The use of artificial intelligence to identify subjects with a positive FOBT predicted to be non-compliant with both colonoscopy and harbor cancer.

Abstract

Subjects with a positive Fecal Occult Blood Test (FOBT) that are non-compliant with colonoscopy are at increased risk for colorectal cancer (CRC). Yet, in clinical practice, many remain non-compliant. To evaluate whether machine learning models (ML) can identify subjects with a positive FOBT predicted to be both non-compliant with colonoscopy within six months and harbor CRC (defined as the "target population"). We trained and validated ML models based on extensive administrative and laboratory data about subjects with a positive FOBT between 2011 and 2013 within Clalit Health that were followed for cancer diagnosis up to 2018. Out of 25,219 included subjects, 9,979(39.6%) were non-compliant with colonoscopy, and 202(0.8%) were both non-compliant and harbored cancer. Using ML, we reduced the number of subjects needed to engage from 25,219 to either 971 (3.85%) to identify 25.8%(52/202) of the target population, reducing the number needed to treat (NNT) from 124.8 to 19.4 or to 4,010(15,8%) to identify 55.0%(52/202) of the target population, NNT=39.7. Machine learning technology may help healthcare organizations to identify subjects with a positive FOBT predicted to be both non-compliant with colonoscopy and harbor cancer from the first day of a positive FOBT with improved efficiency.