OBJECTIVES : In real-world oncology research, algorithms may be employed to establish line of therapy (LOT) by utilizing treatment information from electronic health record (EHR) structured data. It is important to validate how well an algorithm-based LOT matches with a clinically classified LOT. We evaluated concordance of LOT identified by algorithm (structured) and clinical inputs (unstructured).

METHODS : This retrospective study included 300 randomly selected adult patients diagnosed with stage IV small cell lung cancer (SCLC) and non-small cell lung cancer (NSCLC) between 01-Jan-2017 and 30-Sep-2019 within The US Oncology Network. Study data including both structured data (programmatically extracted) and unstructured data (requiring manual chart review) were captured from the EHR. Treatment start/stop dates and death dates were sourced from structured and unstructured data, while reasons for treatment initiation and discontinuation came from unstructured data. 1L and 2L were assigned utilizing a structured data algorithm employing dates of treatment. This was evaluated against the LOT assigned by employing the clinical inputs of reasons for treatment initiation and discontinuation.

RESULTS : Of the 150 patients with SCLC, 148 initiated 1L by both structured and unstructured data (98.6% percentage-agreement); all reported the same regimen (100% percentage-agreement). By algorithm and clinical inputs, 33 patients initiated 2L having identical regimens (kappa-statistic: 0.81, 95%CI: 0.69-0.92). There were 11 discordant patients for 2L: 1 and 10 patients by unstructured and structured data, respectively.

Of the 150 patients with NSCLC, 147 initiated 1L by both structured and unstructured data (98% percentage-agreement); 135/147 reported the same regimen (91.8% percentage-agreement). By algorithm and clinical inputs, 29 patients initiated 2L having identical regimens (kappa-statistic: 0.56, 95%CI: 0.42-0.70). There were 27 discordant patients for 2L: 4 and 23 patients by unstructured and structured data, respectively.

CONCLUSIONS : Algorithm-based utilization of structured data could appropriately approximate LOT. However, approximations were more robust with intravenous treatments.