Determining Line of Therapy from Real-World Data in Non-Small Cell Lung Cancer

Introduction

When using real-world data to answer clinical questions in oncology, accurate determination of lines of therapy (LOT) using treatment records is crucial to conducting meaningful analyses. In the absence of detailed information on progression and toxicity, investigators rely on treatment records alone to infer changes in clinical disease status and determine LOTs. This non-uniform methodology may cause significant variation in how LOTs are defined depending on malignancy type, data source, and treatment patterns. Standardized rules and algorithms for determining LOTs are needed to improve the reproducibility and comparability of real-world data analyses.

There is tremendous opportunity for impactful analysis of real-world data from patients with metastatic non-small cell lung cancer (mNSCLC), but standardized LOT determination has been difficult. Systemic therapies for mNSCLC now include immunotherapy, chemotherapy, and targeted therapies, which can be used alone or in combination. Because of this, treatment patterns and sequencing are complex and heterogeneous¹. This is particularly true for the subset of patients with mNSCLC and rare molecular characteristics and patients with co-morbidities where complex treatment patterns and outcomes are only captured outside of clinical trials.

Previously, others have proposed algorithms for determining LOTs for patients with solid tumors, lung cancer, and mNSCLC.² However, these algorithms do not specifically incorporate common treatment patterns for patients with targetable alterations, such as EGFR, ALK, and ROS1 alterations. In this report, we aim to propose new rules for determining LOT from real-world data of patients with mNSCLC, focusing on treatment patterns among patients receiving targeted therapies.

Methods

We defined a regimen as any systemic anticancer agent (i.e., chemotherapy, immunotherapy, or targeted therapy) or combination of systemic agents administered at the same time. First, we developed rules for determining LOT (“LOT rules”) by expert consensus and review of existing literature. Next, these rules were applied to treatment records from 550 patients with mNSCLC and sensitizing ALK or ROS1 fusions in the ATOMIC Driver Mutation Registry. ATOMIC is a multi-institutional collaborative of twelve academic centers across North America. Trained data abstractors used a standardized digital form to record patients’ regimen agent(s) and start/end dates. Finally, thoracic oncologist members of ATOMIC reviewed the LOT determination and revised the rules to reflect practice patterns most indicative of changes to clinical disease status. These modified rules were then applied to a separate ATOMIC cohort of 1,133 patients with mNSCLC and EGFR sensitizing mutations. Rules were also developed to identify LOTs that are most likely applicable to early-stage disease (“ESD rules”) since data sources of patients with mNSCLC may inadvertently include treatments for this disease stage.

In this report, we demonstrate the use of LOT and ESD rules on treatment records from the ATOMIC EGFR cohort. We compared our LOT determination with those of an existing algorithm for mNSCLC proposed by Hess et al using the SAS macro %mnsclc_lot². Analyses were performed in R version 4.2.1³ and SAS software version 9.4. This study was approved by the University of Pennsylvania Institutional Review Board (IRB #829009).

Results

We propose six rules for determining LOT in real-world data of patients with mNSCLC (Table 1, Figure 1). Among 1,133 patients in ATOMIC with EGFR alterations and mNSCLC, a total of 3168 regimens were recorded with a median of 2 regimens per patient (IQR, 1-4; range, 1-13). After applying our rules, there were 2834 total LOTs with a median of 2 LOTs per patient (IQR, 1-3; range, 1-11). Rules 1-3 prevented 334 regimen changes from advancing the LOT (Table 1). Thirty-seven patients switched regimens within 28 days of the start of the first LOT and had a line 0.5 recorded.

• Download figure
• Open in new tab

Figure 1: Algorithm for line of therapy determination

Comparatively, after applying an algorithm proposed by Hess et al. to the same treatment records, there were 2771 total LOTs with a median of 2 per patient (IQR 1-3; range 1-9). Among our method’s 3999 agent-LOT assignments (e.g., osimertinib as 1L, carboplatin as 2L, etc.), ours differ from the Hess method in 5.7% of assignments (Table 2).

Additionally, we developed two rules to remove systemic therapy agents likely used in the treatment of early-stage disease (ESD) from the LOT determination. (1) If a LOT was durvalumab monotherapy administered prior to November 11, 2022 (i.e., prior to Food and Drug Administration approval for durvalumab in stage IV NSCLC)⁴ or (2) if the LOT was chemotherapy with a cumulative cycle duration of 30-140 days and no subsequent LOT for at least four months of follow-up, the LOT should be considered as a treatment for early-stage disease and excluded from the patient’s LOT enumeration for mNSCLC. Among patients in ATOMIC with EGFR alterations and mNSCLC, the first LOT for 81 patients met the criteria for early-stage disease treatment (ESD Rule 1 met for two patients and ESD Rule 2 met for 79). These LOTs were removed, and the remaining LOTs were renumbered as LOTs for mNSCLC.

Discussion

We propose six simple rules to determine whether a change in treatment for mNSCLC should or should not advance the LOT in real-world data. Our LOT assignments led to different classifications from previous work in 5.7% of cases, primarily in characterizing targeted therapy-containing regimens. Since LOT assignment can affect outcome measures such as real-world progression-free survival or time to treatment discontinuation, even small differences may lead to different conclusions in an observational study. Therefore, it is critical that the LOT determination accurately reflects clinical practice.

There are several ways that our proposed rules better characterize LOTs in mNSCLC populations treated with targeted therapy compared to previously reported methods (Supplemental table 1).^2,4–7 First, we distinguish between biologics and therapies targeting oncogenic drivers in mNSCLC. This distinction is important because biologics, such as angiogenesis inhibitors, may have little anti-tumor activity alone and are often added or removed from treatment regimens without a clear change in disease status, while targeted therapies have tremendous anti-tumor activity and are used in the setting of a change in disease status. Second, there are instances where targeted therapies are added to or replace a previous regimen because of a newly identified targetable mutation and not because of a change in disease status. The rules proposed allow for this approach by allowing for 1 cycle of chemotherapy to be given while waiting for molecular testing results prior to switching to a targeted therapy without advancing the LOT. Third, we allow for a treatment pause and re-initiation within 60 days to be considered the same LOT, which may be more appropriate for targeted therapies typically taken daily by mouth that are paused and restarted for toxicity.^5–7

Retrospective determination of whether a treatment was given in the setting of metastatic disease versus early-stage disease is critical to determining real-world clinical endpoints. In this historical dataset, we used rules based on the approved therapies at the time of the dataset, but even while completing the analysis for this study the indications for several systemic therapies expanded to include early-stage disease, making the agent itself an unreliable indicator of early-stage treatment. Given this complexity, extracting the date of metastatic diagnosis at the time of data abstraction will become even more important to evaluate the practice patterns contemporary to the time of the dataset.

We believe wider use of this framework for LOT determination could lead to better method harmonization of observational studies in mNSCLC and allow for comparison and aggregation across datasets in the future.

Tables and figures