Evaluation of algorithm development approaches: Development of biomarker panels for early detection of colorectal lesions

Abstract

Introduction: Colorectal cancer (CRC) is the third most common cancer in the U.S. Early detection of CRC can substantially increase survival rates. Test compliance may be improved by offering a blood-based test option.

Methods: Endoscopy II trial specimens were tested for AFP, CA19-9, CEA, hs-CRP, CyFra 21-1, Ferritin, Galectin-3, and TIMP-1 levels. These biomarkers, as well as patient demographic information (e.g., age, gender), were included in algorithm development. Six statistical methods were utilized to develop algorithms that would discriminate cancer vs. noncancers. Statistical methods included logistic regression, adaptive index modeling, partial least-squares discriminant analysis, feature vector (weighted and unweighted), and random forest. The performance of these algorithms was compared against benchmark criteria established for stool-based tests.

Results: Using several statistical methods, the presence of CRC and high-risk adenomas was detected with an AUCs of at least 0.65-0.76, with a few of models approaching the stool-based tests benchmark performance. Further, common markers were utilized across the different statistical techniques, with model complexities ranging from 3 to 9 markers.

Conclusions: Predictive models identified subjects with CRC and high-risk adenomas with the similar levels of statistical accuracy. Clinical performance differences were minimal across the statistical techniques, although the intuitive interpretations, model complexity, clinical adoption and implementation varied.

Keywords: Algorithm methodologies; Biomarkers; Colon cancer; Early detection.