Version of Record online: 4 DEC 2017
2017) Post-diagnosis weight loss as a prognostic factor in non-small cell lung cancer. Journal of Cachexia, Sarcopenia and Muscle, doi: 10.1002/jcsm.12253., , and (
Cachexia and its most visible manifestation, weight loss, represent important poor prognostic factors for patients with non-small cell lung cancer. This work examines how severity of weight loss as an indicator of cachexia affects outcomes.
In a retrospective observational study of electronic medical records, patients with non-small cell lung cancer were monitored for weight loss from an initial assessment (within 2 months of index diagnosis) to a landmark at 5 months (at least 3 months after initial assessment). Patients who survived to the landmark were then followed to determine the association of baseline body mass index (BMI) and weight loss during the assessment period with outcomes. Patients were clustered to determine how BMI and weight loss related to survival as approximated by time of last appearance in the database, a strong proxy for time of death.
Twelve thousand one hundred and one patients were divided into 5 cachexia risk groups based on a combination of weight loss and initial BMI. More severe groups demonstrated progressively worse outcomes, with the most severe group surviving for a median of 263 days (95% CI 254–274) from index and having a 1-year survival rate of 31%. The least severe group survived for a median of 825 days from index (95% CI 768–908) and had a 1-year survival rate of 78%. Cachexia risk group was a stronger predictor of survival than any baseline variable, including disease stage, performance status, or age.
In this study, we showed that increasing weight loss and, to a lesser extent, decreasing BMI, led to substantially worse outcomes for non-small cell lung cancer patients independent of other variables. We suggest risk score groups that provide an improved approach for identifying poor prognosis patients with the greatest need.
Cancer-associated muscle wasting is a common and debilitating symptom of late-stage cancer. It is commonly associated with cancer cachexia, a clinical syndrome that includes loss of appetite, unintended weight loss, and fatigue but is distinguished from a caloric deficit by the inability to reverse its consequences with nutritional support alone and the frequent presence of concomitant abnormalities in metabolism and inflammation. Cancer-associated muscle wasting is seen in most late-stage cancer populations but is particularly common in lung and gastrointestinal cancers and may be the presenting symptom at diagnosis. Patients with substantial cachexia have worse prognoses independent of other factors, and patients with low body mass index (BMI) have worse outcomes independent of weight loss.
Historically, involuntary weight loss was the most visible manifestation of cachexia and was typically used as a key qualitative factor to diagnose its presence. However, patients can have significantly different levels of weight loss, and varying standards have been proposed for defining a cachectic population based on weight loss and potentially additional factors such as food intake, inflammation, BMI, and muscle measures.[3, 5] In 2011, an international panel of experts established a consensus definition of cachexia to help standardize clinical trials and clinical management of cachexia, identifying patients with cancer cachexia as those meeting at least one of the following three definitions:
While this consensus has helped establish a starting point for discussing cancer cachexia and its consequences, it does not distinguish among patients with differing severity of cachexia or identify those with particularly poor prognoses. Martin et al. explored whether patients could be subdivided into groups based on BMI and reported weight loss at diagnosis to determine whether patients with greater weight loss had reduced survival; they identified five groups with differing expected survival that could be identified based on combinations of BMI and weight loss.
In this study, we present a complementary approach to that used by Martin et al. While they examined the consequences of cachexia in clinical datasets that included patients with heterogeneous tumours and self-reported weight loss, we identified a cohort of patients with non-small cell lung cancer (NSCLC) in an electronic medical records (EMR) database who had repeated measures of weight. For these patients, we could estimate the consequences of weight loss and BMI post-diagnosis on persistence in the database, which we have validated as a strong proxy for duration of survival (last interaction is within 60 days of death date at least 91% of the time; manuscript in preparation).
Electronic medical records data were obtained from the IMS Health™ Oncology Database, which is an integrated database consisting of oncology EMR and additional data (medical/pharmacy claims and hospital charge data master records) for a subset of patients. The database contains de-identified biomedical data from more than 600 000 US cancer patients who received care from approximately 550 providers/facilities in 37 states. For this study, data were used from 2005 to December 2012.
The initial study cohort comprised patients aged ≥18 years with a diagnosis of lung cancer [International Classification of Diseases, 9th Revision, Clinical Modification (ICD-9-CM) 162.1-162.9] as their only primary cancer diagnosis in a derived tumour type field. Those receiving treatment with a drug most commonly used for small-cell lung cancer (etoposide, irinotecan, or topotecan) were excluded in order to get a cleaner population that was primarily NSCLC, as were those with hospital records indicating surgical treatment that might influence weight (Current Procedural Terminology (CPT) codes 32035-32036, 32100-32160, 32200-32201, 32220-32225, 32310-32320, 32440-32540, 32651-32699, or any ICD-9 Vol. 3 Code 32.xx) for patients with linkage to hospital data.
The initial diagnosis of lung cancer was a patient's index date. Index dates were no later than 18 months before the end of the database. To be able to measure weight changes independently of survival, patients were required to have an initial weight measure within 2 months of index and at least one additional measure in the subsequent 3 months to show potential weight change; all patients were required to survive for at least 5 months after index to avoid immortal time bias. Patients with a single health system contact in their last 3 months in the database were excluded from the primary analysis because the survival proxy was deemed less reliable for these patients.
Patients' baseline weight was their first weight measure after their index date. Patients' maximum change from baseline was the difference between baseline and the lowest subsequent measure within 5 months of index. Patients with extreme baseline weights (below 100 lbs or above 250 lbs) were excluded as likely to display atypical weight change patterns. Patients with likely erroneous data (such as successive gains and losses of 20% of weight over a short period) were eliminated if not resolvable based on the overall record (<0.5% of patients). Patients in the database had on average 11 measures of weight.
Records indicating height <48 inches or ˃84 inches were excluded from analysis. Patients' height was determined based on the average of heights within 2 inches of the modal height (median if no mode existed), as long as no more than 25% of measures were outside of that window.
Time of last appearance in the database was used as a proxy for time of death. For patients with last appearances close to the end of available data, typical interaction frequency was calculated for the 3 months before last appearance, and patients with less than twice their typical interaction frequency before the end of available data were censored.
A combination of analytical methods and judgement were used to determine weight loss and BMI categories. Initially, the population was divided along each dimension into 13 categories that balanced spreading the overall population, traditional breakpoints, and apparent inflection points in each curve. The 13 categories were clustered into six provisional groups for each dimension. All possible cluster combinations were tested in a Cox model with survival as the dependent variable and categorized BMI, categorized weight loss, and their interactions as dependent variables. The top 1000 combinations based on score and Akaike information criterion were used to determine most common strong breakpoints. The provisional groups were tested for potential improvements, such as by holding one dimension fixed and testing finer gradations along the other dimension. Final group determination included judgement regarding materiality of differences between closely related models and preferences for convenient breakpoints.
After determining estimates for median overall survival and 1-year survival for each weight loss/BMI combination, the 36 combinations were ordered by survival, and seven groups were created at natural breakpoints. When assembled into a grid, two of these groups that had scattered members were merged into their neighbours to create contiguous groups. Ambiguous choices were resolved by comparing alternatives using a log-rank test to evaluate similarity within and between cachexia risk groups.
Baseline characteristics were described by groups. Median survival time and 1-year survival rate with 95% confidence interval (CI) were estimated by the Kaplan–Meier method. Multivariate Cox models were built to evaluate the effect of cachexia risk groups, and generalized likelihood-based R2s were reported. Where data were missing, an ‘Unknown’ categorical variable was used. The impact of explanatory variables was assessed by comparing likelihoods of models with all or a subset of variables based on Type III results from sas proc phreg. Results were considered significant at the P < 0.05 level. Analysis was done in sas (version 9.2, sas Institute Inc., Cary, NC, USA).
This study focused on patients in the IMS Health EMR with NSCLC who had a baseline weight measure in the first 2 months, additional weight measures in the subsequent 3 months, and survived for at least 5 months. While requiring 5 months of survival reduces generalizability, it permits multiple observations of weight while avoiding bias wherein longer survivors have more opportunity to lose weight. Table 1 shows cohort attrition for this study. 15 369 patients met the primary requirements to be included in this study; 14 979 of these also had valid BMI measures. The cohort was further limited to the 12 101 patients who had more than one database contact in their final 3 months, which was important for accuracy of the death date proxy.Cohort attrition
|Patients excluded||% Excluded||Patients remaining|
|Lung cancer as sole primary cancer during study period||—||—||44 482|
|No use of drugs primarily associated with small-cell lung cancer||7107||16.0||37 375|
|No major surgery codes||1635||4.4||35 740|
|Survived for at least 5 months post-index||13 831||38.7||21 909|
|Patients at least 18 years old||2||0.0||21 907|
|Patients with weight measures||4965||22.7||16 942|
|Patients without uninterpretable weights||73||0.4||16 869|
|Patients with valid baseline weight (100–250 lb)||1497||8.9||15 372|
|Patients without other data abnormalities||3||0.0||15 369|
|Patients with valid height||390||2.5||14 979|
|Patients with >1 database interactions in last 3 months||2878||19.2||12 101|
Table 2 shows patient characteristics for the study cohort. Most patients in this study with known stage had advanced or metastatic cancer at baseline.Cohort demographics
To understand what levels of weight loss led to meaningful differences in outcomes for patients, we clustered patients by baseline BMI and weight loss. Initially, we examined the two variables independently. When patients were divided into 20 successive groups of 606 patients for each variable, there was a nearly continuous increase in survival with each variable, with weight loss having a steeper slope. Median survivals ranged from a low of 387 days to a high of 559 days for BMI groups and from a low of 263 days to a high of 726 days for weight loss groups.
Patients were assigned to six continuous groups for each variable, with breakpoints chosen to provide the best joint fit for survival data. The 36 groups created by combining these two parameters were then merged into five groups that represented different levels of cachexia risk based on outcomes (see methods). Tables 3A and 3B show survival for the 36 individual groups, while Table 3C summarizes values for the five cachexia risk groups. Differences within groups were marginally significant or non-significant, while the difference among groups was highly significant (P < 0.0001). As was suggested by the cachexia consensus definition, baseline BMI is most significant when it is low (<20). Weight loss impacts survival more broadly, and patients with any weight loss have worse prognosis than patients with none.
Table 4 shows patient characteristics for the five groups. Relative to the overall cohort, the higher risk cohorts showed a higher percentage of patients with late-stage disease. Surprisingly, they also tended to be a bit younger, perhaps indicating a tendency towards more aggressive disease in younger patients. While lower baseline BMI was associated with worse outcomes by itself, the stronger influence of weight loss on survival led to clustering of groups with diverse baseline BMI that obscured this effect.Cohort demographics
N = 12 101
N = 918
N = 2209
N = 4125
N = 3711
N = 1138
Figure 1 shows Kaplan–Meier survival curves for the cachexia risk groups; the five groups show consistent well-separated risk profiles across time, indicating that the cachexia risk groups are important prognostic factors. We next built Cox models for survival using baseline variables and/or cachexia risk groups as explanatory variables. As shown in Table 5, a model including all variables shows that the cachexia risk groups are highly significant variables. The overall model R2 is 12.88%, while reduced models with only the cachexia risk groups or baseline variables have R2 of 7.51% and 6.82% respectively, showing that the cachexia groups appear to have comparable and relatively independent explanatory power to the combination of age, disease stage, gender, performance status, and index year in this dataset. By comparison, a similar model using baseline covariates and the weight loss criteria from Fearon et al. (5% weight loss or 2% for patients with a BMI less than 20)  would give an overall model R2 of 10.34%, 4.65% for the cachexia definition alone.
Kaplan–Meier survival curves for the cachexia risk groups.Impact of cachexia risk and baseline covariates on survival
|Parameter||Parameter estimate||Standard error||χ2||Pr > χ2||Hazard ratio|
|Cachexia risk group|
|ECOG performance Status|
In this retrospective analysis of 12 101 NSCLC patients from an EMR database, patients with greater weight loss during the 5 months post-baseline were found to have worse outcomes based on a strong proxy, time of last appearance of patient in the database. Lower BMI at baseline was found to influence survival, but to a lesser extent than weight loss. While the current definition of cachexia focuses on patients with at least 5% weight loss (or 2% with low BMI), patients with any weight loss were found to have decreased survival in this study, and patients with low BMI had worse outcomes than patients with higher BMI even without weight loss. Weight loss before index visit may have occurred for some of these patients.
This study demonstrates that patient risk from cachexia-associated issues is a continuum and that progressively higher levels of weight loss are associated with worsening prognosis. These results were statistically significant and substantial in magnitude, with median survival for patient cohorts ranging from 263 days to 825 days from index, with all patients required to survive 150 days to be included in this study.
These results were generally consistent with those reported by Martin et al. but suggested a much stronger impact of weight loss than baseline BMI. This may be a consequence of using actual weight measures instead of self-reported historical weight loss, reducing variability in weight-loss measures. In addition, historical weight loss is correlated with post-weight-loss BMI, obscuring the relative impact of these two variables.
This study has limitations. First, data were derived from an EMR database and were often incomplete. Second, the patient cohort may have been biased by the selection criteria, including proxy measures that may not have completely removed small-cell lung cancer patients, may have selectively removed some types of patients (high weight and radiation treatment with etoposide) and the requirement that all patients survive for at least 5 months from index. Third, the requirement for patients to have more than one visit in the last 3 months in the database may limit generalizability. Fourth, patient weight loss before index was unknown, as were any efforts to provide supplemental nutrition to mitigate weight loss. Finally, time of death was estimated based on a proxy, time of last appearance in the database.
Patients with NSCLC were divided into five groups based on initial BMI and early weight loss, with boundaries based on clustering groups of patients with similar outcomes. Increasing weight loss and, to a lesser extent decreasing BMI, is significantly associated with worse outcomes independent of other variables. These factors provide an improved approach for characterizing cachexia risk over the initial presence/absence of cachexia described by Fearon et al.
The authors thank Jonathan Swain and the IMS staff for assistance with this study. Support for this article was provided by Eli Lilly and Company.
The authors certify that they comply with the ethical guidelines for authorship and publishing of the Journal of Cachexia, Sarcopenia and Muscle.
Karin Benoit, Li Li, and Daniel Mytelka are employees and minor shareholders of Eli Lilly and Company.