Study Design and Data Sources

This was a retrospective observational study conducted using administrative claims data linked with abstracted medical chart data to create an individual-level analytic database to support the study objectives.

We previously developed an algorithm [21] utilizing the Optum Research Database (ORD), a deidentified database containing medical and pharmacy claims data with linked enrollment information for individuals enrolled in health plans across the US. It is generally similar to the US insured population regarding age, race, sex, and geographic distribution. The ORD includes data from approximately 8% of US commercial health plan enrollees and 18% of Medicare Advantage enrollees. Medical claims in the ORD include diagnosis and procedure codes from the International Classification of Diseases, Ninth and Tenth Revisions, Clinical Modification (ICD-9-CM and ICD-10-CM); Current Procedural Terminology or Healthcare Common Procedure Coding System codes; site of service codes; paid amounts; and other information. Pharmacy claims in the ORD include drug name, National Drug Code, dosage form, drug strength, fill date, and financial information for outpatient pharmacy services.

For each patient, one healthcare provider was identified from the administrative claims data and subsequently targeted for medical chart abstraction using a hierarchical process. Specifically, eligible providers were required to be oncologists, with priority given to those with the highest number of medical claims (with a diagnosis of epithelial ovarian, fallopian tube, or primary peritoneal cancer) for the given patient, with a claim date most proximal to the date of the first chemotherapy treatment. Upon procurement, some charts may not have been eligible for abstraction. To be included, charts were required to indicate diagnosis of and treatment for ovarian cancer and to have at least 365 days of data available to abstract. Medical chart data were collected from a convenience sample of patients with providers willing to participate in the chart review process.

The study protocol was reviewed and approved by the New England Institutional Review Board (15 June 2020, #20,201,707). The study obtained a waiver of informed consent and a waiver of authorization under 45 CFR 164.512(i).

Study Population and Cohort Assignment

The study population included adult female commercial and Medicare Advantage health plan enrollees who had a diagnosis of ovarian cancer (at least 2 nondiagnostic medical claims ≥ 30 days apart for epithelial ovarian, fallopian tube, or primary peritoneal cancer [ICD-9-CM 183.0x, 183.2x; ICD-10-CM C56*, C570*]) and initiated a systemic chemotherapy (Supplemental Table S1) during the period of December 1, 2014, through September 15, 2017 (initial identification period, Supplemental Fig. S1). This initial identification period was used to identify initial LOTs for each patient. The date of the first chemotherapy treatment for ovarian cancer during the initial identification period was designated as the index date. Subsequent LOTs were identified during the remainder of the full identification period (September 15, 2017, through November 30, 2019; Supplemental Fig. S1). Included patients were also required to have continuous health plan enrollment for the 6 months prior to the index date and at least 6 months after the index date; patient identifying information available to support medical chart abstraction; and medical charts eligible for abstraction. Patients were excluded if they had any of the following: a claim for pregnancy during the full identification period; a claim for systemic chemotherapy (Supplemental Table S1) during the baseline period; at least 2 non-diagnostic medical claims with a diagnosis for any cancer excluding epithelial ovarian, fallopian tube, primary peritoneal, genitourinary, external genitalia, or intra-abdominal at least 30 days apart during the baseline period; medical chart evidence of any distal cancers that were not a metastasis of the ovarian cancer; or claims or medical chart evidence of clinical trial participation involving an active intervention such as medication or surgery.

Patients were assigned to study cohorts based on claims for ovarian cancer treatment during the full identification period. Patients with a claim for PARPi (olaparib, niraparib, rucaparib) during the full identification period were assigned to the PARPi cohort; these patients may also have had a claim for bevacizumab during the identification period. Patients with a claim for bevacizumab but no claim for PARPi during the full identification period were assigned to the bevacizumab cohort. Finally, patients with no claims for PARPi or bevacizumab during the full identification period were assigned to the chemotherapy-only cohort.

Study Variables

Claims-based study variables were assessed during the 6 months prior to the index date and for at least 6 months post-index (claims baseline and follow-up periods, respectively; Supplemental Fig. S1). Chart-based study variables were assessed during the 3 months prior to the index date and for at least 12 months post-index (chart baseline and follow-up periods, respectively; Supplemental Fig. S1). For both claims and chart data, the duration of the variable follow-up period was a minimum of 6 months, a mean of 22 months, and a maximum of 33 months post-index. Baseline variables assessed from claims data included age, sex, insurance type, geographic region, baseline Charlson comorbidity index score [22], and comorbid conditions identified using Clinical Classifications Software from the Agency for Healthcare Research and Quality [23]. Cancer stage at index date and cancer history and surgical history during the 3-month baseline period were assessed from chart data. LOTs during follow-up were identified separately for claims and chart data, as described below.

Line of Therapy Identification

The claims algorithm was designed to identify LOTs based on the timing of events, therapies, and treatment gaps in a three-step process: (1) identify discrete treatment episodes of systemic therapy using treatment start dates and stop dates calculated from days’ supply; (2) identify neoadjuvant and adjuvant therapy among patients with evidence of surgery; (3) identify active and maintenance treatment. Detailed descriptions of each algorithm step are presented in Fig. 1. After treatment episodes had been identified using the three-step algorithm, all adjuvant and active treatment LOTs were categorized as Active LOT 1, Active LOT 2, etc., for up to five LOTs per patient. Maintenance lines following an active LOT were classified as Maintenance LOT X, with X corresponding to the number of the active LOT.

Fig. 1

Three-step line of therapy algorithm. After treatment episodes are identified using the algorithm, adjuvant and active treatment episodes (lines of therapy; LOTs) are categorized as Active LOT 1, Active LOT 2, etc. Maintenance lines following an active LOT are classified as Maintenance LOT X, with X corresponding to the number of the active LOT. aSystemic chemotherapies are listed in Supplemental Table S1. Leucovorin was not identified as a chemotherapy, even though it may be given in conjunction with chemotherapy. bSwitching between platinum agents (cisplatin, oxaliplatin, and carboplatin) or switching between docetaxel and paclitaxel was not considered to constitute the end of a treatment episode. cPoly (ADP-ribose) polymerase inhibitor (olaparib, niraparib, or rucaparib)

LOTs from chart data were identified similarly, with the following differences:

Treatment start and stop dates were identified from medical charts. For dates that had a missing value for day but known month and year, the day was assigned as the 15th of the month.

Identification of active and maintenance treatment was based on mention of the words “active” and “maintenance” in medical charts to describe treatment status or from other chart information sufficient to deduce treatment status if “active” or “maintenance” was not mentioned.

Statistical Analysis

Baseline patient characteristics were analyzed descriptively and stratified by treatment cohort. To validate the algorithm, LOT results generated using claims data vs chart data were compared at the patient level by calculating the percent agreement between several characteristics, including total number of active and maintenance LOTs, type of therapy (neoadjuvant vs adjuvant classification), and type of regimen (individual drugs). The magnitude of agreement between total LOT results and type of therapy for claims vs charts was assessed using Cohen’s kappa statistic (κ) or weighted k (to account for the presence of off-diagonal counts [24]), both of which were interpreted as follows: ≤ 0, no agreement; 0.01–0.20, slight agreement; 0.21–0.40, fair agreement; 0.41–0.60, moderate agreement; 0.61–0.80, substantial agreement; 0.81–1.00, almost perfect agreement. P-values ≤ 0.05 were considered to indicate statistical significance. All analyses were performed using SAS 9.4 (SAS Institute, Cary, NC, USA).