Availability and use of computer-aided detection during colonoscopy: A real-world observational study at an Australian tertiary center

Abstract

BACKGROUND

One of the main aims of colonoscopy is to detect and remove precancerous polyps. Multiple studies have shown that computer-aided detection (CADe) technology enhances key metrics, including adenoma detection rate (ADR), among endoscopists with low baseline detection rates. However, these findings largely come from controlled prospective studies where CADe is systematically used and endoscopists are aware they are being monitored for such metrics, bringing inherent biases. In Australia, CADe implementation is not yet standard practice, and its availability varies across endoscopy centers. We hypothesized that greater endoscopist use of CADe would be associated with higher ADR in real-world clinical settings.

AIM

To evaluate how varying levels of endoscopist use of CADe affect adenoma detection and other colonoscopy quality metrics in a tertiary Australian center, where CADe was available for all elective procedures from 2023.

METHODS

A single-center retrospective cohort study was conducted at a tertiary Australian center after introduction of the Olympus Endo-AID^® CADe module in July 2023, available for all elective procedures. Colonoscopy reports from six months before and after implementation were reviewed. Endoscopists were grouped by observed CADe usage. The primary outcome was change in ADR by group. Secondary outcomes included sessile serrated lesion detection rate (SSL-DR), adenomas per patient (APP), and sessile serrated lesions per patient (SPP).

RESULTS

Seven endoscopists performed 636 pre-CADe and 386 post-CADe colonoscopies. Two endoscopists used CADe 100% of the time, four used it 50%-99%, and one did not use CADe. No endoscopists used CADe 1%-50% of the time. ADR significantly improved from 29% to 41.9% in the 50%-99% group (odds ratio 1.77, 95%CI: 1.13-2.75, P = 0.01). No ADR change was observed in the 100% group, which had a baseline ADR above 60%, although APP increased from 0.90 to 2.08 (relative risk 2.31, 95%CI: 1.97-2.73, P < 0.0001). SSL-DR and SPP were not significantly affected by CADe.

CONCLUSION

In this real-world study, the availability of CADe was associated with an uptake by the majority of the endoscopists and led to significant improvement in ADR even when not being used in all procedures. Similarly to previous studies, no such benefit was observed for endoscopists who had a high baseline ADR. However, endoscopists with high baseline ADR did improve their APP after introduction of CADe. In addition, CADe availability did not improve SSL-DR across the cohort. This real-world significant increase from moderate baseline ADR reinforces its benefit when adopted in routine clinical practice.

Key Words: Computer-aided detection; Adenoma detection rate; Colonoscopy; Artificial intelligence; Endoscopy; Screening, surveillance; Colorectal neoplasm; Polyp; Adenoma

Core Tip: This study evaluates the real-world impact of varying computer-aided detection (CADe) usage frequency on colonoscopy quality metrics in a tertiary Australian endoscopy center. Endoscopists were stratified by CADe usage, revealing that even partial use (50%-99% of procedures) significantly improved adenoma detection rate (ADR). High-frequency users maintained high ADR, but total adenomas per procedure increased, highlighting additional quality benefits. The findings emphasize that CADe effectiveness depends not only on technology availability but also on endoscopist level of usage, offering insights into optimizing implementation and improving patient outcomes through enhanced polyp detection and risk stratification.

INTRODUCTION

Colorectal cancer (CRC) is highly prevalent worldwide[1]. It is the fourth most common cancer in Australia, but the second in number of cancer-related deaths behind lung cancer[2]. Colonoscopy remains the gold standard for CRC screening, playing a pivotal role in prevention of CRC by detecting and removing precancerous and early cancerous polyps. Adenoma detection rate (ADR), defined as the percentage of colonoscopies that detect at least one adenomatous polyp, is an important marker of colonoscopy quality. ADR has shown to be inversely correlated to the risk of post-colonoscopy CRC and cancer mortality/morbidity[3]. Many Gastrointestinal Societies are increasingly advocating for higher minimum standards for ADRs, reflecting the demonstrated benefits of elevated ADRs in multiple observational studies[4]. The American College of Gastroenterology recently raised their recommended ADR from 25% to 35%, advising physicians with lower ADRs to pursue measures to improve their detection rates[5]. Several factors affect ADRs, including recency of training[6], operator fatigue[7], bowel preparation quality[8], the use of distal attachment devices[9] and artificial intelligence (AI).

Since the early 2000s, AI has been gradually integrated into gastroenterology practice with the most prominent example being computer-aided detection (CADe). CADe aims to increase the detection of suspicious colorectal lesions, usually highlighting any suspicious lesions on screen with a green box around the area of interest. Multiple multicenter randomized controlled trials (RCTs) have demonstrated the efficacy of CADe in increasing ADR[10-13], with studies predominantly conducted in China, United States of America and Europe.

Despite the promising results of RCTs, real world studies have revealed more variable outcomes. A recent meta-analysis by Wei et al[14] revealed that CADe only modestly improved ADR compared to conventional colonoscopy, with this effect only noted in prospective studies and not in retrospective studies. The discrepancy between RCTs and real-world studies may be attributed to numerous factors, including more effective bowel preparation in RCTs, heightened operator awareness of being observed in controlled settings, variations in endoscopy proficiency, and differences in the familiarity and uptake of CADe.

The availability of CADe in Australian settings is inconsistent, and even where it is accessible, its uptake amongst endoscopists is variable. The only observational study conducted in Australia to date in this area, by Goetz et al[15], observed the effect of CADe in a single center comparing endoscopist ADRs from colonoscopies before and after implementation of CADe. The study found no statistically significant improvement in ADR following the implementation of CADe technology (52.1% for CADe vs 52.6% for conventional colonoscopy). This result may be related to the high baseline ADR in this study. As shown in previous studies, the impact of CADe may be more significant for endoscopists with lower baseline ADR, with Biscaglia et al[16] comparing the ADRs between trainee endoscopists using CADe with those of experienced endoscopists using conventional colonoscopy and noting no significant difference between the two groups. This is important as the real-world setting encompasses a wide range of experience and usage levels among endoscopists.

This study evaluated the real-world impact of the degree of endoscopist use of CADe on adenoma detection in an Australian tertiary endoscopy center where CADe was available for all procedures.

MATERIALS AND METHODS

Study design and patients

This single-center retrospective observational study, conducted at a tertiary hospital in Melbourne, Australia, evaluated the impact of CADe availability on ADR according to endoscopists’ levels of CADe use during screening, diagnostic, and surveillance colonoscopies. The study was granted ethics approval by the Peninsula Health Human Research Ethics Committee prior to commencement. Patient data was handled in accordance with institutional privacy requirements; all extracted data was de-identified prior to analysis to ensure confidentiality and protect patient privacy. All colonoscopies performed six months prior and six months after the introduction of CADe were considered. Inclusion criteria included: (1) Patients aged 40 years and older; (2) Colonoscopies performed between January 2023 and February 2024; and (3) Procedures conducted for screening, surveillance, or diagnostic purposes. Exclusion criteria included: (1) Procedures not documented in ENDOBASE; (2) Emergency procedures (e.g., for acute colitis, foreign body extraction, suspected active bleeding, or acute obstruction); (3) Planned procedures for polyp removal; or (4) Endoscopists who did not have eligible colonoscopies reported in both the pre- and post-CADe time periods. Endoscopists received no formal training or orientation on CADe usage as part of the study.

We included colonoscopies performed for both screening/surveillance and diagnostic indications in our primary analysis. This decision was guided by evidence demonstrating that ADR is comparable across different procedural indications.

Data collection

A search query was run via the hospital’s endoscopy report documentation software ENDOBASE looking for procedures performed in the six months before and after the introduction of CADe on July 31, 2023. A repeat search query was run with the assistance of the hospital health and information services team to ensure no procedures were missed. The Endoscopy Nurse Unit Manager, in collaboration with the regular endoscopy nursing team, evaluated the frequency of CADe use for each endoscopist based on recalling their collective direct observation during procedures. Endoscopists were classified into four usage categories: (1) CADe used for all (100%) of procedures; (2) CADe used for most (between 50% and 99%) of the procedures; (3) CADe used for some/Less than half (between 1% and 50%) of the procedures; and (4) CADe not used at all (0%). These thresholds were selected to reflect differences in real-world usage patterns. For each procedure, data collected included patient age, sex, indication for colonoscopy, use of CADe, intubation and total procedure time (including time for polypectomy and any additional procedures), number of polyps sent for histopathological analysis, Boston Bowel Preparation Scale (BBPS), and number of histologically confirmed adenomas. Total procedure time was used as a surrogate for withdrawal time, as withdrawal time was not consistently recorded across all colonoscopies, whereas procedure start and end times were reliably documented.

Outcomes and definitions

Colonoscopy reports from January 2023 to February 2024 were reviewed, with data collected on adenomas, sessile serrated lesions (SSLs), and overall polyps detected for each procedure. Additional procedural factors, including the BBPS score and total procedure time, were also recorded. The primary outcome was the change in ADR before and after CADe introduction, stratified by the level of CADe use. Secondary outcomes included polyp detection rate (PDR), SSL detection rate (SSL-DR), adenoma per patient (APP) and SSL per patient (SPP).

Statistical analysis

SPSS Statistics software was used for statistical analysis, with a significance level set at P < 0.05. The comparison of the basal clinical and demographic characteristics was performed by the χ² test of categorical variables and paired t test for continuous variables.

To examine the independent association of CADe availability and adenoma detection, we performed a multivariate regression analysis using a logistic model with mixed-effects modeling. Clustering by endoscopist was addressed by including a random intercept for endoscopist in the model. This allowed us to estimate the odds of adenoma detection after adjusting for the following potential confounders: Patient age, procedure timing (AM vs PM), total procedure duration, colonoscopy indication (screening, surveillance, or diagnostic), and quality of bowel preparation (BBPS score).

A sensitivity analysis was performed to assess whether the effect of CADe differed according to procedural indication. Mixed-effects logistic regression models were repeated separately for screening, surveillance and diagnostic colonoscopies. Each stratified model included the same covariates as the primary analysis (age, procedure timing, bowel preparation quality, and total procedure duration) and incorporated a random intercept for endoscopist to account for clustering.

RESULTS

Patient characteristics at baseline

A total of 1374 colonoscopies were identified, of which 352 were excluded, leaving 1022 procedures for analysis. Among these, 636 were performed before the introduction of CADe. Procedures were stratified based on endoscopist CADe usage patterns (Figure 1):

Figure 1 Flow chart of the inclusion/exclusion process in the two computer-aided detection periods. Colonoscopies for planned polyp removal and emergencies were excluded. CADe: Computer-aided detection.

100% CADe use group: Two endoscopists performed 493 colonoscopies (226 pre-CADe, 267 post-CADe).

50%-99% CADe use group: Four endoscopists performed 498 colonoscopies (393 pre-CADe, 105 post-CADe).

0% CADe use group: One endoscopist performed 31 colonoscopies (17 pre-CADe, 14 post-CADe).

The mean age of patients was significantly higher in the post-CADe group compared to the pre-CADe group (59.4 years vs 63.2 years, P < 0.001). The proportion of male patients was nearly identical between groups (47.9% vs 47.8%, P = 0.986). The most common indications for colonoscopy included iron deficiency anaemia, positive faecal occult blood test, and polyp surveillance. While most indications were comparable between study periods, iron deficiency (with or without anaemia) occurred at a significantly higher proportion in the post-CADe group (Table 1). As well, median total procedure time was significantly longer in the post-CADe group compared to the pre-CADe group (23 minutes vs 18 minutes, P < 0.05).

Table 1 Patient baseline characteristics, n (%).

Characteristic	Pre-CADe	Post-CADe	P value
Age (years)	59.4 (40-89)	63.2 (41-87)	< 0.001a
Male	299 (47.8)	173 (47.9)	0.986
Indication: Iron deficiency ± anaemia	117 (17.5)	94 (19.0)	< 0.05b
Indication: Faecal occult blood positive	72 (10.8)	70 (14.1)	0.087
Indication: Polyp surveillance	76 (11.4)	63 (12.7)	0.497
Average total BBPS score	7.3 (3-9)	7.4 (3-9)	0.744
Median total procedure time (minute)	18 (1-169)	23 (1-408)	< 0.05

^aP < 0.001.

^bP < 0.05.

CADe: Computer-aided detection; BBPS: Boston Bowel Preparation Scale.

Effect of CADe on PDRs

For endoscopists who used CADe in 100% of cases, PDR had no statistically significant increase: From 82.3% pre-CADe to 78.7% post-CADe [odds ratio (OR) = 0.79, 95%CI: 0.51-1.24, P = 0.31], as well as ADR: From 64.2% to 58.1% (OR = 0.77, 95%CI: 0.54-1.11, P = 0.17), and SSL-DR (19.5% pre vs 18.7% post, OR = 0.83, 95%CI: 0.61-1.50, P = 0.83). Among endoscopists who adopted CADe in 50% to 99% of procedures, there was a significant increase in both PDR (46.3% pre vs 63.8% post, OR = 2.04, 95%CI: 1.31-3.19, P < 0.01) and ADR (29.0% pre vs 41.9% post, OR = 1.77, 95%CI: 1.13-2.75, P = 0.01). SSL-DR also increased numerically from 9.7% to 11.4%, but this was not statistically significant (OR = 1.2, 95%CI: 0.61-2.40, P = 0.59). For the single endoscopist who did not use CADe, PDR (52.9% pre vs 50.0% post, OR = 0.88, 95%CI: 0.22-3.66, P = 0.87, ADR (23.6% to 14.3% (OR = 0.54, 95%CI: 0.083-3.51, P = 0.52) and SSL-DR (11.8% to 21.4% (OR = 2.04, 95%CI: 0.29-14.3, P = 0.47), did not show statistical difference (Table 2).

Table 2 Impact of computer-aided detection use frequency on polyp detection rate, adenoma detection rate, and sessile serrated lesion detection rate.

CADe use (n)	PDR % (No. of procedures)				ADR % (No. of procedures)				SSL-DR % (No. of procedures)
	Pre	Post	OR (CI)	P value	Pre	Post	OR (CI)	P value	Pre	Post	OR (CI)	P value
100% (2)	82.3 (226)	78.7 (267)	0.79 (0.51-1.24)	0.31	64.2 (226)	58.1 (267)	0.77 (0.54-1.11)	0.17	19.5 (226)	18.7 (267)	0.83 (0.61-1.50)	0.83
50%-99% (4)	46.3 (393)	63.8 (105)	2.04 (1.31-3.19)	< 0.01b	29.0 (393)	41.9 (105)	1.77 (1.13-2.75)	0.01b	9.7 (393)	11.4 (105)	1.20 (0.61-2.40)	0.59
0% (1)	52.9 (17)	50.0 (14)	0.88 (0.22-3.66)	0.87	23.6 (17)	14.3 (14)	0.54 (0.083-3.51)	0.52	11.8 (17)	21.4 (14)	2.04 (0.29-14.3)	0.47

^bP < 0.05.

CADe: Computer aided detection; PDR: Polyp detection rate; ADR: Adenoma detection rate; SSL-DR: Sessile serrated lesion detection rate; CI: Confidence interval; OR: Odds ratio.

Effect of CADe on APP

APP and SPP were analyzed across the three CADe usage groups to assess whether CADe availability influenced the number of adenomas and SSLs detected per procedure.

Endoscopists who used CADe in 100% of procedures demonstrated a significant increase in APP, rising from 0.90 pre-CADe to 2.08 post-CADe [relative risk (RR) = 2.31, 95%CI: 1.97-2.73, P < 0.0001]. This suggests a more than twofold increase in adenomas detected per patient with consistent CADe use. SPP rose from 0.137 pre-CADe to 0.187 post-CADe but this change was not significant (RR = 1.36, 95%CI: 0.85-2.21, P = 0.1729).

For the 50%-99% CADe use group, APP did not increase significantly (1.15 pre-CADe to 1.31 post-CADe, RR = 1.14, 95%CI: 0.94-1.39, P = 0.1728). SPP did not change significantly (0.132 pre-CADe to 0.11 post-CADe, RR = 0.86, 95%CI: 0.42-1.64, P = 0.6680).

The 0% CADe use group showed no significant change in APP from 0.29 pre-CADe to 0.21 post-CADe, with a rate ratio of 0.73 (95%CI: 0.11-3.74, P = 0.6906).

Analysis of adenoma detection by location

To assess whether CADe influenced the detection of adenomas in specific anatomical locations, we analyzed the average number of adenomas per colonoscopy in different segments of the colon before and after CADe implementation.

Right-sided adenoma detection (caecum, ascending colon, and hepatic flexure)

100% CADe use group: The right sided average adenoma per colonoscopy increased significantly from 0.28 pre-CADe to 0.89 post-CADe (RR = 3.22, 95%CI: 2.44-4.25, P < 0.0001).

50%-99% CADe use group: The right sided average adenoma per colonoscopy rate increased from 0.49 to 0.52 but this change did not reach significance (RR = 1.07, 95%CI: 0.79-1.44, P = 0.68).

0% CADe use group: No notable improvements in right-sided adenoma detection were observed in this group.

Subgroup analysis excluding diagnostic colonoscopies

A sensitivity analysis was performed excluding all colonoscopies performed for diagnostic indications, to isolate the impact of CADe on ADR and related metrics in screening and surveillance procedures alone. This exclusion greatly reduced the number of procedures available for analysis, particularly in endoscopist groups 2 and 3, significantly limiting statistical power.

In this restricted cohort, no statistically significant differences were observed in PDR, ADR, or SSL-DR across any endoscopist group following CADe implementation (Table 3). For example: In the 100% use group, ADR decreased from 78.4% to 70.6% (P = 0.26), and SSL-DR increased from 19.6% to 22.5% (P = 0.73). In the 50%-99% use group, ADR rose marginally from 47.0% to 50.0% (P = 0.86), while SSL-DR remained essentially unchanged (16.4% vs 15.9%, P = 1.00). The 0% use group sample size was too small to support meaningful comparisons.

Table 3 Sensitivity analysis with diagnostic colonoscopies excluded.

CADe use (n)	PDR % (No. of procedures)			ADR % (No. of procedures)			SSL-DR % (No. of procedures)
	Pre	Post	P value	Pre	Post	P value	Pre	Post	P value
100% of time (2)	93.1 (102)	89.2 (102)	0.46	78.4 (102)	70.6 (102)	0.26	19.6 (102)	22.5 (102)	0.73
50%-99% of time (4)	68.2 (134)	72.7 (44)	0.71	47.0 (134)	50 (44)	0.86	16.4 (134)	15.9 (44)	1.00
0% of time (1)	66.7 (3)	66.7 (3)	1.00	66.7 (3)	0 (3)	0.40	33.3 (3)	33.3 (3)	1.00

CADe: Computer aided detection; PDR: Polyp detection rate; ADR: Adenoma detection rate; SSL-DR: Sessile serrated lesion detection rate; OR: Odds ratio.

Multivariate analysis

A mixed-effects multivariable logistic regression model was performed to evaluate the independent association between CADe availability and adenoma detection, accounting for clustering by individual endoscopist through inclusion of a random intercept. After adjustment for patient age, indication, procedure timing, bowel preparation quality, and total procedure duration, CADe use remained significantly associated with increased adenoma detection (OR = 1.47, 95%CI: 1.07-2.01, P = 0.02), indicating an independent effect of the technology beyond measured procedural and patient-level factors (Table 4).

Table 4 Multivariable logistic regression with mixed effects - assessing factors associated with adenoma detection rate.

Variable	OR	95%CI	P value
CADe era	1.47	1.07-2.01	0.02b
Patient age (years)	1.00	0.99-1.01	0.41
Procedure time of day (PM vs AM)	0.92	0.67-1.26	0.61
Procedure time (minutes)	1.01	1.00-1.02	0.02b
Indication: Surveillance compared to screening	2.19	1.45-3.30	< 0.001a
Indication: Diagnostic compared to screening	1.53	1.08-2.19	0.02b
BBPS score (total)	0.97	0.88-1.06	0.50

^aP < 0.001.

^bP < 0.05.

Random intercept included for endoscopist to account for clustering by endoscopist. Reference group for indication is screening. CADe: Computer-aided detection; OR: Odds ratio; CI: Confidence interval; BBPS: Boston Bowel Preparation Scale.

To assess whether the effect of CADe differed by procedural indication, a multivariate sensitivity analysis was performed stratified into screening, surveillance, and diagnostic colonoscopies. CADe availability did not significantly influence adenoma detection in screening or surveillance procedures; however, in diagnostic colonoscopies, CADe use was independently associated with higher adenoma detection (OR = 1.76, 95%CI: 1.17-2.63, P = 0.006; Table 5).

Table 5 Mixed-effects multivariable regression stratified by procedure indication.

Variable	OR	95%CI	P value
Screening
CADe era	1.30	0.59-2.87	0.52
Patient age (years)	1.02	0.99-1.04	0.23
Procedure time of day (PM vs AM)	0.40	0.18-0.85	0.02
BBPS score (total)	0.87	0.67-1.12	0.29
Procedure time	1.01	0.99-1.04	0.22
Surveillance
CADe era	1.27	0.70-2.29	0.43
Patient age (years)	0.99	0.98-1.01	0.73
Procedure time of day (PM vs AM)	1.11	0.61-2.01	0.72
BBPS score (total)	1.10	0.93-1.32	0.32
Procedure time	1.00	0.99-1.02	0.74
Diagnostic
CADe era	1.76	1.17-2.63	0.006b
Patient age (years)	1.01	0.99-1.02	0.35
Procedure time of day (PM vs AM)	1.10	0.72-1.67	0.67
BBPS score (total)	0.94	0.82-1.07	0.34
Procedure time	1.01	0.99-1.02	0.08

^aP < 0.001.

^bP < 0.05.

Reference group for indication is screening. CADe: Computer-aided detection; CI: Confidence interval; BBPS: Boston Bowel Preparation Scale.

DISCUSSION

The use of CADe has been found to improve certain colonoscopy quality indicators such as ADR in several RCTs[10,11,17]. Currently, it is not established whether this is reflected in clinical practice, with mixed results emerging from real-world studies.

Promisingly, this retrospective cohort study demonstrated that CADe availability significantly improved ADR from a baseline of 29% in the group of endoscopists who used CADe in the majority (50%-99%) of procedures. In comparison the group of endoscopists who used CADe 100% of the time had no statistically significant change in ADR. This is likely attributable to their already high baseline ADR of 64.2%, which exceeds established quality benchmarks and leaves limited room for measurable improvement. This suggests a ceiling effect, whereby endoscopists with strong baseline performance gain proportionally less from adjunctive technologies. Additional factors may also have contributed to the absence of further ADR improvement despite 100% CADe use, such as greater case complexity (both endoscopists using CADe 100% of time were interventional luminal endoscopists).

Consistent with this, other recent observational studies that found endoscopists with no change in ADR post CADe availability tended to have a high baseline ADR, suggesting that a “ceiling effect” may account for the lack of improvement with CADe in these highly proficient endoscopists[18]. For instance, in the only other published Australian retrospective observational study on this topic, Goetz et al[15] found no improvement in ADR following the introduction of CADe, though their endoscopists had a high baseline ADR of 52%. Similarly, in a retrospective trial in the United States, Nehme et al[19] found no significant difference in ADR with CADe availability but also had a high baseline ADR of 49.3%. Interestingly, a retrospective pragmatic trial performed in the United States, by Ladabaum et al[20], found no change to ADR with CADe availability even among endoscopists with baseline performance in the middle and lower tertiles. Conversely, our study suggests that endoscopists with a baseline ADR below the recommended standard of 35% may benefit the most from CADe availability when used in the majority of procedures, although we did not specifically stratify endoscopists by baseline ADR.

Although current CRC screening guidelines focus on individuals aged 45-75 years[21], evidence suggests that polyp transformation via the chromosomal instability pathway occurs over an average of 10 years, supporting the inclusion of patients ≥ 40 years for surveillance and hence this age threshold was used for the study[22,23]. The age difference between groups was statistically significant, however, the four-year gap is unlikely to have resulted in a meaningful increase in adenoma or polyp prevalence, as the risk of colorectal neoplasia typically rises over longer timeframes.

In our initial sensitivity analysis, we excluded diagnostic colonoscopies to isolate the effect of CADe on ADR in a screening and surveillance cohort. This greatly reduced the sample size, particularly among endoscopists in groups 2 and 3, limiting statistical power to detect meaningful differences. Within this smaller dataset, no significant changes were observed in ADR following CADe implementation, and the low case numbers made interpretation difficult. We therefore performed a second, more robust sensitivity analysis using mixed-effects logistic regression stratified by indication. In this analysis, CADe did not significantly affect ADR in screening or surveillance procedures but remained independently associated with higher adenoma detection in diagnostic colonoscopies. This may reflect both methodological and clinical factors: Diagnostic procedures comprised most of the cases in this real-world cohort, providing greater statistical power, and they typically involve higher clinical suspicion and lesion prevalence, giving CADe more opportunity to add incremental benefit.

These limitations reinforce the rationale for our primary analytic approach, which included diagnostic, surveillance and screening colonoscopies. This decision aligns with real-world clinical practice and reflects the design of major CADe studies and guideline recommendations. Moreover, prior evidence has demonstrated that ADR is consistent across both screening and non-screening indications, supporting its use as a valid quality measure irrespective of indication[24]. Including diagnostic procedures allowed us to preserve statistical power, maintain generalizability, and evaluate the performance of CADe in a setting more representative of routine endoscopic practice.

Compared to adenomas, SSLs are more easily missed given their indistinct borders, bland color and flat morphology. In examining SSL-DR as a secondary outcome, there was no statistically significant change in SSL-DR for all three groups regardless of CADe use. This is similar to previous research, such as a 2023 meta-analysis which showed no statistically significant improvement in SSL detection with CADe compared to a clear improvement in adenoma detection[24]. A potential explanation for this may be that most CADe systems are not focused on detecting SSLs, with only a few describing its use in their training dataset[25]. Another possibility is the role that endoscopist discretion may play in interpreting the lesion highlighted by CADe and determining whether to proceed with polypectomy.

Importantly, in the group of endoscopists who used CADe in 100% of cases, the total number of APP significantly increased, effectively doubling detection. Therefore, even when ADR did not further improve due to an already high baseline, CADe still contributed to identifying a greater number of APP. This may reflect subtle quality benefits beyond just ADR, including detection of additional adenomas that may otherwise be missed, identification of smaller or flat lesions, and improved risk stratification to guide interval surveillance colonoscopy. While there was an absolute increase in APP in the 50%-99% CADe-use group (from 1.15 to 1.31), this change was not statistically significant.

The median total procedure time for colonoscopies was longer in the post-CADe era compared to pre-CADe procedures, a difference that was statistically significant. This increase in time may reflect additional time taken to remove the greater number of polyps and adenomas detected with the aid of CADe. However, it may also indicate that the presence of CADe acts as a visual prompt for endoscopists to examine the mucosa carefully, thus improving detection rates. Importantly, our multivariate analysis confirmed that the observed increase in adenoma detection associated with CADe use persisted even after adjusting for procedure duration, as well as other potential confounders including patient age, indication, time of day, and bowel preparation quality. This suggests that the improvement in ADR likely reflects a true additive effect of CADe technology.

Although scopes done for iron-deficiency were more frequent in the post-CADe period, this alone is unlikely to explain the observed improvement in detection. Iron-deficiency anaemia is a heterogeneous indication, and while it may increase the baseline likelihood of finding pathology, it does not uniformly predict higher adenoma yield. As well, our multivariate model, which adjusted for indication and other key confounders, showed that CADe use remained independently associated with higher adenoma detection.

A major strength of this study is that the effects of CADe availability were examined with endoscopists unaware that colonoscopy quality metrics would be assessed and compared. This contrasts with previous RCTs performed where endoscopists were aware of being observed and may have been more vigilant in their procedures. Furthermore, as a real-world study, where CADe use is left to the discretion of endoscopists, these results are more generalizable to clinical practice and may provide practical input to hospital management and heads of units/departments on the value of CADe even though it may not be adopted 100% of the time.

Another useful insight from this study relates to the endoscopist who did not utilise CADe for any procedures during the study period. Having the lowest baseline ADR of the groups, this endoscopist ostensibly would have benefited the most from using CADe. This demonstrates that despite availability and integration into departmental workflow, variation in attitudes and willingness to engage with new technology can significantly influence usage rates. The non-use observed in this case underscores that barriers to CADe uptake are not solely technical, but also practitioner dependent. Understanding these factors will be essential for improving real-world implementation and maximising the potential quality benefits of CADe.

We acknowledge there are several limitations applicable to this study. Various potential confounding procedural factors were not controlled for, which can impact quality metrics, such as the use of a distal attachment cap or a second pass/retroflexion in the right colon.

Moreover, there was an uneven distribution of procedures pre- and post-CADe among endoscopists, which limits interpretation. In addition, as no endoscopists in this study utilised CADe between 1% and 49% of procedures, we were unable to assess whether use of CADe in the minority of cases had the potential to improve ADR. As well, while a reasonable cohort of patients were examined in this study, there was only a limited number of endoscopists involved.

Withdrawal time was not consistently recorded; as a result, total procedure time was used as a surrogate, which incorporates insertion and therapeutic time in addition to withdrawal observation. Because CADe use may increase polyp detection and therefore therapeutic time, longer total procedure times in the CADe group may partly reflect time spent on polypectomy rather than solely longer withdrawal observation, representing a key potential confounder. A follow-up prospective study using consistently recorded withdrawal time would help to clarify the independent impact of CADe on withdrawal observation time.

Finally, this study included diagnostic as well as screening and surveillance colonoscopies, which may introduce heterogeneity in baseline ADR. Our subsequent sensitivity analyses demonstrated that the effect of CADe was not uniform across indications. While CADe did not significantly change ADR in screening or surveillance procedures, it remained independently associated with higher adenoma detection in diagnostic colonoscopies. This likely reflects the greater prevalence of pathology and higher clinical suspicion in diagnostic procedures, as well as the larger sample size within this subgroup, providing greater statistical power to detect differences. Nevertheless, these findings support a real-world benefit of CADe in improving lesion detection, particularly in diagnostic settings.

CONCLUSION

In this real-world study, the implementation of CADe was associated with a significant improvement in ADR, even when not being used in all procedures. The ADR improvement was not shown in endoscopists with a high baseline ADR, however in this subgroup there was significant increase in APP. SSL-DR and SPP were not improved with the addition of CADe. CADe was associated with increased total procedure time, however on multivariate analysis, CADe was independently associated with increase ADR. Overall, these findings suggest that CADe holds benefits even without 100% endoscopist use.