A protocol for three observational cohort studies evaluating adverse outcomes, excess costs and repeat procedures after surgery for breast cancer in the USA

This study will be a retrospective cohort analysis of adult early breast cancer patients using administrative claims, enrollment information, and linked clinical data from Optum's de-identified Market Clarity Data (2007–2023). Optum^® Market Clarity Data is an integrated, multi-source dataset comprising medical claims, pharmacy claims, and electronic health records. Optum Market Clarity links electronic health record data – including lab results, vital signs, measurements, diagnoses, procedures, and information derived from unstructured clinical notes using natural language processing – with historical, linked administrative claim data, including pharmacy claims, physician claims, clinical information facility claims, and medications prescribed and administered. Optum Market Clarity is statistically de-identified under the HIPAA Privacy Rule’s Expert Determination method and managed according to Optum customer data use agreements (27,28). Multi-source medical claims data integrates administrative medical and pharmacy claims, along with linked enrollment information. This dataset encompasses a broad spectrum of US insurance types, including commercial plans, Medicare Advantage, Medicaid, and others. Administrative medical data is sourced from both inpatient and outpatient electronic health records (EHRs) across the US. These records are processed, normalized, and standardized to ensure consistency. The structured EHR data includes demographics, prescribed medications, immunizations, contraindications, lab results (including microbiology), vital signs, clinical observations, inpatient stay details, and coded diagnoses and procedures. In addition, Optum employs advanced natural language processing (NLP) technology to extract clinical concepts from unstructured physician notes. This enhances the dataset with valuable insights such as signs and symptoms, family medical history, biomarker data, and oncology-specific variables – many of which may not be explicitly recorded in structured EHR fields. In total, Market Clarity contains data provided from 150 different commercial and government insurance payors on over 70 million patients [29].

The EHR data are compiled into both structured and unstructured data fields. Structured data fields include common variables like patient characteristics (e.g., BMI, race/ethnicity and smoking history), clinical details (e.g., laboratory results and diagnoses) and prevalidated cancer-specific characteristics (e.g., cancer stage and BRCA1/2 mutation status). Some patient characteristics, like race, ethnicity and smoking status were self-reported. The unstructured data are comprised of physician free text documentation, which are generated during the course of patient care, including office visits and operative notes. Public involvement is not planned as part of this database analysis.

Patients will be included if they appear in the Optum Market Clarity dataset from 1/1/2012 through 12/31/2022. All relevant procedures and diagnoses will be identified using International Classification of Diseases 9th revision (ICD-9), ICD-10, Current Procedural Terminology and Healthcare Common Procedure Coding System codes (Supplementary Table 1) [30–32]. For procedures and diagnoses without a clear coding framework or guidance, we will query the unstructured provider notes with relevant keywords to identify patients of interest. Breast cancer patients will be identified based on the presence of ICD-9 or ICD-10 diagnosis codes and will be excluded in the presence of evidence of metastatic disease. Lumpectomy and mastectomy patients will be identified using ICD procedure and Current Procedural Terminology codes. Males or patients with a diagnosis of a male breast cancer are excluded from analysis as male breast cancer may have unique disease characteristics and, as such, have treatment pathways that differ from female patients [33].

To be further considered for inclusion, patients must have at least two nondiagnostic medical claims with a diagnosis code for breast cancer on two separate dates during the study period. The earliest breast cancer claim during this period will be considered as the diagnosis date. Once the breast cancer cohort has been identified, patients will be assigned to both the lumpectomy and mastectomy cohorts based on receipt of either lumpectomy or mastectomy. If a patient received both a lumpectomy and mastectomy and is eligible for inclusion, they will be assigned to both the lumpectomy and mastectomy study cohorts. All patients must have 6 months of continuous insurance enrollment prior to the index, with 12–60 months of postoperative continuous enrollment in the lumpectomy cohort – allowing for sufficient follow-up to detect postlumpectomy re-operations – and 6 months in the mastectomy cohort – allowing for sufficient follow-up to detect postmastectomy complications.

Assignment to the nipple necrosis cohort will be defined by receipt of an NSM. NSM does not have a clearly defined procedure code in the US, so we will utilize the unstructured physician notes to define this cohort. As above, patients will be required to have the same diagnosis codes for inclusion, as well as a mastectomy procedure code. Patients then must have the presence of physician notes indicating receipt of a NSM (text string queries in Supplementary Table 2). Nipple necrosis cases will be further defined from the set of patients who were identified as NSM patients and then had the presence of a NAC ischemia/necrosis text string search in their free text notes (text string queries in Supplementary Table 2). Patients will be excluded from the mastectomy cohort if they are identified as NSM patients.

After patients are allocated into the different analytical cohorts, separate index dates and inclusion/exclusion criteria will be applied as previously described. The index date of each analysis will be the earliest claim for that specific breast cancer surgery procedure (lumpectomy or mastectomy). Patients will be followed until the earliest of death, end of continuous enrollment, or end of the study period. Data will be analyzed descriptively, accounting for the variable follow-up period as necessary.

Patient demographics will be identified from insurance enrollment information. These will include age, gender, race, ethnicity, insurance type (e.g., commercial, Medicare, etc.) and geographic region. Geographic region will be reported in alignment with one of five of the US Census Regions (i.e., Northeast, Midwest, South, West and other).

Baseline health will be comprised of comorbidities, BMI, smoking status, medication use, laboratory results and clinical details about the breast cancer. Comorbidities evaluated will include both those comorbidities used to construct both the Elixhauser Comorbidity and Charlson Comorbidity Indexes [34,35]. As nipple necrosis could be related to poor vascularization, the nipple necrosis cohort will also identify conditions that could impact vasculature, including the presence of diabetes and connective tissue disorders. Patients with missing data on covariates will not be excluded and the prevalence of missing data will be reported.

BMI is derived from structured data fields in the EHR, which can either be stored as a precalculated BMI field or calculated from a patient’s height and weight. BMI will also be categorized as underweight (0–18.5), low normal (18.5–25), overweight (25–30) and obese (greater than 30). Smoking status will also come from structured EHR data and will be defined as current, never or former smoker. For both smoking and BMI, their status will be selected from observations on or before the index surgical date, selecting the closest available observation to the index surgical date.

Use of immunologic, anticoagulant, steroids, antidiabetic and GLP-1 agonist medications during the baseline period will be indicated and identified from claims data using National Drug Codes. Laboratory results for HbA1c, platelet count, white blood cell count, albumin, prothrombin time test, cotinine and thyroid stimulating hormone will be extracted from EHR data. These results will be selected from tests administered on or before and closest to the index surgical date.

Disease details are collected from the EHR, of which some are stored in prebuilt structured fields and others are collected from unstructured notes and converted into structured fields using validated NLP algorithms. The NLP process begins with a random sample of data that is double annotated by two annotators with conflicts being resolved by a third review by a curator. This data becomes the gold standard which is partitioned into training, validation and testing subsets for NLP model development. Supervised machine learning models are trained to identify broader patterns not explicitly and manually created to enable the system to generalize to relevant contexts. The models are designed to identify the occurrences of the desired oncology concepts, their relations and the surrounding contexts such as temporal status and assertion (qualifiers) and apply labeling predictions to incoming data. The models are evaluated against the testing data and performance is quantified using precision, recall and F1 (the harmonic mean of precision and recall). Once models are finalized, they are run at scale, extracted concepts are normalized, de-duplicated and verified before being integrated into the database. Prebuilt disease details will include tumor grade, tumor size, presence of microcalcifications, nipple involvement and tumor location. NLP-derived disease details will include estrogen receptor/progesterone receptor (ER/PR) status and BRCA mutation status.

Information about the index surgical encounter will include procedure details, peri- and postoperative complications and surgeon/hospital profiles. Use of intra-operative adjuncts (i.e., needle localization, specimen radiograph and/or frozen pathology section review) and surgical margins will provide context about the index surgical encounter. Complications to be collected include hematoma, seroma, infections, hemorrhage, wound dehiscence, lymphedema, nerve injury, necrosis and implant loss. Surgeon specialty (e.g., breast surgeon, general surgeon, etc.) and the hospital’s number of beds, and community versus academic center status will additionally be reported.

Patient outcomes will be derived from a combination of claims and EHR data. Main outcomes to be identified include mortality, postoperative procedure and costs. Data on mortality is captured from a variety of sources in order to provide the most complete information on death available including the Social Security Administration SSA public use Death Master File (SSADF). Postoperative procedures for the lumpectomy cohort include the performance of a repeat lumpectomy or subsequent mastectomy after the index lumpectomy. The mastectomy cohort will also identify subsequent cases breast reconstruction and implant removal. The nipple necrosis cohort will quantify the necrosis treatments (e.g., hyperbaric oxygen therapy, debridement and skin graft) after an NSM. All costs during the follow-up period will be standardized and stratified into pharmacy and medical costs. Medical costs will be further divided into costs incurred in the ambulatory, inpatient, emergency and other settings. More details on variables to be analyzed are available in Supplementary Table 3.

In all cohorts, patient demographics and clinical characteristics will be described for patients undergoing lumpectomies, mastectomies and NSM using means for continuous variables and percentages for categorical variables. These will further be compared between subgroups (e.g., patients with and without repeat procedures in the lumpectomy cohort or patients with and without nipple necrosis in the Nipple Necrosis cohort) using t-tests for continuous variables and chi-square tests for categorical variables. Among patients in the lumpectomy cohort, prevalence and time to repeated lumpectomy or subsequent mastectomy will be calculated, with 5-years of follow-up. 4- to 1-year follow-up times will also be presented in sensitivity analyses. Time-to-event (i.e., additional repeated surgery) also be examined using Kaplan–Meier analyses. Patient survival (i.e., no repeated surgery), including median and interquartile range, will be reported at 3, 6, 9, 12, 24, 36, 48 and 60 month time points. Log-rank tests will be used to identify differences in survival between demographic (e.g., age and race/ethnicity) and clinical subgroups (e.g., tumor grade and ER/PR status). Patients will be classified by the total number and combination of surgeries received (e.g., one lumpectomy only vs one lumpectomy and one repeat lumpectomy vs one lumpectomy and one subsequent mastectomy). Median healthcare costs, including the previously mentioned cost subgroups, will be compared between the groups, using Wilcoxon Signed Rank tests to identify differences. p-values less than 0.05 will be considered statistically significant and all tests will be two-sided.

Patients in the mastectomy cohort will be evaluated to construct an overall patient journey, including pre-, peri-, and post-operative descriptive analyses. In addition to baseline health variables, preoperative characteristics will include receipt of neoadjuvant therapy and previous breast surgery. Perioperative characteristics will include the distribution of surgeon specialties, hospital sizes and types and complications. Post-operative characteristics will include total healthcare cost, breast reconstruction patterns and 5-year all-cause mortality for the lumpectomy cohort and 1-year all-cause mortality for the mastectomy cohort.

Nipple necrosis patients will be analyzed to describe the cost and consequences of this adverse outcome. We will describe the frequency and distribution of interventions required to resolve the necrosis and summarize the median cost these interventions. We will also conduct a similar analysis of a larger subset of patients who have had either a nipple or skin necrosis post-operatively, to evaluate the impact of broader necrotic events on healthcare resource utilization.