A Machine-Learned Predictor of Colonic Polyps Based on Urinary Metabolomics

Abstract EN

We report an automated diagnostic test that uses the NMR spectrum of a single spot urine sample to accurately distinguish patients who require a colonoscopy from those who do not.

Moreover, our approach can be adjusted to tradeoff between sensitivity and specificity.

We developed our system using a group of 988 patients (633 normal and 355 who required colonoscopy) who were all at average or above-average risk for developing colorectal cancer.

We obtained a metabolic profile of each subject, based on the urine samples collected from these subjects, analyzed via 1H-NMR and quantified using targeted profiling.

Each subject then underwent a colonoscopy, the gold standard to determine whether he/she actually had an adenomatous polyp, a precursor to colorectal cancer.

The metabolic profiles, colonoscopy outcomes, and medical histories were then analysed using machine learning to create a classifier that could predict whether a future patient requires a colonoscopy.

Our empirical studies show that this classifier has a sensitivity of 64% and a specificity of 65% and, unlike the current fecal tests, allows the administrators of the test to adjust the tradeoff between the two.