Survival curves for 1094 patients according to severity of comorbidity (log rank = 77.17; P<.001). WUHNCI indicates Washington University Head and Neck Comorbidity Index.
Customize your JAMA Network experience by selecting one or more topics from the list below.
Piccirillo JF, Lacy PD, Basu A, Spitznagel EL. Development of a New Head and Neck Cancer–Specific Comorbidity Index. Arch Otolaryngol Head Neck Surg. 2002;128(10):1172–1179. doi:10.1001/archotol.128.10.1172
Most patients with head and neck squamous cell carcinoma are older and may have coexistent or comorbid diseases.
To determine the prognostic impact of individual comorbid conditions in patients with head and neck cancer, to combine the individual comorbid conditions to form a new a head and neck–specific comorbidity instrument, and to compare it with the Modified Kaplan-Feinstein Index to determine if the new disease-specific instrument offers any improvement in survival prediction over a general comorbidity index.
Retrospective review of medical records.
The study population comprised 1153 patients with biopsy-proven, newly diagnosed squamous cell carcinoma of the oral cavity, oropharynx, or larynx.
Seven comorbid conditions (congestive heart disease, cardiac arrhythmia, peripheral vascular disease, pulmonary disease, renal disease, cancer controlled, and cancer uncontrolled) were significantly related to survival. These comorbid conditions were assigned integer weights to reflect their relative prognostic importance and combined to create the new Washington University Head and Neck Comorbidity Index (WUHNCI). Survival was significantly related to levels of comorbidity severity as defined by the WUHNCI. The WUHNCI predicted survival better than the Modified Kaplan-Feinstein Index despite containing far fewer ailments.
Comorbidity is an important feature of the patient with head and neck cancer. The WUHNCI can be used for retrospective review or prospective outcomes research.
MOST PATIENTS with head and neck squamous cell carcinoma are older and many have coexistent, nonneoplastic (comorbid) diseases.1 These conditions may be mild and may not affect survival rates. Examples of mild comorbidity include hypertension that is controlled by medication or a diagnosis of peptic ulcer disease. However, some comorbidities or combinations of comorbidities may be so severe that they can affect survival rates. These comorbidities are referred to as prognostic comorbidity2 and include recent myocardial infarction or ventricular arrhythmia, severe hypertension, severe hepatic disease, and recent severe stroke. These conditions may affect the selection of initial treatment and patient outcome. For example, a patient may not be offered a supraglottic laryngectomy because his or her preexisting lung disease is so severe that the aspiration associated with a partial larynx may lead to life-threatening postoperative pneumonia. Another patient who is "too sick" to tolerate a preferred treatment may be given a less effective or even palliative treatment. In several studies of cancer prognosis, the presence of comorbidity was found to dramatically affect survival and the evaluation of treatment effectiveness, even after controlling for TNM stage.2-6
A number of validated instruments have been developed to classify comorbidity. The Kaplan-Feinstein Comorbidity Index (KFI)2 was developed from a study of the impact of comorbidity on 5-year survival for a cohort of male patients with diabetes mellitus. Another validated comorbidity instrument is the Charlson Comorbidity Index (CCI).7 This instrument was created from a study of patients admitted to a general medical unit of a teaching hospital.
These 2 comorbid instruments were developed from the study of outcomes for general medical patients, not patients with cancer. While various comorbid ailments are common to all populations, the frequency distribution and the relative prognostic impact of each condition to the primary disease process may vary. Furthermore, within a cohort of patients with cancer, the relative impact of individual comorbidities may vary across different cancer types. For example, head and neck cancer is primarily a disease of the elderly, and the pattern of comorbidities in such a patient population may be quite different than that of patients with cervical cancer.
The purposes of the present study were to (1) determine the prognostic impact of individual comorbid conditions in patients with head and neck cancer, (2) combine the individual comorbid conditions to form a new a head and neck–specific comorbidity instrument, and (3) compare it with the KFI to determine if the new disease-specific instrument offers any improvement in prediction of survival over a general comorbidity index.
The study population comprised 1153 patients identified from the pathology records of Barnes-Jewish Hospitals (St Louis, Mo), who had biopsy-proven, newly diagnosed squamous cell carcinoma of the oral cavity, oropharynx, or larynx and had received their initial treatment at Washington University Medical Center (St Louis) between January 1, 1980, and December 31, 1991. Medical records from the Washington University Department of Otolaryngology–Head and Neck Surgery, the Mallinckrödt Institute of Radiology (St Louis), and Barnes-Jewish Hospitals were used to collect baseline demographic and clinical information, initial treatment, and follow-up data. Full 5-year follow-up data were available for 1094 patients (95%) who composed the inception cohort.
Initial "zero time" was defined as the date of commencement of first antineoplastic therapy to the primary tumor or the date the decision was made not to treat a patient. If no such decision was made, the date of diagnosis was used as zero time.
Information for each patient was recorded on a specially designed data extraction form. Data collected before treatment or at zero time included basic demographic information; a description of symptoms and classification of symptoms into a symptom severity system; tobacco habits; medical history with a detailed documentation of comorbid conditions; symptom type and duration; presence of synchronous tumors; results of pertinent radiographic and laboratory studies; pathologic description of the biopsy specimen; and complete anatomic description of the tumor. Data collected after zero time included details of initial therapy, complications, duration to last follow-up, survival duration, and tumor status at last follow-up or death.
To ensure scientific accuracy and high-quality data, specific techniques applied in this study were devised and maintained in a coding handbook, which was referenced during coding sessions. Three different coders (all physicians) extracted the data with systematic interobserver consistency checks. Only minor differences in coding were seen, and in general, these did not cause migration of subjects between categories and were believed to be clinically insignificant.
Previous research demonstrated that cancer-related symptom severity is a measure of the biological index of disease.8 In particular, dysphagia, otalgia, neck lump, and weight loss were identified as independent predictors of survival. Therefore, the presence of these 4 symptoms at zero time was noted, and each patient was classified into 1 of 4 symptom-severity stages (none, mild, moderate, or severe). Symptom-severity stage none is defined by the absence of all 4 symptoms; mild, 1 of the 4 symptoms was recorded as present; moderate, 2 of the 4 symptoms were recorded as present; or severe, 3 or 4 symptoms were recorded as present.
The TNM classification assigned to the patient by the attending physician at zero time was used when this was clearly defined in the medical record. Otherwise, all information obtained prior to zero time, including clinical, endoscopic, and radiographic findings, was used for classification based on the 1992 American Joint Committee on Cancer criteria.9 If more than 1 physician staged a tumor differently and the medical record lacked sufficient anatomic data to identify which stage was correct, then the stage assigned by the most senior physician was recorded. Occasionally, a patient's TNM classification was changed from that which was recorded in the medical record, but only when it was clear that the original TNM classification was incorrect.
The grade of the primary tumor was recorded for all patients using the well, moderately, and poorly differentiated grades of the Broder system.10 As undifferentiated implies that the tissue of origin is unknown, no tumors of this histopathologic grade were included. The histopathologic grade was recorded from the biopsy specimens. When 2 or more conflicting reports of histopathologic grade were reported (eg, from multiple biopsy specimens), the most anaplastic grade was recorded. If the grade was not recorded, it was classified as well differentiated.
The Modified Kaplan-Feinstein Index (MKFI) was used to classify the overall severity of comorbidity. The MKFI is based on the original KFI,2 which was first described in 1974. The MKFI has been used in a variety of studies, including head and neck cancer.1 One of us (J.F.P.), along with other health services researchers at Washington University, modified the KFI to reflect additional diseases and conditions not included in the original KFI. These diseases and conditions include dementia, Parkinson disease, human immunodeficiency virus and acquired immunodeficiency syndrome, peripheral vascular disease, and obesity. In addition, because the original KFI was developed from the study of patients with diabetes mellitus, this condition was not listed as a comorbid ailment. The terms "cancer controlled" and "cancer uncontrolled" were defined as follows:
Cancer controlled: in the case of solid tumors, a cancer that has been treated and there is no evidence of residual disease or recurrence; in the case of lymphoma and leukemia, cancer controlled implies the above or the lymphoma or leukemia is present but indolent and the patient is not currently receiving treatment.
Cancer uncontrolled: the situation in which the patient has a synchronous tumor at the time of diagnosis, received treatment for a previous cancer but the tumor has persisted, or the patient's tumor responded to treatment initially but there is recurrence of the previous cancer at the time the index cancer is diagnosed.
The original KFI and the MKFI categorizes patients into 1 of 4 overall comorbidity groups (none, mild, moderate, or severe) based on the existence of specific diseases and conditions. Cogent individual comorbid ailments are classified according to their severity of organ decompensation: grade 1, mild; grade 2, moderate; or grade 3, severe. The overall comorbidity severity score is defined by the grade of the highest ranked single ailment or, in the case in which there are two grade 2 ailments in different organ systems, the score is grade 3 (severe).
At the time of original chart abstraction, an MKFI score was determined by a review of the comorbid ailments. This score was determined without reference to survival or other important end points. The MKFI is available from the Clinical Outcomes Research Office's Web site (http://oto.wustl.edu/clinepi).
At the time of original medical record review, a special 132-item comorbidity form was used to record the presence and severity of individual medical comorbidities. Only conditions present at or before the time of diagnosis were recorded. The comorbidity form and a full description of the coding criteria for these specific comorbid conditions are available from one of us (J.F.P.).
The first step in the development of the WUHNCI was the identification of common comorbid conditions among the 132 conditions listed on the comorbidity form. Any condition affecting less than 1% of the cohort was considered too uncommon and was excluded from further analysis. Next, the prognostic effect of each of the common (ie, prevalence greater than 1%) comorbid conditions was determined. A series of cross-tabulation tables of individual comorbid conditions on 5-year survival was performed, and χ2 analysis was used to assess the statistical significance of the observed relationships. The comorbid ailments that affected survival at a P value of .1 or less were considered potential independently significant prognostic variables. Variables that achieved this level of significance were entered into a multivariable logistic regression model to determine which comorbid conditions, when controlling for the other significant comorbidities, affected survival. Those comorbidities that maintained independent prognostic significance at a P value of .1 or less were then included in the WUHNCI. Because each comorbid condition had a different impact on survival, each condition was weighted in the final WUHNCI according to its prognostic impact. The whole integer weighting for each comorbid condition was determined by the magnitude of the parameter estimate from the multivariable logistic regression model. The parameter estimate is a measure of how much change will result in the dependent variable (eg, 5-year overall survival) with every unit change in the independent variable (eg, specific comorbid ailment). The WUHNCI comorbidity score is calculated as the sum of the weights of each of the comorbid conditions that are present within the patient.
Data collected on each patient's initial and subsequent treatment(s) included type of treatment (ie, radiotherapy, surgery, chemotherapy, or combination), timing of radiotherapy (ie, preoperative or postoperative), and type of surgical procedure. Postoperative radiotherapy was coded as a combined initial treatment when the pretreatment decision to use combined therapy was explicit in the medical record.
Follow-up information was obtained from medical records from the Department of Otolaryngology–Head and Neck Surgery and the Mallinckrödt Institute of Radiology (Washington University School of Medicine) and Barnes-Jewish Hospital, Barnes Hospital Oncology Data Services, and the Equifax National Death Search (Arlington, Va) service. When needed for individual patients, additional information was obtained from other hospitals and consulting or referring physicians. Follow-up was maintained until a patient's death was documented or until the end of the study period (December 1, 1997). The primary outcome measure was 5-year overall survival. Approval from the Human Studies Committee of Washington University School of Medicine was obtained before commencement of the study.
The information contained on the data extraction forms was entered into a Paradox database (Borland International, Scotts Valley, Calif), using a data entry screen that was identical to the extraction form. Equipped with internal validity checks, the specially designed screens facilitated reliable and efficient data entry. Periodic review for internal consistency and comparison of separate databases provided verification of the entered data. Sorting, tabulation, and statistical analyses including the (χ2, bivariate odds ratio and 95% confidence limits, and multivariate analysis were performed using the SAS statistical analysis software system, release 6.12 (SAS Institute Inc, Cary, NC).
The utility of the WUHNCI and comparison of its predictive ability with the MKFI were assessed from 4 different logistic regression models. All models contained age, sex, race, symptom stage, and TNM stage. In model 1, no comorbidity measure was added. In model 2, the WUHNCI alone was added to the model. In model 3, the MKFI alone was added. In model 4, both the WUHNCI and the MKFI were entered into the model. For each of these models, the log-likelihood ratio χ2 test and the c-statistic were used to compare the prognostic performance. Among pairs of patients in which 1 patient lives and 1 dies, the c-statistic11 reflects the proportion in which a higher risk is assigned to the patient who died than to the one who lived. The c-statistic is graphically represented as the area under the receiver-operating characteristic curve.12,13 The values for the c-statistic range from 0.5 (no discrimination) to 1.0 (perfect discrimination). According to Ohman et al,14 a predictive model with a c-statistic less than 0.6 has no clinical value, 0.6 to 0.7 has limited value, 0.7 to 0.8 has modest value, and greater than 0.8 has discrimination adequate for genuine clinical utility.
The validity of the methodology for identifying the cogent comorbid factors and assignment of integer weights was assessed using a split-half analysis approach. In this technique, the database was divided randomly into 2 groups. The cogent comorbid variables and integer weights were determined from the patients contained within one half of the cohort. The model was then tested in the other half of the cohort using the same multivariable analytic techniques as described for the evaluation of the WUHNCI.
The cohort of 1094 patients had a mean ± SD (range) age of 62.1 ± 11.2 (16-100) years. The median age was 62 years, mode age was 65, and the 25% to 75% interquartile range was 55.0 to 69.8 years. There were 785 men (72%), and 918 (84%) were white. Mean ± SD survival was 68 ± 56 months, median survival was 56 months, and the overall 5-year survival rate was 47% (514 of 1094 patients).
Table 1 shows the relationship between baseline features of the patient population and 5-year survival rates. There was a statistically significant relationship between survival and age group, race, symptom severity, TNM stage, and degree of histological differentiation. Sex, history of cigarette smoking, and type of initial treatment did not significantly affect survival.
In Table 2, the prevalence of individual comorbid conditions is given. The 5 most common conditions (percentage frequency) were pulmonary disease (17.9%), other cancer controlled (8.6%), diabetes mellitus (7.9%), myocardial infarction (6.7%), and peptic ulcer disease (5.2%). Interestingly, the frequency of hypertension and rheumatological and psychiatric illness was very low. The comorbid conditions with a prevalence of less than 1% were excluded from further analysis.
The impact of the common comorbid conditions on survival is demonstrated in Table 3. The following 7 comorbid conditions were significantly related to survival: congestive heart disease, cardiac arrhythmia, peripheral vascular disease, pulmonary disease, renal disease, other cancer controlled, and other cancer uncontrolled. Survival was related to myocardial infarction, although this effect did not achieve statistical significance (P = .08). Because multiple comorbid variables affected survival, the next step was to perform a multivariable analysis.
Table 4 demonstrates the results of multivariable analysis of the impact of the 8 comorbid conditions on survival. As can be seen by the P value and 95% confidence limits around the odds ratio, congestive heart failure, cardiac arrhythmia, peripheral vascular disease, pulmonary disease, renal disease, other cancer uncontrolled, and renal disease were all significantly associated with survival. Other cancer controlled approached significance (P = .07) and was retained in the model building. Myocardial infarction was not significant (P = .5) and was dropped from further analysis.
The prognostic importance of each of the 7 significant comorbid conditions is demonstrated by the parameter estimate. The parameter estimates for the 2 comorbid conditions with the weakness impact (pulmonary disease [0.396] and other cancer controlled [0.442]) were selected as the baseline value for the determination of the whole integer weights for all 7 conditions. Pulmonary disease and other cancer controlled were both assigned the integer weight of 1. Whole integer weights were then assigned to the other 5 comorbid conditions based on the relationship of the parameter estimates. For example, the parameter estimate for congestive heart disease was 1.064 and a whole integer value of 2 was selected as the weight for this comorbid condition because 1.064 is approximately twice as large as 0.396 (the parameter estimate of pulmonary disease) and 0.442 (the parameter estimate of other cancer controlled).
In Table 5, the WUHNCI score is given with the whole integer weights for all 7 comorbid conditions. In Table 6, the prognostic impact of each of the values of the WUHNCI and the 4-category consolidated index are given. As demonstrated by the χ2 statistic, there is a strong prognostic impact of comorbidity. As the level of comorbidity increases, the survival rate decreases. This relationship is further demonstrated in Figure 1.
The prognostic impact of the newly created WUHNCI was compared in a multivariable logistic regression model with the other significant demographic, clinical, and tumor factors. As shown in Table 7, all demographic, clinical, and tumor factors were significantly related to 5-year survival.
The c-statistic for the overall model, including the WUHNCI, was 0.754. This compares with a c-statistic of 0.756 for a model that contained the same demographic, clinical, and tumor factors, but included the significant comorbid ailments as individual conditions rather than being weighted and combined as was done to create the WUHNCI. Because the c-statistic for the WUHNCI is very close to the c-statistic for the unrounded multivariate model, replacing parameter estimate values with rounded integer weights captures nearly all the prognostic information of the comorbid conditions individually.
The performance of the WUHNCI was compared with that of the MKFI. In Table 8, the performance of 4 models is given. For all 4 models, age, sex, race, symptom severity stage, and TNM stage were included. For model 1, no comorbid factors were added. For model 2, the WUHNCI was added. For model 3, the MKFI was added, and for model 4, both the WUHNCI and MKFI were added. As shown in Table 8, either the WUHNCI or the MKFI adds significantly to the predictive power of a model that does not contain a comorbidity factor. The likelihood ratio and c-statistic shows that the WUHNCI performs significantly better than the MKFI, since omitting the MKFI from the model containing both terms doesn't drop the likelihood ratio by a statistically significant amount or change the c-statistic. Whereas, omitting the WUHNCI from the model that contains both terms does drop the likelihood ratio χ2 by a significant amount and changes the c-statistic.
The split-half analysis identified the same 7 cogent comorbid factors as the original research, although the integer weights were slightly different (congestive heart failure , cardiac arrhythmia , peripheral vascular disease , pulmonary disease , renal disease , other cancer controlled , and other cancer uncontrolled ). The predictive model containing age, sex, race, symptom severity, TNM stage, and the new comorbidity variable, as tested in the other half of the cohort, performed quite well (likelihood ratio χ2, 116.514; c-statistic, 0.766). The new comorbidity variable performed better than the MKFI.
Patients with head and neck cancers often have other diseases, illnesses, and conditions in addition to their index tumor. These other conditions are often referred to as comorbidities. The present study identifies the 7 important prognostic comorbidities for patients with head and neck cancer. These comorbidities are congestive heart failure, cardiac arrhythmia, peripheral vascular disease, pulmonary disease, renal disease, previous history of cancer now controlled, and previous history of cancer now uncontrolled. This study also describes the creation of a disease-specific composite comorbidity index, the WUHNCI. This index was found to predict survival better than the MKFI despite containing far fewer comorbid ailments.
General comorbidity instruments were developed and are intended to be used across a wide range of clinical conditions. However, disease-specific comorbidity instruments are intended for an individual ailment or cluster of closely related ailments. Despite their apparent lack of specificity, general instruments perform well as predictors of important outcomes for specific diseases. For example, Singh et al15 found that the severity of comorbidity, as defined by the CCI,7 was significantly related to 5-year survival for a cohort of patients with head and neck cancer.
However, some investigators have found that disease-specific instruments perform better in specific disease conditions than a general instrument. For example, Polanczyk and colleagues16 recently reported that a disease-specific index performed better than CCI in predicting in-hospital mortality for patients with congestive cardiac failure. In another study, Fleming et al17 indicated that the weights assigned to comorbid ailments identified within their own data set were different from the CCI. He concluded from this that the impact of comorbidities on survival could be disease specific. And finally, Ghali and coworkers18 found in a study of coronary artery bypass patients in Massachusetts that using weights derived from the study with CCI-defined comorbid conditions substantially improved the ability of the model to predict mortality compared with the CCI.
These 3 studies suggest that disease-specific measures may perform better than a general comorbidity measure. However, there are at least 2 methodological problems with these studies that undermine the conclusion that disease-specific instruments will necessarily do better. First, all 3 studies were based on secondary data analyses of administrative and financial databases. Charlson used primary data to develop and validate the CCI, and it is possible that the CCI may not work as well with secondary data as with primary data. Second, the calibration and testing of the disease-specific comorbidity indexes was done on the same data set. This resulted in a higher χ2 estimate and c-statistic than could possibly be obtained from iterations in which the instruments were developed and tested in separate populations. For this same reason, the WUHNCI must be tested prospectively in a separate cohort of patients with head and neck cancer, ideally not from Barnes-Jewish Hospital.11 Only through testing in a separate population can the validity of the WUHNCI be scientifically assessed.
It is interesting to note that alcohol-related conditions were not identified as a significant comorbid factor in the present study. Clearly, alcohol use and abuse are important features in the development and prognosis of head and neck cancer. Deleyiannis et al19 noted that alcoholism and a history of alcohol-related systemic health problems were associated with an increased risk of death among a cohort of 649 patients with head and neck cancer. In that study, a detailed alcohol history was obtained by trained clinical research interviewers as part of another National Cancer Institute–sponsored research project. The association between alcohol use and mortality was independent of age, site of cancer, anatomic stage, histopathologic grade, smoking, and type of antineoplastic treatment.
There are several possible reasons why alcohol-related conditions were not found to be important in this project. First, the amount of alcohol consumed and the implications of this consumption were all derived from a retrospective review of medical records, not from in-person interviews. It is likely that there was a fair degree of underreporting and misclassification of alcohol use and abuse in the medical records. The classification of alcohol abuse for this project required a fairly high level of alcohol use and abuse (a patient experienced 1 or more episodes of delirium tremens or seizure, recurrent episodes or hospitalizations for alcohol-associated ailments, or nutritionally caused cachexia or anemia). Despite the failure to include alcohol-related conditions in the final comorbidity model, we recognize alcohol use and abuse as important factors for many patients with head and neck cancer.
The WUHNCI can be used for both the retrospective review of the medical records of patients with head and neck cancer or as part of a prospective outcomes research project. In addition, the WUHNCI can be used with administrative data sets, such as hospital discharge records or the Surveillance, Epidemiology, and End Results (SEER)-Medicare linked database. To be used with administrative data sets, the International Classification of Diseases, Ninth Revision codes for each of the 7 comorbid ailments that are part of the WUHNCI must be specified (available from the authors). Overall comorbidity is then determined from an analysis of the primary and secondary diagnoses represented on the hospital discharge "face sheet." Because comorbidity is an important aspect of the patient with cancer and the assessment of the quality of cancer, measures of comorbidity should be included in all ongoing outcomes studies. Since there are general and disease-specific comorbidity indexes available, the continued exclusion of comorbidity adjustment from clinical research can no longer be justified.
Accepted for publication April 22, 2002.
This research was supported in part by the American Cancer Society Junior Clinical Research Award (JCRA-1) and the National Cancer Institute, Bethesda, Md (R01 CA62072).
Corresponding author and reprints: Jay F. Piccirillo, MD, Department of Otolaryngology–Head and Neck Surgery, Washington University School of Medicine, Box 8115, 660 S Euclid Ave, St Louis, MO 63110 (e-mail: firstname.lastname@example.org).
Create a personal account or sign in to: