OBJECTIVES: Roughly 50% of patients with colorectal cancer (CRC) experience recurrence within 2 years of surgical resection. Population-level real-world data often lacks recurrence status, making it difficult to adequately classify patients. Several algorithms have been developed to detect progression, metastasis, and recurrence in structured-health data, but methods and validation vary, thus potentially misclassifying patients and leading to inconsistent results dependent on the algorithm used for outcomes research. This review aimed to identify current approaches to developing and validating structured-health data algorithms used to define disease recurrence in patients with CRC.

METHODS: A targeted literature review was performed in PubMed (MEDLINE) database. Observational or validation studies that utilized administrative claims, EMR/EHR, or cancer registries to define or estimate recurrence status and published within the past 10 years were included if they were the original source of the algorithm. Model type, algorithm components, and performance (specificity, sensitivity, positive predictive value, negative predictive value) were reported. Algorithms were classified by recurrence components, (secondary malignancy codes, treatment/procedure codes, or combination of both) and compared.

RESULTS: A total of 128 papers were identified in PubMed and manual searches and six studies were included in the final review. Two studies did not validate the algorithm against a gold standard (chart review). The specificity was generally high across all validated algorithms (79-100%). Sensitivities were highly variable between and within model type and components (13-83%). The algorithm with the highest sensitivity was a rule-based model with secondary malignancy and treatment/procedure codes.

CONCLUSIONS: Recent algorithms used to estimate CRC recurrence using population-level health data are limited in number and heterogenous. All validated algorithms had high specificities, but sensitivities were variable. The reported low sensitives could result in an underestimation of CRC recurrence. There is a need for standardization of detection algorithms with improved performance to accurately identify this patient population in real-world data.