Influence of Artificial Intelligence on the Adenoma Detection Rate throughout the Day

Abstract

Background

Artificial intelligence systems recently demonstrated an increase in polyp and adenoma detection rate. Over the daytime, the adenoma detection rate decreases as tiredness leads to a lack of attention. It is not clear if a polyp detection system with artificial intelligence leads to constant adenoma detection over the day.

Methods

We performed a database analysis of screening and surveillance colonoscopies with and without the use of AI. In both groups, patients were investigated with the same endoscopy equipment and by the same endoscopists. Only patients with good bowel preparation (BBPS >6) were included. We correlated the daytime, the investigational time, day of the week, and the adenoma and polyp detection.

Results

A total of 303 colonoscopies were analyzed. 163 endoscopies in the AI⁺ group and 140 procedures in the AI⁻ group were included. In both groups, the total adenoma detection rate was equal (AI⁺ 0.39 vs. AI⁻ 0.43). The adenoma detection rate throughout the day had a significant decreasing trend in the group without the use of AI (p = 0.015), whereas this trend was not present in the investigations that have been performed with AI (p = 0.65). The duration of investigation did not show a significant difference between the groups (8.9 min in both groups). No relevant effect was noticed in adenoma detection between single days of the working week with or without the use of AI.

Conclusion

AI helps overcome the decay in adenoma detection over the daytime. This may be attributed to a constant awareness caused by the use of the AI system.

Introduction

Colonoscopy is the gold standard for the detection of colorectal diseases, in particular the detection of precursor lesions and early stages of colorectal cancer [1]. After implementation of national colon cancer screening programs throughout numerous countries, it is clear that the detection and removal of colorectal adenoma leads to a decrease in colorectal cancer incidence [2]. Studies have shown that an increase in adenoma detection rate (ADR) is directly linked to a decrease in colorectal cancer as well.

In the past two decades, high effort has therefore been undertaken to improve adenoma detection. The most important of that was the implementation of high-definition, high-resolution endoscopy units which led to a significant increase in ADR as compared to older systems [3]. Other methods are, among other things, wide-angle endoscopes, mechanical fixation of the colon folds with various instruments, and retroflexing endoscopes [4].

The latest technology is using real-time artificial intelligence systems to detect and/or characterize polyps [5]. This new technology has already been shown to increase the ADR as well as the adenomas detected per patient [6]. This is achieved mostly by the detection of small and diminutive adenomas. Only little is known about the machine-human interaction and how exactly the increase in adenoma detection is caused. Data from large-scale investigations demonstrate a decrease in ADR attributed to the fatigue of the investigator over the day. With this decay of awareness over the day, the optimal point of time for colonoscopy is early in the morning. Differences of 6, 3% up to 20% of detected polyps, were reported [7, 8].

In theory, an artificial intelligence system works as an additional investigator but without the tendency for fatigue. In this study, we aimed to investigate the influence of AI on polyp and adenoma detection throughout the working day and the influence on the investigator’s behavior.

Materials and Methods

We performed a database study including investigations of patients who were scheduled for screening or surveillance colonoscopy at a single university center. The total study period was from February 2, 2020, to February 10, 2021. The group of patients that was investigated without AI (AI⁻) was consecutively collected mainly retrospectively backwards from the time point after inauguration of an AI system into the endoscopy unit. The group of patients with AI (AI⁺) was collected forward from this time point accordingly. Patients gave their written informed consent to participate in data acquisition for clinical evaluation of the AI system (ethics protocol # 152/17).

Data were collected and stored with pseudonyms. The study adheres to the Declaration of Helsinki and the European data protection laws. Exclusion criteria were bad bowel preparation with a BBPS ≤6, inflammatory bowel disease, acute GI infections, GI bleeding of any source, recent major colorectal surgery, age below 18 years, investigations outside the normal working day (7:30 a.m.–4:30 p.m.).

Primary endpoint was the adenoma detection according to daytime. Secondary outcomes were ADR, polyp detection rate (PDR), adenoma per patient (APP), withdrawal time in investigations without polypectomy. Withdrawal time excluding polypectomies was chosen to rule out any potential errors in documentation of pure withdrawal time.

All patients were investigated under conscious sedation with propofol. The participating endoscopists all had an experience of >1,000 colonoscopies and were certified to perform colonoscopies for the local colorectal cancer center.

The endoscopy equipment in both groups was equal (Fujifilm Eluxeo Series, Fujifilm Europe, Düsseldorf, Germany) despite the use of the AI System (CADEye Fujifilm Europe, Düsseldorf, Germany). In both groups, the use of WLE, LCI, and BLI was allowed and performed by the choice of the investigator.

The sample size was chosen according to assumed ADR increase by AI and the difference in daytime adenoma detection in a direct comparison model. Aiming to detect 9% difference in ADR (8–10%) in a population with above 40% ADR in our hospital (see also results), in total, 302 individuals need to be included to show at least non-inferiority. However, the analysis of time trends cannot be planned according to analyses comparing to independent samples.

Data were analyzed using descriptive statistics, Student’s t test, Mann-Whitney U-test, odds ratio calculation, and Mann-Kendall test, which is a test to investigate trends in consecutive data, which do not show periodic changes. To search for periodic changes, the data were analyzed using a short-time Fourier analysis.

Results

In total, 306 colonoscopies were included in the database. Three investigations were excluded because they did not match inclusion criteria. In total, 303 colonoscopies were analyzed. Although we included patients from various age groups, the vast majority were between 50 and 80 years old. 140 and 163 patients in the AI⁻ group and the AI⁺ group, respectively, were investigated.

We found no statistical differences regarding age and sex between the two groups (overall age: 64.1 ± 14.5 years, 185 males, 119 females; AI⁻: 64.4 ± 15.2 years, 90 males, 50 females; AI⁺: 63 ± 13.8 years, 94 males, 69 females). In total (n = 306), the ADR was 0.41, the PDR was 0.48, and the APP was 0.97.

The withdrawal time of investigations without polypectomies was equal in both groups (8.9 min, p = 0.89). If withdrawal time changed over the observation period as a potential adaption of the endoscopist throughout the observation period, the analysis could not identify any significant trend. In the AI⁻ group, the ADR was 0.41 over the complete day, whereas in the AI⁺ group, the ADR was 0.39 over the complete day, showing no statistical differences.

Using the Mann-Kendall test, we could identify a significant decreasing trend in adenoma detection in the investigations performed with AI⁻ (p = 0.015) as sorted by daytime, whereas the adenoma detection did not show any trend in the investigations that were performed with the AI system (p = 0.65) (Fig. 1). Short-time Fourier analysis ruled out periodic effects over the complete period of data acquisition.

Fig. 1.

Gliding averages (of three overlapping investigations) of detected adenoma per the averages also include colonoscopies without adenoma detection. Investigations on the horizontal axis are ordered by daytime. In the left panel (a), showing the investigations without AI, a decreasing trend which is proven with Mann-Kendall test is visible. Using the AI system (b) leads to constant adenoma detection.

We could not identify a single day on which the odds of detecting at least one APP were different depending on AI use. In a model assuming Friday to be the weakest day of performance, we could only identify Tuesday colonoscopies as an independent risk factor for increased adenoma detection (OR 2.92, 1.04–8.24, p = 0.042) in the group without AI use. If AI was used, no day with higher or lower adenoma detection could be identified.

Discussion

Our study confirms a decrease in adenoma detection over the working day. Other studies have shown this before and showed a decrease in AD between 4% and 27% [8–10]. In our study, this decreasing trend was also present but only in the group of patients that were investigated without the use of artificial intelligence. We attribute this to a higher awareness level of the investigator. This might be due to a “second investigator effect” produced by the AI system that does not only run as a passive system but also gives feedback to the investigator directly during the investigation. In case of the used AI system, this feedback is given as a bounding box in case of polyp detection and in addition as an acoustic signal (ping sound) which may raise the attention during endoscopy. However, this can only be assumed, as we did not directly investigate the attention level which may be performed by other methods such as eye tracking and pupillometry, heart rate, and electric skin resistance measurements. Studies investigating the direct physiological or psychological impact of AI diagnostic systems on endoscopists have not been published so far. The use of AI systems may have a similar effect as having additional investigators in the room, as shown by several studies. A meta-analysis clearly demonstrated that the participation of nurses in the detection of polyps of adenoma increased the ADR and PDR significantly [11].

As the investigation time did not show the same trend over time, we assume that the increase in detected adenomas is not caused by longer investigation times in general. Interestingly, in our cohort, the ADR is not directly increased by AI. This effect may be caused by a high general baseline ADR without AI, which is much lower in many studies that showed an increase in ADR by using AI systems [12].

However, the point of time is not the only variable influencing adenoma detection. A meta-analysis by Wu et al. published in 2018 [13] revealed that a lower ADR in the afternoon is not present if endoscopists performed colonoscopies in half-day shifts compared to full shifts, which suggests that permanent tight schedules colonoscopy shifts are of disadvantage for the patient.

In our analysis, we only found the decreasing trend over single days and not over a complete week, which underlines the role of fatigue caused by consecutive investigations in a tight schedule, whereas day to day recovery reverses this effect. A study investigating 1,884 screening endoscopies discovered that the weakest performance in adenoma detection appears on Fridays [14].

Not all available data support a lower adenoma detection over the day. Jaho et al. [15] found opposite results with even higher ADR in the afternoon and stated that other local factors besides daytime may affect the investigator’s performance.

A retrospective study by Lei and colleagues also aimed to investigate the performance of computer-aided diagnosis in colonoscopy and found no difference in ADR in the morning or in the afternoon [11] in patients who were investigated with AI. However, the authors did not include colonoscopies without the use of AI into their analysis as comparators. Therefore, they cannot reliably conclude about the direct performance of the AI system as our data do.

A limitation of our study is the retrospective data acquisition. However, the data for the primary endpoints (detected adenoma) would not have been different if evaluated prospectively. To rule out bias with regards to the investigation time, we only analyzed investigations without polypectomies. Because of the same problem, we concentrated on the adenoma detection because we noticed that in many cases, rectal small hyperplastic polyps have been reported inconsistently.

The difference in data acquisition in both groups may also have an effect on direct comparability. However, we think that only a minor percentage of effect is caused by the fact that the colonoscopies were performed under study conditions but mainly through a higher attention level caused by the use of AI system itself. This effect is in general a possible explanation of increased adenoma detection besides the pure technical detection by the AI system.

In summary, we confirm studies that show a decreasing trend in adenoma detection during the working day. This effect was absent in our group of patients that have been investigated with a commercial AI system for polyp detection and polyp characterization. We recommend implementation of AI systems to daily routine, especially in case of high-workload endoscopy units and in addition to adherence to colonoscopy performance guidelines. Recovery of the endoscopist may be an additional factor influencing colonoscopy performance.

Statement of Ethics

The Local Ethics Committee of the Otto-v.-Guericke University Magdeburg, Ref. No. 152/17, approved the study. The abovementioned ethics committee waived the need for written informed consent for study participation. We performed the study according to the revised Declaration of Helsinki and according to the European data protection laws and regulations.

Conflict of Interest Statement

Jochen Weigt received research grant and lecture honorary from Fujifilm Int. All other authors declare no COI.

Funding Sources

No funding was received for this study.

Author Contributions

Jochen Weigt: planning of the study, endoscopy, data analysis, interpretation, and writing of the manuscript. Rino Richter: data analysis, writing of the manuscript, and data interpretation. Johannes Bruns: planning of the study and data acquisition. Wilfried Obst: interpretation of the results and data analysis. Verena Keitel-Anselmino: data interpretation, correction, and writing of the manuscript.

Funding Statement

No funding was received for this study.

Data Availability Statement

Data are not openly available. All data generated for the study purpose are included in the analysis and in this article. Further inquiries can be directed to the corresponding author.