Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2016 Jun 1.
Published in final edited form as: J Cancer Surviv. 2014 Sep 4;9(2):161–171. doi: 10.1007/s11764-014-0393-z

An Algorithm to Identify the Development of Lymphedema After Breast Cancer Treatment

Tina WF Yen *, Purushuttom W Laud †, Rodney A Sparapani, Jianing Li †, Ann B Nattinger ‡
PMCID: PMC4362809  NIHMSID: NIHMS625633  PMID: 25187004

Abstract

Purpose

Large, population-based studies are needed to better understand lymphedema, a major source of morbidity among breast cancer survivors. One challenge is identifying lymphedema in a consistent fashion. We sought to develop and validate an algorithm using Medicare claims to identify lymphedema after breast cancer surgery.

Methods

From a population-based cohort of 2,597 elderly (65+) women who underwent incident breast cancer surgery in 2003 and completed annual telephone surveys through 2008, two algorithms were developed using Medicare claims from half of the cohort and validated in the remaining half. A lymphedema-positive case was defined by patient report.

Results

A simple two ICD-9 code algorithm had 69% sensitivity, 96% specificity, positive predictive value >75% if prevalence of lymphedema is >16%, negative predictive value >90%, and area under receiver operating characteristic curve (AUC) of 0.82 (95% CI: 0.80 – 0.85). A more sophisticated, multi-step algorithm utilizing diagnostic and treatment codes, logistic regression methods, and a reclassification step performed similarly to the two-code algorithm.

Conclusions

Given the similar performance of the two validated algorithms, the ease of implementing the simple algorithm and the fact that the simple algorithm does not include treatment codes, we recommend that this two-code algorithm be validated in and applied to other population-based breast cancer cohorts.

Implications for Cancer Survivors

This validated lymphedema algorithm will facilitate the conduct of large, population-based studies in key areas (incidence rates, risk factors, prevention measures, treatment and cost/economic analyses) that are critical to advancing our understanding and management of this challenging and debilitating chronic disease.

Keywords: lymphedema, algorithm, breast cancer, Medicare, survivorship

Introduction

Conducting high quality studies on large administrative datasets that address the quality of medical care often require applying validated claims-based algorithms to identify disease processes, treatment patterns, treatment complications and outcomes. While developing such algorithms is not straightforward and validating algorithms can be equally challenging, efforts to develop and validate claims-based algorithms have been successful [1–5]. With respects to cancer care, few validated claims-based algorithms to identify incident cancers exist in the current literature, including three (one from our group) which have been published to identify incident breast cancer [6–10]. Validated claims-based algorithms exist that identify recurrent disease and second primary cancers [11–15]. However, only a few validated algorithms have been published that address complications of cancer treatment [16, 17].

Despite advances in our understanding of breast cancer and its treatments, lymphedema (arm/hand swelling) remains one of the most feared complications of breast cancer treatment, is a major source of treatment-related morbidity, and is one of the most difficult diseases to study. This chronic condition can result in significant physical discomfort and disability, cosmetic deformity and psychological distress that all adversely affect activities of daily living and quality of life and is associated with significant medical costs [18–23]. Women are at risk of developing lymphedema over their lifetime and the time frame from treatment to development of lymphedema is highly variable [24].

The incidence of lymphedema in individual studies ranges from 0 to 94% [25–27]. This wide variation is due to the fact that the majority of studies are relatively small, single-institutional, retrospective reports with marked study heterogeneity with respect to differences in study design, patient populations and breast cancer treatments, diagnostic methods and criteria used to define lymphedema (subjective versus objective methods) [28], and timing and duration of follow-up assessments. Sentinel lymph node biopsy (SLNB), which involves the removal of only a few lymph nodes in eligible women, is associated with lower but still substantial rates of lymphedema than a full axillary node dissection (ALND); a recent systemic review and meta-analysis of studies from 2000–2012 reports a pooled estimated lymphedema incidence rate of 17% – 21% [24].

Given the general limitations of prior studies, larger population-based studies of lymphedema are needed to increase and expand our understanding of this complex and disabling chronic disease. However, a major challenge to population-based studies is identifying patients with lymphedema in a consistent fashion. To our knowledge, three prior studies have utilized a claims-based approach to defining lymphedema [29, 23, 30] [Table 1]. These three studies each used diagnostic codes with or without different treatment codes to define lymphedema but none have validated their approach.

Table 1.

Claims-based definitions of lymphedema in the literature

Study Year Data source Codes used to identify lymphedema claims
Shih YC. 2009 [23] MedStat MarketScan Health and Productivity Management database ICD-9 457.0 and 457.1
Kwan ML. 2010 [30] Kaiser Permanente Northern California electronic medical records ICD-9 457.0 and 457.1
CPT 97140 and 97139
Durable medical equipment (DME) orders associated with breast cancer-related lymphedema
Reiner AS. 2011[29] Surveillance, Epidemiology, and End Results (SEER)-Medicare linked database ICD-9 457.0 and 457.1
Any claims for treatment prescribed for secondary lymphedema (36 CPT and HCPCS codes; see Table 1 of study)
Open in a new tab

As the number of cancer survivors continues to grow, the importance of increasing our awareness and understanding of the many short- and long-term consequences of cancer and its treatment has become increasingly emphasized [31]. We therefore sought to develop and validate an algorithm utilizing Medicare claims to identify women who develop lymphedema after breast cancer treatment in a population-based cohort of older breast cancer survivors.

Methods

Data sources

The study sample was derived from a population-based cohort of elderly breast cancer survivors who participated in a National Cancer Institute-sponsored survey study examining breast cancer care outcomes, which has been previously described [32, 33]. In brief, women residing in four geographically diverse states (California, Florida, Illinois, and New York) were identified from Medicare claims as having had an incident breast cancer surgery between the ages of 65 and 89 years in 2003, using our validated claims-based algorithm [6]. As shown in Figure 1, potential participants (n=8,742) were contacted by mail in September 2005; 5,659 were excluded either due to study ineligibility (n=2,995) or non-participation (n=2,664). Baseline (wave 1) and three subsequent structured telephone interviews were conducted by an experienced survey center. Survey waves 2, 3 and 4 were completed at a median duration of 40 (interquartile range: 6 [36–42 months]), 49 (interquartile range: 5 [46–51 months]) and 60 months (interquartile range: 4 [58–62 months]) after surgery, respectively. A total of 3,083 women confirmed their diagnosis of an incident breast cancer and completed the initial survey, with a participation rate of 70% among those eligible for the study; participation rates remained over 90% for the three follow-up surveys [32]. A multiple regression model predicting survey participation was performed comparing the 3,083 survey participants to the 2,664 non-participants [32]. Older subjects (75+ years) and those residing in NY were less likely to participate but participation did not differ based on socioeconomic status, race/ethnicity, comorbidity or type of breast surgery.

Figure 1.

Figure 1

Open in a new tab

Study diagram. Between 2005 through 2006, a cohort of 8,742 women aged 65 years or older were initially identified as potentially eligible, based on Medicare claims, for a breast cancer study. A total of 5,659 were excluded, either due to ineligibility (n=2,995) or non-participation (n=2,664). A total of 3,083 women completed the baseline/initial survey wave and participation rates remained over 90% for the three follow-up survey waves. The cohort for this study comprises the 2,597 women who completed at least Wave 2 of the survey and had available claims information, generating a total of 6,754 observations for waves 2–4. These observations were randomly sampled and half were placed in the training cohort (3,356 observations; 2,077 patients) and half in the validation cohort (3,398 observations; 2,119 patients).

Tumor characteristics and stage information was provided by the four state cancer registries [34]. Medicare claims information was collected from inpatient, outpatient and carrier Standard Analytical Files through 2008. Three coding manuals (ICD-9 CM International Classification of Diseases, Current Procedural Terminology, and HCPCS Health Care Financing Administration Common Procedure Coding System) were searched to identify diagnostic, procedural, treatment and durable medical equipment codes that could potentially indicate lymphedema [35–37].

Lymphedema definition, case construction, creation of training and validation cohorts

Self-report of lymphedema was ascertained at approximately 3, 4 and 5 years of follow-up (at survey waves 2, 3 and 4). Of the initial 3,083 participants who completed wave 1, 5%–10% of women were either no longer eligible to participate or lost to follow-up due to inability to contact or refusal at each subsequent wave: 295 at wave 2, 136 at wave 3, and 180 at wave 4 [32].

For participants who responded to more than one wave of the survey, we had the opportunity to test the validity of the algorithm with respect to lymphedema onset over multiple time intervals. Taking as the unit of analysis a time interval between surgery and a survey wave, each woman could have between 1 and 3 observations included in the analysis. Of the 2,597 women who had complete survey and claims information through at least survey wave 2, 1,981 (76%) completed waves 3 and 4 and had claims information through wave 4. These 2,597 women who had complete survey and claims information through at least survey wave 2 generated a total of 6,754 observations from waves 2–4 (Figure 1). These observations were randomly sampled and half were placed in the training cohort and half in the validation cohort. Therefore, a patient with more than one observation could have observations in one or both cohorts. There were a total of 3,356 observations (2,077 patients) included in the training cohort and 3,398 observations (2,119 patients) in the validation cohort.

For each of these observations, a case determination was made. A lymphedema-positive case was defined as a “yes” response to the lymphedema survey question: “Since undergoing breast cancer surgery (or last survey), has a doctor ever told you that you have lymphedema or arm edema?” This question was adapted from previously published work [38]. Once a woman responded “yes”, she was not re-asked the lymphedema question at subsequent waves and was deemed a lymphedema-positive case for all subsequent observations. If an observation was not deemed a lymphedema case (i.e., the woman responded “no” to the question), she was asked the lymphedema question at the next survey wave. Case determinations were made for each observation period and included in the algorithm construction training cohort or validation cohort if claims information was available for the observation period.

Lymphedema algorithm performance measures

Algorithm performance was assessed by determining sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and area under the receiver operating characteristic curve (AUC). Self-reported lymphedema was the gold standard. Sensitivity (true positive rate) measures the proportion of self-reported lymphedema cases that were correctly identified to be lymphedema-positive by the algorithm. Specificity (true negative rate) measures the proportion of self-reported non-lymphedema cases that were correctly identified to be lymphedema-negative by the algorithm. The ideal algorithm would have 100% sensitivity (predict all cases from the self-reported lymphedema group as having lymphedema) and 100% specificity (not predict any cases from the non-lymphedema group as having lymphedema).

The PPV and NPV represent the probability that the algorithm gives the correct result. The PPV measures the proportion of algorithm-positive lymphedema cases that were self-reported lymphedema cases; the NPV measures the proportion of algorithm-negative cases that were self-reported non-lymphedema cases. The PPV is dependent on the prevalence of disease [39]. If the prevalence of lymphedema is very low, the PPV will not approach 100% even if both the sensitivity and specificity are high. As lymphedema prevalence rates increase, the PPV increases and the NPV decreases.

The accuracy of the algorithm was measured by the AUC, which is determined by plotting the true positive rate (sensitivity) along the y-axis by the false positive rate (1-specificity) along the x-axis. An AUC of 1.0 represents a perfect test while an AUC of 0.5 represents a worthless test.

Simple lymphedema algorithm

Since the two lymphedema diagnosis codes (ICD-9 457.0, 457.1) were consistently used in all three published studies that have utilized a claims-based approach to defining lymphedema [23, 29, 30] [Table 1], we first developed a simple 2-code algorithm, which required only identifying either ICD-9 457.0 or 457.1 codes at any time point after surgery. Having an algorithm based on diagnosis codes only and excluding lymphedema treatment codes enhances the ability to use this algorithm to study lymphedema treatment outcomes. This simple 2-diagnosis code algorithm in the training cohort yielded 71% sensitivity and 95% specificity (Table 3).

Table 3.

Comparison of sensitivity, specificity and AUC of two lymphedema algorithms in training and validation cohorts

2-code algorithm (ICD 457.0 or 457.1) Multi-step algorithm
Sensitivity (95% CI) Specificity (95% CI) AUC (95% CI) Sensitivity (95% CI) Specificity (95% CI) AUC (95% CI)
Training cohort 71.2% (65.5% – 76.9%) 95.3% (94.6% – 96.1%) 0.83 (0.80 – 0.86) 71.2% (65.5% – 76.9%) 95.8% (95.0% – 96.5%) 0.83 (0.81 – 0.86)
Validation cohort 69.4% (63.9% – 75.0%) 95.7% (95.0% – 96.4%) 0.82 (0.80 – 0.85) 69.4% 63.9% – 75.0%) 96.0% (95.3% – 96.7%) 0.83 (0.80 – 0.86)
Open in a new tab

Abbreviations: CI, confidence interval; AUC area under receiver operating characteristic curve

Multi-step lymphedema algorithm

Next, we sought to determine whether a more sophisticated algorithm could be developed with improved performance compared to the simple two-code algorithm. For this algorithm development process, we included the evaluation of treatment codes to determine if they would enhance algorithm performance, particularly sensitivity.

Library of potential lymphedema codes

We began by conducting an exhaustive search for diagnostic, procedural, treatment and durable medical equipment codes that could potentially indicate lymphedema (Table 2). This search, led by a practicing breast surgeon and researcher (TY), was conducted by all research team members. An initial screen of the number of claims for each code that was present per patient since surgery was performed in the training cohort. The prevalence of these codes was examined separately in women reporting lymphedema and those that did not for differences which would aid in identifying codes potentially reflective of lymphedema. All HCPCS codes and three ICD-9 codes (459.1, 782.8, 782.9) were either never or very rarely identified and were therefore not further assessed. All other codes were further explored.

Table 2.

Diagnostic, procedural and durable medical equipment codes potentially indicating lymphedema in the training cohort

Code Outcome variable definition Medicare coveragea Prev (%) SN (%) SP (%)
Diagnostic Codes (ICD-9) - 2008
457.0 Postmastectomy lymphedema syndrome 3.8 34 98
457.1 Other lymphedema, including acquired (chronic) and secondary 7.9 60 96
457.2 Lymphangitis (chronic, subacute, NOS) 0.2 <1 99
459.1 Post-phlebetic syndrome (chronic venous HTN due to DVT) 0.2 0 99
459.2 Compression of vein, includes vena cava syndrome 0.8 3 99
459.3 Chronic venous HTN (idiopathic) 0.1 0 99
611.0 Inflammatory disease of the breast, includes mastitis 7.3 14 93
682.3 Diffuse cellulitis, acute abscess, acute lymphangitis (upper arm and forearm, axilla, shoulder) 3.2 11 97
682.4 Diffuse cellulitis, acute abscess, acute lymphangitis (hand, wrist) 1.3 2 99
729.81 Swelling of limb (arm, hand) 15.7 35 86
782.3 Edema (anasarca, dropsy, localized edema NOS) 21.9 26 78
782.8 Changes in skin texture (induration/thickening) 0.6 1.2 99
782.9 Other symptoms involving skin and integumentary tissues 0.6 0.8 99
Procedural, Treatment and Durable Medical Equipment Codes
CPT Physical Medicine and Rehabilitation Codes - 2008
97001 Physical therapy evaluation 42.0 66 60
97002 Physical therapy re-evaluation 7.4 15 93
97003 Occupational therapy evaluation 9.1 34 93
97004 Occupational therapy re-evaluation 0.7 4 99
97016 Vasopneumatic devices 1.0 6 99
97124 Massage, including effleurage, petrissage and/or tapotement 5.7 10 95
97140 Manual therapy techniques, including manual lymphatic drainage 30.6 75 73
HCPCS 2008
A6448 Light compression bandage, elastic (width < 3 inches); start 1/1/04 C NF
A6449 Light compression bandage, elastic (width < 3 and >= 5 inches); start 1/1/04 C <0.1 0 100
A6450 Light compression bandage, elastic (width >=5 inches); start 1/1/04 C NF
A6451 Moderate compression bandage, elastic; start 1/1/04 C NF
A6452 High compression bandage, elastic; start 1/1/04 C 0 0 100
A6542 Gradient compression stocking, custom made; start1/1/06 (DELETED January 1, 2010) M NF
A6543 Gradient compression stocking, lymphedema; start1/1/06 (DELETED January 1, 2010) M NF
A6549 Gradient compression stocking, not otherwise specified; start 1/1/06 (Code changed 1/1/2010) M NF
E0651 Pneumatic compressor; segmental home model without calibrated gradient pressure; start 1/1/88 D NF
E0652 Pneumatic compressor; segmental home model with calibrated gradient pressure; start 1/1/88 D NF
E0655 Pneumatic appliance; half arm; start 1/1/86 D NF
E0665 Pneumatic appliance; full arm; start 1/1/86 D NF
E0668 Segmental pneumatic appliance for use with pneumatic compressor; full arm; start 1/1/88 D NF
E0672 Pressure pneumatic appliance; full arm; start 1/1/95 D NF
E0676 Intermittent limb compression device (includes all accessories), NOS; start 1/1/07 D NF
L8010 Mastectomy sleeve; start 1/1/86 D 0.1 0.8 100
L8210 Gradient compression stocking, custom made; start 1/1/86 (DELETED January 1, 2006) M NF
L8220 Gradient compression stocking, lymphedema; start 1/1/86 (DELETED January 1, 2006) M 0 0 100
L8239 Gradient compression stocking, NOS; start 1/1/98 (DELETED January 1, 2006) C NF
S8420 Gradient pressure aid (sleeve and glove combination), custom made; start 1/1/02 I NF
S8421 Gradient pressure aid (sleeve and glove combination), ready made; start 1/1/02 I NF
S8422 Gradient pressure sleeve, custom, medium; start 1/1/02 I NF
S8423 Gradient pressure sleeve, custom, heavy; start 1/1/02 I NF
S8424 Gradient pressure aid (sleeve), ready made; start 1/1/02 I NF
S8427 Gradient pressure aid (glove), ready made; start 1/1/02 I NF
S8428 Gradient pressure aid (gauntlet), ready made; start 1/1/02 I NF
S8431 Compression bandage, roll; start 1/1/02 I NF
S8950 Complex lymphedema therapy; start 1/1/00 I NF
Open in a new tab

Abbreviations: NF, code not found; NOS, not otherwise specified; Prev, prevalence; SN, sensitivity; SP, specificity HCPCS codes list start and, if applicable, stop dates.

a

Medicare coverage status for HCPCS codes[34, 35][34, 35][34, 35][34, 35]34,35: C, carrier judgment; D, special coverage instructions apply; I, not payable by Medicare (no grace period); M, non-covered by Medicare

The four most clinically likely diagnostic codes for lymphedema are in bold (ICD-9 457.0. 457.1, 729.81 and 782.3). The presence of any of these four codes since surgery had a sensitivity of 80% with a specificity of 68%. Individually, the most sensitive code was 457.1 (other lymphedema, including acquired [chronic] and secondary) with a sensitivity of 60% and specificity of 96% (Table 2). The prevalence of either 457.0 (postmastectomy lymphedema syndrome) or 457.1, the two lymphedema-specific diagnosis codes, was 9.5%; the sensitivity and specificity were 71% and 95%, respectively.

The three codes in italics (ICD-9 457.2, 682.3, 682.4) were deemed “sequelae” codes as these diagnoses (lymphangitis, diffuse cellulitis, acute abscess, acute lymphangitis) are more likely to occur in patients with lymphedema than those without lymphedema. These codes would have been considered in the algorithm to rule out lymphedema cases; however, given the very low prevalence and sensitivity of these codes (presence of any one of these 3 codes had 12% sensitivity and 97% specificity), these codes were not considered good candidates and were not further explored.

The three underlined codes (ICD-9 459.2, 459.3, 611.0) were deemed “non-specific” codes as lymphedema could result as a sequelae of these diagnoses (vein compression, idiopathic chronic venous hypertension, inflammatory disease of the breast), although unlikely. These codes would have been considered in the algorithm if they enhanced the algorithm performance; however, due to the very low prevalence and sensitivity of these codes (presence of any one of these 3 codes had 17% sensitivity and 92% specificity), they were also not considered good candidates and were not further explored.

All 7 procedural/treatment (CPT) codes related to lymphedema were further evaluated. As shown in Table 2, the most prevalent codes were for physical therapy (CPT 97001, which are relatively non-specific and could be unrelated to lymphedema) and manual therapy techniques (CPT 97140), which are more specific for lymphedema. Of these 7 codes, sensitivity (75%) and specificity (73%) were highest for the manual therapy techniques code (CPT 97140), which includes manual lymphatic drainage therapy, a treatment that is specific for lymphedema.

Medicare coverage for the HCPCS codes that were evaluated is summarized in Table 2. Since all A, E and L-codes (compression bandages, compression stockings, pneumatic devices) have either C (carrier judgment), D (special coverage instructions apply) or M (non-covered) Medicare coverage status [40, 41] is not surprising that these codes were either not found or rarely identified due to variable or no Medicare reimbursement. As expected, no single claim with any S code, which includes all gradient pressure aids as well as S8950 (complex lymphedema therapy, the most specific therapy for lymphedema), was found as S codes are all I (not payable by Medicare) coverage status codes [40].

Development of 3-step lymphedema algorithm

Lymphedema algorithm construction subsequently focused on the 4 ICD codes (457.0, 457.1, 729.81 and 782.3) and the 7 CPT treatment codes and discovering what combinations of available billing codes were indicative of lymphedema. This process involved a careful and deliberate model search on the test data that was not automated and was refined based on clinical insight. An extensive number of logistic regression models were evaluated to predict self-reported lymphedema using variables constructed from these 11 codes. These constructed variables included: 1) presence/absence of code; 2) if present, the number of times the code was identified at various time intervals (15, 30, 90, 120 and 180 days) after each code was identified; 3) for the 4 diagnosis codes, whether a CPT code was identified at the same time as the ICD-9 code; 4) the presence of any of the 10 discussed ICD-9 codes or seven procedural/treatment codes prior to surgery; 5) the presence of any arm/shoulder/hand ICD-9 codes 710 – 996 prior to surgery; 6) evaluation of any codes present in the false-negative (code negative but survey positive) group that could potentially be used to increase algorithm sensitivity. These analyses revealed that the combination of diagnosis and treatment codes with the highest sensitivity included 6 codes: ICD 457.0, 457.1, 729.81, CPT 97016, 97124 or 97140. Using the presence of any of these 6 codes for screening, the sensitivity was 84% but specificity was only 61% for detecting lymphedema. We therefore sought to develop a multi-step algorithm that would improve specificity without compromising sensitivity significantly.

The first phase of algorithm development involved screening for the presence of any of the above 6 diagnostic and treatment codes within a timeframe of one year prior to surgery to the most recent available claims. A patient with any of the 6 codes present moved on to the second step of the algorithm, which involved applying a logistic regression model, to remove cases of edema which existed prior to incident breast cancer surgery. A model predicting an incident lymphedema case was developed using the presence of these 6 codes (individually and in combinations) as predictor variables. Stepwise regression was used to determine the final logistic model which included the presence of ICD 457.0, 457.1, and/or 729.81 codes since surgery. The third step of the algorithm required reclassifying to algorithm-negative the lymphedema algorithm-positive cases (n=14 for the training cohort) that had ICD-9 codes 782.3 (edema) or 729 (other disorders of soft tissue) one year prior to surgery. In the training cohort, this three-step algorithm yielded an AUC of 0.83, sensitivity of 71%, specificity of 95% and PPV of >75% if the prevalence of lymphedema is 16% or higher (Table 3). The NPV of this algorithm is excellent at >90% for a wide range of lymphedema prevalence rates.

Algorithm Face Validity

To provide face validity of the lymphedema algorithm, risk factors for lymphedema among the 1,148 women in this algorithm cohort who had documented nodal status and wave 3 information were assessed with a multiple logistic regression model and compared with the previously published results in the same breast cancer survey cohort where lymphedema was defined by self-report [33]. The independent variables entered into the model included: hypothesized risk factors (number of lymph nodes removed, surgeon case volume), variables for which univariate associations were significant at the 0.05 level (tumor size, type of breast cancer [DCIS or invasive], lymph node status, type of breast surgery, receipt of chemotherapy), and factors that have been observed as predictors of lymphedema by other studies (patient age, radiation therapy, hormonal therapy).

All data analyses were conducted using SAS statistical software (Version 9.3, SAS Institute; Cary, NC).

Results

In the total cohort of 2,597 patients, the mean age of the women at the time of surgery was 72.0 (SD 5.4) years and 92% were white. Of the women with available pathologic tumor staging information, the majority had early stage disease: 96% had tumors less than 5 cm in size and 78% were lymph node-negative. Two-thirds (66%) underwent breast-conserving surgery; 34% underwent mastectomy. No axillary surgery was performed in 18%, while 26% underwent SLNB only and 56% underwent ALND. Two-thirds of the cohort received radiation therapy, 20% chemotherapy, and 70% hormonal therapy. There were no differences in patient, tumor or treatment characteristics between the training and validation cohort groups. In the training cohort, 7.2% of observations were lymphedema-positive cases. In the validation cohort, 7.8% of observations were lymphedema-positive cases. These proportions were lower than expected, but are likely explained by the strict definition of lymphedema (patient’s recall of being told by her doctor that she had lymphedema since undergoing breast cancer surgery) required for this study.

Algorithm validation

When applied to the training cohort, the simple 2-code algorithm had excellent specificity (95%) and moderate sensitivity (71%). When this simple algorithm was applied to the validation cohort, the results were similar: AUC 0.82 (95% CI: 0.80 – 0.85), sensitivity 69%, specificity 96%, PPV > 75% if prevalence of lymphedema is >16% and NPV >90% (Figure 2, solid lines; Table 3).

Figure 2.

Figure 2

Open in a new tab

Performance of two algorithms. This figure displays the positive and negative predictive value (PPV and NPV) curves in the validation cohort for the multi-step (dashed lines) and simple, 2-code (solid lines) lymphedema algorithms. The x-axis represents the prevalence of lymphedema; the y-axis represents the predictive value. The blue lines represent the PPV curves; the black lines represent the NPV curves.

The multi-step algorithm in the training cohort yielded similar results to the simple algorithm (Table 3). Moreover, this multi-step algorithm in the validation cohort yielded similar sensitivity and specificity results and AUC (0.83; 95% CI: 0.80 – 0.86) compared to the simple algorithm (Table 3). Similar to the simple algorithm, the PPV of this multistep algorithm is >75% if the prevalence rate of lymphedema is at least 16% and the NPV of this simplified algorithm is > 90% for a wide range of lymphedema prevalence rates (Figure 2, dashed lines).

To provide further face validity of the simple lymphedema algorithm, using a multiple logistic regression model, we assessed risk factors for lymphedema (defined by the 2-code algorithm) among the 1,148 women in this algorithm cohort who had documented nodal status and wave 3 information and compared these results to those previously published in the same breast cancer survey cohort where lymphedema was defined by self-report [33]. As shown in Table 4, in the algorithm cohort of 1,148 women, the results of the multiple logistic regression model with the outcome of lymphedema defined by the 2-code claims algorithm are similar to those results with lymphedema defined by self-report, which are similar to previously published results [33].

Table 4.

Comparison of predictors of lymphedema when defined as self-report versus the simple 2-code algorithm

Lymphedema defined by self-reporta Lymphedema defined by two-code algorithma
Variable Category (n) OR 95% CI P Values OR 95% CI P Values
No. lymph nodes examined <0.0001 <0.0001
None (182) 1.00 1.00
1 – 5 (506) 1.51 0.61 – 3.75 1.92 0.68 – 5.48
6 – 10 (201) 3.96 1.54 – 10.20 5.01 1.68 – 14.97
11 – 15 (151) 3.76 1.39 – 10.12 6.74 2.17 – 20.94
≥16 (108) 7.92 2.92 – 21.49 10.87 3.47 – 34.11
Lymph node metastasis 0.0055 0.11
No (946) 1.00 1.00
Yes (202) 2.00 1.23 – 3.27 1.59 0.90 – 2.80
Type of breast cancer 0.26 0.79
Invasive (886) 1.00 1.00
DCIS (182) 0.70 0.30 – 1.59 1.20 0.48 – 2.94
Unknown (80) 0.55 0.25 – 1.21 0.80 0.35 – 1.85
Tumor size (mm) 0.92 0.28
0 – 20 (838) 1.00 1.00
21 – 50 (195) 1.10 0.69 – 1.76 0.91 0.52 – 1.59
> 50 (45) 1.27 0.55 – 2.92 1.16 0.46 – 2.92
Unknown (70) 1.18 0.51 – 2.72 2.17 0.94 – 5.00
Type of breast surgery 0.26 0.44
BCS (752) 1.00 1.00
Mastectomy (396) 1.40 0.78 – 2.53 1.30 0.67 – 2.53
Receipt of radiation therapy 0.70 0.17
No (383) 1.00 1.00
Yes (765) 1.12 0.63 – 1.99 1.58 0.82 – 3.03
Receipt of chemotherapy 0.24 0.86
No (940) 1.00 1.00
Yes (208) 0.73 0.43 – 1.24 1.05 0.58 – 1.91
Receipt of hormonal therapy 0.53 0.42
No (352) 1.00 1.00
Yes (796) 0.88 0.58 – 1.32 0.83 0.52 – 1.32
Patient age (years) 0.21 0.84
65 – 74 (747) 1.00 1.00
75 – 84 (363) 0.71 0.47 – 1.06 0.88 0.55 – 1.39
85 – 89 (38) 0.68 0.22 – 2.04 0.87 0.25 – 3.06
Annual surgeon case volume of Medicare operations 0.91 0.65
< 6 (304) 1.00 1.00
6 – 12 (350) 0.91 0.58 – 1.44 1.03 0.60 – 1.77
> 12 (444) 0.92 0.59 – 1.43 1.24 0.74 – 2.08
Unknown (50) 0.72 0.27 – 1.89 0.68 0.22 – 2.17
Open in a new tab

Abbreviations: ALND, axillary lymph node dissection; BCS, breast-conserving surgery; CI, confidence interval; DCIS, ductal carcinoma in situ; OR, odds ratio; SLNB, sentinel lymph node biopsy

a

The self-reported prevalence of lymphedema was 14%; the simplified algorithm prevalence rate was slightly lower at 10%.

Discussion

In this study, we have developed and validated a simple 2-code algorithm and a more complex 3-step algorithm for the use of Medicare claims data to identify women who develop lymphedema after breast cancer treatment in a population-based cohort of older breast cancer survivors. The more complicated, multi-step algorithm performs similarly to the much simpler 2-code algorithm, which involves only the identification of either lymphedema ICD-9 code 457.0 or 457.1 at any time point after incident breast cancer surgery.

This simple algorithm has an overall sensitivity of 69% and specificity of 96%, using a gold standard of a patient reporting that her doctor told her she had lymphedema or arm edema since undergoing breast cancer surgery. The positive predictive value of this algorithm varies across different lymphedema prevalence rates (Figure 2) but is greater than 75% when the incidence of lymphedema is greater than 16%. The negative predictive value of this algorithm is greater than 90% across a wide rate of lymphedema incidence rates. Given the similar performance of both algorithms, and the ease of implementing the simplified algorithm, we recommend using the simpler two ICD-9 code algorithm. Another reason to favor using the simple algorithm is that it does not include lymphedema treatment codes, thereby enhancing the ability to most effectively study lymphedema treatment and its outcomes in cohorts identified by this algorithm as having lymphedema. The ability to systematically study lymphedema treatment, in particular, in large, population-based cohorts is critical since lymphedema is a life-long, chronic disease with significant treatment costs, which are often out-of-pocket expenses [23]. Overall, there is a lack of high-quality studies evaluating access to and utilization of lymphedema treatment, as well as comparative effectiveness and cost-effectiveness of various lymphedema treatment modalities [42–44].

The algorithm development process illustrates several issues with respect to the use of Medicare claims to identify incident lymphedema cases. First, our gold standard for lymphedema was strictly defined by a patient’s recall of being told by her doctor that she had lymphedema or arm edema since undergoing breast cancer surgery. Review of medical records for documentation of lymphedema was not performed. In addition, by using this strict definition, we acknowledge that our algorithm may not capture women who have mild symptoms of lymphedema and/or do not report their symptoms to their physician. Our algorithm focuses on women with likely more severe, clinically important cases of lymphedema. Second, despite exhaustive attempts to maximize the algorithm sensitivity, we were unable to achieve a sensitivity greater than 70%, which may be partially due to undercoding of lymphedema claims. Therefore, milder cases of lymphedema are likely not captured by this algorithm and incidence rates of lymphedema determined by this algorithm would need to be inflated by approximately 30% to more closely represent true incidence rates.

In an attempt to maximize the positive predictive value of the algorithm, we accepted this moderate sensitivity of 70%. There are limitations to using this algorithm if the prevalence of lymphedema is low [39]. However, based on the recently published meta-analysis of studies from 2000–2012, lymphedema rates are greater than 15% at a certain time point postoperatively even in this era of SLNB [24]. At these incidence rates (Figure 2), the PPV of the algorithm is 75% or higher. Finally, optimal development of the multi-step algorithm may have been affected by the lack of Medicare coverage for essentially all of the HCPCS durable medical equipment codes that were evaluated. Of the 28 codes examined, only 2 were identified with a near-zero prevalence (Table 2); therefore, no HCPCS codes could be evaluated for inclusion during algorithm development. However, as previously discussed, since the simple algorithm does not include codes related to lymphedema treatment, this simple algorithm allows studies focusing on various aspects of lymphedema treatment to be performed.

Despite these limitations, to our knowledge, this is the first claims-based, validated algorithm that identifies incident cases of lymphedema after breast cancer surgery. Face validity of this algorithm is demonstrated in Table 4, which shows that algorithm-identified cases of lymphedema, compared to self-reported cases of lymphedema, are similar with respects to predictors of lymphedema. To assess its generalizability, the algorithm should be further validated in other large, population-based breast cancer cohorts, including younger women and more contemporary cohorts. Results must be interpreted with the knowledge of the algorithm’s performance.

Potential uses of this algorithm include subsequent studies of lymphedema in large administrative databases to further refine studies in several key areas, including changes in incidence rates, risk factors, and prevention measures. Furthermore, as already mentioned, since this algorithm does not include lymphedema treatment codes, studies on treatment options and cost/economic analyses of the burden and treatment of lymphedema options could be explored. By identifying patients with clinically apparent lymphedema after breast cancer treatment by a standard, claims-based definition, this validated algorithm will facilitate the conduct of large, population-based studies of lymphedema in a uniform fashion. The results of these studies will be critical to advancing our understanding and management of this challenging and debilitating chronic disease that is feared by all breast cancer survivors.

Acknowledgments

This research was supported by a career development award and supplement to Dr. Yen from the National Cancer Institute (K07CA125586, K07CA125586-03S1) and two research grants from the National Cancer Institute to Dr. Nattinger (R01CA81379, R01CA127648). The content does not necessarily represent the official views of the National Cancer Institute or the National Institutes of Health.

Footnotes

Conflict of Interest

All authors (Tina Yen, Purushuttom Laud, Rodney Sparapani, Jianing Li and Ann Nattinger) declare that they have no conflicts of interest.

Integrity of Research and Reporting

This study was approved by the Medical College of Wisconsin’s institutional review board. All procedures followed were in accordance with the ethical standards of the responsible committee on human experimentation (institutional and national) and with the Helsinki Declaration of 1975, as revised in 2000. Informed consent was obtained from all patients for being included in the study.

References

  • 1.Efron B. How biased is the apparent error rate of a prediction rule? J Am Stat Assoc. 1986;81:461–70. [Google Scholar]
  • 2.Cooper GS, Virnig B, Klabunde CN, Schussler N, Freeman J, Warren JL. Use of SEER-Medicare data for measuring cancer surgery. Medical care. 2002;40(8 Suppl):IV-43–8. doi: 10.1097/01.MLR.0000020943.21850.F1.. [DOI] [PubMed] [Google Scholar]
  • 3.McClish DK, Penberthy L, Whittemore M, Newschaffer C, Woolard D, Desch CE, et al. Ability of Medicare claims data and cancer registries to identify cancer cases and treatment. Am J Epidemiol. 1997;145:227–33. doi: 10.1093/oxfordjournals.aje.a009095. [DOI] [PubMed] [Google Scholar]
  • 4.Rector TS, Wickstrom SL, Shah M, Thomas Greeenlee N, Rheault P, Rogowski J, et al. Specificity and sensitivity of claims-based algorithms for identifying members of Medicare+Choice health plans that have chronic medical conditions. Health Serv Res. 2004;39:1839–57. doi: 10.1111/j.1475-6773.2004.00321.x.. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Molinaro AM, Simon R, Pfeiffer RM. Prediction error estimation: a comparison of resampling methods. Bioinformatics. 2005;21(15):3301–7. doi: 10.1093/bioinformatics/bti499.. [DOI] [PubMed] [Google Scholar]
  • 6.Nattinger AB, Laud PW, Bajorunaite R, Sparapani RA, Freeman JL. An algorithm for the use of Medicare claims data to identify women with incident breast cancer. Health Serv Res. 2004;39:1733–49. doi: 10.1111/j.1475-6773.2004.00315.x. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Freeman JL, Zhang D, Freeman DH, Goodwin JS. An approach to identifying incident breast cancer cases using Medicare claims data. J Clin Epidemiol. 2000;53:605–14. doi: 10.1016/s0895-4356(99)00173-0. [DOI] [PubMed] [Google Scholar]
  • 8.Warren JL, Feuer E, Potosky AL, Riley GF, Lynch CF. Use of Medicare hospital and physician data to assess breast cancer incidence. Med Care. 1999;37:445–56. doi: 10.1097/00005650-199905000-00004. [DOI] [PubMed] [Google Scholar]
  • 9.Ramsey SD, Scoggins JF, Blough DK, McDermott CL, Reyes CM. Sensitivity of administrative claims to identify incident cases of lung cancer: a comparison of 3 health plans. J Manag Care Pharm. 2009;15:659–68. doi: 10.18553/jmcp.2009.15.8.659. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Goldberg DS, Lewis JD, Halpern SD, Weiner MG, Lo Re V., 3rd Validation of a coding algorithm to identify patients with hepatocellular carcinoma in an administrative database. Pharmacoepidemiol Drug Saf. 2013;22:103–7. doi: 10.1002/pds.3367.. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11.Hassett MJ, Ritzwoller DP, Taback N, Carroll N, Cronin AM, Ting GV, et al. Validating Billing/Encounter Codes as Indicators of Lung, Colorectal, Breast, and Prostate Cancer Recurrence Using 2 Large Contemporary Cohorts. Med Care. 2012 doi: 10.1097/MLR.0b013e318277eb6f.. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.Chubak J, Yu O, Pocobelli G, Lamerato L, Webster J, Prout MN, et al. Administrative data algorithms to identify second breast cancer events following early-stage invasive breast cancer. J Natl Cancer Inst. 2012;104:931–40. doi: 10.1093/jnci/djs233.. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13.Lamont EB, Herndon JE, 2nd, Weeks JC, Henderson IC, Earle CC, Schilsky RL, et al. Measuring disease-free survival and cancer relapse using Medicare claims from CALGB breast cancer trial participants (companion to 9344) J Natl Cancer Inst. 2006;98:1335–8. doi: 10.1093/jnci/djj363.. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.McClish D, Penberthy L, Pugh A. Using Medicare claims to identify second primary cancers and recurrences in order to supplement a cancer registry. J Clin Epidemiol. 2003;56:760–7. doi: 10.1016/s0895-4356(03)00091-x. [DOI] [PubMed] [Google Scholar]
  • 15.Earle CC, Nattinger AB, Potosky AL, Lang K, Mallick R, Berger M, et al. Identifying cancer relapse using SEER-Medicare data. Med Care. 2002;40(8 Suppl):IV-75–81. doi: 10.1097/01.MLR.0000020941.11143.76.. [DOI] [PubMed] [Google Scholar]
  • 16.Allen LA, Yood MU, Wagner EH, Aiello Bowles EJ, Pardee R, Wellman R, et al. Performance of claims-based algorithms for identifying heart failure and cardiomyopathy among patients diagnosed with breast cancer. Med Care. 2014;52:e30–8. doi: 10.1097/MLR.0b013e31825a8c22.. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Sewell JM, Rao A, Elliott SP. Validating a Claims-based Method for Assessing Severe Rectal and Urinary Adverse Effects of Radiotherapy. Urology. 2013;82:335–40. doi: 10.1016/j.urology.2013.02.071.. [DOI] [PubMed] [Google Scholar]
  • 18.Petrek JA, Pressman PI, Smith RA. Lymphedema: current issues in research and management. CA Cancer J Clin. 2000;50:292–307. doi: 10.3322/canjclin.50.5.292. quiz 8–11. [DOI] [PubMed] [Google Scholar]
  • 19.Sparaco A, Fentiman IS. 9. Arm lymphoedema following breast cancer treatment. Int J Clin Pract. 2002;56:107–10. [PubMed] [Google Scholar]
  • 20.Hayes SC, Johansson K, Stout NL, Prosnitz R, Armer JM, Gabram S, et al. Upper-body morbidity after breast cancer: incidence and evidence for evaluation, prevention, and management within a prospective surveillance model of care. Cancer. 2012;118(8 Suppl):2237–49. doi: 10.1002/cncr.27467.. [DOI] [PubMed] [Google Scholar]
  • 21.Maunsell E, Brisson J, Deschenes L. Arm problems and psychological distress after surgery for breast cancer. Can J Surg. 1993;36:315–20. [PubMed] [Google Scholar]
  • 22.Fu MR, Ridner SH, Hu SH, Stewart BR, Cormier JN, Armer JM. Psychosocial impact of lymphedema: a systematic review of literature from 2004 to 2011. Psychooncology. 2013;22:1466–84. doi: 10.1002/pon.3201.. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23.Shih YC, Xu Y, Cormier JN, Giordano S, Ridner SH, Buchholz TA, et al. Incidence, treatment costs, and complications of lymphedema after breast cancer among women of working age: a 2-year follow-up study. J Clin Oncol. 2009;27:2007–14. doi: 10.1200/JCO.2008.18.3517.. JCO.2008.18.3517 [pii] [DOI] [PubMed] [Google Scholar]
  • 24.Disipio T, Rye S, Newman B, Hayes S. Incidence of unilateral arm lymphoedema after breast cancer: a systematic review and meta-analysis. Lancet Oncol. 2013;14:500–15. doi: 10.1016/S1470-2045(13)70076-7.. [DOI] [PubMed] [Google Scholar]
  • 25.Erickson VS, Pearson ML, Ganz PA, Adams J, Kahn KL. Arm edema in breast cancer patients. J Natl Cancer Inst. 2001;93:96–111. doi: 10.1093/jnci/93.2.96. [DOI] [PubMed] [Google Scholar]
  • 26.Petrek JA, Heelan MC. Incidence of breast carcinoma-related lymphedema. Cancer. 1998;83(12 Suppl American):2776–81. doi: 10.1002/(sici)1097-0142(19981215)83:12b+<2776::aid-cncr25>3.0.co;2-v. [DOI] [PubMed] [Google Scholar]
  • 27.Armer JM, Stewart BR. Post-breast cancer lymphedema: incidence increases from 12 to 30 to 60 months. Lymphology. 2010;43:118–27. [PMC free article] [PubMed] [Google Scholar]
  • 28.McLaughlin SA. Lymphedema: separating fact from fiction. Oncology. 2012;26:242–9. [PubMed] [Google Scholar]
  • 29.Reiner AS, Jacks LM, Van Zee KJ, Panageas KS. A SEER-Medicare population-based study of lymphedema-related claims incidence following breast cancer in men. Breast Cancer Res Treat. 2011;130:301–6. doi: 10.1007/s10549-011-1649-1.. [DOI] [PubMed] [Google Scholar]
  • 30.Kwan ML, Darbinian J, Schmitz KH, Citron R, Partee P, Kutner SE, et al. Risk factors for lymphedema in a prospective breast cancer survivorship study: the Pathways Study. Arch Surg. 2010;145:1055–63. doi: 10.1001/archsurg.2010.231.. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.From Cancer Patient to Cancer Survivor: Lost in Translation. 1. Washington, D.C: National Academies Press; 2005. [Google Scholar]
  • 32.Nattinger AB, Pezzin LE, Sparapani RA, Neuner JM, King TK, Laud PW. Heightened attention to medical privacy: challenges for unbiased sample recruitment and a possible solution. Am J Epidemiol. 2010;172:637–44. doi: 10.1093/aje/kwq220.. kwq220 [pii] [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33.Yen TW, Fan X, Sparapani R, Laud PW, Walker AP, Nattinger AB. A contemporary, population-based study of lymphedema risk factors in older women with breast cancer. Ann Surg Oncol. 2009;16:979–88. doi: 10.1245/s10434-009-0347-2.. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34.The North American Association of Central Cancer Registries, Inc. [Accessed 17 February 2014]; http://www.naaccr.org.
  • 35.ICD-9 CM International Classification of Diseases, 9th revision. Millenium edition. 2008. Los Angeles, CA: Practice Management Information Corporation; 2007. [Google Scholar]
  • 36.Current Procedural Terminology: CPT 2004. Chicago, IL: American Medical Association; 2003. [Google Scholar]
  • 37.HCPCS Health Care Financing Administration Common Procedure Coding System. National Level II Medicare Codes. Millennium edition. 2008. Los Angeles, CA: Practice Management Information Corporation; 2007. [Google Scholar]
  • 38.Kwan W, Jackson J, Weir LM, Dingee C, McGregor G, Olivotto IA. Chronic arm morbidity after curative breast cancer treatment: prevalence and impact on quality of life. J Clin Oncol. 2002;20:4242–8. doi: 10.1200/JCO.2002.09.018. [DOI] [PubMed] [Google Scholar]
  • 39.Altman DG, Bland JM. Diagnostic tests 2: Predictive values. BMJ. 1994;309:102. doi: 10.1136/bmj.309.6947.102. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40.HIPAASpace. Medical Data Services. Healthcare Lookup Services. [Accessed 17 February 2014];HCPCS Lookup. http://www.hipaaspace.com/Medical_Billing/Coding/Healthcare_Common_Procedure_Coding_System/HCPCS_Codes_Lookup.aspx.
  • 41.HCPCSdata.com. [Accessed 17 February 2014];HCPCS L Codes. 2006 http://www.hcpcsdata.com/2006/L/default.htm.
  • 42.International Society of Lymphology. The diagnosis and treatment of peripheral lymphedema. 2009 Concensus Document of the International Society of Lymphology. Lymphology. 2009;42:51–60. [PubMed] [Google Scholar]
  • 43.Koul R, Dufan T, Russell C, Guenther W, Nugent Z, Sun X, et al. Efficacy of complete decongestive therapy and manual lymphatic drainage on treatment-related lymphedema in breast cancer. Int J Radiat Oncol Biol Phys. 2007;67:841–6. doi: 10.1016/j.ijrobp.2006.09.024.. [DOI] [PubMed] [Google Scholar]
  • 44.Sayko O, Pezzin LE, Yen TW, Nattinger AB. Diagnosis and treatment of lymphedema after breast cancer: a population-based study. PM R. 2013;5(11):915–23. doi: 10.1016/j.pmrj.2013.05.005.. [DOI] [PMC free article] [PubMed] [Google Scholar]

RESOURCES