CADDIE

CADDIE is an artificial intelligence-driven, cloud-based software designed to enhance colorectal polyp detection during standard white-light colonoscopy by providing real-time visual and auditory cues to gastroenterologists. Its methodology relies on analyzing live endoscopy video feeds to identify potential lesions, aiming to improve adenoma detection rates without increasing unnecessary resections.

CADDIE is a cloud-based, artificial intelligence software system intended to aid gastroenterologists in the real-time detection of suspected colorectal polyps during standard white-light colonoscopy procedures. It functions by interfacing with the endoscopic video processor, analyzing the live video feed through its AI algorithms to identify visual characteristics consistent with colonic mucosal lesions like polyps and adenomas. When a potential lesion is detected within the endoscope's field of view, CADDIE superimposes a green square marker on the video display, accompanied by a short, low-volume sound, to draw the clinician's attention. The system also incorporates a "Cecum AI" feature, which, upon photo-documentation of cecal landmarks by the clinician, provides a reminder to check the activation status of the primary polyp detection function; CADDIE is designed for patients aged 45 and over undergoing screening and surveillance endoscopic mucosal evaluations and is not intended to replace the clinician's full evaluation or primary diagnostic interpretation. Non-clinical bench testing of CADDIE demonstrated high efficacy, with an object-level true positive rate for polyp detection of 98.27% and a frame-level false positive rate of 1.54%, alongside minimal video delay and image degradation. The Cecum AI component also showed strong performance with a frame-level true positive rate of 94.05%. A prospective, multi-center, randomized controlled clinical trial involving 841 patients confirmed CADDIE's clinical benefits, showing a statistically significant 33% increase in adenomas per colonoscopy (APC) compared to standard colonoscopy (0.82 vs 0.62) and a significant improvement in adenoma detection rate (ADR) from 35.9% to 42.9%. Importantly, the system demonstrated non-inferiority in positive percent agreement, indicating no increase in unnecessary biopsies despite heightened detection rates, and increased detection of both diminutive (≤5 mm) and large (≥10 mm) adenomas, with no CADDIE-related adverse events reported.

AI Regulatory Concerns

Powered by

Jev reads the full FDA letter and returns calibrated concern probabilities per regulatory category. These are AI estimates, not regulatory advice.

• Q1. How does the cloud-based architecture of CADDIE impact latency and real-time performance in clinical settings with varying network conditions?; What specific AI model architectures are employed for polyp detection and Cecum AI, and how were they optimized to achieve the reported true positive and false positive rates?; Beyond adenoma detection rate (ADR) and adenomas per colonoscopy (APC), what is the impact of CADDIE on the detection rates of specific challenging polyp morphologies (e.g., flat, depressed lesions) or those in difficult-to-visualize colonic segments?; How does the "Cecum AI" reminder feature specifically influence endoscopist workflow and adherence to quality metrics like cecal intubation documentation, beyond just polyp detection activation status?; What were the key characteristics of the 415 polyps in the development dataset confirmed by resection or photo-documentation but lacking histopathology, and how might their inclusion have biased the training of the AI algorithm?