Gastroenterol Res

Gastroenterology Research, ISSN 1918-2805 print, 1918-2813 online, Open Access

Article copyright, the authors; Journal compilation copyright, Gastroenterol Res and Elmer Press Inc

Journal website https://gr.elmerpub.com

Original Article

Volume 000, Number 000, October 2026, pages 000-000

Artificial Intelligence-Assisted Colonoscopy Improves Adenoma Detection by Trainee and Less-Experienced Endoscopists: A Systematic Review and Meta-Analysis of Randomized Controlled Trials

↓ Figure 1. PRISMA 2020 flow diagram of study selection, showing the number of records retrieved from each database and register searched.

↓ Figure 2. Forest plot for the primary outcome (overall adenoma detection rate), with the 95% prediction interval shown beneath the pooled estimate. CADe: computer-aided detection; CI: confidence interval; IV: inverse variance.

↓ Figure 3. Forest plot for right-colon adenoma detection rate, with the 95% prediction interval shown beneath the pooled estimate. CADe: computer-aided detection; CI: confidence interval; IV: inverse variance.

↓ Figure 4. Forest plot for polyp detection rate, with the 95% prediction interval shown beneath the pooled estimate. CADe: computer-aided detection; CI: confidence interval; IV: inverse variance.

↓ Figure 5. Forest plot for advanced adenoma detection rate, with the 95% prediction interval shown beneath the pooled estimate. Advanced adenoma counts for Yao 2024 are the trainee-performed first-pass values, consistent with the extraction rule stated in Section “Outcomes”. CADe: computer-aided detection; CI: confidence interval; IV: inverse variance.

↓ **Table 1.** Characteristics of Included Studies
Study (year)	Country	Design	CADe system	Operator category	Operator definition	CADe n	Control n	Total n
“Trainee” denotes operators enrolled in formal endoscopic training; “less-experienced” denotes independently practicing operators below an investigator-defined lifetime procedure threshold. Counts are the numbers of patients analyzed in each arm. For the two tandem trials (Yamaguchi 2024, Yao 2024), counts and outcomes refer to the trainee-performed first pass. AID-2: Artificial Intelligence Detection 2 trial; CADe: computer-aided detection; CNN: convolutional neural network; ENDOAID-TRAIN: ENDO-AID in Endoscopists-in-Training trial; PGY: postgraduate year; RCT: randomized controlled trial; USA: United States of America.
Repici 2022 (AID-2) [20]	Italy; Switzerland	Parallel RCT	GI Genius (Medtronic)	Less experienced	Independently practicing non-experts (< 2,000 lifetime colonoscopies)	330	330	660
Lau 2024 (ENDOAID-TRAIN) [21]	Hong Kong	Parallel RCT	ENDO-AID (Olympus)	Trainee	Endoscopists-in-training (< 500 procedures, < 3 years of training)	386	380	766
Yamaguchi 2024 [23]	Japan	Back-to-back tandem RCT	CAD EYE (Fujifilm)	Trainee	Trainees (0–20 prior colonoscopies)	113	118	231
Yao 2024 [24]	China	Multicenter back-to-back tandem RCT	Custom CNN (EndoAngel)	Trainee	Novices (no prior colonoscopy experience)	227	229	456
Chang 2026 [22]	USA	Pragmatic parallel RCT (room- and fellow-level randomization)	GI Genius (Medtronic)	Trainee	Gastroenterology fellows (PGY 4–6)	401	644	1,045
Total						1,457	1,701	3,158

↓ Table 1. Characteristics of Included Studies

Study (year)

Country

Design

CADe system

Operator category

Operator definition

CADe n

Control n

Total n

“Trainee” denotes operators enrolled in formal endoscopic training; “less-experienced” denotes independently practicing operators below an investigator-defined lifetime procedure threshold. Counts are the numbers of patients analyzed in each arm. For the two tandem trials (Yamaguchi 2024, Yao 2024), counts and outcomes refer to the trainee-performed first pass. AID-2: Artificial Intelligence Detection 2 trial; CADe: computer-aided detection; CNN: convolutional neural network; ENDOAID-TRAIN: ENDO-AID in Endoscopists-in-Training trial; PGY: postgraduate year; RCT: randomized controlled trial; USA: United States of America.

Repici 2022 (AID-2) [20]

Italy; Switzerland

Parallel RCT

GI Genius (Medtronic)

Less experienced

Independently practicing non-experts (< 2,000 lifetime colonoscopies)

330

660

Lau 2024 (ENDOAID-TRAIN) [21]

Hong Kong

Parallel RCT

ENDO-AID (Olympus)

Trainee

Endoscopists-in-training (< 500 procedures, < 3 years of training)

386

380

766

Yamaguchi 2024 [23]

Japan

Back-to-back tandem RCT

CAD EYE (Fujifilm)

Trainee

Trainees (0–20 prior colonoscopies)

113

118

231

Yao 2024 [24]

China

Multicenter back-to-back tandem RCT

Custom CNN (EndoAngel)

Trainee

Novices (no prior colonoscopy experience)

227

229

456

Chang 2026 [22]

USA

Pragmatic parallel RCT (room- and fellow-level randomization)

GI Genius (Medtronic)

Trainee

Gastroenterology fellows (PGY 4–6)

401

644

1,045

Total

1,457

1,701

3,158

↓ **Table 2.** Detailed Participant, Operator, Intervention, and Comparator Characteristics of the Included Trials
Characteristic	Repici 2022 (AID-2) [20]	Lau 2024 [21]	Yamaguchi 2024 [23]	Yao 2024 [24]	Chang 2026 [22]
Values are n, n (%), or mean ± standard deviation unless otherwise stated, and are presented as CADe arm/control arm where both are available. Data are as reported in the primary publications; where a trial reported a characteristic only for the whole cohort rather than by arm, this is indicated, and NR denotes a characteristic not reported in the source publication. ADR: adenoma detection rate; AID-2: Artificial Intelligence Detection 2 trial; BBPS: Boston Bowel Preparation Scale; CADe: computer-aided detection; CADx: computer-aided diagnosis; CAQ: computer-aided quality; CNN: convolutional neural network; FDA: United States Food and Drug Administration; FIT: fecal immunochemical test; GI: gastrointestinal; IQR: interquartile range; NR: not reported in the source publication; PGY: postgraduate year.
Participant characteristics
Patients analyzed (CADe/control)	330/330	386/380	113/118	227/229	401/644
Age, years	62.3 ± 10 (whole cohort)	66.00 ± 10.05/65.36 ± 11.33	63.1 ± 10.8/63.3 ± 11.8	50.63 ± 12.25/49.85 ± 11.68	54.6 ± 11.6/54.2 ± 12.2
Female sex	330/660 (50.0%) (whole cohort)	181/386 (46.9%)/169/380 (44.5%)	55/113 (48.7%)/52/118 (44.1%)	110/227 (48.5%)/106/229 (46.3%)	168/401 (41.9%)/287/644 (44.6%)
Colonoscopy indication	Screening, surveillance, or diagnostic	Screening 7.3%/6.1%; surveillance 32.6%/31.8%; symptomatic 60.1%/62.1%	FIT-positive 46.9%/54.2%; detailed examination 19.5%/22.0%; polyp surveillance 13.3%/11.0%; abdominal symptoms 11.5%/8.5%; other 8.9%/4.2%	Screening 63.9%/63.8%; diagnostic 26.4%/28.0%; surveillance 9.7%/8.3%	GI bleeding 52%/57%; symptoms 17%/17%; surveillance for neoplasia 13%/12%; screening 14%/11%; imaging follow-up 3%/2%
History of previous adenoma	Not separately reported; surveillance included as an eligible indication	Not separately reported; surveillance 32.6%/31.8%. Family history of colorectal cancer 19.5%/15.6%	Polyp surveillance 15/113 (13.3%)/13/118 (11.0%); family history of colorectal cancer 11.5%/11.0%	Not separately reported; surveillance 9.7%/8.3%. Family history 8.8%/7.0%	Not separately reported; surveillance for neoplasia 13%/12%. Family history of colon cancer 8%/7%
Bowel preparation quality	NR	BBPS total 7.85 ± 1.17/7.84 ± 1.21; patients with BBPS 0-1 in any segment excluded	Modified Aronchick excellent or good 86.7%/83.9%; BBPS whole colon 8.4 ± 1.2/8.2 ± 1.5	BBPS adequate (sum ≥6.0 and each segment ≥2.0) 86.3%/88.2%	NR by arm
Baseline control-arm ADR	44.5%	44.5%	61.0%	16.2%	34.9%
Operator characteristics
Number of operators	10	22 junior endoscopists	6 trainees	8 novices (plus 10 experts for the second pass)	16 fellows
Specialty	Gastroenterology/endoscopy	Gastroenterologists (78.8%/76.3% of procedures) and surgeons-in-training (21.2%/23.7%)	Gastroenterology	Gastroenterology	Gastroenterology
Formal training status	No, independently practicing	Yes, endoscopists-in-training	Yes, trainees who had started fellowship training	Yes, fellows in training	Yes, GI fellows, PGY 4–6
Years of endoscopy experience	NR; qualified for independent screening colonoscopy and non-complex therapy	< 3 years of training	First year of fellowship training	0 years of independent colonoscopy; > 1 year of gastroenterology fellowship	1st to 3rd year of fellowship
Lifetime colonoscopies performed	< 2,000 per operator	< 500 per operator	0–20 cases per operator	0 prior independent colonoscopies; 20 supervised colonoscopies with expert hands-on training before study entry	Median 84 (IQR 54–127) for PGY 4; 327 (IQR 272–373) for PGY 5; 468 for PGY 6
Experience strata used by investigators	Single stratum (non-expert, < 2,000 procedures)	Beginner (< 200 procedures, n = 12) and intermediate (200–500 procedures, n = 10)	Single stratum	Novice versus expert (expert defined as ≥5,000 colonoscopies)	Fellow year 1 (PGY 4), year 2 (PGY 5), year 3 (PGY 6)
Prior experience with CADe	NR	Yes, all junior endoscopists received training on the CADe system before study initiation	NR	None reported	NR
Intervention characteristics
CADe platform	GI Genius	ENDO-AID	CAD EYE	EndoAngel (custom CNN)	GI Genius
Manufacturer	Medtronic, Dublin, Ireland	Olympus, Tokyo, Japan	Fujifilm, Tokyo, Japan	Renmin Hospital of Wuhan University, Wuhan, China	Medtronic, Dublin, Ireland
Software/version	NR	Preinstalled on the EVIS X1 CV-1500 processor	NR	Custom deep-learning system on a dedicated host computer	US FDA-approved system; version NR
Visual alerts	Yes, real-time on-screen marking of candidate lesions	Yes, alert flag with picture-in-picture display in normal mode; a target mode is also available	Yes, real-time on-screen marking of candidate lesions	Yes, real-time on-screen marking of candidate lesions	Yes, real-time on-screen marking of candidate lesions
Auditory alerts	NR	NR	NR	NR	NR
CADe active during insertion	No	No	No, colonoscope inserted without CAD EYE in both groups	No	No
CADe active during withdrawal	Yes	Yes	Yes	Yes	Yes
Additional computer-aided diagnosis (CADx)	No	No	Yes, CADx used to characterize lesions when measuring polyps	No	No
Additional computer-aided quality (CAQ) or withdrawal-speed feedback	No	No	No	Yes, computer-aided quality system providing withdrawal-speed and mucosal-exposure feedback	No
Other co-interventions	None	On-site or next-door supervisor available for safety	Expert second pass without CAD EYE (tandem design)	Expert second pass with AI assistance (tandem design)	Usual attending supervision, not standardized
Comparator characteristics
Endoscopy platform	High-definition colonoscopy	Olympus EVIS X1 CV-1500 processor with CF-HQ290L/I (97.4%/98.4%) or CF-EZ1500DL/I colonoscopes	Fujifilm EC-L600ZP7 colonoscope	Fujinon EC-590WM or EC-600WM; Olympus EVIS LUCERA CV260 or CV290	Standard colonoscopy in a non-AI room
High definition versus standard definition	High definition	High definition	High definition	High definition	High definition
White-light imaging	Yes	Yes	Yes	Yes	Yes
Permitted image-enhancement modalities	NR	Narrow-band imaging or texture and color enhancement imaging permitted for polyp characterization only; magnification and chromoendoscopy not allowed	Narrow-band-type imaging available through the CAD EYE platform for characterization	NR	Not mandated or tracked
Distal attachment devices	NR	Prohibited (transparent cap, Endocuff Vision)	NR	NR	Mucosal exposure devices such as distal caps were neither mandated nor tracked
Other detection-enhancing techniques	None reported	None reported	None reported	None reported	Right-colon examination techniques such as retroflexion and second view were neither mandated nor tracked

↓ Table 2. Detailed Participant, Operator, Intervention, and Comparator Characteristics of the Included Trials

Characteristic

Repici 2022 (AID-2) [20]

Lau 2024 [21]

Yamaguchi 2024 [23]

Yao 2024 [24]

Chang 2026 [22]

Values are n, n (%), or mean ± standard deviation unless otherwise stated, and are presented as CADe arm/control arm where both are available. Data are as reported in the primary publications; where a trial reported a characteristic only for the whole cohort rather than by arm, this is indicated, and NR denotes a characteristic not reported in the source publication. ADR: adenoma detection rate; AID-2: Artificial Intelligence Detection 2 trial; BBPS: Boston Bowel Preparation Scale; CADe: computer-aided detection; CADx: computer-aided diagnosis; CAQ: computer-aided quality; CNN: convolutional neural network; FDA: United States Food and Drug Administration; FIT: fecal immunochemical test; GI: gastrointestinal; IQR: interquartile range; NR: not reported in the source publication; PGY: postgraduate year.

Participant characteristics

Patients analyzed (CADe/control)

330/330

386/380

113/118

227/229

401/644

Age, years

62.3 ± 10 (whole cohort)

66.00 ± 10.05/65.36 ± 11.33

63.1 ± 10.8/63.3 ± 11.8

50.63 ± 12.25/49.85 ± 11.68

54.6 ± 11.6/54.2 ± 12.2

Female sex

330/660 (50.0%) (whole cohort)

181/386 (46.9%)/169/380 (44.5%)

55/113 (48.7%)/52/118 (44.1%)

110/227 (48.5%)/106/229 (46.3%)

168/401 (41.9%)/287/644 (44.6%)

Colonoscopy indication

Screening, surveillance, or diagnostic

Screening 7.3%/6.1%; surveillance 32.6%/31.8%; symptomatic 60.1%/62.1%

FIT-positive 46.9%/54.2%; detailed examination 19.5%/22.0%; polyp surveillance 13.3%/11.0%; abdominal symptoms 11.5%/8.5%; other 8.9%/4.2%

Screening 63.9%/63.8%; diagnostic 26.4%/28.0%; surveillance 9.7%/8.3%

GI bleeding 52%/57%; symptoms 17%/17%; surveillance for neoplasia 13%/12%; screening 14%/11%; imaging follow-up 3%/2%

History of previous adenoma

Not separately reported; surveillance included as an eligible indication

Not separately reported; surveillance 32.6%/31.8%. Family history of colorectal cancer 19.5%/15.6%

Polyp surveillance 15/113 (13.3%)/13/118 (11.0%); family history of colorectal cancer 11.5%/11.0%

Not separately reported; surveillance 9.7%/8.3%. Family history 8.8%/7.0%

Not separately reported; surveillance for neoplasia 13%/12%. Family history of colon cancer 8%/7%

Bowel preparation quality

BBPS total 7.85 ± 1.17/7.84 ± 1.21; patients with BBPS 0-1 in any segment excluded

Modified Aronchick excellent or good 86.7%/83.9%; BBPS whole colon 8.4 ± 1.2/8.2 ± 1.5

BBPS adequate (sum ≥6.0 and each segment ≥2.0) 86.3%/88.2%

NR by arm

Baseline control-arm ADR

44.5%

61.0%

16.2%

34.9%

Operator characteristics

Number of operators

22 junior endoscopists

6 trainees

8 novices (plus 10 experts for the second pass)

16 fellows

Specialty

Gastroenterology/endoscopy

Gastroenterologists (78.8%/76.3% of procedures) and surgeons-in-training (21.2%/23.7%)

Gastroenterology

Formal training status

No, independently practicing

Yes, endoscopists-in-training

Yes, trainees who had started fellowship training

Yes, fellows in training

Yes, GI fellows, PGY 4–6

Years of endoscopy experience

NR; qualified for independent screening colonoscopy and non-complex therapy

< 3 years of training

First year of fellowship training

0 years of independent colonoscopy; > 1 year of gastroenterology fellowship

1st to 3rd year of fellowship

Lifetime colonoscopies performed

< 2,000 per operator

< 500 per operator

0–20 cases per operator

0 prior independent colonoscopies; 20 supervised colonoscopies with expert hands-on training before study entry

Median 84 (IQR 54–127) for PGY 4; 327 (IQR 272–373) for PGY 5; 468 for PGY 6

Experience strata used by investigators

Single stratum (non-expert, < 2,000 procedures)

Beginner (< 200 procedures, n = 12) and intermediate (200–500 procedures, n = 10)

Single stratum

Novice versus expert (expert defined as ≥5,000 colonoscopies)

Fellow year 1 (PGY 4), year 2 (PGY 5), year 3 (PGY 6)

Prior experience with CADe

Yes, all junior endoscopists received training on the CADe system before study initiation

None reported

Intervention characteristics

CADe platform

GI Genius

ENDO-AID

CAD EYE

EndoAngel (custom CNN)

GI Genius

Manufacturer

Medtronic, Dublin, Ireland

Olympus, Tokyo, Japan

Fujifilm, Tokyo, Japan

Renmin Hospital of Wuhan University, Wuhan, China

Medtronic, Dublin, Ireland

Software/version

Preinstalled on the EVIS X1 CV-1500 processor

Custom deep-learning system on a dedicated host computer

US FDA-approved system; version NR

Visual alerts

Yes, real-time on-screen marking of candidate lesions

Yes, alert flag with picture-in-picture display in normal mode; a target mode is also available

Yes, real-time on-screen marking of candidate lesions

Auditory alerts

CADe active during insertion

No, colonoscope inserted without CAD EYE in both groups

CADe active during withdrawal

Yes

Additional computer-aided diagnosis (CADx)

Yes, CADx used to characterize lesions when measuring polyps

Additional computer-aided quality (CAQ) or withdrawal-speed feedback

Yes, computer-aided quality system providing withdrawal-speed and mucosal-exposure feedback

Other co-interventions

None

On-site or next-door supervisor available for safety

Expert second pass without CAD EYE (tandem design)

Expert second pass with AI assistance (tandem design)

Usual attending supervision, not standardized

Comparator characteristics

Endoscopy platform

High-definition colonoscopy

Olympus EVIS X1 CV-1500 processor with CF-HQ290L/I (97.4%/98.4%) or CF-EZ1500DL/I colonoscopes

Fujifilm EC-L600ZP7 colonoscope

Fujinon EC-590WM or EC-600WM; Olympus EVIS LUCERA CV260 or CV290

Standard colonoscopy in a non-AI room

High definition versus standard definition

High definition

White-light imaging

Yes

Permitted image-enhancement modalities

Narrow-band imaging or texture and color enhancement imaging permitted for polyp characterization only; magnification and chromoendoscopy not allowed

Narrow-band-type imaging available through the CAD EYE platform for characterization

Not mandated or tracked

Distal attachment devices

Prohibited (transparent cap, Endocuff Vision)

Mucosal exposure devices such as distal caps were neither mandated nor tracked

Other detection-enhancing techniques

None reported

Right-colon examination techniques such as retroflexion and second view were neither mandated nor tracked

↓ **Table 3.** Pooled Outcome Summary
Outcome	Trials (n)	Patients (n)	Pooled RR (95% CI)	P value	I² (%)	95% PI	Hartung–Knapp 95% CI	Pooled RD (95% CI); NNT
Risk ratios and risk differences were pooled with DerSimonian–Laird random-effects models. Event counts and denominators for every trial are shown in Figures 2 to 5 and in Supplementary Material 3 (gr.elmerpub.com). Prediction intervals describe the range within which the true effect of a future comparable trial would be expected to lie. Hartung–Knapp confidence intervals are presented as a more conservative alternative to DerSimonian–Laird inference given the small number of trials. NNT is derived from the pooled risk difference and is reported only where that difference excluded zero (overall ADR 95% CI 9 to 37; right-colon ADR 95% CI 9 to 22; polyp detection rate 95% CI 7 to 27). Advanced adenoma detection rate values were revised during peer review to use the trainee-performed first pass counts reported by Yao 2024, for consistency with the extraction rule stated in Section “Outcomes”. ADR: adenoma detection rate; CI: confidence interval; NA: not applicable; NNT: number needed to treat; PI: prediction interval; RD: risk difference; RR: risk ratio.
Overall ADR (primary)	5	3,158	1.17 (1.07–1.29)	0.0012	28.8	0.95–1.46	1.02–1.35	+6.9% (+2.7 to +11.0); NNT 15
Right-colon ADR	3	2,471	1.31 (1.14–1.50)	0.0002	26.6	0.86–1.97	0.96–1.77	+8.2% (+4.6 to +11.7); NNT 12
Polyp detection rate	4	2,498	1.16 (1.08–1.25)	0.0001	27.8	0.97–1.39	1.02–1.32	+8.9% (+3.7 to +14.2); NNT 11
Advanced adenoma detection rate	3	1,882	0.97 (0.73–1.29)	0.8240	0.0	0.52–1.81	0.56–1.68	+0.2% (−2.2 to +2.6); NNT NA

↓ Table 3. Pooled Outcome Summary

Outcome

Trials (n)

Patients (n)

Pooled RR (95% CI)

P value

I² (%)

95% PI

Hartung–Knapp 95% CI

Pooled RD (95% CI); NNT

Risk ratios and risk differences were pooled with DerSimonian–Laird random-effects models. Event counts and denominators for every trial are shown in Figures 2 to 5 and in Supplementary Material 3 (gr.elmerpub.com). Prediction intervals describe the range within which the true effect of a future comparable trial would be expected to lie. Hartung–Knapp confidence intervals are presented as a more conservative alternative to DerSimonian–Laird inference given the small number of trials. NNT is derived from the pooled risk difference and is reported only where that difference excluded zero (overall ADR 95% CI 9 to 37; right-colon ADR 95% CI 9 to 22; polyp detection rate 95% CI 7 to 27). Advanced adenoma detection rate values were revised during peer review to use the trainee-performed first pass counts reported by Yao 2024, for consistency with the extraction rule stated in Section “Outcomes”. ADR: adenoma detection rate; CI: confidence interval; NA: not applicable; NNT: number needed to treat; PI: prediction interval; RD: risk difference; RR: risk ratio.

Overall ADR (primary)

3,158

1.17 (1.07–1.29)

0.0012

28.8

0.95–1.46

1.02–1.35

+6.9% (+2.7 to +11.0); NNT 15

Right-colon ADR

2,471

1.31 (1.14–1.50)

0.0002

26.6

0.86–1.97

0.96–1.77

+8.2% (+4.6 to +11.7); NNT 12

Polyp detection rate

2,498

1.16 (1.08–1.25)

0.0001

27.8

0.97–1.39

1.02–1.32

+8.9% (+3.7 to +14.2); NNT 11

Advanced adenoma detection rate

1,882

0.97 (0.73–1.29)

0.8240

0.0

0.52–1.81

0.56–1.68

+0.2% (−2.2 to +2.6); NNT NA

↓ **Table 4.** Risk of Bias Assessment (Cochrane RoB 2)
Study	D1: Randomization process	D2: Deviations from intended interventions	D3: Missing outcome data	D4: Measurement of the outcome	D5: Selection of the reported result	Overall
Judgments in domain 2 reflect the fact that operator exposure to the CADe display cannot be blinded in any trial of this intervention. The domain 1 judgment for Chang 2026 reflects randomization at the level of the procedure room and fellow day rather than the individual patient. D: domain; RoB 2: Cochrane Risk of Bias 2 tool.
Repici 2022 [20]	Low	Some concerns	Low	Low	Some concerns	Some concerns
Lau 2024 [21]	Low	Low	Low	Low	Low	Low
Yamaguchi 2024 [23]	Low	Some concerns	Low	Low	Low	Some concerns
Yao 2024 [24]	Low	Some concerns	Low	Low	Low	Some concerns
Chang 2026 [22]	Some concerns	Some concerns	Low	Low	Low	Some concerns

↓ Table 4. Risk of Bias Assessment (Cochrane RoB 2)

Study

D1: Randomization process

D2: Deviations from intended interventions

D3: Missing outcome data

D4: Measurement of the outcome

D5: Selection of the reported result

Overall

Judgments in domain 2 reflect the fact that operator exposure to the CADe display cannot be blinded in any trial of this intervention. The domain 1 judgment for Chang 2026 reflects randomization at the level of the procedure room and fellow day rather than the individual patient. D: domain; RoB 2: Cochrane Risk of Bias 2 tool.

Repici 2022 [20]

Low

Some concerns

Low

Some concerns

Lau 2024 [21]

Low

Yamaguchi 2024 [23]

Low

Some concerns

Low

Some concerns

Yao 2024 [24]

Low

Some concerns

Low

Some concerns

Chang 2026 [22]

Some concerns

Low

Some concerns

↓ **Table 5.** GRADE Summary of Findings: CADe-Assisted Versus Standard Colonoscopy Performed by Trainee and Less-Experienced Endoscopists
Outcome (trials; patients)	Anticipated absolute effects	Relative effect (95% CI)	Certainty of evidence	Reasons for downgrading
All outcomes began as high-certainty evidence because all included studies were randomized controlled trials; (−1) and (−2) denote the number of levels by which certainty was downgraded in that domain. Doi plots and LFK indices were generated for all outcomes but are exploratory with three to five trials and were not used as a basis for downgrading. Certainty definitions: Moderate, the true effect is likely close to the estimate but may be substantially different; Low, the true effect may be substantially different from the estimate; Very low, the estimate is very uncertain. CADe: computer-aided detection; CI: confidence interval; GRADE: Grading of Recommendations, Assessment, Development and Evaluation; NNT: number needed to treat; RoB 2: Cochrane Risk of Bias 2 tool; RR: risk ratio.
Overall adenoma detection rate (5 trials; 3,158 patients)	38.2% with standard colonoscopy versus 46.2% with CADe; risk difference +6.9% (+2.7 to +11.0); NNT 15	RR 1.17 (1.07–1.29)	Moderate	Risk of bias (−1): operator exposure to the CADe display cannot be blinded and 4 of 5 trials were rated at some concerns using RoB 2. Inconsistency, indirectness, and imprecision not serious; publication bias undetected.
Right-colon adenoma detection rate (3 trials; 2,471 patients)	25.3% versus 35.5%; risk difference +8.2% (+4.6 to +11.7); NNT 12	RR 1.31 (1.14–1.50)	Low	Risk of bias (−1): as above. Imprecision (−1): only 3 contributing trials, and the Hartung-Knapp confidence interval (0.96–1.77) crosses the null. Inconsistency and indirectness not serious; publication bias undetected.
Polyp detection rate (4 trials; 2,498 patients)	56.5% versus 66.5%; risk difference +8.9% (+3.7 to +14.2); NNT 11	RR 1.16 (1.08–1.25)	Low	Risk of bias (−1): as above. Indirectness (−1): polyp detection rate counts lesions of any histology rather than neoplasia, and CADe substantially increased the non-neoplastic resection rate in the single trial that measured it. Inconsistency and imprecision not serious; publication bias undetected.
Advanced adenoma detection rate (3 trials; 1,882 patients)	9.3% versus 9.0%; risk difference +0.2% (−2.2 to +2.6)	RR 0.97 (0.73–1.29)	Very low	Risk of bias (−1): as above. Imprecision (−2): 3 trials contributing 172 events, with a confidence interval compatible with both appreciable benefit and appreciable harm. Inconsistency and indirectness not serious; publication bias undetected.

↓ Table 5. GRADE Summary of Findings: CADe-Assisted Versus Standard Colonoscopy Performed by Trainee and Less-Experienced Endoscopists

Outcome (trials; patients)

Anticipated absolute effects

Relative effect (95% CI)

Certainty of evidence

Reasons for downgrading

All outcomes began as high-certainty evidence because all included studies were randomized controlled trials; (−1) and (−2) denote the number of levels by which certainty was downgraded in that domain. Doi plots and LFK indices were generated for all outcomes but are exploratory with three to five trials and were not used as a basis for downgrading. Certainty definitions: Moderate, the true effect is likely close to the estimate but may be substantially different; Low, the true effect may be substantially different from the estimate; Very low, the estimate is very uncertain. CADe: computer-aided detection; CI: confidence interval; GRADE: Grading of Recommendations, Assessment, Development and Evaluation; NNT: number needed to treat; RoB 2: Cochrane Risk of Bias 2 tool; RR: risk ratio.

Overall adenoma detection rate (5 trials; 3,158 patients)

38.2% with standard colonoscopy versus 46.2% with CADe; risk difference +6.9% (+2.7 to +11.0); NNT 15

RR 1.17 (1.07–1.29)

Moderate

Risk of bias (−1): operator exposure to the CADe display cannot be blinded and 4 of 5 trials were rated at some concerns using RoB 2. Inconsistency, indirectness, and imprecision not serious; publication bias undetected.

Right-colon adenoma detection rate (3 trials; 2,471 patients)

25.3% versus 35.5%; risk difference +8.2% (+4.6 to +11.7); NNT 12

RR 1.31 (1.14–1.50)

Low

Risk of bias (−1): as above. Imprecision (−1): only 3 contributing trials, and the Hartung-Knapp confidence interval (0.96–1.77) crosses the null. Inconsistency and indirectness not serious; publication bias undetected.

Polyp detection rate (4 trials; 2,498 patients)

56.5% versus 66.5%; risk difference +8.9% (+3.7 to +14.2); NNT 11

RR 1.16 (1.08–1.25)

Low

Risk of bias (−1): as above. Indirectness (−1): polyp detection rate counts lesions of any histology rather than neoplasia, and CADe substantially increased the non-neoplastic resection rate in the single trial that measured it. Inconsistency and imprecision not serious; publication bias undetected.

Advanced adenoma detection rate (3 trials; 1,882 patients)

9.3% versus 9.0%; risk difference +0.2% (−2.2 to +2.6)

RR 0.97 (0.73–1.29)

Very low

Risk of bias (−1): as above. Imprecision (−2): 3 trials contributing 172 events, with a confidence interval compatible with both appreciable benefit and appreciable harm. Inconsistency and indirectness not serious; publication bias undetected.