Engine Technical
Design Document

The flagship engine. Sentinel Endo analyzes every frame of the colonoscopy video in real time, highlighting suspected polyps with a visual overlay on the endoscopy monitor. The system detects diminutive polyps (≤5mm), flat and sessile lesions, and serrated polyps that account for the majority of missed lesions — acting as an always-alert second observer that never fatigues, never looks away during instrument exchanges, and maintains detection sensitivity regardless of procedure duration or time of day. A meta-analysis of 28 RCTs involving 23,861 patients demonstrated a 20% increase in adenoma detection rate and 55% decrease in adenoma miss rate with AI-assisted colonoscopy. A 2024 Annals of Internal Medicine analysis of 44 RCTs with 36,201 cases confirmed higher average adenomas per colonoscopy (0.98 vs. 0.78, IRD 0.22) and higher ADR (44.7% vs. 36.7%, RR 1.21). The ENDOANGEL model-assisted colonoscopy achieved the highest detection efficacy at 97.8% for colorectal adenomas and polyps across a network meta-analysis of 64 studies. Multicenter prospective RCTs confirmed CADe sensitivity of 95.19% with specificity of 98.44%, with the system specifically excelling at diminutive lesion detection where subgroup analysis showed statistically significant improvement (IRR 1.46; 95% CI 1.19–1.80).

Once a polyp is detected, the critical question is: is it neoplastic? Sentinel Endo's CADx engine analyzes surface patterns, vascular architecture, and pit morphology to predict histology in real time. A systematic review of 20 studies demonstrated pooled CADx sensitivity of 0.94 (95% CI 0.92–0.95), specificity of 0.87 (95% CI 0.83–0.90), and accuracy of 0.91 (95% CI 0.88–0.93) for predicting polyp histology. For diminutive polyps (≤5mm), this enables the ASGE-endorsed resect-and-discard strategy: adenomas are resected and discarded without pathological examination, while hyperplastic polyps in the rectosigmoid are diagnosed and left in situ. A 6-center ASC practice saved $1.2M annually in pathology costs by reducing submissions by 42% with concordance rates above 90%, while a network meta-analysis of 50,834 patients confirmed that real-time CADx-assisted optical diagnostic sensitivity for adenomas was 88% with specificity 78% — performance that meets the ASGE PIVI threshold for clinical adoption.

Quality in colonoscopy is measurable — but rarely measured comprehensively. Engine 03 monitors every procedural quality indicator in real time: withdrawal time (≥6 minutes per international guidelines), mucosal exposure percentage, luminal distension adequacy, blind spot coverage (behind folds, near the ileocecal valve, at flexures), and inspection technique quality. The system generates a per-procedure quality score and identifies specific segments where inspection was inadequate — enabling the endoscopist to re-examine missed areas before completing the procedure. At deployed sites, AI-guided withdrawal achieved 94% mucosal surface coverage versus 76% with unassisted colonoscopy, with an 18% improvement in blind spot coverage.

The single most important quality metric in colonoscopy — adenoma detection rate — varies sevenfold across endoscopists. A community-based cohort of 314,872 colonoscopies by 136 gastroenterologists showed ADR variability ranging from 7.4% to 52.5%, with patients in the highest ADR quintile experiencing 48% lower interval cancer risk. Engine 04 automatically tracks ADR, adenomas per colonoscopy (APC), sessile serrated lesion detection rate (SSLDR), cecal intubation rate, bowel preparation adequacy rate, and withdrawal time for every endoscopist. The system generates monthly performance dashboards, identifies trends, benchmarks against national standards (GIQuIC registry), and flags endoscopists who fall below minimum thresholds — enabling targeted training, peer comparison, and quality improvement. At deployed practices, ADR variability across endoscopists decreased by 48% within 12 months of implementation — the floor was raised, not just the ceiling.

Sentinel Endo extends AI detection to esophagogastroduodenoscopy (EGD), detecting Barrett's esophagus and its dysplastic progression, early esophageal squamous cell carcinoma, gastric intestinal metaplasia, early gastric cancer, and celiac disease–associated villous atrophy. Barrett's surveillance is particularly impactful: dysplastic Barrett's is notoriously difficult to detect endoscopically, with miss rates as high as 40% for high-grade dysplasia on random biopsy protocols. Engine 05 highlights suspicious mucosal areas in real time, guiding targeted biopsies that increase dysplasia detection yield while reducing the total number of biopsies required.

Sentinel Endo automatically scores bowel preparation using the Boston Bowel Preparation Scale (BBPS) with AI precision, segment by segment, in real time. The system achieves 95% agreement with expert BBPS scoring, identifies segments with inadequate preparation, recommends additional washing or suctioning before completing the segment examination, and documents preparation quality for the procedure report — ensuring standardized, auditable, and objective quality reporting. At deployed sites, AI-guided preparation assessment reduced repeat procedures due to inadequate preparation by 30% — because the system identifies inadequately prepared segments during the procedure, enabling real-time correction rather than requiring the patient to return for a repeat colonoscopy.

Withdrawal technique is the single most modifiable factor in adenoma detection rate. Engine 07 monitors withdrawal speed in real time, alerts when the endoscopist is moving too quickly through segments, identifies areas behind folds that have not been adequately inspected, and provides segment-by-segment coverage maps that ensure the entire mucosal surface has been visualized. The CADe system itself acts as a "supervisor" for the endoscopist — as multicenter RCTs demonstrated, the average withdrawal time was longer in the CADe group (430 seconds vs. 421 seconds), likely because the lesion detection function with blue boxes helps the endoscopist focus on suspicious lesions and slow down when indicated. Engine 07 makes this effect explicit and intentional rather than incidental.

Colorectal cancer screening is a population health challenge — and the FIT-positive pathway is where the most lives are saved or lost. A meta-analysis of 10 RCTs on 5,421 FIT-positive patients showed that CADe-assisted colonoscopy increased ADR from 52% to 62% (RR 1.19; 95% CI 1.08–1.31) in this high-prevalence population, with higher detection of both adenomas and serrated lesions. Engine 08 integrates with FIT-based screening programs to optimize the entire pathway: prioritizing FIT-positive patients by predicted advanced neoplasia risk, calculating personalized screening and surveillance intervals based on prior findings (using USMSTF guidelines augmented with AI risk scoring), tracking screening compliance across populations, and identifying patients who are overdue for colonoscopy. The system manages the tension between over-screening (unnecessary procedures with associated costs and risks) and under-screening (missed cancers) by providing personalized, evidence-based interval recommendations.