Artificial intelligence for kinematic (procedural motion) analysis in gastrointestinal endoscopy: A systematic review

Abstract

BACKGROUND

Artificial intelligence (AI) in gastrointestinal endoscopy has focused on computer-aided detection (CADe) for recognition errors; AI addressing endoscope motion, coverage, workflow, and skill – collectively kinematic analysis – has not been systematically synthesized.

AIM

To map kinematic AI evidence across eight predefined domains and assessed its quality relative to CADe and surgical AI.

METHODS

MEDLINE/PubMed, EMBASE-OVID, Cochrane, and IEEE Xplore were searched following Preferred Reporting Items for Systematic reviews and Meta-Analyses checklist (CRD420261320279). Studies applying AI to kinematic data in gastrointestinal endoscopy were eligible. Risk of bias used Cochrane Risk of Bias 2, Quality Assessment of Diagnostic Accuracy Studies-2, and PROBAST + AI; certainty of evidence used GRADE.

RESULTS

Fifty-eight studies were included. Evidence generation was asymmetric: (1) Randomized controlled trial ratio approximately 5:1 (> 40 vs 8); (2) Patient ratio approximately 4:1 (> 27000 vs approximately 6200); (3) Annual publication ratio approximately 12:1; and (4) ≥ 6 vs 0 regulatory approvals favoring diagnostic AI. Withdrawal speed monitoring and coverage/blind-spot mapping reached moderate GRADE certainty, but all 8 randomized controlled trials were from China and 6 of 8 used the ENDOANGEL platform. The remaining six domains have not reached moderate certainty: (1) Workflow recognition and skill assessment are approaching clinical readiness; and (2) Autonomous navigation, robotic control, simultaneous localization and mapping/three-dimensional reconstruction, and instrument tracking remain at the engineering stage. Comparison with surgical AI indicated that the gap reflects ecosystem-level factors – absence of open datasets, benchmarking challenges, and regulatory pathways for motion-acting AI – rather than technical immaturity.

CONCLUSION

Kinematic AI addresses exposure errors that CADe cannot correct and has shown additive benefit alongside CADe. Two domains are ready for geographically diverse validation; six remain at the engineering stage. Priorities are open datasets, benchmarking infrastructure, and dedicated regulatory pathways for motion-acting AI.

Key Words: Artificial intelligence; Kinematics; Endoscopy; Exposure error; Quality monitoring

Core Tip: Artificial intelligence (AI) in gastrointestinal (GI) endoscopy has been dominated by computer-aided detection (CADe) of lesions – targeting recognition errors – with over 40 randomized controlled trials (RCTs) and multiple regulatory approvals, whereas AI addressing endoscope motion, coverage, and procedural workflow has developed along a separate, largely engineering-focused trajectory. No prior systematic review has mapped the distribution, translational maturity, and certainty of evidence for kinematic AI across GI endoscopy domains. Compared with diagnostic AI, kinematic AI has produced approximately one-fifth the number of RCTs, one-quarter the number of enrolled patients, about one-twelfth the annual publication output, and no standalone regulatory approvals. Only two of eight domains reached moderate GRADE certainty; all eight RCTs were conducted in Chinese centers and six used the ENDOANGEL platform, indicating pronounced geographic and platform concentration. The remaining six domains are at the pre-clinical/engineering stage; the gap with surgical AI is best explained by ecosystem-level factors rather than by technical immaturity alone. The four-arm RCT demonstrated that computer-aided quality and CADe address independent failure modes with additive benefit on adenoma detection rate, supporting integration of kinematic monitoring into existing CADe platforms as the most direct translational pathway. Priority investments include building open kinematic datasets, establishing GI-specific benchmarking challenges analogous to the EndoVis series, conducting colonoscopy three-dimensional coverage RCTs outside China, and defining regulatory pathways for AI that acts on motion rather than on images.

INTRODUCTION

When a gastric cancer is missed during gastroscopy [esophagogastroduodenoscopy (EGD)] or a colorectal adenoma escapes detection during colonoscopy, two distinct failure mechanisms may be responsible: (1) The endoscopist may have seen the lesion but failed to recognize it (a recognition error); and (2) The endoscopist may never have adequately examined the relevant mucosal area (an exposure error)[1,2]. This distinction has important implications for how artificial intelligence (AI) can improve endoscopy.

The past decade of AI development has focused predominantly on recognition errors. Computer-aided detection (CADe) and computer-aided diagnosis (CADx) systems analyze endoscopic images to highlight or classify lesions, with more than 40 randomized controlled trials (RCTs) enrolling over 27000 patients, multiple regulatory approvals, and real-time clinical decision support systems covering esophageal, gastric, and colorectal neoplasia[3-7]. However, these systems cannot detect exposure errors: No polyp detector, however sensitive, can find lesions behind folds that are never visualized or in gastric areas examined too briefly[8].

A separate category of AI applications addresses this gap by analyzing endoscope kinematics. In this systematic review, we define kinematic AI as any AI system whose primary function is to analyze or act upon the motion, spatial trajectory, coverage, procedural phase, or motor performance of the endoscope or its instruments. This definition is deliberately broader than pure motion-sensor analysis; it encompasses: (1) Direct motion/speed measurement (e.g., withdrawal speed quantification, kinematic skill scoring); (2) Spatial coverage estimation derived from visual scene reconstruction [e.g., simultaneous localization and mapping (SLAM)-based three-dimensional (3D) colon coverage, depth-based coverage mapping]; (3) Proxy kinematic measures inferred from anatomical landmark recognition (e.g., WISENSE-style blind-spot monitoring, which uses learned landmark categories to infer whether the endoscope has traversed each mapped region); and (4) Higher-level procedural intelligence derived from motion-aware video understanding [e.g., endoscopic submucosal dissection (ESD) workflow-phase recognition]. We acknowledge that [proxy kinematic measures inferred from anatomical landmark recognition (e.g., WISENSE-style blind-spot monitoring, which uses learned landmark categories to infer whether the endoscope has traversed each mapped region)] and [higher-level procedural intelligence derived from motion-aware video understanding (e.g., ESD workflow-phase recognition)] rely in part on image-classification models; they are included as kinematic AI because the variable they ultimately quantify – whether each anatomical region was adequately traversed, or which procedural motion phase is underway – is kinematic rather than diagnostic in nature. To operationalize this distinction at the algorithmic level for borderline cases, we treat an image-based model as kinematic when its per-frame output is a label of anatomical position, procedural phase, or coverage state rather than a lesion identity, and when those frame-level labels are aggregated over time to yield a procedure-level metric of trajectory, coverage, or workflow. A model whose per-frame output is a lesion characterization is classified as diagnostic even when it shares the same convolutional backbone; the decisive criterion is what the algorithm ultimately quantifies and whether that quantity is integrated across the procedure, not how it processes each individual frame. The implications of alternative, narrower classification boundaries are examined in the Discussion.

Yao et al[9] provided direct proof of concept that this category of AI is clinically non-redundant with CADe: In a four-arm RCT, adding computer-aided quality (CAQ) monitoring to CADe improved adenoma detection rate (ADR) by 26% beyond what CADe alone achieved (30.6% vs 21.3%), demonstrating that kinematic and diagnostic AI address independent error mechanisms with additive benefit.

Despite this importance, kinematic AI has developed across disconnected disciplines – gastroenterology, computer science, and robotics – with no unified synthesis. In the parallel field of minimally invasive surgery, AI-driven kinematic analysis has matured far more rapidly[10-14]: Standardized skill-assessment datasets (JIGSAWS, 2014), annual benchmarking competitions (the EndoVis series, 9 editions from 2015 to 2024), workflow-recognition benchmarks exceeding 91% accuracy (Cholec80), and autonomous robotic procedures in living animals have all been established, leading to Food and Drug Administration (FDA)-cleared surgical AI systems[15-19]. Whether gastrointestinal endoscopy (GIE) can replicate this ecosystem – and what structural barriers have prevented equivalent progress – has not been systematically examined. This systematic review provides the first systematic mapping and certainty assessment of kinematic AI evidence in GIE, quantifies the evidence imbalance relative to diagnostic AI and to surgical AI, and identifies the translational bottlenecks that must be addressed for clinical adoption.

MATERIALS AND METHODS

Protocol and registration

This systematic review followed the Preferred Reporting Items for Systematic reviews and Meta-Analyses 2020 statement (Supplementary material) and was prospectively registered in PROSPERO (CRD420261320279)[20].

Search strategy and information sources

MEDLINE/PubMed, EMBASE-OVID, the Cochrane Central Register of Controlled Trials, and IEEE Xplore were searched from January 2010 to January 2026. Controlled vocabulary and free-text terms covering AI, deep learning (DL), machine learning, reinforcement learning (RL), kinematics, motion analysis, coverage, trajectory, workflow, skill assessment, SLAM, and GIE were combined; the full search strategy is provided in Supplementary material. Reference lists of included studies and relevant reviews were hand-searched to identify additional records.

Eligibility criteria

We included studies that applied AI or machine learning to endoscope kinematic or procedural-motion data in GIE, operationalized through the eight predefined domains above. Studies whose primary output was lesion detection, characterization, or histological prediction (pure CADe/CADx) without a motion, coverage, workflow, or skill component were excluded, as were studies using non-AI methods, studies in non-gastrointestinal (GI) settings, narrative reviews and editorials, conference abstracts without full text, and non-English publications. For hybrid systems (e.g., ENDOANGEL) that combine CADe and CAQ modules, inclusion was determined at the study level based on the primary function evaluated; studies that evaluated only the CADe component were excluded even if produced on a platform whose other modules are kinematic. Because several engineering-oriented domains – such as autonomous navigation, SLAM/3D reconstruction, robotic control, and instrument tracking – publish predominantly through preprint servers and indexed conference proceedings rather than peer-reviewed journals, full-text arXiv preprints were eligible provided they reported sufficient methodological detail for data extraction and formal risk-of-bias appraisal; conference abstracts and preprints lacking such detail were excluded. Eight included studies met this threshold as preprints. These were not exempted from quality appraisal: Each was assessed with PROBAST + AI in the same way as published technical studies, and their non-peer-reviewed status is acknowledged as a limitation.

Study selection, data extraction and outcomes

Two reviewers independently screened titles and abstracts, followed by full-text assessment, with disagreements resolved by discussion. Extracted data included study design, sample size, AI method, domain, outcomes, and performance metrics. Each study was assigned a translational maturity level (TML): (1) A (RCT with patient outcomes); (2) B (prospective non-randomized clinical study); (3) C (retrospective clinical study); (4) D (algorithm with external validation); (5) E (algorithm with internal validation only); and (6) F (simulation, phantom, animal, or dataset study). The TML scheme was not a previously validated instrument but an operational classification defined for the present review; it adapts the principle of staged technology-readiness assessment – articulated for machine-learning systems in the machine learning technology readiness levels framework[21] – to the mixed clinical and engineering literature of kinematic AI, in which conventional clinical-evidence hierarchies cannot accommodate simulation, phantom, and dataset studies. Levels A-C correspond to conventional clinical study designs and levels D-F to pre-clinical engineering maturity. Two reviewers independently assigned a TML to each of the 57 primary studies; initial agreement was reached for 52 of 57 studies (91.2%), and the 5 discrepant assignments – arising at the C/D and E/F boundaries where a study combined clinical and algorithmic elements – were resolved by discussion. The primary outcome was the mapping of AI applications for kinematic analysis in GIE across the eight predefined domains, including the number of RCTs, total patient enrollment, meta-analyses, and regulatory approvals, compared with those of CADe. Secondary outcomes were: (1) GRADE certainty per domain; (2) Risk of bias; (3) Translational maturity distribution; (4) Identification of research gaps defined as domains lacking patient-level clinical outcomes; and (5) Comparison of the kinematic AI ecosystem in GIE with that of surgical AI.

Risk of bias and certainty of evidence assessment

Risk of bias was assessed using the Cochrane Risk of Bias 2 (RoB 2) tool for RCTs[22], the Quality Assessment of Diagnostic Accuracy Studies (QUADAS)-2 tool for diagnostic accuracy studies[23], and PROBAST + AI for technical studies including algorithm-development, simulation, and dataset studies[24]. Certainty of evidence per domain was rated high, moderate, low, or very low using the GRADE approach[25], considering risk of bias, inconsistency, indirectness, imprecision, and publication bias. For engineering-stage domains without clinical outcome data, GRADE was applied to the closest clinically relevant surrogate (e.g., phantom/animal performance), and these domains were explicitly labelled as pre-clinical to avoid conflating low certainty with negative evidence. Applying GRADE to engineering-stage domains requires justification, because GRADE was developed for clinical outcomes. We retained it for two reasons: (1) GRADE explicitly permits rating certainty for surrogate outcomes, provided the surrogate is identified and the rationale for downgrading is made transparent; and (2) Applying one certainty framework across all eight domains allows the clinical-versus-engineering evidence gap to be expressed on a single comparable scale. The pre-clinical label signifies that low certainty in these domains reflects an early translational stage by design, not refuted or negative evidence.

Statistical analysis

Given the heterogeneity of study designs (RCTs, prospective and retrospective clinical studies, algorithm-development and dataset studies, and simulation or phantom experiments), AI methods, outcome measures, and clinical domains, quantitative meta-analysis was not feasible. A narrative synthesis was therefore performed following the Synthesis Without Meta-analysis reporting guideline[26]. Studies were grouped by the eight kinematic domains and synthesized within each domain. The comparative evidence gap between CADe and kinematic AI was reported as a set of individual ratio estimates (number of RCTs, enrolled patients, meta-analyses, regulatory approvals, and annual publication output) rather than as a single composite number, because individual indicators provide more interpretable comparisons to readers.

RESULTS

Study selection and characteristics

The systematic search identified 1524 records. After removal of 460 duplicates, 1064 records were screened by title and abstract, and 384 full-text articles were assessed for eligibility. Three hundred and twenty-six records were excluded (15 narrative reviews, 2 systematic reviews, and 309 studies with insufficient data or outside the eligibility criteria). Fifty-eight studies[1,2,9,27-81] met the final inclusion criteria (Figure 1 and Table 1). Publication volume increased sharply over time, with 95% (55/58) of studies published from 2019 onward and 2024 as the peak year (13 studies). The detailed year-by-year distribution was 1 study in 2010[48], 1 in 2015[38], 1 in 2018[64], 3 in 2019[30,49,59], 7 in 2020[1,2,32,39,41,51,60], 9 in 2021[31,44,45,50,52,61,62,65,69], 6 in 2022[9,53,55,66,67,70], 10 in 2023[28,29,33,43,56,71-74,76], 13 in 2024[34,40,46,47,57,58,63,68,75,77-79,81], 6 in 2025[27,35-37,54,80], and 1 in 2026[42]. Study designs were heterogeneous: (1) 8 RCTs[1,2,9,27,30,31,40,60]; (2) 4 prospective non-randomized clinical studies[28,38,39,55]; (3) 10 retrospective clinical studies[29,33-35,53,56-58,62,63]; (4) 35 algorithm-development, simulation, phantom, or dataset studies[32,36,37,41-52,54,59,61,64-78,80,81]; and (5) 1 prior systematic review[79] retained for cross-reference in the skill-assessment domain. The eight RCTs were concentrated in three domains – withdrawal speed monitoring (n = 4), coverage/blind spot mapping (n = 3), and skill assessment (n = 1) – and all eight originated from Chinese academic centers; six of eight used the ENDOANGEL platform (Wuhan University; the remaining two were conducted by the same group under the WISENSE designation and by an independent Chinese center). Sample sizes ranged from simulation experiments[42] to 170297 images[61] for technical studies, and from 100 colonoscopy videos[56] to 1780 patients[29] for clinical studies. AI methods spanned convolutional neural networks[1,9,27,32,41], transformers[36,52,74], RL[42,43,66-68], neural radiance fields[70,71,73], 3D Gaussian splatting[47,74,75], and state-space models (Mamba)[78,80]. A single AI platform, ENDOANGEL, contributed 8 studies across 3 domains[1,9,27,29-31,40,59], representing the most extensively validated kinematic AI system in the field.

Figure 1  Preferred Reporting Items for Systematic reviews and Meta-Analyses flow diagram.

Table 1 Characteristics of 58 included studies.

Number	Ref.	Domain	Study design	Sample size	AI method	TML	Key finding
1	Gong et al[1], 2020	Withdrawal speed monitoring	RCT	704 patients	CNN (ENDOANGEL)	A	ENDOANGEL ADR 16.3% vs 7.7% control (intention-to-treat)
2	Su et al[2], 2020	Withdrawal speed monitoring	RCT	659 patients	CNN (real-time quality control)	A	ADR 28.9% vs 16.5%; polyp detection rate 383% vs 25.4%
3	Yao et al[9], 2022	Withdrawal speed monitoring	RCT	1076 patients	CNN (ENDOANGEL CAQ)	A	COMBO ADR 30.6% vs CADe-only 21.3% vs CAQ-only 24.5% vs control 148%
4	Liu et al[27], 2025	Withdrawal speed monitoring	RCT	1254 patients	CNN (ENDOANGEL)	A	6-center ADR improved 22.6% to 32.7% in moderate/Low detectors
5	Barua et al[28], 2023	Withdrawal speed monitoring	Prospective	332 patients	CNN (speedometer)	B	No benefit at high-ADR centers (45.8% vs 45.2%); ceiling effect
6	Lu et al[29], 2023	Withdrawal speed monitoring	Retrospective	1780 patients	CNN (ENDOANGEL)	C	AI-assisted arm (CADe + CAQ + combined) eliminated time-of-day quality decline (13.7%-5.7% unassisted; stable 22%-23% with AI)
7	Liu et al[55], 2022	Withdrawal speed monitoring	Prospective	103 colonoscopies	CNN	B	AI-based fold examination quality correlated with ADR (r = 0.852)
8	Lux et al[56], 2023	Withdrawal speed monitoring	Multicenter	100 colonoscopy videos	DL (documentation)	C	AI prototype for automated withdrawal time measurement and photo-documentation (5 centers)
9	Lui et al[57], 2024	Withdrawal speed monitoring	Retrospective	350 videos	DL (real-time)	C	AI real-time monitoring of effective withdrawal time
10	Li et al[58], 2024	Withdrawal speed monitoring	Retrospective	472 videos	YOLOv5	C	Novel withdrawal time indicator based on YOLOv5
11	Wu et al[30], 2019	Coverage/blind spot mapping	RCT	324 patients	CNN (WISENSE)	A	EGD blind spots 5.9% vs 22.5% control
12	Wu et al[31], 2021	Coverage/blind spot mapping	RCT	1050 patients	CNN (ENDOANGEL)	A	5-hospital blind spots 5.38 vs 9.82
13	Chen et al[60], 2020	Coverage/blind spot mapping	RCT	437 patients	DNN (ENDOANGEL)	A	AI reduced blind spots; sedated 3.4% vs 22.4% in EGD
14	Freedman et al[32], 2020	Coverage/blind spot mapping	Algorithm dev	Synthetic + real videos	CNN (depth-based)	E	Coverage quantification; algorithm 0.075 vs expert 0.177 MAE; 93% agreement
15	Wu et al[59], 2019	Coverage/blind spot mapping	Algorithm dev	3170 gastric cancer + 5981 benign images	DNN (ENDOANGEL)	E	EGC detection 92.5% accuracy, 94.0% sensitivity
16	Li et al[61], 2021	Coverage/blind spot mapping	Algorithm dev	170297 images + 5779 videos	DL (IDEA)	D	Real-time 31-site gastric anatomical recognition; 95.3% video accuracy in EGD
17	Cao et al[33], 2023	Workflow recognition	Retro + animal	201026 labeled frames	CNN (AI-Endo)	D	83.5% real-time ESD phase recognition across centers
18	Furube et al[34], 2024	Workflow recognition	Retrospective	94 videos	CNN	C	Esophageal ESD phase recognition 90% accuracy
19	Liu et al[35], 2025	Workflow recognition	Multicenter	195 videos	CNN	C	International 7-center esophageal ESD workflow recognition
20	Chen et al[36], 2025	Workflow recognition	Dataset	66656 frames	Transformer	F	Renji ESD benchmark dataset
21	Biffi et al[37], 2025	Workflow recognition	Dataset	2.7M frames	TCN	F	REAL-colon temporal segmentation benchmark
22	Ward et al[62], 2021	Workflow recognition	Retrospective	50 videos	CNN (LSTM)	D	Automated POEM phase identification
23	Zhang et al[78], 2026	Workflow recognition	Algorithm dev	385 videos	Mamba (SPRMamba)	E	State-space model for ESD phase recognition
24	Nerup et al[38], 2015	Skill assessment	Prospective	10 experienced + 11 trainees in colonoscopy	ML (kinematic)	D	MEI-based kinematic scoring discriminated expert vs trainee
25	Vilmann et al[39], 2020	Skill assessment	Prospective	24 endoscopists	ML (simulation)	D	Computerized assessment validated in simulation
26	Yao et al[40], 2024	Skill assessment	RCT	685 patients	CNN (ENDOANGEL)	A	AI novice miss rate 188% vs control novice 437% vs expert 27.0%
27	Wittbrodt et al[63], 2024	Skill assessment	Retrospective	50 colonoscopies	ML (random forest)	D	ML colonoscopy skill; withdrawal time r = 0.99
281	Cold et al[79], 2024	Skill assessment	Prior systematic review (cross-reference)	13 studies	Various	-	Systematic review of computer-aided colonoscopy competence assessment
29	Martin et al[41], 2020	Autonomous navigation	Animal study	2 pigs; 10 novices	CNN	E	Completion rates 58% direct vs 96% intelligent teleop vs 100% semi-autonomous
30	Hwang et al[42], 2026	Autonomous navigation	Simulation	50 experiments	Supervised DL	F	Autonomous robotic colonoscopy; 90% success rate
31	Corsi et al[43], 2023	Autonomous navigation	Simulation	Simulation	Constrained RL	F	Safe colonoscopy navigation with formal verification
32	Prendergast et al[64], 2018	Autonomous navigation	Phantom	Phantom	ML (localization)	F	Autonomous haustral fold detection for robotic endoscopy
33	Huang et al[65], 2021	Autonomous navigation	Simulation	Simulation	ML (magnetic)	F	Autonomous navigation of magnetic colonoscope
34	Lazo et al[66], 2022	Autonomous navigation	Simulation	Simulation	DL visual servoing	F	Autonomous soft-robot intraluminal navigation
35	Pore et al[67], 2022	Autonomous navigation	Simulation	Simulation	End-to-end RL	F	Deep visuomotor control for colonoscopy
36	Tan et al[68], 2025	Autonomous navigation	Simulation	Simulation	Human-intervention RL	F	Safe navigation via human-intervention-based RL
37	Ma et al[44], 2021	SLAM/3D reconstruction	Algorithm dev	Real + phantom sequences	RNN-SLAM	E	38%-46% drift reduction in 3D colon reconstruction
38	Ozyoruk et al[45], 2021	SLAM/3D reconstruction	Dataset + algo	42700 frames	Self-supervised	E	EndoSLAM dataset; unsupervised depth estimation
39	Bonilla et al[46], 2024	SLAM/3D reconstruction	Algorithm dev	Endoscopy videos	3DGS	E	Gaussian Pancakes for endoscopic reconstruction
40	Wang et al[47], 2024	SLAM/3D reconstruction	Algorithm dev	Endoscopy videos	3DGS + SLAM	E	EndoGSLAM: Real-time dense reconstruction (> 100 fps)
41	Recasens et al[69], 2021	SLAM/3D reconstruction	Algorithm dev	Endoscopy videos	DL (depth + motion)	E	Endo-Depth-and-Motion tracking
42	Shao et al[70], 2022	SLAM/3D reconstruction	Algorithm dev	Endoscopy videos	Self-supervised	E	Self-supervised monocular depth in endoscopy
43	Hayoz et al[71], 2023	SLAM/3D reconstruction	Algorithm dev	Endoscopy videos	DL (pose est)	E	Robust camera pose estimation in endoscopy
44	Azagra et al[72], 2023	SLAM/3D reconstruction	Dataset	96 procedures	SLAM baseline	F	Endomapper dataset of calibrated procedures
45	Bobrow et al[73], 2023	SLAM/3D reconstruction	Dataset	10015 frames	NeRF (2D-3D reg)	E	Colonoscopy 3D dataset with paired depth
46	Shi et al[74], 2023	SLAM/3D reconstruction	Algorithm dev	Colonoscopy videos	NeRF (ColonNeRF)	E	High-fidelity long-sequence colonoscopy reconstruction
47	Elvira et al[75], 2024	SLAM/3D reconstruction	Algorithm dev	Full procedures	CudaSIFT-SLAM	E	Multiple-map SLAM for full procedure mapping
48	Guo et al[80], 2025	SLAM/3D reconstruction	Algorithm dev	Endoscopy videos	NeRF (UC-NeRF)	E	Uncertainty-aware NeRF from sparse endoscopic views
49	Kaleta et al[81], 2025	SLAM/3D reconstruction	Algorithm dev	Endoscopy videos	3DGS (PR-ENDO)	E	Physically based relightable Gaussian Splatting
50	Trovato et al[48], 2010	Robotic control/instrument tracking	Simulation	Simulation	Q-learning	F	First RL-based colon-endoscope-robot locomotion
51	Turan et al[49], 2019	Robotic control/instrument tracking	Simulation	Simulation	Deep RL	F	Learning to navigate endoscopic capsule robots
52	İncetan et al[50], 2021	Robotic control/instrument tracking	Simulation	Simulation	Deep RL (VR-caps)	F	Virtual environment for capsule endoscopy
53	Jha et al[51], 2020	Robotic control/instrument tracking	Dataset	590 frames	DL (segmentation)	F	Kvasir-Instrument: GIE tool-segmentation dataset
54	Ali et al[52], 2021	Robotic control/instrument tracking	Challenge	2531 detection + 643 segmentation frames	Transformer/CNN	F	DL for artefact and disease detection and segmentation
55	Brand et al[53], 2022	Robotic control/instrument tracking	Multicenter retro	580 videos	DL	D	DL model to improve polyp-detection usability
56	Jha et al[54], 2025	Robotic control/instrument tracking	Challenge	Multiple datasets	Various DL	F	Polyp and instrument segmentation benchmark
57	Zhang et al[76], 2022	Robotic control/instrument tracking	Simulation	Simulation	Deep RL	F	RL-based stomach coverage scanning of WCE
58	Ng et al[77], 2024	Robotic control/instrument tracking	Simulation	Simulation	Deep RL	F	RL navigation of tendon-driven flexible robotic endoscope

¹Cold et al[79] (2024) is a prior systematic review retained for cross-reference in the skill-assessment domain; it is not counted within the design-level synthesis (8 randomized controlled trials + 4 prospective + 10 retrospective + 35 algorithm/simulation/phantom/dataset studies = 57 primary studies).

Note on heterogeneous sample size units: Because the included studies span clinical and engineering literatures, the sample size column reports the unit native to each study type, and these units are not directly comparable as indicators of translational maturity. For interpretation: (1) ‘Patients’ denotes prospective clinical enrolment in randomized controlled trials or prospective cohorts [translational maturity level (TML) A and B, direct clinical evidence]; (2) ‘Videos’, ‘colonoscopies’, or ‘cases’ denote clinical material analysed retrospectively offline (TML C and D, clinical-but-non-prospective data); (3) ‘Images’, ‘frames’, or annotated-dataset counts denote algorithm-development studies with no patient-level outcomes (TML D and E, engineering-stage evidence); and (4) ‘Phantom’, ‘simulation’, or ‘animal’ entries denote pre-clinical evaluation furthest from clinical translation (TML F). Readers should therefore weigh entries by reading the sample size column together with the TML column. A large frame or image count does not imply greater clinical maturity than a smaller patient cohort, and conversely a small patient enrolment may carry greater translational weight than a multi-million-frame dataset. TML: Translational maturity level. A: Randomized controlled trial with patient outcomes; B: Prospective non-randomized clinical study; C: Retrospective clinical study; D: Algorithm development with external validation; E: Algorithm development with internal validation only; F: Simulation/phantom/dataset study; EGD: Esophagogastroduodenoscopy; ESD: Endoscopic submucosal dissection; ADR: Adenoma detection rate; CNN: Convolutional neural network; DL: Deep learning; ML: Machine learning; RL: Reinforcement learning; NeRF: Neural radiance fields; 3DGS: Three-dimensional gaussian splatting; SLAM: Simultaneous localization and mapping; POEM: Peroral endoscopic myotomy; WCE: Wireless capsule endoscopy; MEI: Magnetic endoscopic imaging; RCT: Randomized controlled trial; AI: Artificial intelligence; CAQ: Computer-aided quality; CADe: Computer-aided detection; IDEA: Intelligent detection endoscopic assistant; TCN: Temporal convolutional network; LSTM: Long short-term memory; RNN: Recurrent neural network; MAE: Mean absolute error; GIE: Gastrointestinal endoscopy; 3D: Three-dimensional.

Evidence imbalance between diagnostic and kinematic AI

Across the indicators pre-specified in methods (Table 2), kinematic AI has produced substantially less clinical evidence than CADe/CADx despite emerging in a comparable timeframe (2015-2019). Specific ratios were approximately 5:1 for RCTs (> 40 vs 8), 4:1 for enrolled patients (> 27000 vs approximately 6200), > 10:1 for systematic reviews/meta-analyses, and approximately 12:1 for annual PubMed publication output (approximately 150 vs approximately 12 studies per year, 2020-2024). CADe/CADx has received ≥ 6 regulatory approvals (FDA/Conformité Européenne/Ministry of Food and Drug Safety), compared with no standalone regulatory approvals for kinematic AI. The certainty gap is equally pronounced: CADe for ADR improvement carries high GRADE certainty from more than 20 concordant RCTs, whereas only 2 of 8 kinematic domains (withdrawal monitoring, coverage mapping) reached moderate certainty. Within kinematic AI itself, 3 domains accounted for all 8 RCTs (withdrawal speed monitoring, n = 4; coverage/blind spot mapping, n = 3; skill assessment, n = 1), while 5 domains had zero patient-level clinical data despite technical progress.

Table 2 Multi-dimensional evidence imbalance between diagnostic artificial intelligence (computer-aided detection/ computer-aided diagnosis) and kinematic artificial intelligence (computer-aided quality) in gastrointestinal endoscopy.

Evidence indicator	Diagnostic AI (CADe/CADx)	Kinematic AI (CAQ)	Ratio or qualitative gap
RCTs published	> 40	8	Approximately 5:1
Patients enrolled in RCTs	> 27000	Approximately 6200	Approximately 4:1
Systematic reviews/meta-analyses	> 10	0 (this is the first)	> 10:1
Regulatory approvals (FDA/CE/MFDS)	≥ 6 devices	0 standalone	Qualitative
GRADE certainty for primary outcome	High (ADR improvement)	Moderate (2 of 8 domains)	2 levels lower
Domains with ≥ 1 clinical trial	3/3 (detection, classification, characterization)	2/8 (withdrawal speed, coverage)	Qualitative
Domains with zero patient-level data	0/3	6/8 (75%)	Qualitative
Annual research output (2020-2024 PubMed)	Approximately 150 studies/year	Approximately 12 studies/year	Approximately 12:1
Industry investment in device development	Multiple companies (Medtronic, Fujifilm, NEC, Olympus, etc.)	Single academic platform dominates (ENDOANGEL)	Qualitative

Ratios are reported per indicator. No single composite number is used because indicators differ in magnitude (approximately 5:1 for randomized controlled trials, 4:1 for enrolled patients, approximately 12:1 for annual publication output, and qualitatively larger gaps for regulatory approvals and industry participation). CADe: Computer-aided detection; CADx: Computer-aided diagnosis; CAQ: Computer-aided quality; ADR: Adenoma detection rate; FDA: Food and Drug Administration; CE: Conformité Européenne; MFDS: Korean Ministry of Food and Drug Safety; AI: Artificial intelligence; RCT: Randomized controlled trial.

Clinically validated domains: Withdrawal monitoring and coverage mapping

Withdrawal speed monitoring (10 studies, 4 RCTs, > 5900 patients; GRADE: Moderate): ENDOANGEL-based systems provide real-time feedback on colonoscope withdrawal speed and mucosal-fold inspection. The Yao et al[9] four-arm trial established that CAQ and CADe have additive effects on ADR (COMBO 30.6% vs CAQ-only 24.5% vs CADe-only 21.3% vs control 14.8%). A six-center RCT (Liu et al[27], 2025; 1254 patients) confirmed an ADR improvement from 22.6% to 32.7% in moderate-detectors and low-detectors. Barua et al[28] (2023) found no benefit at high-baseline-ADR centers (approximately 45%), establishing a ceiling effect. In the Lu et al[29] (2023) retrospective study, AI-assisted colonoscopy was associated with elimination of the time-of-day decline in quality (unassisted ADR fell from 13.7% in the morning to 5.7% in the afternoon, whereas AI-assisted ADR remained at 22%-23%). We note explicitly that the AI-assisted arm in that study combined three subgroups (CADe, CAQ, and CADe + CAQ). The observed effect therefore cannot be attributed to kinematic AI alone but reflects the contribution of the combined AI intervention, and we interpret it accordingly in the Discussion. GRADE was downgraded from high to moderate owing to structurally unavoidable endoscopist unblinding in real-time AI trials and to single-platform dominance (ENDOANGEL).

Coverage/blind-spot mapping (6 studies, 3 RCTs, > 1800 patients; GRADE: Moderate): In this review, blind spots during EGD are defined as anatomical regions of the upper GI tract that endoscopy guidelines recommend be photographically documented (e.g., 22 or 31 predefined gastric sites in the Japanese and Chinese systematic-examination protocols) but that were not actually visualized during a given procedure. The WISENSE system (Wu et al[30], 2019) reduced EGD blind-spot rates from 22.5% to 5.9% (P < 0.001), and a similar benefit was confirmed in a five-hospital RCT (5.38 vs 9.82 mean blind spots)[31]. For colonoscopy, Freedman et al[32] (2020) developed depth-based coverage quantification. This method estimates per-pixel colonic wall depth from the monocular video feed and integrates it over the withdrawal phase to compute the fraction of mucosa that passed within the camera’s field of view. The system achieved 93% agreement with expert coverage labeling[32], although no colonoscopy-coverage RCT yet exists. GRADE was moderate owing to unblinding and to the absence of clinical trials specifically targeting colonoscopy coverage.

Emerging domains approaching clinical readiness

Workflow recognition (7 studies; GRADE: Low): ESD phase recognition is the automated identification, from endoscopic video, of the canonical procedural steps of ESD (marking, submucosal injection, mucosal incision, submucosal dissection, and hemostasis). AI-Endo (Cao et al[33], 2023) achieved 83.5% real-time ESD phase recognition with external validation on 15 independent cases. Furube et al[34] (2024) extended this to esophageal ESD with 90% phase accuracy, and Liu et al[35] (2025) performed the first international multicenter esophageal ESD workflow-recognition study with an educational RCT component. Two 2025 benchmark datasets (Renji ESD, 66656 frames; REAL-Colon, 2.7 million frames) have been released, suggesting growing methodological maturity[36,37]. GRADE was low because no prospective patient-outcome trial has been conducted and because phase-recognition accuracy, measured against expert annotation, is a surrogate rather than a clinical outcome.

Skill assessment (5 studies; GRADE: Very low): Kinematic skill scoring based on magnetic endoscopic imaging of the colonoscope tip trajectory (Nerup et al[38], 2015; Vilmann et al[39], 2020) discriminated expert from trainee colonoscopists. AI-assisted novice colonoscopy (Yao et al[40], 2024) reached adenoma miss rates comparable to experts in a tandem RCT (AI-novice 18.8% vs control novice 43.7% vs expert 27.0%). However, the combined paradigm of DL applied to concurrent kinematic-sensor data – DL-kinematic sensor integration, defined here as the joint analysis of video, force/torque, electromagnetic tracker, or robotic-encoder signals by a deep neural network to derive automated skill metrics, as popularized by JIGSAWS in robotic surgery[10] – has not been attempted for flexible GI endoscopy. GRADE was very low owing to the single available RCT, the absence of DL-kinematic sensor integration, and the lack of a standardized kinematic-measurement infrastructure for flexible endoscopes. The clinical importance of optimal technique selection and standardised skill assessment is established in adjacent surgical evidence. Comparative data after totally laparoscopic total gastrectomy show that the choice of esophagojejunal anastomotic technique directly influences anastomotic complication rates, with strictures concentrated in the more technically demanding circular-stapler arm[82] and in colorectal cancer surgery anastomotic leakage confers an approximately fivefold increase in 30-day mortality, with additional perioperative risk factors further amplifying complication and mortality risk in geriatric patients[83]. Translating this principle to flexible endoscopy provides a direct clinical rationale for AI-based kinematic skill metrics in endoscopic training, where standardised objective motion assessment is a plausible upstream intervention for the procedural failure modes that determine patient survival, leakage rates, and overall complication burden.

Engineering-stage (pre-clinical) domains

The four domains below consist predominantly of algorithm-development, simulation, phantom, or dataset studies. Their principal outputs are mathematical accuracy, phantom performance, or formally verified safety properties, rather than patient outcomes. In the IDEAL framework and the FDA Software-as-a-Medical-Device (SaMD) Total Product Lifecycle, these studies sit at stages 0-1 (idea, development, pre-clinical) by design. Accordingly, we describe them here as pre-clinical and map their translational maturity rather than evaluate them against a patient-outcome endpoint, which would constitute a category error.

Autonomous navigation (8 studies; GRADE: Very low – pre-clinical): Technical progress has been substantial, ranging from convolutional neural network-based semi-autonomous magnetic colonoscopy in pigs[41] to full-procedure autonomous robotic colonoscopy simulation[42]. RL dominates the recent literature (Proximal Policy Optimization, Soft Actor-Critic, constrained RL with formal safety verification)[43]. All studies to date used simulation, phantom, or animal models; no first-in-human clinical trial and no regulatory pathway has been established. GRADE is very low; the domain is at an engineering stage.

SLAM and 3D reconstruction (13 studies; GRADE: Very low – pre-clinical): This is the most technically active domain, with rapid methodological progression from classical SLAM to recurrent neural network (RNN)-SLAM (38%-46% drift reduction)[44], self-supervised depth estimation[45], neural radiance fields, and 3D Gaussian Splatting (> 100 fps at approximately two minutes of training)[46,47]. Real-time clinical deployment is now technically feasible, yet no study has prospectively tested reconstruction outputs in patient endoscopy. The bottleneck is not computational speed but the absence of validated clinical metrics linking coverage completeness to patient outcomes.

Robotic control and instrument tracking (9 studies; GRADE: Very low – pre-clinical): Robotic capsule control ranged from Q-learning[48] to deep RL controllers[49,50], all in simulation. For GI instrument tracking, the principal public dataset is Kvasir-Instrument (590 images)[51], compared with thousands of annotated frames in laparoscopic EndoVis challenges. No ESD-knife tracking, endoscopic retrograde cholangiopancreatography instrument segmentation, or hemostasis-clip detection study was identified[51-54]. GRADE is very low; the domain is at an engineering stage.

Risk of bias assessment

Using RoB 2, all eight RCTs received a “high risk” judgment for Domain 2 (deviations from intended interventions) because endoscopist blinding to a real-time AI display is structurally impossible. Because this limitation is inherent to all real-time AI-assisted trials rather than a methodological flaw of individual studies, the overall judgment was rated “some concerns”, consistent with prior systematic reviews of CADe. Future RCT designs could potentially mitigate this structural unblinding by employing sham AI interfaces (non-functional overlays indistinguishable from active feedback to the endoscopist) or by relying on retrospective kinematic analysis of blinded video recordings as the primary outcome measure rather than on real-time intervention; either approach would shift the bias profile from Domain 2 (deviations from intended interventions) toward Domain 4 (measurement of outcome), where blinded outcome assessment is methodologically feasible. Randomization and outcome measurement were adequate in 7 of 8 trials; the remaining trial (Yao et al[40], 2024) had some concerns for outcome measurement owing to a possible learning effect in the tandem design (Table 3)[1,2,9,27,30,31,60]. QUADAS-2 assessment of 12 diagnostic accuracy studies showed patient-selection concerns in 67% of studies (convenience sampling), while reference standard and flow/timing were generally low risk (Table 4)[28,29,33-35,38,39,55,57,58,61,62]. PROBAST + AI assessment of 38 technical studies revealed overall high risk of bias in 25 studies (66%), primarily driven by Domain 4 (Analysis; 21/38, 55%) owing to the absence of external validation. Domain 1 (Participants) was rated high in 15 studies (39%), predominantly those conducted entirely in simulation or phantom environments; Domain 2 (Predictors) was uniformly low risk, as deep-learning architectures extract features automatically (Supplementary Table 1)[32,36,37,41-54,56,59,63-81].

Table 3 Risk of bias assessment (Cochrane Risk of Bias 2 for randomized controlled trials).

Ref.	D1: Randomization	D2: Deviations from intended interventions	D3: Missing outcome data	D4: Measurement of outcome	D5: Selection of reported result	Overall
Gong et al[1], 2020	Low (computer-generated, sealed envelopes)	High (real-time AI display; inherently unblindable)	Low (ITT; 2.1% attrition)	Low (histopathology reference; blinded pathologist)	Low (pre-registered; all outcomes reported)	Some concerns (unblinding inevitable)
Su et al[2], 2020	Low (computer-generated randomization)	High (real-time AI overlay visible to endoscopist)	Low (ITT; < 3% missing)	Low (histopathology reference; blinded assessment)	Low (pre-registered; primary/secondary endpoints reported)	Some concerns (unblinding inevitable)
Yao et al[9], 2022	Low (computer-generated; 4-arm parallel design)	High (CAQ display visible; endoscopist aware of AI arm)	Low (ITT; complete follow-up)	Low (histopathology; blinded pathologist)	Low (pre-registered 4-arm design; all outcomes reported)	Some concerns (unblinding inevitable)
Wu et al[30], 2019	Low (computer-generated; concealed allocation)	High (WISENSE display visible during EGD)	Low (complete data; no attrition)	Low (blind spot count by independent reviewer)	Low (pre-registered; all endpoints reported)	Some concerns (unblinding inevitable)
Chen et al[60], 2020	Low (computer-generated randomization)	High (AI display visible during EGD)	Low (< 2% missing; ITT)	Low (blind spot assessment by blinded reviewer)	Low (all pre-specified outcomes reported)	Some concerns (unblinding inevitable)
Wu et al[31], 2021	Low (centralized randomization; 5-hospital)	High (real-time AI feedback visible to endoscopist)	Low (ITT; minimal attrition)	Low (blind spot mapping by independent reviewer)	Low (pre-registered multicenter; all outcomes reported)	Some concerns (unblinding inevitable)
Yao et al[40], 2024	Low (computer-generated; tandem design)	High (AI assistance visible during colonoscopy)	Low (complete follow-up; tandem design ensures paired data)	Some concerns (miss rate depends on tandem sequence; learning effect possible)	Low (pre-registered tandem RCT; all outcomes reported)	Some concerns (unblinding + potential tandem sequence effect)
Liu et al[27], 2025	Low (centralized randomization; 6-center)	High (real-time AI quality control visible)	Low (ITT; < 3% attrition across 6 centers)	Low (histopathology; blinded pathologist)	Low (pre-registered multicenter; all outcomes reported)	Some concerns (unblinding inevitable)

RCT: Randomized controlled trial; AI: Artificial intelligence; ITT: Intention-to-treat; CAQ: Computer-aided quality; EGD: Esophagogastroduodenoscopy.

Table 4 Risk of bias assessment (Quality Assessment of Diagnostic Accuracy Studies-2 for diagnostic accuracy studies).

Ref.	Patient selection	Index test	Reference standard	Flow and timing	Overall
Barua et al[28], 2023	Low (consecutive patients in implementation trial)	Low (AI speedometer threshold pre-defined)	Low (withdrawal time by independent timer)	Low (all patients received same assessment)	Low (all domains low risk)
Lu et al[29], 2023	Unclear (single-center convenience sample; time-of-day subgroups)	Low (ENDOANGEL AI output pre-specified)	Low (histopathology for ADR; blinded pathologist)	Low (all patients received colonoscopy and pathology)	Unclear (patient selection concern)
Liu et al[55], 2022	Unclear (single-center; convenience sampling)	Low (AI withdrawal assessment pre-defined)	Low (expert endoscopist consensus)	Low (all patients assessed by both AI and experts)	Unclear (patient selection concern)
Lui et al[57], 2024	Unclear (retrospective video selection; single center)	Low (AI withdrawal monitoring pre-specified)	Low (manual review by experienced endoscopist)	Low (all videos analyzed by both methods)	Unclear (patient selection concern)
Li et al[58], 2024	Unclear (retrospective convenience sample)	Low (YOLOv5 threshold pre-specified)	Low (manual withdrawal time measurement)	Low (all videos assessed)	Unclear (patient selection concern)
Li et al[61], 2021	Unclear (single-center convenience sample)	Low (IDEA system output pre-defined)	Low (expert annotation of anatomical landmarks)	Low (all patients received same EGD protocol)	Unclear (patient selection concern)
Cao et al[33], 2023	Unclear (retrospective video collection; single center + animal)	Low (AI-endo phase output pre-specified)	Low (expert surgeon frame-level annotation)	Low (all videos fully annotated)	Unclear (patient selection concern)
Furube et al[34], 2024	Unclear (retrospective single-center video selection)	Low (AI phase recognition threshold pre-defined)	Low (expert endoscopist annotation)	Low (all videos received complete annotation)	Unclear (patient selection concern)
Liu et al[35], 2025	Low (multicenter; prospective + retrospective cohorts)	Low (AI workflow recognition pre-specified)	Low (expert panel annotation consensus)	Low (all cases assessed by AI and experts)	Low (all domains low risk)
Ward et al[62], 2021	Unclear (retrospective single-center video selection)	Low (AI POEM phase output pre-defined)	Low (expert surgeon phase annotation)	Low (all videos fully annotated)	Unclear (patient selection concern)
Nerup et al[38], 2015	High (small convenience sample; 12 endoscopists only)	Unclear (MEI kinematic thresholds derived from same cohort)	Low (expert/trainee classification pre-defined)	Low (all endoscopists assessed)	High (small sample + index test concern)
Vilmann et al[39], 2020	High (small convenience sample; 27 endoscopists in simulation)	Unclear (computerized metrics derived from training data overlap)	Low (GRS expert assessment as reference)	Low (all participants completed assessment)	High (small sample + index test concern)

AI: Artificial intelligence; ADR: Adenoma detection rate; POEM: Peroral endoscopic myotomy; MEI: Magnetic endoscopic imaging; IDEA: Intelligent detection endoscopic assistant; EGD: Esophagogastroduodenoscopy; GRS: Global Rating Scale.

GRADE certainty of evidence

Table 5 summarizes the GRADE assessment across all eight domains. Moderate certainty was achieved only for withdrawal monitoring and coverage mapping (downgraded from high owing to unblinding and single-platform dominance). Workflow recognition was rated low (no patient-outcome trial). The remaining five domains were rated very low: (1) Skill assessment (downgraded for indirectness because its only RCT used adenoma miss rate rather than a kinematic outcome, and for imprecision) and the four engineering-stage domains (autonomous navigation, SLAM/3D reconstruction, robotic control, and instrument tracking); and (2) Downgraded for absence of clinical trials, indirectness from simulation/phantom settings, and imprecision from small sample sizes. The contrast with diagnostic AI – for which GRADE certainty for CADe improving ADR is high, based on more than 20 concordant RCTs – illustrates the scale of the evidence imbalance.

Table 5 GRADE certainty of evidence assessment across 8 kinematic artificial intelligence domains.

Domain	n	Risk of bias	Inconsistency	Indirectness	Imprecision	Pub. bias	GRADE
Withdrawal speed	10	Not serious1	Not serious	Not serious	Not serious	Unlikely	Moderate
Coverage mapping	6	Not serious1	Not serious	Not serious	Serious2	Unlikely	Moderate
Workflow recognition	7	Serious3	Not serious	Serious4	Serious	Undetected	Low
Skill assessment	5	Serious	Serious5	Very serious6	Very serious	Undetected	Very low
Navigation (pre-clinical)	8	Serious	Not serious	Very serious7	Very serious	Undetected	Very low
SLAM/3D (pre-clinical)	13	Serious8	Not serious	Very serious7	Very serious	Suspected9	Very low
Robotic control (pre-clinical)	5	Serious	Serious	Very serious7	Very serious	Undetected	Very low
Instrument tracking (pre-clinical)	4	Serious	Not serious	Very serious7	Very serious	Undetected	Very low

¹Not serious despite inherent unblinding, which is structurally unavoidable in real-time artificial intelligence trials and does not selectively affect kinematic studies.

²No colonoscopy coverage randomized controlled trial; small total sample.

³Only 1 randomized controlled trial component; predominantly retrospective.

⁴Phase-recognition accuracy is a surrogate, not a patient outcome.

⁵Only 2 ML-based studies with different methods.

⁶No study uses DL–kinematic sensor integration for flexible endoscopy.

⁷All studies in simulation, phantom, or animal models only.

⁸37% lacked external validation.

⁹Positive-result publication bias suspected given absence of negative/null technical studies.

SLAM: Simultaneous localization and mapping; 3D: Three-dimensional.

Sensitivity analysis under a narrower kinematic definition

To address the concern that our operational definition of kinematic AI is broader than a pure motion-signal-based definition, we conducted a pre-specified sensitivity analysis using a four-tier hierarchical taxonomy (Figure 2)[9]. Tier 1 (true kinematic AI) comprised studies that analyze direct motion signals from sensors or encoders (withdrawal-speed monitoring, magnetic endoscopic imaging-based skill assessment, and encoder-based robotic control). Tier 2 (proxy kinematic inference) comprised studies that derive motion variables from monocular video (instrument tracking, SLAM/3D reconstruction, and autonomous navigation). Tier 3 (workflow intelligence) comprised ESD phase-recognition studies, which use video-frame classification to infer procedural-motion phase. Tier 4 [procedural quality-assurance (QA) systems] comprised landmark-classification studies (WISENSE blind-spot mapping and gastric anatomical landmark recognition) in which kinematic adequacy is inferred from whether predefined anatomical regions were photographed rather than from explicit motion variables.

Figure 2 Hierarchical taxonomy of artificial intelligence domains in gastrointestinal endoscopy. Four tiers are arranged along the horizontal axis by their proximity to direct motion-signal analysis: (1) Tier 1 [true kinematic artificial intelligence (AI), direct motion signal]; (2) Tier 2 (proxy kinematic inference, vision-derived motion); (3) Tier 3 (workflow intelligence, phase recognition); and (4) Tier 4 (procedural quality-assurance systems, landmark classification). The vertical axis represents translational maturity (regulatory approval, clinical randomized controlled trial, prospective/retrospective, algorithm/external validation, engineering/simulation). The dashed-line box denotes the strict kinematic definition (Tiers 1-2) used in the pre-specified sensitivity analysis. Boxes with bold black borders mark domains with clinical-trial (randomized controlled trial) evidence. Adjacent AI ecosystems – surgical AI (left) and diagnostic AI/computer-aided detection/computer-aided diagnosis (right) – are shown for translational reference. AI: Artificial intelligence; FDA: Food and Drug Administration; RCT: Randomized controlled trial; SLAM: Simultaneous localization and mapping; 3D: Three-dimensional; ADR: Adenoma detection rate; CADe: Computer-aided detection; CADx: Computer-aided diagnosis; CAQ: Computer-aided quality; CE: Conformée Européenne; ESD: Endoscopic submucosal dissection; MFDS: Ministry of Food and Drug Safety; QA: Quality-assurance.

Under a strict kinematic definition restricted to Tiers 1-2, four of the eight original domains were retained, comprising 5 RCTs (4 withdrawal-speed and 1 skill-assessment) and approximately 4400 enrolled patients; the kinematic-to-CADe RCT ratio became approximately 8:1 (vs 5:1 under the broad definition), and the patient ratio became approximately 6:1 (vs 4:1). Moderate GRADE certainty was retained only for withdrawal-speed monitoring; coverage/blind-spot mapping was reclassified to Tier 4 and would therefore be excluded under the strict definition. Restricting still further to Tier 1 alone (3 domains) yielded an RCT ratio of approximately 13:1. Across all three definitions – broad (Tiers 1-4), strict (Tiers 1-2), and strictest (Tier 1 only) – the core conclusion of a substantial multi-fold evidence gap between kinematic and diagnostic AI was preserved, indicating that the principal finding of this review is robust to the boundary of the kinematic definition.

DISCUSSION

Principal findings

This systematic review provides four principal findings. First, kinematic AI and diagnostic AI address independent error mechanisms with additive clinical benefit but receive markedly asymmetric research investment. The multi-fold evidence imbalance (RCT ratio approximately 5:1, patient ratio approximately 4:1, meta-analysis ratio > 10:1, annual publication output approximately 12:1, regulatory approvals ≥ 6 vs 0) quantifies this asymmetry across interpretable dimensions. For device manufacturers, the direct practical implication is that integrating kinematic monitoring into existing CADe platforms – rather than further sensitivity optimization of detection modules – is a tractable translational pathway for reducing miss rates.

Second, within kinematic AI, a two-to-three-level GRADE gap separates the two validated domains from the six non-validated domains, and the dominant explanation for this gap is the absence of translational infrastructure rather than algorithmic immaturity. SLAM achieves over 100 fps real-time rendering; autonomous robotic controllers complete simulated colonoscopies; ESD workflow recognition reaches 83%-90% accuracy against expert annotation. These demonstrate technical feasibility. Their translational bottlenecks are: (1) The absence of validated clinical-outcome metrics linking kinematic measures (coverage completeness, motion smoothness, phase-transition timing) to patient outcomes beyond ADR; (2) The absence of a regulatory framework for AI that directly controls endoscope manipulation; and (3) The absence of standardized kinematic-measurement infrastructure for flexible endoscopes, which – unlike rigid laparoscopes with known geometric constraints and embedded encoders – lack equivalent sensing. Each bottleneck requires interdisciplinary collaboration between engineers, clinicians, and regulators. We justify the infrastructure framing by referencing the explicit contrast with surgical AI below, where infrastructure (datasets, challenges, regulatory precedents) demonstrably preceded clinical adoption.

Third, all eight clinical RCTs of kinematic AI were conducted in Chinese centers and six of eight used the ENDOANGEL platform. This pronounced geographic and platform concentration, which was highlighted by both handling editor and reviewers, limits external generalizability and introduces the possibility that regional health-system factors (workload, baseline ADR, documentation culture, training pathways) influenced the observed effects. External replication of CAQ concepts in Western, Japanese, and other health-system contexts, and validation of alternative (non-ENDOANGEL) platforms, should be considered prerequisites for broad clinical adoption.

Fourth, the six non-validated domains should not be interpreted as having produced “low-quality” evidence. Workflow recognition and skill assessment provide video-annotation and clinical-pilot evidence, while the four engineering-stage domains (autonomous navigation, SLAM/3D reconstruction, robotic control, and instrument tracking) produce the type of evidence their design permits – mathematical validation, phantom/animal performance, or simulation-based safety verification – and sit at early IDEAL/SaMD translational stages. The value of the current review for these domains is mapping rather than clinical evaluation.

Comparison with previous literature

The underdevelopment of kinematic AI in GIE relative to surgical AI is not primarily a natural consequence of field age but an ecosystem gap, as direct comparison with surgery demonstrates (Supplementary Table 2, Supplementary Figures 1-3)[10-18,33,51]. We make this comparison cautiously and explicitly acknowledge its limits, because the physical substrates of the two domains differ in important ways.

Laparoscopic instruments are rigid, have well-defined geometric constraints, and – in robotic platforms (e.g., da Vinci) – are equipped with built-in encoders that provide direct kinematic readouts of instrument position, velocity, and force. Flexible GIEs, by contrast, exhibit complex non-linear tip deformation, are subject to loop formation and redundant insertion configurations, and currently lack equivalent embedded kinematic sensing. Consequently, kinematic AI in flexible endoscopy must in most cases infer motion indirectly from the monocular video feed, using optical flow, depth estimation, pose reconstruction, or anatomical-landmark recognition, rather than from direct sensor readouts. The computational task of scene understanding from monocular video is comparable between the two domains, but the physical sensing substrate is not identical, and the challenge of deriving kinematic information from video alone is genuinely harder for flexible endoscopy in several respects.

Our ecosystem comparison with surgery is therefore intended as a translational – not a purely technical – parallel. Surgery established its first kinematic dataset (JIGSAWS, 2014)[10], organized annual instrument-segmentation challenges (EndoVis, 9 editions from 2015 to 2024)[11], released the Cholec80 workflow-recognition benchmark (2016)[12], and achieved autonomous in vivo intestinal anastomosis with outcomes comparable to experienced surgeons (STAR, 2016)[13]. By 2022, STAR had performed fully autonomous laparoscopic surgery on living pigs[14], and Moon Surgical received FDA 510(k) clearance for a robotic-assisted surgical system with AI-enabled camera positioning[15]. The first remote robotic surgery (Operation Lindbergh) was performed in 2001[16], more than two decades before any comparable GIE milestone. The quantitative infrastructure gap is pronounced: (1) Approximately 30000 annotated instrument-segmentation frames vs 590 in GIE (approximately 50:1); (2) 50-84 machine-learning-based skill-assessment studies vs approximately 2 (approximately 30:1)[17,18]; and (3) Six or more public kinematic datasets vs one. The key point of the analogy is not that the underlying technical problems are equally difficult, but that the combination of open datasets, annual benchmarking competitions, substantial industry investment, and dedicated regulatory pathways – which surgery has built – has not yet been replicated for GIE kinematic AI[19].

This contrast, however, conflates two barriers of fundamentally different kinds, and distinguishing them is essential before the surgical analogy can be applied. Some limitations concern the data ecosystem surrounding the field, whereas others are intrinsic sensor and physical modelling constraints of the flexible endoscope itself.

Data ecosystem limitations

The first class concerns the research infrastructure surrounding the field rather than the endoscope itself: The absence of open annotated datasets, of recurring benchmarking challenges, and of sustained industry investment and dedicated regulatory pathways. These are the gaps quantified above (one public instrument dataset vs six or more; no GI-specific challenge vs nine EndoVis editions) and they are, in principle, remediable by coordinated community action – exactly as surgery demonstrated. The surgical analogy is valid for this class of barrier, because data-ecosystem maturity is largely independent of the physical platform and can be transferred across fields.

Sensor and physical modelling limitations

The second class is intrinsic to flexible endoscopy and is not resolved by data sharing alone. Robotic laparoscopes are rigid, kinematically constrained, and instrumented with joint encoders that yield direct position, velocity, and force readouts; their forward kinematics are exactly modellable. Flexible endoscopes have no equivalent embedded sensing: The shaft deforms non-linearly, forms loops, and adopts redundant insertion configurations, so the same tip pose can arise from many shaft states and motion must be inferred indirectly from the monocular video stream by optical flow, depth estimation, or landmark recognition. This inference problem is ill-posed in ways that encoder-based surgical kinematics is not, and it would persist even with abundant data. For this class of barrier the surgical analogy is therefore one of trajectory, not of technical equivalence: Progress will require advances in endoscope-specific sensing hardware and physical modelling, not only a larger data ecosystem.

The clinical importance of preoperative anatomical mapping has been illustrated in recent laparoscopic experience. Anatomical anomalies such as adult intestinal malrotation can necessitate non-standard reconstruction techniques to avoid limb torsion, obstruction, and anastomotic leak[84], and postoperative dynamic intestinal obstruction occurs in approximately one in five patients after laparoscopic colorectal surgery, independently predicted by preoperative obstruction and prior abdominal surgery[85]. Accurate preoperative spatial mapping – the function that endoscopic SLAM/3D reconstruction is designed to support – is therefore a clinical motivation for the sensor and physical modelling advances called for above.

Implications for clinical practice, regulation, and ecosystem development

Beyond ESD, kinematic AI may also become relevant to therapeutic endoscopy in urgent settings. Endoscopic self-expanding stent placement is already an established, effective and safer alternative to emergency surgery for acute malignant colorectal obstruction, with substantially lower complication rates than emergency laparotomy in recent comparative data[86]. Workflow-recognition and phase-segmentation models analogous to those validated in ESD could, in principle, support such time-critical interventions by identifying procedural phases and flagging deviations during stent deployment.

At the clinical level, the additive benefit of kinematic monitoring is now established: The Yao et al[9] (2022) four-arm trial demonstrated that CADe alone improved ADR from 14.8% to 21.3%, but adding CAQ further increased ADR to 30.6%, confirming that exposure errors and detection errors are independent failure modes. This benefit is not uniform. Barua et al[28] (2023) showed that withdrawal-speed monitoring provided no additional benefit at high-baseline-ADR centers (baseline approximately 45%), establishing a ceiling effect with practical implications for resource allocation. Kinematic AI should be prioritized where baseline quality indicators are moderate or low, and where the marginal gain is greatest. The Lu et al[29] (2023) study observed that time-of-day decline in ADR was eliminated in the AI-assisted arm. As the reviewers correctly emphasized, the AI-assisted arm in Lu et al[29] combined CADe, CAQ, and CADe + CAQ subgroups, so the observed effect cannot be attributed to kinematic AI alone. It does, however, support the broader hypothesis that AI-assisted monitoring may attenuate the clinical impact of endoscopist fatigue; dedicated CAQ-only and CADe-only comparative trials are required to isolate the contribution of each modality. Even at high-baseline-ADR centers where the ceiling effect limits incremental ADR gains, kinematic AI may retain non-detection clinical and administrative value: Examples include sustaining inspection quality across long shifts to mitigate endoscopist fatigue, consistent with the time-of-day stabilisation reported by Lu et al[29], and providing automated time-stamped documentation of withdrawal time, photodocumentation, and inspection completeness for quality-audit and medicolegal purposes, as prototyped by Lux et al[56]. At such centers, kinematic AI is therefore better framed as a QA and procedural-audit layer than as a detection augmenter.

At the regulatory level, a structural barrier persists. Current AI-as-a-medical-device frameworks are optimized for diagnostic outputs (e.g., per-frame polyp probability), where the clinical decision pathway is well defined: (1) The AI flags; and (2) The physician decides. Kinematic AI operates on a different paradigm. Autonomous navigation requires the AI to act on physical endoscope control; SLAM-based coverage mapping must define what counts as “adequate” examination in real time; and workflow recognition must trigger alerts at procedurally critical moments. None of these functions fit neatly into existing SaMD classification tiers. The absence of any standalone regulatory approval for kinematic AI – despite 8 RCTs and technically advanced algorithms – reflects not only an evidence gap but a regulatory framework gap. Defining kinematic AI-specific regulatory pathways, analogous to what the FDA established for AI-enabled robotic surgical systems [e.g., Moon Surgical 510(k) clearance][15], is a prerequisite for clinical translation. As of late 2025, no FDA, European Medicines Agency, or European Commission instrument specifically targets motion-acting AI in flexible GIE. The FDA’s final guidance on Predetermined Change Control Plans for AI-enabled device software functions (December 2024)[87], article 6 (1) of the European AI act (Regulation 2024/1689, in force August 2024)[88], and the joint Medical Device Coordination Group/AI-Board guidance Medical Device Coordination Group 2025-6 (June 2025)[89] cover medical-device AI generically and do not name endoscopic motion as a distinct subdomain. The closest authorised motion-acting precedent – Moon Surgical’s ‘Physical AI’ ScoPilot, FDA-cleared in 2025 with a Predetermined Change Control Plan – is indicated for rigid-laparoscope robotic control, not flexible endoscopy[90], and the only regulatory authorisation globally that explicitly bundles withdrawal-speed and blind-spot quality-control functions with polyp CADe is the Chinese NMPA approval of the ENDOANGEL platform in May 2023[91]. No FDA or European Medicines Agency workshop, request for information, or task force naming kinematic endoscopic AI as a regulatory focus has been identified to date, indicating that motion-acting endoscopic AI remains a regulatory void addressed only by analogy to diagnostic AI and surgical-robotics autonomy frameworks, awaiting the first dedicated instrument.

At the ecosystem level, the comparison with surgery suggests that the rate-limiting factor is infrastructure, not algorithmic capability. Surgery’s translational trajectory was built on a self-reinforcing cycle: Open datasets (JIGSAWS, 2014)[10] enabled benchmarking competitions (EndoVis, 2015–2024)[11], which attracted industry investment[19], which funded clinical trials, which generated regulatory precedents, which in turn attracted further investment. GIE kinematic AI has not entered this cycle. The field has one public instrument dataset (590 images)[51] vs surgery’s six or more; no annual GI-specific benchmarking challenge vs nine EndoVis editions; and a single academic platform (ENDOANGEL) dominating the clinical literature. Breaking into this cycle requires coordinated, simultaneous action on multiple fronts – open dataset creation, challenge organization, and industry engagement – rather than incremental algorithm improvement.

Future perspectives

Five priority research gaps warrant focused effort: (1) Colonoscopy 3D coverage RCT: SLAM technology can now quantify mucosal coverage in real time, a trial testing whether coverage feedback reduces adenoma miss rates is feasible with current hardware and should be conducted outside China to address external validity; (2) GI instrument tracking: 590 annotated images vs thousands in surgical endoscopy, a multi-center annotation initiative for ESD knives, ERCP instruments, and hemostasis clips is needed; (3) Flexible endoscopy skill assessment: Combined DL and electromagnetic kinematic-sensor analysis should replicate the JIGSAWS paradigm; (4) Laparoscopic Endoscopic Cooperative Surgery procedure AI: Zero studies were identified despite clinical use of Laparoscopic Endoscopic Cooperative Surgery since 2008; and (5) AI-based colonoscope loop detection and correction remains unexplored. Future open datasets for flexible GIE should prioritise synchronised multimodal recordings – endoscopic video time-locked to magnetic endoscope imaging[38], force–torque traces from the insertion port, and electromagnetic-tracker pose, to enable JIGSAWS-style cross-platform skill and trajectory benchmarking comparable to that established for robotic surgery[10].

Strengths and limitations

This is the first systematic review to systematically map AI applications targeting endoscope kinematics across GIE domains. Key strengths include the coverage of four databases spanning both clinical and engineering literature (capturing the 62% of studies published in non-medical venues), an eight-domain classification enabling cross-domain translational comparison, risk-of-bias assessment across heterogeneous designs using RoB 2, QUADAS-2, and PROBAST + AI, the first GRADE certainty rating for kinematic AI, and quantitative benchmarking against surgical AI.

Limitations warrant emphasis. First, all eight RCTs of kinematic AI originated from Chinese centers and six used the ENDOANGEL platform; geographic and platform generalizability is therefore unproven, and regional health-system factors may have contributed to the observed effects. Several non-exclusive factors may account for this concentration. High colonoscopy volumes at large Chinese tertiary centres provide both the data scale needed to train motion-recognition models and the throughput needed to recruit pragmatic single-system trials rapidly. In addition, the dominant platform was developed and trialled by a single vertically integrated academic group, which compresses the conventional separation between algorithm development, clinical validation, and trial sponsorship and accelerates iteration. These same factors, however, reinforce rather than resolve the platform-dependence concern, since the observed effects remain entangled with one platform and one health-system context, and independent multi-platform Western trials are needed before the magnitude of benefit can be regarded as generalizable. Second, heterogeneous study designs precluded quantitative meta-analysis, and certainty estimates for engineering-stage domains should be interpreted as pre-clinical rather than as indicative of weak evidence per se. Third, eight included studies were arXiv preprints without peer review, reflecting the predominance of engineering literature in several domains. As detailed in the Methods, these preprints were eligible only where methodological detail was sufficient for data extraction and formal risk-of-bias appraisal with PROBAST + AI, and were not exempted from quality assessment; nonetheless, their non-peer-reviewed status remains a limitation. Fourth, our definition of kinematic AI is broader than pure motion-sensor analysis and includes image-recognition-based proxy measures (e.g., anatomical-landmark blind-spot monitoring, workflow-phase recognition); under a narrower, purely motion-based definition, coverage-mapping and workflow-recognition domains would be reclassified as primarily image-recognition tasks, which would reduce the number of domains considered “kinematic” from eight to four but would not alter the core conclusion of a substantial evidence imbalance. To make this scope explicit, we organised the included domains into a four-tier hierarchical taxonomy (Figure 2)[9] distinguishing true kinematic AI (Tier 1, direct motion signal), proxy kinematic inference (Tier 2, vision-derived motion), workflow intelligence (Tier 3, phase recognition), and procedural QA systems (Tier 4, landmark classification), and conducted a pre-specified sensitivity analysis under this Tier 1-2-only definition; the kinematic-to-CADe RCT ratio rose from approximately 5:1 to approximately 8:1, but the core conclusion of a substantial evidence imbalance was preserved, indicating robustness of the principal finding to the boundary of the kinematic definition. Fifth, the comparison with surgical AI is offered as a translational parallel; direct extrapolation of technical performance between rigid-instrument surgery and flexible endoscopy is constrained by differences in embedded kinematic sensing, as discussed above.

CONCLUSION

AI-based kinematic analysis in GIE addresses exposure errors that CADe cannot correct, yet has produced only a fraction of the clinical evidence generated for diagnostic AI. Withdrawal-speed monitoring and coverage/blind-spot mapping have reached moderate GRADE certainty and are ready for geographically diverse validation; workflow recognition and skill assessment are approaching clinical readiness; the remaining four domains are at the engineering stage and require ecosystem-level investment – open datasets, benchmarking challenges, and dedicated regulatory pathways – before meaningful clinical trials can be designed. Looking forward, the demonstrated additive benefit of co-deployed CADe and CAQ implies that integration should extend beyond software: Next-generation flexible endoscopes may need to embed dedicated kinematic sensing – built-in shaft-pose encoders, insertion-port force-torque measurement, and a unified data bus carrying synchronised video and motion streams to a single onboard detection-plus-quality AI module – so that exposure-error and recognition-error AI converge into one physically integrated procedural-quality system rather than two parallel software overlays running on the same scope.