An algorithmic approach to deriving line of therapy in a real-world data set for non-small cell lung cancer (NSCLC).

Abstract

e18099 Background: Real World Data (RWD) is being used for outcomes research and regulatory submissions. A key variable needed to understand treatment outcomes is Line of Therapy (LoT). However, LoT is generally not captured in RWD sources such as electronic health records (EHR) or claims data, and is typically derived using manual abstraction. To determine whether an automated approach to LoT derivation is possible, we created an algorithm and applied it to patients (pts) in the Syapse Learning Health Network. Methods: We selected confirmed NSCLC pts from 4 health systems in the RWD set, verifying diagnosis using ICD-9/10/O3 topography and morphology codes. We analyzed the EHR-derived medication list using a regimen-independent algorithm that classified antineoplastic drugs (AD), as defined by ATC L01, into LoT. Within each LoT, we compared the top 80% of AD prescribed (by volume of pts) to the LoT as indicated on each drug’s FDA label. We then used descriptive statistical summaries to outline the alignment between automated algorithmic results and indicated usage within that LoT. Results: In a set of 10,842 NSCLC pts, a total of 106 unique AD were prescribed in the first line as identified by our algorithm, and 13 drugs were prescribed as first line for 80% of the pts. Of those, 9 (69%) of those are indicated for first line, 3 are not indicated for NSCLC, and 1 is indicated for a subsequent NSCLC line, per FDA labels. 82 unique AD were prescribed in the second line as identified by our algorithm, and 15 drugs were prescribed as second line for 80% of the pts. Of those, 11 (73%) are indicated for treatment/continuation therapy for recurrent, advanced or metastatic disease, 3 are not indicated for NSCLC, and 1 is indicated for first line NSCLC per FDA labels. 36 unique AD were prescribed in subsequent line as identified by our algorithm, and 18 drugs were prescribed as subsequent line for 80% of the pts. Of those, 12 (67%) are indicated for treatment of recurrent, advanced or metastatic disease or subsequent systemic therapy, 5 are not indicated for NSCLC and 1 is indicated for first line per FDA labels. Conclusions: An automated algorithmic approach for deriving lines of therapy may be a viable solution to scalably calculate LoT in RWD sets. A deeper analysis using statistical sensitivity and specificity assessment of such algorithms is needed to validate the potential of an algorithmic approach.