A novel, noninvasive, multimodal screening test for the early detection of precancerous lesions and colorectal cancers using an artificial intelligence–based algorithm.

3627 Background: Colorectal cancer (CRC) ranks as the second leading cause of cancer-related mortality worldwide, and its incidence is increasing in younger populations. Detection of early-stage CRC and its precursor lesions, such as advanced adenomas (AA), is crucial for successful treatment and reduces CRC-related mortality. Although non-invasive methods for early detection are available and increase screening compliance, their performance with respect to detection of advanced adenomas and early-stage cancers is limited. Here, we describe a novel and non-invasive stool-based approach combining diagnostic biomarkers with an algorithm generated by machine learning/artificial intelligence (ML/AI) resulting in significantly improved diagnostic performance for the detection of not only CRC, but especially precancerous lesions like AA. Methods: Data were generated from a combined cohort of stool samples collected at 10 clinical sites in Europe (COLOFUTURE study) and 21 clinical sites in the US (eAArly DETECT study). The evaluable study cohort consists of 690 subjects, including 78 CRC, 146 AA, 147 with non-advanced adenoma (AD) and 319 normal colonoscopy negative control (NC) subjects (49% female, 51% male, average age 61.8 years). Results were compared with colonoscopy and pathology findings to determine sensitivity and specificity for detection of early-stage CRC and AA vs. AD and NC. Nucleic acids were extracted from stabilized stool samples using a silica bead-based extraction method. Expression of mRNA biomarkers was analyzed utilizing Real-Time PCR. Human hemoglobin was quantified using FIT. The Emerge Quantitative Evolution ML/AI platform was leveraged to develop classifiers capable of distinguishing CRC and AA from AD and NC. Results: Applying the combined approach of non-invasive diagnostic testing with an AI/ML generated algorithm, the sensitivity for detection of CRC overall was 92.3%. In addition, this method enabled AA detection with a sensitivity of 82.2%. Specificity turned out to be 90.1 % (AD+NC). Conclusions: This innovative, non-invasive, multimodal screening strategy based on self-collected stool samples which combines mRNA expression patterns and FIT analysis with an AI/ML-generated algorithm represents a substantial improvement in the effectiveness of non-invasive CRC screening. Such improvements in usability and performance leading to reliable detection of early-stage CRC and precursor lesions, such as advanced adenomas, are required for wide-spread availability and adaption of non-invasive screening tests to accepted clinically relevant usage and to prevent the development of CRC effectively.