Development of an Algorithm to Classify Colonoscopy Indication from Coded Health Care Data

Abstract

Introduction: Electronic health data are potentially valuable resources for evaluating colonoscopy screening utilization and effectiveness. The ability to distinguish screening colonoscopies from exams performed for other purposes is critical for research that examines factors related to screening uptake and adherence, and the impact of screening on patient outcomes, but distinguishing between these indications in secondary health data proves challenging. The objective of this study is to develop a new and more accurate algorithm for identification of screening colonoscopies using electronic health data.

Methods: Data from a case-control study of colorectal cancer with adjudicated colonoscopy indication was used to develop logistic regression-based algorithms. The proposed algorithms predict the probability that a colonoscopy was indicated for screening, with variables selected for inclusion in the models using the Least Absolute Shrinkage and Selection Operator (LASSO).

Results: The algorithms had excellent classification accuracy in internal validation. The primary, restricted model had AUC= 0.94, sensitivity=0.91, and specificity=0.82. The secondary, extended model had AUC=0.96, sensitivity=0.88, and specificity=0.90.

Discussion: The LASSO approach enabled estimation of parsimonious algorithms that identified screening colonoscopies with high accuracy in our study population. External validation is needed to replicate these results and to explore the performance of these algorithms in other settings.

Keywords: LASSO; ROC; classification; cohort identification; colonoscopy; data use and quality; health information technology; screening.

Figure 1.

Legend: Area under the Receiver Operating Characteristics (ROC) Curves for the Restricted and Extended Algorithms. The primary, restricted algorithm did not include ICD-9 screening V-codes (V76.41, V76.50, V76.51) and the HCPCS preventative examination code (G0344) as candidate variables, whereas the secondary, extended algorithm included these variables.