Evaluation of the accuracy of algorithms to identify soft tissue sarcoma (STS) in administrative claims
Nicole Princic1, Donna McMorrow1, Philip Chan1
1IBM Watson Health, 61 Summer Avenue, Reading, Cambridge, MA 01867 USA.
Background:
Lack of using a validated algorithm to select patients is a source of selection bias in oncology studies using administrative claims. The objective of this study to evaluate published algorithms to identify patients with soft tissue sarcoma (STS) in administrative claims and to evaluate new algorithms to improved performance.
Methods:
Two cancer populations including STS cases and non-STS controls were selected from the MarketScan Explorys Linked Claims-Electronic Medical Record (EMR) Database between January 1, 2000 and July 31, 2018. Eligible cases had a diagnosis on a clinical record for STS in the EMR while controls had no evidence of STS on any EMR records. Both cases and controls were enrolled in administrative claims during a period of observation and were aged ≥ 18 years. A split sample was used to test and validate algorithms using data from administrative claims. Values for sensitivity, specificity, and positive predictive value (PPV) were calculated for 14 algorithms. Prior literature validating algorithms in administrative claims across other cancer types report both sensitivity and specificity ranging from as low as 73% to as high as 95%. This was used as a benchmark for defining algorithm success.
Results:
There were 784 STS cases and 249,062 non-STS cancer controls eligible for analysis. Requiring at least two claims with an ICD-CM diagnosis code for STS achieved a sensitivity of 67% but had a specificity of 72%. Algorithms that required NCCN-recommended systemic treatment for STS improved the specificity to over 90% but dropped the sensitivity to below 20%. Other combinations of diagnostic tests, symptoms, and procedures did not improve performance.
Conclusions:
The algorithms tested in this study sample did not achieve sufficient performance and suggest the ability to accurately identify the STS population in administrative data is problematic. Difficulties are likely due to the origin of STS in a variety of locations, the non-specific symptoms of STS, and the common diagnostic tests recommended to diagnose the disease. Future research applying machine learning to examine timing and patterns of variables that comprise the diagnostic process may further investigate the ability to accurately identify STS cases in claims databases.
