statistical test was performed to compare the two index scores by computing an index of concordance, the c-statistic (C). This statistic, derived by Harrell et al.,16 estimates the probability that, of two randomly chosen pairs of patients, the patient with the lower index score will outlive the patient with the higher index score. Values of C near 0.5 indicate that the index score is no better than chance in determining which patient will live longer. Values of C near 1.0 indicate the index score virtually always determines that the patient with the lower score has better survival. Certain pairs of observations were excluded from the calculation because of (1) pairs who had equivalent index scores (38%), (2) subsequent pairs who were both censored at the time of analysis (7%) or (3) further pairs in which one patient was censored before the event of the other patient (1%). For NRM, patients were censored at the time of relapse or disease progression. The c-statistic was computed for survival and NRM based on time to event analysis over the first 2 years post transplant. Standard errors for the difference in the c-statistics for the two index scores were estimated by applying a bootstrap procedure to the dataset using 100 bootstrap samples.17
Patient and treatment characteristics for 444 patients are shown in Table 1. The median age at transplant was 47 years (range; 18–69). Diagnoses included acute leukemia (39%), non-Hodgkin lymphoma or Hodgkin lymphoma (31%), myelodysplastic syndrome (8%), chronic myeloid leukemia (7%), severe aplastic anemia (2%) or other malignant disorders (14%). The donor type was divided between UCB (52%), MRD (43%) and MUD (4%). The median follow-up among surviving patients was 4.0 years (range; 0.4–8.3 years).
Predictive importance of specific comorbidities.
Results of an initial exploratory analysis are shown in Figure 1. We measured the impact on the HR for the intermediate and high-risk scores on NRM after removing individual comorbidities from the HCT-CI. A negative percentage change in the HR indicates that removal of the comorbidity decreased the correlation between the HCT-CI and NRM and that the factor is probably an important part of the index. The factors that showed a negative percent change included cardiac function, diabetes, psychiatric disturbance, infection, pulmonary abnormalities and prior malignancy. A positive percentage change indicates that removal of the comorbidity has increased the correlation between the HCT-CI and NRM and thus the factor may not be an important component of the index. Comorbidities that showed such a positive change and had a prevalence of >1% included obesity and cerbrovascular disease. A change ⩽1% indicates that removal of the comorbidity from the HCT-CI has minimal effect, although this result may simply indicate low prevalence of the comorbidity.
Modified comorbidity index
Given that some comorbidities show specific effects and that others show no effect on the overall index score (Figure 1), we proposed a weighting scheme in which each comorbidity is assigned the original specific weight from the regression analysis without taking the additional step of assigning a restricted set of four possible weights of 0–3 after the regression analysis as proposed by the HCT-CI. The modified CI used the β coefficients or weights and adjusted relative risks shown in Table 2.
For example based on these results, if a patient presented with all comorbidities excluding moderate pulmonary abnormality at transplant, the MCI would be exp(0.39+0.38+0.56+0.46+0.07+0.07+0.75+0.35+1.81+0.49)= exp(5.33)=206.44.
If a patient presented with no comorbidities, the MCI would be: exp(0)=1.0.
As shown in Table 2, typically the factors showing no impact on the final score, or where the β coefficients have a value of 0 or less, were comorbidities that were uncommon. The only factors that showed no or minimal impact on the final score but had sufficient prevalence for analysis were obesity, psychiatric disturbance, mild hepatic function and prior solid tumor. Some of the factors showing more impact and having sufficient prevalence were cardiac disorders, diabetes, moderate or severe hepatic function and moderate or severe pulmonary abnormalities.
Comparison of the MCI to the HCT-CI.
Correlation between the MCI and the HCT-CI was high. Evaluating the continuous measure of the final overall scores, the Spearman correlation coefficient was 0.83 (P<0.01). In evaluating risk categories, the percentage of patients in the low, intermediate and high-risk cohorts using the HCT-CI was 21, 30 and 49%, respectively. For the MCI score, the corresponding percentages were 24, 52 and 23%, respectively. The χ2 test showed a statistically significant association between these two categorical measures (P<0.01).
The 2-year cumulative incidence of NRM for the HCT-CI and the MCI are shown in Figures 2a and b. Using the HCT-CI, the 2-year NRM for the low, intermediate and high-risk groups were 18% (95% CI; 10–28%), 23% (95% CI; 16–30%) and 27% (95% CI; 21–33%), (P=0.12). Using the MCI, the 2-year NRM for the low, intermediate and high-risk groups were 15% (95% CI; 8–22%), 23% (95% CI; 18–28%) and 34% (95% CI; 24–44%), (P<0.01), with similar competing risks.
Two-year OS for the HCT-CI and the MCI are shown in Figures 3a and b. The 2-year OS rates for the low, intermediate and high-risk groups in the HCT-CI were 62% (95% CI; 51–71%), 58% (95% CI; 49–65%) and 50% (95% CI; 43–56%), (P=0.08) compared with 63% (95% CI; 54–70%), 55% (95% CI; 48–61%) and 32% (95% CI; 20–44%), (P<0.01) with the MCI, respectively.
A comparison of the two index scores after adjusting for age, disease risk, donor type, conditioning intensity and baseline Karnofsky score is shown in Table 3 by three different methods: the relative risk from the regression models, which measures the degree of risk for each score; the adjusted likelihood ratio from the respective models in which a higher score indicates better prediction; and finally the c-statistic that is used to obtain a P-value for the comparison between the two risk scores.
Using the MCI classification, the relative risk for NRM increased 23 and 69% for the intermediate and high-risk scores, respectively. The high-risk category also gained statistical significance (P<0.01) indicating better discrimination. The adjusted likelihood ratio increased 54% from 14.6 to 22.5. Finally, the c-statistic increased from 0.547 to 0.595 for NRM (P=0.02).
Analysis for OS showed similar results. For the MCI, the relative risk of overall mortality increased 27 and 80% for the intermediate and high-risk scores, respectively. The adjusted likelihood ratio increased 87% from 15.6 to 29.1 and the c-statistic increased from 0.557 to 0.600 (P=0.04); again suggesting improved discrimination between groups using the MCI.
Investigation into the effect of MCI and the different strata within the factors of age and baseline Karnofsky score on NRM did not show an interaction between these variables. The relative risks for the intermediate and high-risk cohorts for the MCI were similar between the different strata of age (<50, ⩾50) and baseline Karnofsky (⩾90, <90).
We believe that owing to a more efficient weighting of the comorbidities, our revised index score—the product of a pure multiplicative model—has higher discriminating and predictive power for the outcomes of NRM and OS. This is likely because of the fact that the HCT-CI loses some power in the early and unnecessary step of categorizing the adjusted HR. The primary modification in the MCI was removal of the step of assigning categorical weights and instead using a pure multiplicative model to calculate the final index.
Our analysis highlights that certain comorbidities in the HCT-CI may be uninformative. The initial exploratory analysis and the equation for MCI showed that arrhythmia, heart valve disease, inflammatory bowel disease, obesity, prior solid tumors and rheumatologic disorders showed no positive association (HR >1.0) with NRM. This may have been because of low prevalence of these comorbidities or owing to a truly low predictive power of the specific comorbidities on NRM. This highlights that collection of such factors may be unnecessary and their elimination may reduce the burden of CI assignments. Alternatively, the prevalence of the condition on the population in which the index was developed may affect the model.
Although our analysis showed that the MCI had more predictive power than the HCT-CI, the major limitation of this study was that the results were not based on a comparison of the methods in a different cohort of patients from which the MCI was developed. Our patient numbers did not provide enough power to create a separate and independent validation cohort. The primary purpose for this study was to introduce a new measure that we believe to be a more predictive measure of comorbidity risk. Efforts are underway for a future analysis to increase patient numbers so that a more precise MCI can be created as well as to provide an independent validation cohort.
Another limitation was the choice of the cut point between the intermediate and high-risk groups. The choice for the low-risk group was equivalent to the HCT-CI in that it was primarily the patients without any comorbidities. The choice of the cut point between the intermediate and high-risk groups was chosen because of the natural cut point of 2.0 and to give similar sample sizes for the two risk categories. However, our proposed thresholds are not definitive and can be modified. Validation of the MCI in larger and more diverse cohorts of transplant recipients will help identify optimal cut points for classifying patients into the risk groups.
A disadvantage of the MCI as compared to the HCT-CI is that the latter score can be relatively easily calculated once the presence of comorbidities is known. Calculation of the MCI can be simplified, however, by using the ‘exp’ function in a software spreadsheet such as Microsoft Excel®. Another alternative would be to use an application online such as the one at http://bmt.ahc.umn.edu:8082/hct/. Using either of these methods, calculation of the MCI could be quickly generated at any time such as at the pre-HCT workup evaluation.
In summary, the MCI will probably have higher discriminating and predictive power on the endpoints of NRM and OS by use of the actual coefficients in defining the index. Given that more patients are coming to transplant with pre-existing comorbidities and different subgroups of patients may have different prevalences of certain comorbidities, this improved discrimination of assigning patient comorbidity will better inform future studies among HCT recipients by better adjusting for these important risk factors.
Sorror ML, Maris MB, Storer B, Sandmaier BM, Diaconescu R, Flowers C et al. Comparing morbidity and mortality of HLA-matched unrelated donor hematopoietic cell transplantation after nonmyeloablative and myeloablative conditioning: influence of pretansplantation comorbidities. Blood 2004; 204: 960–968.
Artz AS, Pollyea DA, Kocherginsky M, Stock W, Rich E, Odenike O et al. Performance status and comorbidity predict transplant-related mortality after allogeneic hematopoietic cell transplantation. Biol Blood Marrow Transplant 2006; 12: 954–964.
Alamo J, Shahjahan M, Lazarus HM, de Lima M, Giralt SA . Comorbidity indices in hematopoietic stem cell transplantation: a new report card. Bone Marrow Transplant 2005; 36: 473–474.
Sorror ML, Maris MB, Storb R, Baron F, Sandmaier BM, Maloney DG, Storer B . Hematopoietic cell transplantation (HCT)-specific comorbidity index: a new tool for risk assessment before allogeneic HCT. Blood 2005; 106: 2911–2919.
Sorror ML, Giralt S, Sandmaier BM, De Lima M, Shahjahan M, Maloney DG, Deeg HJ . Hematopoietic cell transplantation specific comorbidity index as an outcome predictor for patients with acute myeloid leukemia in first remission: combined FHCRC and MDACC experiences. Blood 2007; 110: 4606–4613.
Sorror ML, Sandmaier BM, Storer BE, Maris MB, Baron F, Maloney DG et al. Comorbidity and disease status based risk stratification of outcomes among patients with acute myeloid leukemia or myelodysplasia receiving allogeneic hematopoietic cell transplantation. J Clin Oncol 2007; 25: 4246–4254.
Majhail NS, Brunstein CG, McAvoy S, DeFor TE, Al-Hazzouri A, Setubal D et al. Does the hematopoietic-cell transplantation specific comorbidity index predict transplant outcomes? A validation study in a large cohort of umbilical cord blood and matched related donor transplants. Biol Blood Marrow Transplant 2008; 14: 985–992.
Brunstein CG, Barker JN, Weisdorf DJ, DeFor TE, Miller JS, Blazar BR et al. Umbilical cord blood transplantation after nonmyeloablative conditioning: impact on transplant outcomes in 110 adults with hematological disease. Blood 2007; 8: 3064–3070.
Barker JN, Weisdorf DJ, DeFor TE, Blazar BR, McGlave PB, Miller JS et al. Transplantation of 2 partially HLA-matched umbilical cord blood units to enhance engraftment in adults with hematologic malignancy. Blood 2005; 105: 1343–1347.
Flynn CM, Hirsch B, Defor T, Barker JN, Miller JS, Wagner JE et al. Reduced intensity compared with high dose conditioning for allo transplantation in acute myeloid leukemia and myelodysplastic syndrome: a comparative clinical analysis. Am J Hematol 2007; 82: 867–872.
Majhail NS, Weisdorf DJ, Wagner JE, Defor TE, Brunstein CG, Burns LJ . Comparable results of umbilical cord blood and HLA-matched sibling donor hematopoietic stem cell transplantation after reduced-intensity preparative regimen for advanced Hodgkin lymphoma. Blood 2006; 107: 3804–3807.
Kaplan EL, Meier P . Nonparametric estimation from incomplete observations. J Am Stat Assoc 1958; 53: 457–481.
Lin DY . Non-parametric inference for cumulative incidence functions in competing risks studies. Stat Med 1997; 16: 901–910.
Cox DR . Regression models and life tables. J R Stat Soc B 1972; 34: 187–220.
Fine JP, Gray RJ . A proportional hazards model for the subdistribution of a competing risk. J Am Stat Assoc 1999; 94: 496–509.
Harrell FE, Caoliff RM, Pryor DB, Lee LK, Rosati RA . Evaluating the yield of medical tests. J Am Med Assoc 1982; 247: 2543–2546.
Efron E, Tibshirani RJ . An Introduction to the Bootstrap. Chapman & Hall, Inc.: New York, NY, USA, 1993.
About this article
Cite this article
DeFor, T., Majhail, N., Weisdorf, D. et al. A modified comorbidity index for hematopoietic cell transplantation. Bone Marrow Transplant 45, 933–938 (2010) doi:10.1038/bmt.2009.275
Risks and benefits in a personalized application of allogeneic transplantation in patients with AML in first CR
Seminars in Hematology (2019)
Der Hautarzt (2017)
Validation of the acute leukemia‐EBMT score for prediction of mortality following allogeneic stem cell transplantation in a multi‐center GITMO cohort
American Journal of Hematology (2017)
Current Dermatology Reports (2017)
Prognostic Value of the Hematopoietic Cell Transplantation Comorbidity Index for Patients Undergoing Reduced-Intensity Conditioning Cord Blood Transplantation
Biology of Blood and Marrow Transplantation (2017)