Predicting the risk of chemotherapy toxicity in older patients: The Chemotherapy Risk Assessment Scale for High-Age Patients (CRASH) score†
Version of Record online: 9 NOV 2011
Copyright © 2011 American Cancer Society
Volume 118, Issue 13, pages 3377–3386, 1 July 2012
How to Cite
Extermann, M., Boler, I., Reich, R. R., Lyman, G. H., Brown, R. H., DeFelice, J., Levine, R. M., Lubiner, E. T., Reyes, P., Schreiber, F. J. and Balducci, L. (2012), Predicting the risk of chemotherapy toxicity in older patients: The Chemotherapy Risk Assessment Scale for High-Age Patients (CRASH) score. Cancer, 118: 3377–3386. doi: 10.1002/cncr.26646
We thank Anna Davis, Jyoti Dubey, Tammy Green, and Ellie O'Neil, for their support in creating and managing the study database.
- Issue online: 18 JUN 2012
- Version of Record online: 9 NOV 2011
- Manuscript Accepted: 14 SEP 2011
- Manuscript Revised: 13 SEP 2011
- Manuscript Received: 15 AUG 2011
- chemotherapy side-effects;
- Common Toxicity Criteria for Adverse Events version 3.0;
- predictive score;
- Chemotherapy Risk Assessment Scale for High-Age Patients (CRASH) score
Tools are lacking to assess the individual risk of severe toxicity from chemotherapy. Such tools would be especially useful for older patients, who vary considerably in terms of health status and functional reserve.
The authors conducted a prospective, multicentric study of patients aged ≥70 years who were starting chemotherapy. Grade 4 hematologic (H) or grade 3/4 nonhematologic (NH) toxicity according to version 3.0 of the Common Terminology Criteria for Adverse Events was defined as severe. Twenty-four parameters were assessed. Toxicity of the regimen (Chemotox) was adjusted using an index to estimate the average per-patient risk of chemotherapy toxicity (the MAX2 index). In total, 562 patients were accrued, and 518 patients were evaluable and were split randomly (2:1 ratio) into a derivation cohort and a validation cohort.
Severe toxicity was observed in 64% of patients. The Chemotherapy Risk Assessment Scale for High-Age Patients (CRASH) score was constructed along 2 subscores: H toxicity and NH toxicity. Predictors of H toxicity were lymphocytes, aspartate aminotransferase level, Instrumental Activities of Daily Living score, lactate dehydrogenase level, diastolic blood pressure, and Chemotox. The best model included the 4 latter predictors (risk categories: low, 7%; medium-low, 23%; medium-high, 54%; and high, 100%, respectively; Ptrend < .001). Predictors of NH toxicity were hemoglobin, creatinine clearance, albumin, self-rated health, Eastern Cooperative Oncology Group performance, Mini-Mental Status score, Mini-Nutritional Assessment score, and Chemotox. The 4 latter predictors provided the best model (risk categories: 33%, 46%, 67%, and 93%, respectively; Ptrend < .001). The combined risk categories were 50%, 58%, 77%, and 79%, respectively; Ptrend < .001). Bootstrap internal validation and independent sample validation demonstrated stable risk categorization and Ptrend < .001.
The CRASH score distinguished several risk levels of severe toxicity. The split score discriminated better than the combined score. To the authors' knowledge, this is the first score systematically integrating both chemotherapy and patient risk for older patients and has a potential for future clinical application. Cancer 2011. © 2011 American Cancer Society.
Half of all malignancies occur in patients aged ≥70 years. Many such patients will be candidates for chemotherapy, whether adjuvant, curative in intent, or palliative. These patients vary considerably in health status and functional reserve, and the challenge is objectively evaluating the risk of treatment complications. In other domains, such as anesthesia, risk scores, such as the American Society of Anesthesiologists (ASA) score and the Physiologic and Operative Severity Score for the Enumeration of Mortality and Morbidity (POSSUM) score (http://www.asahq.org/clinical/physicalstatus.htm1; accessed October 19, 2011), have been in use for decades. No such scores exist for chemotherapy. We hypothesized that a predictive score could be built and validated across several chemotherapy regimens, taking into account both chemotherapy and patient variables. In a pilot study, we developed and validated a method for comparing the average risk of severe toxicity from various chemotherapy regimens: the MAX2 index.2, 3 We observed an independent contribution of several patient-specific variables.
Patients who develop severe toxicities often experience them during the first cycle of chemotherapy.4-7 Therefore, the search for patient-specific predictors allowing individualization of prediction beyond average risk has generated several studies.5, 8-13 Unfortunately, those studies assessed few parameters, rarely targeted elderly patients, and none had a systematic adjustment for various types of chemotherapy. Furthermore, studies conducted without a prospective collection of toxicity data likely underestimate side effects. However, they are helpful in providing a list of potential risk factors. The objective of the current prospective trial was to offer an analysis of a large number of suspected variables and to create and validate in independent patients a predictive score with potential for clinical application: the Chemotherapy Risk Assessment Scale for High-Age Patients (CRASH) score.
MATERIALS AND METHODS
Choice of the Endpoint and Candidate Variables
Our conceptual approach was as follows: Given the wide variety of regimens used in clinical practice, we wanted an index that would be valid across regimens. We hypothesized that we could identify a subgroup of risk factors common to multiple treatments. Because few publications specifically targeted elderly patients, we also included variables from untargeted studies. Our primary endpoint was the first occurrence of grade 4 hematologic (H) toxicity and/or grade 3/4 nonhematologic (NH) toxicity (“severe toxicity”). This endpoint was chosen because it is associated with instructions to modify the dosages of a regimen in most clinical trials. The time to first severe toxicity also was evaluated. We limited variable selection to baseline measures that can be obtained routinely in geriatric oncology practice: 1) the clinical variables included age, sex, body mass index (BMI),3 diastolic blood pressure,3 comorbidity measured with the Cumulative Illness Rating Scale-Geriatric (CIRS-G),14-16 and polypharmacy (drug counts).3 2) Laboratory variables included white blood cell count,17 hemoglobin,16 lymphocyte count,18 aspartate aminotransferase,3 creatinine clearance,11, 13, 19 albumin,8 and lactate dehydrogenase (LDH).3, 10 3) Geriatric and functional assessment variables included self-rated health; Eastern Cooperative Oncology Group (ECOG) performance status (PS),20 the Lawton 9-item Instrumental Activities of Daily Living (IADL),21 nutrition assessed with the Mini-Nutritional Assessment (MNA),22 cognition assessed with the Folstein Mini-Mental Status (MMS),23 and depression assessed with the Geriatric Depression Scale (GDS)-short form.9, 24 4) Cancer-specific variables included disease stage, bone marrow invasion (histologically proven or radiologic presence of bone metastases), prior chemotherapy, and response (rated as an ordinal variable with adjuvant setting rated best and disease progression rated worst). We did not include tumor type because 1) in developing the MAX2 index, we noted that, for a given regimen, tumor type has minimal influence on the patient risk of toxicity3; 2) no good grouping system of tumor types exists for this purpose; 3) the main effects of the tumor on the patient's general condition are likely captured by stage, functional and nutritional scores, and laboratory values; and 4) toxicity of the chemotherapy regimen (Chemotox), calculated using the MAX2 method.2, 3 The MAX2 index summarizes the overall risk of severe toxicity of a regimen for an average study patient using data from published clinical trials. Briefly, the MAX2 index is the average of the highest frequency of both grade 4 H toxicity and grade 3/4 NH toxicity (the “maximum 2” toxicities). It is reproducible across cancer types and studies, and it is sensitive to toxicity differences between regimens: For example, gemcitabine 1.25 g/m2 on days 1, 8, and 15 every 4 weeks has a MAX2 of 0.03 and 0.09 in nonsmall cell lung cancer, 0.065 in non-Hodgkin lymphoma, and 0.07 in Hodgkin disease.3 Docetaxel 75 mg/m2 every 3 weeks has a MAX2 of 0.253; and combined cyclophosphamide, doxorubicin, vincristine, and prednisone (CHOP) has a MAX2 of 0.341. The MAX2 score correlates well with the overall risk of severe toxicity in clinical trials, as tested in 2526 patients from ECOG trials2 (Fig. 1). The correlative equation is Chemotox = exp(−0.96 + 8.30*MAX2)/1 + exp(−0.96 + 8.30*MAX2).
Patients had a baseline assessment within 2 weeks before starting chemotherapy. Toxicity was followed through weekly complete blood counts, comprehensive metabolic panels, and medical evaluations at the beginning of each cycle and at the end of treatment, and identification of toxicities through medical records. Patients were followed throughout chemotherapy until a maximum of 1 month after last cycle. If chemotherapy was continued beyond 6 months, then follow-up was ended at 6 months to avoid skewing the data through long chemotherapy durations.
Patients aged ≥70 years were recruited at Moffitt Cancer Center (a National Cancer Institute-designated Comprehensive Cancer Center), Fawcet Memorial Hospital, Leesburg Regional Hospital, Morton Plant Hospital, Sarasota Memorial Hospital, Space Coast Associates, and Watson Clinic. These institutions are part of the Moffitt Affiliate Research Network, a formal network with institutional quality control on data and procedures. Patients had histologically proven cancer and were initiating a new first-line to fourth-line course of chemotherapy. Measurable disease was not required; but, when present, response was assessed according to Response Evaluation Criteria in Solid Tumors (RECIST). Patients had to be able to answer the questionnaires and to sign an informed consent form. Exclusion criteria were dementia and planned concomitant radiation therapy. Aplasia-inducing chemotherapy, such as acute leukemia treatment or bone marrow transplantation, was not allowed.
On the basis of our pilot, we hypothesized a 50% incidence of severe toxicity with 10 evaluable patients planned per variable with outcome in the derivation sample (n = 460). We defined the minimum threshold for a potentially clinically useful model as the ability to discriminate a group of patients with at least a 15% greater absolute risk than the remaining patients. On the basis of reasonable assumptions concerning variance and test performance, we targeted 150 patients for the validation sample for an 80% confidence of measuring a significant discrimination of patients with 15% or greater risk at 1 standard deviation above the mean of predictive covariates. Our intent was to accrue the whole cohort sequentially. However, because of slower than planned accrual on grant funding, 500 patients were accrued as planned (primary cohort), and 62 patients were recovered from previous Moffitt Cancer Center study cohorts that combined baseline geriatric assessment and prospective evaluation of toxicity using a methodology similar to that used for the primary cohort: chiefly, the pilot for the current trial (secondary cohort). We accrued 381 patients from Moffitt Cancer Center and 181 patients from affiliate sites. Forty-patients were ineligible because of screening failures (11 patients) withdrawal of consent (5 patients), the decision to receive chemotherapy elsewhere (5 patients), refusal by insurance (1 patient), disease progression/death before chemotherapy (2 patients), never received chemotherapy (2 patients), early withdrawal from chemotherapy (not toxicity related; 2 patients), missing laboratory assessments (12 patients), missing clinical assessments (1 patient), and loss to follow-up (3 patients).
To derive the CRASH model, approximately two-thirds of the original sample were selected randomly (n = 331 patients) using the random function generator from the SPSS statistical software package (SPSS version 17.0; SPSS Inc., Chicago, Ill). The remaining 187 patients constituted the independent validation cohort. This random split approach was chosen to neutralize accrual time bias.
Identification of individual predictors
The outcome variables of grade 3/4 or grade 4 H toxicities were treated as binary. Continuous measures were categorized into quartiles. For ECOG PS, because of a nonlinear progression, a dummy variable with scores of 0, 1 or 2, and 3 or 4 was created. Because these initial building steps were designed to catch all relevant variables, logistic regression was used to identify predictors with relatively liberal P values < .25. Because the chemotherapy regimen is determined independently by the oncologist, it was forced as an adjustment variable in all analyses instead of simply being considered a random covariate.
Multivariate identification of predictors
The identified individual predictors were tested in several logistic regression models, always including Chemotox. The final choice of predictors was based on a forward-selection approach, with the predictors being selected based on P values < .10. Feasibility was part of the selection method with the goal of identifying a model with a manageable number of predictors while not compromising greatly the amount of accounted variance (improvement in the Nagelkerke correlation coefficient [R2]) and model discrimination based on the area under the receiver operating characteristic curve (c-statistic). Model calibration was assessed using the Hosmer and Lemeshow goodness-of-fit test.25
Point values for the CRASH variables were assigned as follows: 1) We used a 3-point scale from 0 to 2 to offer a user-friendly format; 2) odds ratios (ORs) for severe event from the logistic regression analysis (stronger ORs [>1.5] warranted the full 3-point scale, whereas weaker ORs spanned only 2 points); 3) whenever a statistically significant linear trend was observed across the categories of a predictor, linear variables received points at each level of the 3-point scale. For this last factor, a linear trend analysis was conducted for each predictor. The CRASH score was developed by summing the points. Predictive variables by individual analysis were integrated according to their incremental contribution to the model performance. Items that added minimally to the model (Nagelkerke R2 improvement <0.01) were discarded. The raw total scores were then grouped into 4 risk categories using a best-split approach.
Validation of the model was performed with 2 methods. Bootstrapping according to the method of Steyerberg et al was conducted based on logistic regression predicting toxicities in 200 randomly selected samples with replacement from the derivation cohort of 331 patients.26 Then, validation was performed on a separate cohort of 187 patients. To pass validation, the model had to maintain a span of at least 15% between the lowest and highest risk categories, maintain progressivity among categories, and maintain discrimination, although some loss of performance was expected in the validation samples. This study was approved by the Institutional Review Boards of the University of South Florida and of each affiliate institution.
Baseline patient characteristics are listed in Table 1. Patients who received treatment at Moffitt Cancer Center and the affiliate sites had very similar characteristics overall. Moffitt Cancer Center patients were slightly more likely to have an ECOG PS of 2 and to screen positive for depression, whereas the affiliate patients were slightly more likely to be at risk of malnutrition.27 In total, 121 different chemotherapy regimens and schedules were received. At both Moffitt Cancer Center and the affiliate sites, the most frequent regimens used were carboplatin-based combinations. No regimen schedule was used in >10% of patients. Prophylactic growth factors were received by 155 patients (29.9%). Because these growth factors reduce the risk of grade 4 neutropenia by an average of 12%, the MAX2 score of their chemotherapy was adjusted downward by 0.06.
|Characteristic||Derivation Cohort, n = 331||Validation Cohort, n = 187||Overall, n = 518|
|Age, y||76 (70-92)||75 (70-89)||75.5 (70-92)|
|BMI, kg/m2||25.9 (14.8-51)||26.1 (16.4-43.7)||26 (14.8-51)|
|Diastolic BP, mm Hg||73 (40-101)||71 (50-95)||72 (40-101)|
|CIRS-G Severity Index scorea||1.70 (0-3)||1.75 (1-4)||1.71 (0-4)|
|No. of medications||6 (0-20)||6 (0-23)||6 (0-23)|
|WBC, ×103/μL||7.3 (2.28-91.54)||7.0 (2.9-83.3)||7.17 (2.28-91.54)|
|Hemoglobin, g/dL||12.3 (7.9-17.4)||12.2 (7.7-16.4)||12.3 (7.7-17.4)|
|Platelets, H×103/μL||241.5 (49-723)||231.5 (20-595)||236 (20-723)|
|AST, U/L||25 (10-393)||25.5 (8-271)||25 (8-393)|
|LDH, U/L||517 (98 -9394)||502 (139-5620)||514 (98-9394)|
|Creatinine clearance, mL/min||60.35 (12.57-133)||60.15 (17-130)||60.3 (12.57-133)|
|Albumin, g/dL||3.8 (1.7-4.8)||3.7 (2.4-4.4)||3.8 (1.7-4.8)|
|Self-rated healthb||3.0 (1-5)||3.0 (1-5)||3.0 (1-5)|
|ECOG PS, %|
|IADL scorec||28 (14-29)||28 (11-29)||28 (11-29)|
|MNA scored||25 (11-30)||25 (8-30)||25 (8-30)|
|MMS scoree||28 (15-30)||28 (21-30)||28 (15-30)|
|GDS scoref||2 (0-12)||2 (0-12)||2 (0-12)|
|Cancer type, %|
|Cancer stage, %|
|Bone marrow invasion, %||16.3||27.8||20.5|
|Received prior treatment, %|
|Treatment intent, %|
|Tumor response, %|
|Toxicity of chemotherapy regimen: MAX2 scoreh||0.17 (0.02-0.78)||0.17 (0.04-0.55)||0.17 (0.02-0.78)|
Overall, 254 patients (49.1%) completed their assigned treatment, moved to another modality, or were continuing at 6 months. Seventy-four patients (14.3%) stopped for disease progression, 121 (23.4%) stopped for toxicity, 52 (10%) stopped chemotherapy or study follow-up because of patient withdrawal/refusal/seasonal migration, and 11 (2.1%) stopped for miscellaneous other reasons. Seventeen patients died during chemotherapy, including 9 (1.7%) who died from complications and 8 (1.5%) who died of progressive disease.
Sixty-four percent of patients experienced severe toxicity, 32% had grade 4 H toxicity, and 56% had grade 3 or 4 NH toxicity. The median time to first toxicity was 22 days (interquartile range, 9-51 days). All variables had <3% missing values with the exception of LDH (20% missing, mostly in the secondary sample). Missing values were imputed using the median value of the variable. Sensitivity analyses were conducted with alternate imputation methods for LDH, with median substitution performing best. Such an approach leads to a conservative estimate of our model's predictive value.
Only hemoglobin (OR, 0.92; P = .24) and creatinine clearance (OR, 0.87; P = .17) were correlated with severe toxicity as individual variables. Combining these predictors into a logistic regression model yielded a weak risk model (Nagalkerke R2, <0.05). Because various data from work conducted in parallel with the CRASH trial pointed toward different predictors for H and NH toxicity,28 a 2-subcomponent approach was tested with an H subscore and an NH subscore. In individual predictor analysis, diastolic blood pressure, IADL, aspartate aminotransferase, lymphocytes, and LDH were associated with grade 4 H toxicity. ECOG PS, hemoglobin, creatinine clearance, albumin, MMS, self-rated health, and MNA were correlated with grade 3/4 NH toxicity. ORs and P values are listed in Table 2. The best performing model for H toxicity included diastolic blood pressure, IADL, and LDH along with Chemotox (Tables 3 and 4; Fig. 2). The c-statistic was 0.76. The best performing model for NH toxicity included ECOG PS, MMS, MNA, and Chemotox. The c-statistic was 0.66. The combination of the 2 subscores (counting Chemotox only once) yielded a model with a c-statistic of 0.65. The lower c-statistic in the NH and combined models may be partially the result of range restriction because of high base rates of toxicities. The c-statistic is correlated with the OR and, thus, reflects a residual risk of 33% and 50% in each lowest category, which only allows the scores to span a risk ratio of 2 to 3 times. In all models, there was a regular progression of the risk (trend test; P < .001). The Hosmer and Lemeshow test did not reveal any significant divergence in calibration. Bootstrap validation yielded very stable results with no significant shrinkage of the data.
|Hematologic Toxicity||Nonhematologic Toxicity|
|Biomarker||OR (95% CI)||P||OR (95% CI)||P|
|Age||0.99 (0.94-1.05)||.83||1.01 (0.96-1.06)||.76|
|Sex||0.68 (0.37-1.30)||.25||1.28 (0.83-1.98)||.27|
|BMI||0.03 (0.97-1.08)||.33||0.99 (0.95-1.04)||.84|
|Diastolic BP||1.30 (1.02-1.65)||.03b||1.00 (0.98-1.02)||.96|
|CIRS Severity Index||1.08 (0.61-1.91)||.78||1.09 (0.66-1.81)||.73|
|Polypharmacy||1.00 (0.94-1.07)||.93||1.03 (.97-1.08)||.35|
|WBC||1.02 (0.98-1.06)||.33||0.98 (0.94-1.03)||.45|
|Hemoglobin||0.99 (0.85-1.15)||.91||0.90 (0.79-1.03)||.12b|
|Lymphocytes||1.05 (0.98-1.12)||.20b||0.95 (0.85-1.05)||.32|
|AST||1.01 (1.00-1.01)||.14b||1.00 (1.00-1.01)||.43|
|CrCL||1.0 (0.99-1.01)||.74||0.84 (0.69-1.03)||.09|
|Albumin||0.76 (0.44-1.30)||.39||0.74 (0.45-1.20)||.22b|
|LDH||1.41 (1.17-1.86)||.004b||1.00 (1.00-1.00)||.69|
|Self-rated health||1.02 (0.80-1.29)||.89||0.87 (0.71-1.07)||.19b|
|ECOG PS||1.13 (0.81-1.57)||.47||1.41 (1.05-1.89)||.03b|
|IADL||0.77 (0.58-1.03)||.08b||0.98 (0.91-1.06)||.58|
|MNA||0.99 (0.92-1.06)||.69||0.73 (0.60-.90)||.003b|
|MMS||0.97 (0.87-1.07)||.51||0.77 (0.63-93)||.008b|
|GDS||1.00 (0.91-1.12)||.94||1.04 (0.95-1.14)||.44|
|Tumor stage||1.00 (0.77-1.30)||.99||1.10 (0.87-1.39)||.65|
|Bone marrow invasion||1.19 (0.61-2.33)||.62||1.46 (0.78-2.72)||.26|
|Prior chemotherapy||1.30 (0.75-2.24)||.35||0.85 (0.53-1.36)||.50|
|Tumor response||0.96 (0.66-1.4)||.83||1.16 (0.84-1.63)||.44|
|Chemotox||2.20 (1.72-2.81)||<.001b||1.13 (0.93-1.37)||.23b|
|Step||Predictor||−2 Log Likelihood||R2b||C- Statistic||P (R2 Change)|
|LDH (if ULN 618 U/L; otherwise, 0.74 /L*ULN)||0-459||>459|
Independent sample validation supported the robustness of the models. Quantitative results were stable with the exception of the high-risk H category, which included only 2 patients: 1 with and 1 without toxicity. There was some subsequent decrease in the c-statistic. In the other categories, the decrease in c-statistic was minimal, and the trend test remained P < .001 (Table 5).5
|Optimal Model Statistics||Derivation||Bootstrap Validation||Independent Sample Validation|
|C-statistic (95% CI)||0.76 (0.71-0.82)||0.77 (0.72-0.83)||0.65 (0.57-0.73)|
|Hosmer and Lemeshow Pa||.14||.19||.71|
|Linear trend P||<.001||<.001||<.001|
|C-statistic (95% CI)||0.66 (0.59-0.70)||0.65 (0.58-0.70)||0.62 (0.54-0.70)|
|Hosmer and Lemeshow P||.85||.70||.32|
|Linear trend P||<.001||<.001||<.001|
|C-statistic (95% CI)||0.65 (0.59-0.71)||.65 (.59-.71)||0.64 (0.56-0.72)|
|Hosmer and Lemeshow P||.85||.70||.32|
|Linear trend P||<.001||<.001||<.001|
|Capecitabine 2g||Capecitabine 2.5 g||5-FU/LV (Roswell-Park)|
|Cisplatin/pemetrexed||Carboplatin/gemcitabine AUC 4-6/1 g d1,d8||5-FU/LV (Mayo)|
|Dacarbazine||Carboplatin/pemetrexed||5-FU/LV and bevacizumab|
|Docetaxel weekly||Carboplatin/paclitaxel q3w||CAF|
|FOLFIRI||Cisplatin/gemcitabine d1,d8||Carboplatin/docetaxel q3w|
|Gemcitabine 1 g 3/4 wk||ECF||CHOP|
|Gemcitabine 1.25 g 3/4 wk||Fludarabine||Cisplatin/docetaxel 75/75|
|Paclitaxel weekly||FOLFOX 85 mg||Cisplatin/etoposide|
|Pemetrexed||Gemcitabine 7/8 wk then 3/4 wk||Cisplatin/gemcitabine d1,d8,d15|
|Gemcitabine/irinotecan||Cisplatin/paclitaxel 135-24 h q3w|
|PEG doxorubicin 50 mg q4w||CMF classic|
|Topotecan weekly||Doxorubicin q3w|
|XELOX||FOLFOX 100-130 mg|
Our objective was to design a score capable of distinguishing at least 2 patient groups with an absolute risk difference ≥15%. Although this could not be achieved directly by a global set of predictors, it could be accomplished with a score composed of 2 subscales. With 2 subscales, our risk range was wide, especially in the H toxicity category. The CRASH score allows us to stratify patients into 4 risk categories (low, medium-low, medium-high, and high), and the overall stability of our results over 2 validations indicates that our model is reliable. Nevertheless, the combined risk of severe toxicity remains significant even in the lowest category, and this is a reminder of the need for careful monitoring of any older patient on chemotherapy. We demonstrated that patient differences contribute 2 to 3 times more than chemotherapy differences to the risk of toxicity. Our study confirmed that H and NH toxicities largely are associated with different predictors, which may prove important for designing future trials. According to need, the CRASH score offers the flexibility of being used either as a predictor of overall severe toxicity or as a predictor of H toxicity versus NH toxicity. Either approach may be useful in stratifying patients for clinical trials and assisting clinical decision making. It is noteworthy that the CRASH score was developed in both the academic and community setting. This enabled us to compare the treated populations from both settings, which were very similar. Although we cannot exclude an unmeasured selection bias, this finding bodes well for widespread reliability in its application. Our adjustment method for toxicity of chemotherapy, MAX2, has the advantage of being validated, but it may appear cumbersome for a clinical application. However, this can be addressed by making a list of scores available online (www.moffitt.org/saoptools; accessed October 19, 2011) (See also Table 6).
Despite a high rate of combined toxicity, older patients generally appear to be capable of tolerating and completing planned chemotherapy.29 Our treatment-related mortality was <2%, which compares very well with results from randomized studies. Therefore, severe toxicity according to Common Toxicity Criteria for Adverse Events, version 3.0 (CTCAE) should be understood as “severe” in the sense of considering changes to the chemotherapy but does not necessarily impact the patient's function or well being. An analysis by our group suggests that about 1 in 3 of these events may not warrant dose modification to avoid recurrence.30 This opens an important avenue of research: How can the CTCAE criteria be refined to better identify which toxicities warrant changes in treatment, especially for older patients? The CRASH score can be a very helpful stratification tool for such studies.
Our results also highlight the power of geriatric instruments in predicting the outcome of chemotherapy. Three of those instruments were retained as dominant variables in the analysis despite being compared with several common oncologic predictors. Patients with excellent performance on these instruments appeared to be at low risk. MMS alterations may be surrogates for inflammatory status31 or physiologic and psychological stress. Patients with cancer often present with mild cognitive impairment at baseline.32 Elevated diastolic blood pressure also may reflect an inflammatory or stress effect.33 The MNA performed better than albumin level. The 2 factors are correlated,34 and the MNA incorporates other potential influencers. The item “more than 3 prescription medications,” for example, frequently was checked. There was no difference in risk between patients who had an ECOG PS of 1 and patients who had an ECOG PS of 2, highlighting the difficulty of distinguishing the 2 levels in the elderly. However, because an ECOG PS of 2 was associated with more toxicity in younger patients (eg, see Sweeney et al35), formal validation of the CRASH score should be conducted in patients aged <70 years. The low impact of creatinine clearance in the multivariate model is most likely because oncologists already adjust for it when planning treatment dosing.
To put the CRASH score into context, it provided an integrated risk score that was stable over 2 validations. This goes beyond previous research that assessed a limited number of factors and had no validation samples. A similar score was presented with ours at the 2010 American Society of Clinical Oncology meeting but has yet to be validated.36 Such scores are an important step in individualizing risk prediction in the elderly. Further research should build on this model-development study. For example, the CRASH score could be tested in other cohorts, or the most discriminative questions could be extracted from geriatric instruments to shorten the evaluation time needed and to sharpen the predictivity of the CRASH score. Estimating the risk of severe toxicity from chemotherapy is only 1 aspect of treatment planning in older patients and should be integrated into a multidisciplinary oncogeriatric approach to the decision. A lesser level of specific toxicities also may be important for a given patient.
In conclusion, we developed a predictive instrument with a performance level that may be useful in assisting the clinical decision making of oncologists by providing an objective risk estimate. An instrument that is valid across a wide range of chemotherapies, such as the CRASH score, clearly is needed given the impressive variety of regimens used by oncologists. Further validation in other populations of older patients is recommended.
This work was supported by grant RSG-03-151-01-CCE from the American Cancer Society.
CONFLICT OF INTEREST DISCLOSURES
Dr. Extermann has received honoraria from Amgen and Sanofi. Dr. Lyman has received research support from Amgen. Dr. Balducci has received honoraria from Amgen, Cephalon, Novartis, and Sanofi.
- 6Risk and timing of neutropenic events in adult cancer patients receiving chemotherapy: the results of a prospective nationwide study of oncology practice. J Natl Compr Canc Netw. 2008; 6: 109-118., , , et al.
- 13Analysis of age, estimated creatinine clearance and pretreatment hematologic parameters as predictors of fludarabine toxicity in patients treated for chronic lymphocytic leukemia: a CALGB coordinated intergroup study. Cancer Chemother Pharmacol. 2002; 50: 37-45., , , et al.
- 20Appraisal of methods for the study of chemotherapy in man: comparative therapeutic trial of nitrogen mustard and triethylene thiophosphoramide. J Chron Dis. 1960; 11: 7-33., , , et al.
- 22Mini Nutritional Assessment: a practical assessment tool for grading nutritional state of elderly patients. In: Vellas BJ, Guigoz Y, Garry PJ, Albarede JL, eds. Facts Research and Intervention in Geriatrics. 3rd ed. New York: Serdi Publishing Company; 1997: 15-60., , .
- 24Geriatric Depression Scale (GDS). Recent evidence and development of a shorter version. Clin Gerontol. 1986; 5( 1/2): 165-171., .
- 25Applied Logistic Regression. 2nd ed. New York: John Wiley & Sons, Inc.; 2000., .
- 28The impact of polypharmacy on toxicity from chemotherapy in elderly patients: focus on cytochrome P-450 inhibition and protein binding effects [abstract]. J Clin Oncol. 2008; 26( 15S). Abstract 9505., , , .
- 30Chemotherapy dose adjustment after severe toxicity in older cancer patients [abstract]. J Clin Oncol. 2008; 26(May 20 suppl). Abstract 9627., .