This cohort study evaluates the role of artificial intelligence in the association between time of day and colonoscopy quality.
Key Points
Question
Can artificial intelligence (AI) systems eliminate time-related declines in adenoma detection during colonoscopy?
Findings
In this cohort study including 1780 colonoscopy procedures, in the unassisted group, the adenoma detection rate (ADR) at late sessions was significantly higher compared with that of early sessions, while with the assistance of AI systems, no such statistically significant difference was found.
Meaning
These findings suggest AI systems may be a potential tool for minimizing time-related degradation of colonoscopy quality.
Abstract
Importance
Time of day was associated with a decline in adenoma detection during colonoscopy. Artificial intelligence (AI) systems are effective in improving the adenoma detection rate (ADR), but the performance of AI during different times of the day remains unknown.
Objective
To validate whether the assistance of an AI system could overcome the time-related decline in ADR during colonoscopy.
Design, Setting, and Participants
This cohort study is a secondary analysis of 2 prospective randomized controlled trials (RCT) from Renmin Hospital of Wuhan University. Consecutive patients undergoing colonoscopy were randomly assigned to either the AI-assisted group or unassisted group from June 18, 2019, to September 6, 2019, and July 1, 2020, to October 15, 2020. The ADR of early and late colonoscopy sessions per half day were compared before and after the intervention of the AI system. Data were analyzed from March to June 2022.
Exposure
Conventional colonoscopy or AI-assisted colonoscopy.
Main Outcomes and Measures
Adenoma detection rate.
Results
A total of 1780 patients (mean [SD] age, 48.61 [13.35] years, 837 [47.02%] women) were enrolled. A total of 1041 procedures (58.48%) were performed in early sessions, with 357 randomized into the unassisted group (34.29%) and 684 into the AI group (65.71%). A total of 739 procedures (41.52%) were performed in late sessions, with 263 randomized into the unassisted group (35.59%) and 476 into the AI group (64.41%). In the unassisted group, the ADR in early sessions was significantly higher compared with that of late sessions (13.73% vs 5.70%; P = .005; OR, 2.42; 95% CI, 1.31-4.47). After the intervention of the AI system, as expected, no statistically significant difference was found (22.95% vs 22.06%, P = .78; OR, 0.96; 95% CI; 0.71-1.29). Furthermore, the AI systems showed better assistance ability on ADR in late sessions compared with early sessions (odds ratio, 3.81; 95% CI, 2.10-6.91 vs 1.60; 95% CI, 1.10-2.34).
Conclusions and Relevance
In this cohort study, AI systems showed higher assistance ability in late sessions per half day, which suggests the potential to maintain high quality and homogeneity of colonoscopies and further improve endoscopist performance in large screening programs and centers with high workloads.
Introduction
Colorectal cancer (CRC) is the third most common cancer worldwide.1 Prevention strategies—especially screening colonoscopy—can substantially reduce CRC incidence and mortality via detection and removal of precancerous polyps.2,3 Adenoma detection rate (ADR) is a critical quality measure of screening colonoscopy and each 1.0% increase in ADR is associated with a 3.0% decrease in the risk of interval CRC (hazard ratio, 0.97; 95% CI, 0.96-0.98).4 Ensuring high ADR is hence paramount in the prevention of CRC.
Time of day has been identified as an indispensable factor related to suboptimal ADR. Sanaka et al5 first reported performance time of the day as an independent risk factor of colonoscopy quality. Subsequently, substantial evidence has shown that morning colonoscopies are associated with improved ADR.6,7,8,9 According to a previous study,7 the afternoon ADR was significantly lower than in the morning (21.0% vs 26.1%; P = .02) for endoscopists working the full day. In addition, Lee et al10 reported that each elapsed hour in the day was associated with a 4.6% reduction in polyp detection (P = .005). Continuous and repetitive visual stimuli may lead to weaker reliable response and poorer judgment, which jeopardize diagnostic abilities and efficiency. Unfortunately, an objective approach to minimize time-related degradation of colonoscopy quality is still lacking.
According to available evidence, the incorporation of artificial intelligence (AI) as an aid for colonoscopy results in a significant increase in ADR.11 There are 2 main reasons for missed polyps: polyps that are unrecognized inside the visual field due to repetitive movement and monotonous vision and missed polyps outside the visual field because endoscopists are more rushed through the course of the day, so blind spots increase during the withdrawal phase.12,13 In our previous studies, we have developed a deep-learning based computer-aided detection (CADe) system to augment detection of polyps inside the visual field and a computer-aided quality improvement (CAQ) system to improve withdrawal visualization quality.14,15 Through randomized clinical trials (RCT), we have demonstrated its positive association with increasing ADR.14,15 However, the extra benefit of AI systems in eliminating the time-related decline of ADR remains unknown.
Therefore, we hypothesized that AI systems may be an effective tool to address the previously stated problem. The aim of this study was to validate whether the assistance of AI systems could overcome the time-related degradation of colonoscopy quality according to the data from our previous randomized clinical trials.
Method
Study Cohort
This study is a secondary analysis of 2 RCTs. Patients undergoing screening, diagnosis, and surveillance colonoscopy were included in this study, but not those undergoing therapeutic colonoscopy. Details of the interventions are provided in the eMethods in Supplement 1. RCT 1 was a prospective, randomized parallel group study to evaluate whether the CAQ system could improve ADR during colonoscopy. Consecutive patients who were able to give informed consent were randomly assigned (1:1) to either the CAQ colonoscopy group (355 patients) or the unassisted (control) colonoscopy group (349 patients) between June 18, 2019, and September 6, 2019. The random allocation sequence was a computer-generated random number series, and randomization was done in blocks of 4. All the colonoscopies were performed by 6 endoscopists (including J.Z. and X.H.) with endoscopy experience of 1 to 3 years and a total colonoscopy volume of 1500 to 4000.14 RCT 2 was a prospective, single-center, 4-group, randomized parallel study to evaluate the interaction effect of improving the ADR between a CADe system for colorectal polyp detection and a CAQ system for real-time withdrawal speed monitoring. Between July 1, 2020, and October 15, 2020, 1076 participants were randomly allocated into 4 treatment groups (control: 271; CADe: 268; CAQ: 269; and CADe plus CAQ [combination]: 268). An electronic digital capture system was used to generate a random number of each patient before the procedure, and randomization was done in blocks of 16. All the colonoscopies were performed by 4 endoscopists (L.W., J.Z., W.Z., and C.H.) with more than 2000 screening colonoscopies in experience.15 The ethics committees of Renmin Hospital of Wuhan University approved both protocols, and all participants provided written informed consent. We followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline.
Definition
In this study, the AI group was defined as the combination of all intervention groups, including the CAQ group in RCT 1 and the CAQ group, CADe group, and combination group in RCT 2. The combined control (control-c) group was defined as the combination of control groups in RCT 1 and RCT 2. For single computer-aided functions, the combined CAQ (CAQ-c) group was defined as the combination of the CAQ group in RCT 1 and the CAQ group in RCT 2.
Full-day colonoscopy procedures were included in this study. A half day was defined as morning (earlier than 1 pm) and afternoon (1 pm or later) according to previous studies.16 Since this study was focused on the process of withdrawal, to avoid the influence of long-time insertion on the research group, we used the end time of the procedure to define early and late sessions, which is closer to the process of examination. All colonoscopy procedures were divided into 2 groups according to the end time of procedure. The first group consisted of procedures commenced in the early session per half day (8:00 am to 10:59 am or 1:00 pm to 2:59 pm), which was defined as the early group, and the second group consisted of procedures commenced in the later session per half day (11:00 am to 12:59 pm or 3:00 pm to 4:59 pm), which was defined as the late group. We divided endoscopists into 3 groups according to their experience, high experience (≥3000 colonoscopies), middle experience (2000-3000 colonoscopies), and low experience (<2000 colonoscopies). Detailed definitions of ADR, polyp detection rate (PDR), proximal colon, distal colon and advanced adenoma were provided in the eMethods in Supplement 1.
Statistical Analysis
Baseline characteristics are reported with early sessions and late sessions. Continuous variables were reported as mean with SD, while categorical variables were reported as frequencies and percentages. Comparison of baseline clinical and demographic characteristics between the early and the late session was performed using a χ2 test for categorical variables and using a 2-sample t test for continuous variables. Logistic regression was applied to assess the association between the intervention and the ADR/PDR. A covariate-adjusted logistic regression model was built by adding group differences into the models as covariates to address possible confounding factors, including withdrawal time, endoscopist experience, bowel preparation, age, sex, body mass index (BMI; calculated as weight in kilograms divided by height in meters squared), indication for colonoscopy, inpatient or outpatient status, and endoscope vendors. A Poisson regression was performed to assess the association between the intervention and the mean number of adenomas and polyps per colonoscopy. Odds ratios (ORs) with 95% CIs and corresponding P values were calculated. All reported P values are 2-sided, and P values of less than .05 were considered statistically significant. All calculations were performed using SPSS version 26 (IBM). Data were analyzed from March to June 2022.
Results
Characteristics of the Study Cohort
A total of 1780 patients (mean [SD] age, 48.61 [13.35] years, 837 [47.02%] women) were enrolled. Overall, 58.48% of procedures (1041 of 1780) were performed in early sessions, with 34.29% (357 of 1041) randomized into the control-c group and 65.71% (684 of 1041) into the AI group. A total of 41.52% of procedures (739 of 1780) were performed in late sessions, with 35.59% (263 of 739) randomized into the control-c group and 64.41% (476 of 739) into the AI group. There were 863 polyps and 326 adenomas detected in the study cohort. Baseline characteristics are presented in Table 1.
Table 1. Baseline Characteristics.
| Characteristic | Patients, No. (%) | P value | |
|---|---|---|---|
| Early session (n = 1041) | Late session (n = 739) | ||
| Age, mean (SD), y | 50.14 (13.21) | 48.53 (13.59) | .01 |
| Sex | |||
| Male | 497 (47.74) | 340 (46.01) | .47 |
| Female | 544 (52.26) | 399 (53.99) | |
| Body mass indexa | |||
| <25 | 791 (75.98) | 594 (80.38) | .05 |
| 25 to <30 | 223 (21.42) | 124 (16.78) | |
| ≥30 | 27 (2.59) | 21 (2.84) | |
| Indication of colonoscopy | |||
| Screening | 670 (64.36) | 414 (56.02) | <.001 |
| Diagnostic | 283 (27.19) | 276 (37.35) | |
| Surveillance | 88 (8.45) | 49 (6.63) | |
| Inpatient | 190 (18.25) | 158 (21.38) | .10 |
| Outpatient | 851 (81.75) | 581 (78.62) | |
| Endoscope vendor | |||
| Olympus | 686 (65.90) | 482 (65.22) | .09 |
| Fujifilm | 304 (29.20) | 203 (27.47) | |
| Sonoscape | 51 (4.90) | 54 (7.31) | |
| Withdrawal time, mean (SD), min | 8.21 (4.33) | 7.81 (4.38) | .07 |
| Bowel preparation | |||
| Boston bowel preparation scale | |||
| <2 In any segment | 125 (12.01) | 80 (10.83) | .44 |
| ≥2 In all segments | 916 (87.99) | 659 (89.17) | |
| Endoscopist experience | |||
| High experience | 465 (44.67) | 347 (46.96) | .42 |
| Mid experience | 472 (45.34) | 312 (42.22) | |
| Low experience | 104 (9.99) | 80 (10.83) | |
Body mass index is calculated as weight in kilograms divided by height in meters squared.
Association of Time of Day and ADR/PDR
In the control-c group, the ADR at early sessions was significantly higher compared with that of late sessions (early vs late, 13.73% vs 5.70%; P = .005; OR, 2.42; 95% CI, 1.31-4.47). After the intervention of the AI system, as expected, no statistically significant difference was found in ADR between early and late sessions (early vs late, 22.95% vs 22.06%; P = .78; OR, 0.96; 95% CI; 0.71-1.29) or PDR (early vs late, 55.99% vs 52.31%; P = .57; OR, 1.08; 95% CI, 0.83-1.40). Also, the CADe group (early vs late, 25.30% vs 14.71%; P = .08; OR, 1.97; 95% CI, 0.93-4.17), CAQ-c group (early vs late, 19.48% vs 20.00%; P = .69; OR, 0.92; 95% CI, 0.60-1.39), and combination group (early vs late, 27.81% vs 35.35%; P = .13; OR, 0.63; 95% CI, 0.34-1.15) showed no statistically significant difference in adenoma detection between early and late sessions. Subgroup analysis of different bowel preparation, withdrawal time, and endoscopist experience showed the same association (Table 2). Other detection rates, such as advanced adenoma and analysis in different RCTs, are shown in eTable 1 and eTable 2 in Supplement 1. Detailed analysis among individual endoscopists who performed more than 150 colonoscopies in this study is shown in eTable 3 in Supplement 1.
Table 2. Comparison of Detection Rates Between Early and Late Session.
| Characteristic | ADR, No. (%) | P value | OR (95% CI) | PDR, No. (%) | P value | OR (95% CI) | ||
|---|---|---|---|---|---|---|---|---|
| Early | Late | Early | Late | |||||
| Unassisted group | 49 (13.73) | 15 (5.70) | .005 | 2.42 (1.31-4.47) | 148 (41.46) | 83 (31.56) | .12 | 1.34 (0.93-1.92) |
| BBPS | ||||||||
| ≥2 In all segments | 45 (14.24) | 14 (5.81) | .006 | 2.45 (1.29-4.66) | 128 (40.51) | 75 (31.12) | .14 | 1.34 (0.91-1.95) |
| <2 In any segment | 4 (9.76) | 1 (4.55) | .35 | 4.71 (0.18-124.63) | 20 (48.78) | 8 (36.36) | .96 | 0.97 (0.25-3.83) |
| Withdrawal time, min | ||||||||
| ≥6 | 33 (17.46) | 8 (7.08) | .02 | 2.68 (1.17-6.14) | 92 (48.68) | 45 (39.82) | .32 | 1.29 (0.78-2.15) |
| <6 | 16 (9.52) | 7 (4.67) | .14 | 2.03 (0.79-5.25) | 56 (33.33) | 38 (25.33) | .14 | 1.51 (0.88-2.59) |
| Endoscopist experience | ||||||||
| High experience | 9 (16.07) | 5 (12.20) | .52 | 1.51 (0.43-5.25) | 29 (51.79) | 14 (34.15) | .22 | 1.85 (0.70-4.91) |
| Mid experience | 19 (15.08) | 3 (3.06) | .006 | 6.42 (1.71-24.10) | 47 (37.30) | 27 (27.55) | .41 | 1.30 (0.69-2.45) |
| Low experience | 21 (12.00) | 7 (5.65) | .18 | 1.88 (0.75-4.72) | 72 (41.14) | 42 (33.87) | .49 | 1.20 (0.71-2.01) |
| AI group | 157 (22.95) | 105 (22.06) | .78 | 0.96 (0.71-1.29) | 383 (55.99) | 249 (52.31) | .57 | 1.08 (0.83-1.40) |
| CADe | 42 (25.30) | 15 (14.71) | .08 | 1.97 (0.93-4.17) | 102 (61.45) | 47 (46.08) | .12 | 1.58 (0.89-2.81) |
| Combined CAQ group | 68 (19.48) | 55 (20.00) | .69 | 0.92 (0.60-1.39) | 171 (49.00) | 140 (50.91) | .52 | 0.89 (0.63-1.26) |
| Combined CADe and CAQ group | 47 (27.81) | 35 (35.35) | .13 | 0.63 (0.34-1.15) | 110 (65.09) | 62 (62.63) | .65 | 1.15 (0.63-2.08) |
| BBPS | ||||||||
| ≥2 In all segments | 135 (22.50) | 91 (21.77) | .63 | 0.93 (0.67-1.27) | 345 (57.50) | 219 (52.39) | .41 | 1.12 (0.85-1.48) |
| <2 In any segment | 22 (26.19) | 14 (24.14) | .50 | 1.35 (0.56-3.26) | 38 (45.24) | 30 (51.72) | .48 | 0.76 (0.36-31.63) |
| Withdrawal time, min | ||||||||
| ≥6 | 144 (25.90) | 91 (23.51) | .88 | 1.02 (0.74-1.41) | 336 (60.43) | 215 (55.56) | .51 | 1.10 (0.83-1.47) |
| <6 | 13 (10.16) | 14 (15.73) | .28 | 0.62 (0.26-1.48) | 47 (36.72) | 34 (38.20) | .99 | 1.00 (0.54-1.85) |
| Endoscopist experience | ||||||||
| High experience | 6 (12.50) | 12 (30.77) | .06 | 0.34 (0.11-1.06) | 19 (39.58) | 22 (56.41) | .21 | 0.55 (0.22-1.39) |
| Mid experience | 91 (26.30) | 47 (21.96) | .43 | 1.20 (0.77-1.86) | 200 (57.80) | 109 (50.93) | .29 | 1.23 (0.84-1.81) |
| Low experience | 60 (20.69) | 46 (20.63) | .74 | 0.93 (0.58-1.47) | 164 (56.55) | 118 (52.91) | .80 | 1.05 (0.71-1.56) |
Abbreviations: ADR, adenoma detection rate; BBPS, Boston bowel preparation scale; CADe, computer-aided detection; CAQ, computer-aided quality improvement; OR, odds ratio; PDR, polyp detection rate.
Association of Time of Day and AI Assistance Capability
The assistance capability of AI systems was compared through OR values.17 For adenoma detection, the ADR was significantly improved after the assistance of AI systems in both early sessions (22.95% vs 13.73%; P = .01; OR, 1.60; 95% CI, 1.10-2.34) and late sessions (22.06% vs 5.70%; P < .001; OR, 3.81; 95% CI, 2.10-6.91). Furthermore, the AI systems had higher assistance capability in late sessions compared with early sessions (OR, late vs early, 3.81; 95% CI, 2.10-6.91 vs 1.60; 95% CI, 1.10-2.34). Subgroup analysis of different bowel preparation, withdrawal time, and endoscopist experience showed the same trend (Table 3).
Table 3. Comparison of AI Assistance Capabilities Between Early and Late Session.
| Outcome | No. | Early session | Late session | ||||
|---|---|---|---|---|---|---|---|
| Detection rate, AI vs unassisted, % | P value | OR (95% CI) | Detection rate, AI vs unassisted, % | P value | OR (95% CI) | ||
| Adenoma detection rate | |||||||
| All procedures | 1780 | 22.95 vs 13.73 | .01 | 1.60 (1.10-2.34) | 22.06 vs 5.70 | <.001 | 3.81 (2.10-6.91) |
| BBPS | |||||||
| ≥2 In all segments | 1575 | 22.50 vs 14.24 | .07 | 1.46 (0.98-2.17) | 21.77 vs 5.81 | <.001 | 3.63 (1.94-6.76) |
| <2 In any segment | 205 | 26.19 vs 9.76 | .08 | 3.39 (0.89-12.98) | 24.14 vs 4.55 | .10 | 9.06 (0.66-123.65) |
| Withdrawal time, min | |||||||
| ≥6 | 1245 | 25.90 vs 17.46 | .006 | 1.85 (1.19-2.87) | 23.51 vs 7.08 | <.001 | 4.24 (1.95-9.19) |
| <6 | 535 | 10.16 vs 9.52 | .79 | 1.12 (0.48-2.61) | 15.73 vs 4.67 | .02 | 3.41 (1.26-9.23) |
| Endoscopist experience | |||||||
| High experience | 184 | 12.50 vs 16.07 | .80 | 0.86 (0.26-2.85) | 30.77 vs 12.20 | .12 | 2.65 (0.78-8.97) |
| Mid experience | 784 | 26.30 vs 15.08 | .03 | 1.91 (1.07-3.42) | 21.96 vs 3.06 | .003 | 6.32 (1.85-21.58) |
| Low experience | 812 | 20.69 vs 12.00 | .25 | 1.41 (0.79-2.54) | 20.63 vs 5.65 | .001 | 4.03 (1.62-10.02) |
| Polyp detection rate | |||||||
| All procedures | 1780 | 55.99 vs 41.46 | .005 | 1.53 (1.14-2.06) | 52.31 vs 31.56 | .001 | 1.90 (1.32-2.73) |
| BBPS | |||||||
| ≥2 In all segments | 1575 | 57.50 vs 40.51 | .001 | 1.67 (1.22-2.28) | 52.39 vs 31.12 | <.001 | 1.99 (1.36-2.93) |
| <2 In any segment | 205 | 45.24 vs 48.78 | .54 | 0.73 (0.27-1.98) | 51.72 vs 36.36 | .99 | 1.01 (0.28-3.61) |
| Withdrawal time | |||||||
| ≥6 | 1245 | 60.43 vs 48.68 | .001 | 1.87 (1.30-2.69) | 55.56 vs 39.82 | .004 | 1.97 (1.24-3.10) |
| <6 | 535 | 36.72 vs 33.33 | .75 | 1.09 (0.64-1.88) | 38.20 vs 25.33 | .04 | 2.02 (1.04-3.91) |
| Endoscopist experience | |||||||
| High experience | 184 | 39.58 vs 51.79 | .38 | 0.68 (0.29-1.62) | 56.41 vs 34.15 | .10 | 2.59 (0.84-7.95) |
| Mid experience | 784 | 57.80 vs 37.30 | .002 | 2.10 (1.31-3.37) | 50.93 vs 27.55 | .04 | 1.91 (1.04-3.52) |
| Low experience | 812 | 56.55 vs 41.14 | .07 | 1.50 (0.96-2.34) | 52.91 vs 33.87 | .03 | 1.78 (1.05-3.02) |
Abbreviations: AI, artificial intelligence; BBPS, Boston bowel preparation scale; OR, odds ratio.
Further analysis was performed to examine the assistance capability of different AI systems. All AI systems showed higher assistance capability in late sessions compared with early sessions for ADR (OR, CADe: 2.75; 95% CI, 0.98-7.68 vs OR, 1.49; 95% CI, 0.87-2.56; CAQ: OR, 3.56; 95% CI, 1.89-6.72 vs OR, 1.51; 95% CI, 0.98-2.32; combination: OR, 10.40; 95% CI, 3.85-28.05 vs OR, 1.89; 95% CI, 1.12-3.21) (eTable 4 in Supplement 1). AI assistance capability in different RCTs were shown in eTable 5 in Supplement 1.
AI Assistance Capability Over Different Hours
Among the control-c group, a significant decline in ADR was noted for hour passage of the half day with the lowest ADR in the last hour per half day at 2.99% (noon) and 0.00% (4 pm), respectively (Table 4; eFigure in Supplement 1). Correspondingly, AI systems showed the highest assistance capability in the last hour per half day (noon, 20.22% vs 2.99%; P = .009; OR, 8.71; 95% CI, 1.72-44.07; 4 pm, 24.32% vs 0.00%).
Table 4. Comparison of AI Assistance Capabilities Over Different Hours.
| Hour | ADR, No. (%) | P value | OR (95% CI) | PDR, No. (%) | P value | OR (95% CI) | ||
|---|---|---|---|---|---|---|---|---|
| Unassisted | AI | Unassisted | AI | |||||
| 8 am | 6 (8.96) | 17 (22.08) | .07 | 2.98 (0.91-9.76) | 24 (35.82) | 44 (57.14) | .02 | 2.63 (1.17-5.91) |
| 9 am | 17 (15.18) | 63 (27.27) | .15 | 1.63 (0.84-3.14) | 49 (43.75) | 128 (55.41) | .20 | 1.43 (0.83-2.49) |
| 10 am | 20 (17.09) | 53 (22.94) | .19 | 1.56 (0.81-3.01) | 55 (47.01) | 135 (58.44) | .12 | 1.54 (0.89-2.65) |
| 11 am | 11 (11.22) | 53 (26.24) | .006 | 2.98 (1.36-6.54) | 39 (39.80) | 113 (55.94) | .03 | 1.88 (1.06-3.31) |
| Noon | 2 (2.99) | 36 (20.22) | .009 | 8.71 (1.72-44.07) | 21 (31.34) | 92 (51.69) | .03 | 2.23 (1.07-4.66) |
| 1 pm | 4 (8.70) | 16 (15.69) | .83 | 1.16 (0.30-4.52) | 15 (32.61) | 54 (52.94) | .35 | 1.51 (0.64-3.53) |
| 2 pm | 2 (13.33) | 8 (18.60) | .54 | 2.16 (0.18-26.53) | 5 (33.33) | 22 (51.16) | .63 | 1.57 (0.25-9.74) |
| 3 pm | 2 (4.55) | 7 (11.86) | .49 | 1.96 (0.29-13.35) | 12 (27.27) | 25 (42.37) | .91 | 1.08 (0.33-3.49) |
| 4 pm | 0 | 9 (24.32) | NA | NA | 11 (20.37) | 19 (51.35) | .04 | 3.68 (1.07-12.67) |
Abbreviations: AI, artificial intelligence; ADR, adenoma detection rate; NA, not applicable; OR, odds ratio; PDR, polyp detection rate.
AI Assistance Capability Among Different Lesion Sizes and Locations
For adenoma detection, the AI assistance capability increase in the late sessions compared with early sessions was mainly associated with an increase in diminutive (OR, 3.52; 95% CI, 1.90-6.52 vs OR, 1.42; 95% CI, 0.96-1.08), small adenomas (OR, 7.32; 95% CI, 0.93-57.63 vs OR, 3.35; 95% CI, 1.14-9.82) and distal colon adenomas (OR, 4.51; 95% CI, 1.94-10.49 vs OR, 1.40; 95% CI, 0.89-2.23) (Table 5). With the assistance of the CADe system, more diminutive adenomas (OR, 2.42; 95% CI, 0.97-5.99 vs OR, 1.59; 95% CI, 0.95-2.68) were detected during late sessions compared with early sessions, while more small adenomas (OR, 9.20; 95% CI, 1.10-76.85 vs OR, 1.09; 95% CI, 0.26-4.63) were detected by the CAQ system. Different AI system assistance capabilities among different lesion sizes and locations are shown in eTable 6 and eTable 7 in Supplement 1.
Table 5. Comparison of AI Assistance Capabilities Among Different Sizes and Locations.
| Capability | Early session | Late session | ||||||
|---|---|---|---|---|---|---|---|---|
| No. (%) | P value | OR (95% CI) | No. (%) | P value | OR (95% CI) | |||
| Unassisted (n = 357) | AI (n = 684) | Unassisted (n = 263) | AI (n = 476) | |||||
| Adenoma detection rate | ||||||||
| Diminutive (≤5 mm) | 47 (13.17) | 139 (20.32) | .08 | 1.42 (0.96-2.08) | 14 (5.32) | 96 (20.17) | <.001 | 3.52 (1.90-6.52) |
| Small (>5 to <10 mm) | 4 (1.12) | 30 (4.39) | .03 | 3.35 (1.14-9.82) | 1 (0.38) | 18 (3.78) | .06 | 7.32 (0.93-57.63) |
| Large (≥10 mm) | 2 (0.56) | 21 (3.07) | .03 | 5.16 (1.16-22.98) | 1 (0.38) | 12 (2.52) | .10 | 5.71 (0.70-46.76) |
| Proximal colon | 22 (6.16) | 85 (12.43) | .01 | 1.95 (1.17-3.25) | 8 (3.04) | 55 (11.55) | .006 | 3.03 (1.37-6.72) |
| Distal colon | 29 (8.12) | 93 (13.60) | .15 | 1.40 (0.89-2.23) | 7 (2.66) | 57 (11.97) | <.001 | 4.51 (1.94-10.49) |
| Polyp detection rate | ||||||||
| Diminutive (≤5 mm) | 142 (39.78) | 364 (53.22) | .01 | 1.47 (1.10-1.98) | 80 (30.42) | 238 (50.00) | .001 | 1.84 (1.28-2.65) |
| Small (>5 to <10 mm) | 10 (2.80) | 41 (6.00) | .06 | 2.00 (0.96-4.18) | 6 (2.28) | 27 (5.67) | .35 | 1.57 (0.61-4.02) |
| Large (≥10 mm) | 3 (0.84) | 28 (4.09) | .01 | 4.62 (1.36-15.75) | 3 (1.14) | 15 (3.15) | .31 | 1.95 (0.53-7.19) |
| Proximal colon | 73 (20.45) | 182 (26.61) | .31 | 1.19 (0.85-1.67) | 37 (14.07) | 120 (25.21) | .04 | 1.60 (1.02-2.51) |
| Distal colon | 113 (31.65) | 293 (42.84) | .03 | 1.39 (1.03-1.87) | 60 (22.81) | 201 (42.23) | <.001 | 2.00 (1.36-2.94) |
Abbreviations: AI, artificial intelligence; OR, odds ratio.
Discussion
In this cohort study, we found that later colonoscopy sessions per half day were associated with a decline in adenoma detection (early vs late, 13.73% vs 5.70%; P = .005; OR, 2.42; 95% CI, 1.31-4.47) during all day shift schedules. Furthermore, with the assistance of AI systems, no such association was found in both ADR (early vs late, 22.95% vs 22.06%; P = .78; OR, 0.96; 95% CI, 0.71-1.29) and PDR (early vs late, 55.99% vs 52.31%; P = .57; OR, 1.08; 95% CI, 0.83-1.40), which suggests that AI systems could overcome the time-related degradation of colonoscopy quality. To the best of our knowledge, this is the first study evaluating the association between AI systems and time of day through a prospective clinical trial. We believe that AI systems will be an effective tool for minimizing the time-related degradation of colonoscopy quality.
Extensive studies have explored the adverse outcome of later time of day on clinical performance in other domains, such as increasing omission errors and false positives among radiologists and decreasing abnormal cytology detection rates on gynecologic cytology samples among cytotechnologists.18,19 Radiologists reported eyestrain and blurred vision with an increased number of images read. They hold an opinion that time of day significantly impacts reading competence-related cognitive functions. This phenomenon is not only manifested for the previously mentioned clinical departments, but also for endoscopists. Korenblit et al20 reported a decline in sensitivity and diagnostic accuracy of endoscopic ultrasound-guided fine needle aspiration for solid pancreatic lesions with progressively later starting times of endoscopic ultrasounds, which combines the technical aspects of endoscopy, real-time interpretation of moving images, and fine needle aspiration sampling. The result was explicitly interpreted as the accumulation of fatigue due to prolonged concentration, comprehensive image review, and demands to read quickly.
Among endoscopists, increasing amounts of colonoscopies posed a sharply increased burden to endoscopists, leading to a decrease of ADR.21 According to a national CRC screening program, ADR reduction was associated with the second half of a session (11 am to 2 pm or 4 pm to 6 pm).9 Our study showed the same trend in colonoscopies without AI intervention. Due to the reliance on stereotypic movements during colonoscopies, as well as the immediate analysis of images, we proposed that increasing procedures may result in associated fatigue and distraction, which in turn leads to cognitive errors and misdiagnosis.6,8,10,22 Furthermore, subgroup analysis was performed to explore whether the conclusion was influenced by confounding factors, such as withdrawal time, bowel preparation, and endoscopist experience.23,24,25 We also found a decline in ADR in late sessions compared with early sessions among different colonoscopy settings. For these aforementioned reasons, we proposed that the accrual of endoscopist fatigue may be an independent factor of time-related degradation of colonoscopy quality.
To address the previously mentioned problem, some studies revealed that by applying a 3-hour colonoscopy shift schedule, an assigned time of 45 minutes per colonoscopy, or a schedule of 6 patients (for an esophagogastroduodenoscopy, colonoscopy, or both) per 4-hour block, the decline in ADR associated with time of day could be eliminated, suggesting that shorter shifts and fewer colonoscopy procedures may be a potential approach to overcome the time-related problem.16,26,27 However, the existing approach had limitations as it is not applicable to many institutions with high workloads and different working schedules. Our study proposed 2 real-time computer-aided systems based on deep learning and demonstrated that the intervention of AI systems was associated with an increase in ADR all-day, especially later sessions per half day, even in different colonoscopy settings. In the future, AI systems have the potential to ensure the quality of colonoscopies in large–workload clinical practices.
Adenomas are missed for 2 main reasons; one is failure to recognize polyps within the visual field due to cognitive limitations, and the other is the existence of blind spots due to technical defects.28 Furthermore, as the day progresses, repetitive stereotypic maneuvering and visual monotony could lead to decreased cognitive levels and increased medical error.8,29 To augment cognition, we developed a CADe system with real-time visual alerts when identifying polyps on the monitor. The ADR was significantly higher in the CADe group compared with that of the control group in our previous study. In addition, the CADe system showed higher assistance capability at late sessions compared with early sessions (OR, 2.75; 95% CI, 0.98-7.68 vs OR, 1.49; 95% CI, 0.87-2.56), which suggests that the CADe system is an ideal approach to address unrecognized polyps, especially at later sessions per half day.
However, the quality of mucosa exposure determined the limits of the CADe system, as polyps may be missed outside of the visual field. To avoid such nonvisualized lesions due to technical defects, we developed a CAQ system for colonoscopy quality improvement. Shorter mean annual withdrawal times were independently associated with lower ADR;25 however, endoscopists tended to slow down withdrawal speed at the end of the procedure to achieve standard time.14 Therefore, our CAQ system not only times the whole withdrawal process, but also monitors real-time withdrawal speed during procedures. In addition, avoiding endoscope slipping might be difficult for some endoscopists, particularly in the hepatic flexure.30 Accordingly, our CAQ system reminds endoscopists of blind spots caused by endoscope slipping to better expose mucosa and apply sufficient examination. Prior studies had estimated the validity of our CAQ system with significantly higher ADR in the CAQ group compared with that of the control group.14,15 Moreover, the CAQ system showed higher assistance capability at late sessions compared with early sessions (OR, 3.56; 95% CI, 1.89-6.72 vs OR, 1.51; 95% CI, 0.98-2.32). In addition, the combination system combining the CADe system and CAQ system showed the highest assistance capability during late sessions (OR, 10.40; 95% CI, 3.85-28.05 vs OR, 1.89; 95% CI, 1.12-3.21).
Regarding the lesion size and locations, the AI assistance capability increase during late sessions was mainly associated with an increase in diminutive, small adenomas and distal colon adenomas. With the assistance of the CADe system, more diminutive adenomas (OR, 2.42; 95% CI, 0.97-5.99 vs OR, 1.59; 95% CI, 0.95-2.68) were detected during late sessions compared with early sessions, while more small adenomas (OR, 9.20; 95% CI, 1.10-76.85 vs OR, 1.09; 95% CI, 0.26-4.63) were detected by the CAQ system. Generally, diminutive polyps are more likely to be missed within the visual field. To counter this case, the CADe system applies real-time visual alerts when identifying polyps. Additionally, the result that more small adenomas were detected with CAQ system assistance may be attributed to more adequate mucosal exposure, which resulted from low and uniform withdrawal speed and the reminder regarding blind spots.
Among the analysis of hourly detection rates in the control-c group, we found an increase in adenoma and polyp detection after the first hour of colonoscopy procedures per half day, which has been termed as the warming-up effect in previous studies,10,16 which is the need for endoscopists to perform a few endoscopy procedures to operate more comfortably before reaching better performance. In addition, a previous study31 about time of day in radiology expressed the opinion that clinician performance of the visual search task can be impacted by circadian rhythms, leading to different clinical performance during different times of the day; some also noted a decrease in performance after lunch. Since movements while performing colonoscopies are repetitive and stereotypical, this warming-up effect may result from increasing concentration due to improvement in endoscopists’ working state. On the other hand, we found the AI system showed the highest assistance capability in the last hour per half day (noon, 20.22% vs 2.99%; P = .009; OR, 8.71; 95% CI, 1.72-44.07; 4 pm, 24.32% vs 0.00%), which demonstrated the potential benefit of using AI to maintain high quality colonoscopy procedures.
Limitations
There are some limitations in our study. First, the result was only validated at one center; whether the conclusion of this study is generalizable in clinical centers with different work schedules remains unknown, and further study is needed. Second, the ADR in the control-c group was lower than that in Western populations. Due to differences in environment, diet and lifestyle, the morbidity of colon polyps/adenomas could be different among regions and populations.32 Multiple regional samples are needed to further investigate the adaptability and effectiveness of our findings, especially regions of high morbidity of colon adenoma.
Conclusions
In conclusion, our results suggest that later sessions per half day were associated with a decline adenoma detection. Furthermore, AI systems could eliminate the time-related degradation of colonoscopy quality. In the future, the application of AI systems has the potential to maintain high quality and homogeneity of colonoscopies and further improve endoscopist performance in large screening programs and centers with high workloads.
eMethods.
eFigure. ADR/PDR at Each Hour of the Day
eTable 1. Comparison of Detection Rates Between Early and Late Sessions
eTable 2. Detection Rates Between Early and Late Session in Different RCTs
eTable 3. Detection Rates Among Individual Endoscopists
eTable 4. Comparison of AI Assistance Capabilities Between Early and Late Sessions
eTable 5. AI Assistance Capabilities Between Early and Late Sessions in Different RCTs
eTable 6. Different AI Systems Assistance Capabilities Among Different Sizes
eTable 7. Different AI Systems Assistance Capabilities Among Different Locations
Data Sharing Statement
References
- 1.Bray F, Ferlay J, Soerjomataram I, Siegel RL, Torre LA, Jemal A. Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2018;68(6):394-424. doi: 10.3322/caac.21492 [DOI] [PubMed] [Google Scholar]
- 2.Atkin WS, Edwards R, Kralj-Hans I, et al. ; UK Flexible Sigmoidoscopy Trial Investigators . Once-only flexible sigmoidoscopy screening in prevention of colorectal cancer: a multicentre randomised controlled trial. Lancet. 2010;375(9726):1624-1633. doi: 10.1016/S0140-6736(10)60551-X [DOI] [PubMed] [Google Scholar]
- 3.Doubeni CA, Corley DA, Quinn VP, et al. Effectiveness of screening colonoscopy in reducing the risk of death from right and left colon cancer: a large community-based study. Gut. 2018;67(2):291-298. doi: 10.1136/gutjnl-2016-312712 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Corley DA, Jensen CD, Marks AR, et al. Adenoma detection rate and risk of colorectal cancer and death. N Engl J Med. 2014;370(14):1298-1306. doi: 10.1056/NEJMoa1309086 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Sanaka MR, Shah N, Mullen KD, Ferguson DR, Thomas C, McCullough AJ. Afternoon colonoscopies have higher failure rates than morning colonoscopies. Am J Gastroenterol. 2006;101(12):2726-2730. doi: 10.1111/j.1572-0241.2006.00887.x [DOI] [PubMed] [Google Scholar]
- 6.Sanaka MR, Deepinder F, Thota PN, Lopez R, Burke CA. Adenomas are detected more often in morning than in afternoon colonoscopy. Am J Gastroenterol. 2009;104(7):1659-1664. doi: 10.1038/ajg.2009.249 [DOI] [PubMed] [Google Scholar]
- 7.Gurudu SR, Ratuapli SK, Leighton JA, Heigh RI, Crowell MD. Adenoma detection rate is not influenced by the timing of colonoscopy when performed in half-day blocks. Am J Gastroenterol. 2011;106(8):1466-1471. doi: 10.1038/ajg.2011.125 [DOI] [PubMed] [Google Scholar]
- 8.Teng TY, Khor SN, Kailasam M, Cheah WK, Lau CCL. Morning colonoscopies are associated with improved adenoma detection rates. Surg Endosc. 2016;30(5):1796-1803. doi: 10.1007/s00464-015-4448-7 [DOI] [PubMed] [Google Scholar]
- 9.Lee TJ, Rees CJ, Blanks RG, et al. Colonoscopic factors associated with adenoma detection in a national colorectal cancer screening program. Endoscopy. 2014;46(3):203-211. doi: 10.1055/s-0033-1358831 [DOI] [PubMed] [Google Scholar]
- 10.Lee A, Iskander JM, Gupta N, et al. Queue position in the endoscopic schedule impacts effectiveness of colonoscopy. Am J Gastroenterol. 2011;106(8):1457-1465. doi: 10.1038/ajg.2011.87 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Spadaccini M, Iannone A, Maselli R, et al. Computer-aided detection versus advanced imaging for detection of colorectal neoplasia: a systematic review and network meta-analysis. Lancet Gastroenterol Hepatol. 2021;6(10):793-802. doi: 10.1016/S2468-1253(21)00215-6 [DOI] [PubMed] [Google Scholar]
- 12.Marcondes FO, Gourevitch RA, Schoen RE, Crockett SD, Morris M, Mehrotra A. Adenoma detection rate falls at the end of the day in a large multi-site sample. Dig Dis Sci. 2018;63(4):856-859. doi: 10.1007/s10620-018-4947-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Wang P, Xiao X, Glissen Brown JR, et al. Development and validation of a deep-learning algorithm for the detection of polyps during colonoscopy. Nat Biomed Eng. 2018;2(10):741-748. doi: 10.1038/s41551-018-0301-3 [DOI] [PubMed] [Google Scholar]
- 14.Gong D, Wu L, Zhang J, et al. Detection of colorectal adenomas with a real-time computer-aided system (ENDOANGEL): a randomised controlled study. Lancet Gastroenterol Hepatol. 2020;5(4):352-361. doi: 10.1016/S2468-1253(19)30413-3 [DOI] [PubMed] [Google Scholar]
- 15.Yao L, Zhang L, Liu J, et al. Effect of an artificial intelligence-based quality improvement system on efficacy of a computer-aided detection system in colonoscopy: a four-group parallel study. Endoscopy. 2022;54(8):757-768. doi: 10.1055/a-1706-6174 [DOI] [PubMed] [Google Scholar]
- 16.Lee A, Jensen CD, Marks AR, et al. . Endoscopist fatigue estimates and colonoscopic adenoma detection in a large community-based setting. Gastrointest Endosc. 2017;85(3):601-610.e2. doi: 10.1016/j.gie.2016.09.033 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Ndounga Diakou LA, Trinquart L, Hróbjartsson A, et al. Comparison of central adjudication of outcomes and onsite outcome assessment on treatment effect estimates. Cochrane Database Syst Rev. 2016;3(3):MR000043. doi: 10.1002/14651858.MR000043.pub2 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Elsheikh TM, Kirkpatrick JL, Fischer D, Herbert KD, Renshaw AA. Does the time of day or weekday affect screening accuracy? A pilot correlation study with cytotechnologist workload and abnormal rate detection using the ThinPrep Imaging System. Cancer Cytopathol. 2010;118(1):41-46. doi: 10.1002/cncy.20060 [DOI] [PubMed] [Google Scholar]
- 19.Stec N, Arje D, Moody AR, Krupinski EA, Tyrrell PN. A systematic review of fatigue in radiology: is it a problem? AJR Am J Roentgenol. 2018;210(4):799-806. doi: 10.2214/AJR.17.18613 [DOI] [PubMed] [Google Scholar]
- 20.Korenblit J, Tholey DM, Tolin J, et al. Effect of the time of day and queue position in the endoscopic schedule on the performance characteristics of endoscopic ultrasound-guided fine-needle aspiration for diagnosing pancreatic malignancies. Endosc Ultrasound. 2016;5(2):78-84. doi: 10.4103/2303-9027.180470 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Joseph DA, Meester RG, Zauber AG, et al. Colorectal cancer screening: estimated future colonoscopy need and current volume and capacity. Cancer. 2016;122(16):2479-2486. doi: 10.1002/cncr.30070 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Chan MY, Cohen H, Spiegel BM. Fewer polyps detected by colonoscopy as the day progresses at a Veteran’s Administration teaching hospital. Clin Gastroenterol Hepatol. 2009;7(11):1217-1223. doi: 10.1016/j.cgh.2009.07.013 [DOI] [PubMed] [Google Scholar]
- 23.Jover R, Zapater P, Bujanda L, et al. ; COLONPREV Study Investigators . Endoscopist characteristics that influence the quality of colonoscopy. Endoscopy. 2016;48(3):241-247. doi: 10.1055/s-0042-100185 [DOI] [PubMed] [Google Scholar]
- 24.Park JH, Kim SJ, Hyun JH, et al. Correlation between bowel preparation and the adenoma detection rate in screening colonoscopy. Ann Coloproctol. 2017;33(3):93-98. doi: 10.3393/ac.2017.33.3.93 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Shaukat A, Rector TS, Church TR, et al. Longer withdrawal time is associated with a reduced incidence of interval cancer after screening colonoscopy. Gastroenterology. 2015;149(4):952-957. doi: 10.1053/j.gastro.2015.06.044 [DOI] [PubMed] [Google Scholar]
- 26.Munson GW, Harewood GC, Francis DL. Time of day variation in polyp detection rate for colonoscopies performed on a 3-hour shift schedule. Gastrointest Endosc. 2011;73(3):467-475. doi: 10.1016/j.gie.2010.07.025 [DOI] [PubMed] [Google Scholar]
- 27.Freedman JS, Harari DY, Bamji ND, et al. The detection of premalignant colon polyps during colonoscopy is stable throughout the workday. Gastrointest Endosc. 2011;73(6):1197-1206. doi: 10.1016/j.gie.2011.01.019 [DOI] [PubMed] [Google Scholar]
- 28.Mahmud N, Cohen J, Tsourides K, Berzin TM. Computer vision and augmented reality in gastrointestinal endoscopy. Gastroenterol Rep (Oxf). 2015;3(3):179-184. doi: 10.1093/gastro/gov027 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Tawfik DS, Profit J, Morgenthaler TI, et al. Physician burnout, well-being, and work unit safety grades in relationship to reported medical errors. Mayo Clin Proc. 2018;93(11):1571-1580. doi: 10.1016/j.mayocp.2018.05.014 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Pickhardt PJ, Nugent PA, Mysliwiec PA, Choi JR, Schindler WR. Location of adenomas missed by optical colonoscopy. Ann Intern Med. 2004;141(5):352-359. doi: 10.7326/0003-4819-141-5-200409070-00009 [DOI] [PubMed] [Google Scholar]
- 31.Alshabibi AS, Suleiman ME, Tapia KA, Brennan PC. Effects of time of day on radiological interpretation. Clin Radiol. 2020;75(2):148-155. doi: 10.1016/j.crad.2019.10.006 [DOI] [PubMed] [Google Scholar]
- 32.Gu M, Thapa S. Colorectal cancer in the United States and a review of its heterogeneity among Asian American subgroups. Asia Pac J Clin Oncol. 2020;16(4):193-200. doi: 10.1111/ajco.13324 [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
eMethods.
eFigure. ADR/PDR at Each Hour of the Day
eTable 1. Comparison of Detection Rates Between Early and Late Sessions
eTable 2. Detection Rates Between Early and Late Session in Different RCTs
eTable 3. Detection Rates Among Individual Endoscopists
eTable 4. Comparison of AI Assistance Capabilities Between Early and Late Sessions
eTable 5. AI Assistance Capabilities Between Early and Late Sessions in Different RCTs
eTable 6. Different AI Systems Assistance Capabilities Among Different Sizes
eTable 7. Different AI Systems Assistance Capabilities Among Different Locations
Data Sharing Statement