A Novel Convolutional Neural Network Model as an Alternative Approach to Bowel Preparation Evaluation Before Colonoscopy in the COVID-19 Era: A Multicenter, Single-Blinded, Randomized Study

Abstract

INTRODUCTION: 

Adequate bowel preparation is key to a successful colonoscopy, which is necessary for detecting adenomas and preventing colorectal cancer. We developed an artificial intelligence (AI) platform using a convolutional neural network (CNN) model (AI-CNN model) to evaluate the quality of bowel preparation before colonoscopy.

METHODS: 

This was a colonoscopist-blinded, randomized study. Enrolled patients were randomized into an experimental group, in which our AI-CNN model was used to evaluate the quality of bowel preparation (AI-CNN group), or a control group, which performed self-evaluation per routine practice (control group). The primary outcome was the consistency (homogeneity) between the results of the 2 methods. The secondary outcomes included the quality of bowel preparation according to the Boston Bowel Preparation Scale (BBPS), polyp detection rate, and adenoma detection rate.

RESULTS: 

A total of 1,434 patients were enrolled (AI-CNN, n = 730; control, n = 704). No significant difference was observed between the evaluation results (“pass” or “not pass”) of the groups in the adequacy of bowel preparation as represented by BBPS scores. The mean BBPS scores, polyp detection rate, and adenoma detection rate were similar between the groups. These results indicated that the AI-CNN model and routine practice were generally consistent in the evaluation of bowel preparation quality. However, the mean BBPS score of patients with “pass” results were significantly higher in the AI-CNN group than in the control group, indicating that the AI-CNN model may further improve the quality of bowel preparation in patients exhibiting adequate bowel preparation.

DISCUSSION: 

The novel AI-CNN model, which demonstrated comparable outcomes to the routine practice, may serve as an alternative approach for evaluating bowel preparation quality before colonoscopy.

INTRODUCTION

Colorectal cancer (CRC) was the second leading cause of cancer death worldwide in 2020 (9.4%) (¹). In developed countries, there are more than 600,000 patients dying of CRC every year (²). Colonoscopy is a surveillance tool used for CRC screening (³). Endoscopic screening has reduced the incidence and mortality rates of CRC in the past decade because precancerous lesions can be removed during the procedure (^4,5).

A high-quality colonoscopy is decisive for the maximum detection of precancerous lesions (^3,6). The Boston Bowel Preparation Scale (BBPS) is a rating scale used to evaluate the adequacy of bowel preparation (^7,8). Several studies have demonstrated the potential of using artificial intelligence (AI)-assisted BBPS scoring during colonoscopy (^9–11). However, none of the AI techniques explored in these studies were designed to evaluate the adequacy of bowel preparation before colonoscopy. In routine practice, the evaluation of bowel preparation before colonoscopy is mostly performed by the patients themselves, by caregivers, or by medical staff; however, close contact between patients and medical staff can increase the possibility of disease transmission during the Coronavirus Disease 2019 (COVID-19) pandemic. In addition, the COVID-19 pandemic has led to a sustained reduction in the number of patients referred for and diagnosed using colonoscopy and treated for CRC (¹²). Improved bowel preparation quality may reduce the need for repeat colonoscopies during the COVID-19 era (¹³) because inadequate bowel preparation has been reported to shorten the interval between repeat colonoscopies to 270 days, which is much shorter than the 3-year interval recommended by the guidelines stipulated by the Cancer Council Australia (¹⁴).

To increase the quality of bowel preparation, reduce close person-to-person contact, and prevent repeat colonoscopies, we developed the first state-of-the-art AI platform using a convolutional neural network (CNN) model to help patients evaluate the adequacy of bowel cleansing without the aid of caregivers or medical staff. A clinical trial was conducted to investigate its feasibility for use in routine practice. Here, we report the results of the trial of this AI-CNN model by patients undergoing colonoscopy.

MATERIALS AND METHODS

Study design

This was a prospective, single-blinded (colonoscopist-blinded), randomized controlled trial conducted at Xiamen Chang Gung Hospital (XCGH) and Shenzhen Hospital of Southern Medical University (SHSMU) from November 2020 to December 2021. Consecutive patients undergoing colonoscopy were enrolled between May 2021 and August 2021. This study was approved by the institutional review board (IRB) of the participating hospitals (IRB number: XMCGIRB2021005) and was conducted in accordance with the Declaration of Helsinki. A waiver of informed consent was approved by the IRB. This study has been registered on the Chinese Clinical Trial Registry (ChiCTR), with a registration number of ChiCTR2100052134. All authors had access to the study data and reviewed and approved the final manuscript. Patients and the public were not involved in the design, analyses, and interpretation of this study.

Population

Outpatients, inpatients, or individuals undergoing a physical examination aged 18–60 years with a colonoscopy scheduled for screening and diagnosis and able to use smartphone to scan a QR code were eligible to participate. Exclusion criteria were (i) contraindications for colonoscopy, (ii) familial adenomatous polyposis, (iii) known colorectal polyps, (iv) a history of gastrointestinal surgery, (v) inability to provide informed consent, and (vi) refusal to join the study.

Instructions for bowel preparation

Enrolled patients were equally randomized into either the AI-CNN group or the control (self-evaluation) group. We recommended a low-fiber diet on the day (24 hours) preceding colonoscopy to all patients. At the time they scheduled their colonoscopy appointment, each of the patients received standard bowel preparation instructions from a well-trained medical practitioner who explained the instructions verbally. All the patients were given a leaflet with general guidelines pertaining to bowel preparation, such as the recommended low-fiber diet before colonoscopy, the importance of bowel preparation, how to prepare the purgative solution, and the evaluation of bowel preparation adequacy. The leaflet displayed photographic examples of bowel preparation quality: Stool should eventually be a yellowish clear liquid; if any cloudiness (including turbid liquid, particles, or small amounts of feces) is observed in the liquid stool, the bowel preparation is not complete.

To be taken 4–6 hours before their colonoscopy, all patients were prescribed polyethylene glycol (PEG) electrolyte powder (PEG electrolyte powder [III], Beaufour Ipsen) for bowel cleansing. Patients were provided with 4 packs of PEG powder (73.56 g/pack) and instructed to mix it with water to obtain 2 L of solution. They were instructed to first consume 1.5 L of PEG at a rate of 500 mL hourly and to consume the remaining 0.5 L of PEG mixed with 20 mL of dimethicone at a constant rate. A leaflet with clear written instructions on how to consume the laxatives was given to patients.

After taking the laxatives, all the patients were asked to scan a QR code using smartphone for randomization (Supplementary Figure 1, Supplementary Digital Content 1, https://links.lww.com/AJG/C600). The system displayed instructions for using the application, taking photographs of their feces, and uploading images to the system. After uploading their images, the patients assigned to the AI-CNN group automatically received an evaluation result of “pass” or “not pass,” which indicated whether their bowel preparation was adequate or not (Supplementary Figure 2 and Supplementary Figure 3, Supplementary Digital Contents 2 and 3, https://links.lww.com/AJG/C601, https://links.lww.com/AJG/C602). Patients assigned to the control group evaluated the adequacy of bowel preparation themselves according to the instructions from the leaflet after uploading their images. For patients in both groups who received results of “not pass,” the system displayed tips instructing the patients to gently rub their lower abdomen, drink more water, or walk to improve the bowel preparation quality.

Data collection and colonoscopy

Each patient's demographics and clinical characteristics were recorded at the time of their colonoscopy appointment. The colonoscopies were performed by 9 and 15 experienced colonoscopists at XCGH and SHSMU, respectively, who were blinded to the bowel evaluation method that each patient used (AI or self-evaluation). The nurses working in the endoscopic centers were also blinded to patients' bowel evaluation methods (AI or self-evaluation) before, during, and after the procedures. Olympus CV-290 (XCGH and SHSMU) and Fujifilm ELUXEO 7000/EPX4450 (SHSMU) endoscopes were used to conduct the colonoscopies. Each colonoscopist recorded the quality of bowel preparation and colonoscopy findings. All the colonoscopists and nurses at both medical centers received training on the use of the BBPS, a validated tool for assessing colon cleansing quality (¹⁵). Each segment of the colon (right, transverse, and left) is scored from 0 to 3, where “0” represented a poorly prepared colon segment and “3” represented a successfully prepared colon segment with visible mucosa. The scores of the 3 segments are added to obtain the total BBPS score, ranging from 0 to 9. Adequate bowel preparation is defined as a total BBPS score of ≥ 6 (¹⁶).

AI-CNN model

A CNN model was used to evaluate the images of the patients' rectal effluent (Figure 1). To train the model, 4,302 images were collected from Zhongshan Hospital Xiamen University. The number of trained images was determined according to the analysis of accuracy. In the training set, 250 images were added each time. The performance of the model trained up under different sizes of training sets was evaluated by using 10-fold cross-validation. With the increased number of training samples, the accuracy of the model increased rapidly in the beginning. When the training set contained approximately 2,500 images, the accuracy of the model reached 97% and the slope of the curve gradually converged to 0, indicating that the accuracy of the model was improved with increasing training samples (Supplementary Figure 4, Supplementary Digital Content 4, https://links.lww.com/AJG/C603). Accordingly, a training set containing 3,000 images or more would yield an accuracy rate higher than 95%.

Figure 1.: Architecture of the convolutional neural network model.

All the images were labeled with “pass” or “not pass” by an experienced nurse from the endoscopy center of Chang Gung Memorial Hospital in Taiwan. Initially, the parameters of the CNN model were set at random. During the training process, the labeled images were input into the CNN model to iteratively adjust the parameters of the model. When the training process was finished, the CNN model was able to assign either a “pass” or “not pass” designation to a newly uploaded bowel preparation image.

The CNN model was developed as a web service by using the MobileNet architecture; this was performed to enhance convenience because we aimed to serve many patients from different hospitals simultaneously and to provide immediate feedback. Compared with other CNN architectures, MobileNet addresses potential problems with classification by using a smaller scale network with fewer parameters. The MobileNet architecture was, therefore, selected for its lower computational resource requirements and its fast inference speed.

Outcome measures

The primary outcome considered was the consistency between the AI-CNN and control groups in the adequacy of bowel preparation as measured using the AI-CNN or self-evaluation results and the total BBPS scores. The secondary outcomes were the total BBPS score, rate of adequate bowel preparation (total BBPS score ≥ 6), polyp detection rate (PDR), ADR, and number of polyps.

Statistical analysis

The sample size was calculated based on the BBPS difference between groups. According to our previous feasibility study, the mean (± SD) BBPS scores of the AI-CNN and control groups were 6.2 ± 0.73 and 6.1 ± 0.75, respectively (unpublished data). Based on a significance level of 0.025, an expected intergroup difference of 0.12 in the BBPS score, and an equivalence margin of −0.120, a sample size of 681 in each group would achieve 80% power to detect the noninferiority between groups by using a 1-sided, 2-sample Student t test. Considering a drop-out rate of 5%, a total of 1,434 patients were planned to be recruited in this study. The sample size was estimated using PASS 15 (Power Analysis and Sample Size Software, NCSS Statistical Software, Kaysville, UT).

The data of the intention-to-treat (ITT) population, which included all the eligible patients undergoing colonoscopy, were used to evaluate the study outcomes. The consistency between the AI-CNN and control groups was analyzed using a Breslow-Day test. We analyzed the results of the AI-CNN model and self-evaluation (“pass” or “not pass”) and the corresponding total BBPS scores (adequate [≥ 6] or inadequate [< 6]) determined by the blinded colonoscopists to determine whether the BBPS scores differed between the study groups. Continuous variables are expressed as mean with SD and were analyzed using a Wilcoxon rank sum test or 2-sample Student t test as appropriate. Categorical variables are presented as frequencies or percentages and were analyzed using the Pearson χ² test or Fisher exact test as appropriate. Statistical analysis was performed using SAS version 9.4 (SAS Institute, Cary, NC). A P value of < 0.05 was considered statistically significant.

RESULTS

Patient characteristics

A total of 1,454 consecutive patients eligible for study inclusion were enrolled and prospectively assigned into the AI-CNN group (733 cases) or control group (721 cases). All 1,454 patients completed a colonoscopy; of these, 3 patients in the AI-CNN group and 17 patients in the control group were excluded because of missing data or unqualified results (Figure 2). Eventually, the data of 730 patients in the AI-CNN group and 704 patients in the control group were analyzed. In the AI-CNN group, 698 of the 730 cases and 698 of the 704 cases in the AI-CNN and control groups, respectively, were judged as “pass” by the AI-CNN system.

Figure 2.: Patient disposition.

No significant differences were observed between the groups in baseline characteristics, including demographics, medical histories, and time of colonoscopy (Table 1). The mean ± SD age of all the patients was 42.4 ± 11.1 years, and 57.5% were male.

Table 1.: Baseline characteristics of study patients

Primary outcome—consistency between AI-CNN and control groups

The test of consistency indicated no significant difference in the evaluation results between the “adequate” and “inadequate” BBPS categories in the AI-CNN and control groups. Among the patients who received a “pass” result, 90.7% (633 of 698) and 91.5% (639 of 698) had a total BBPS score of ≥ 6 in the AI-CNN and control groups, respectively. Among the patients who received a “not pass” result, similar patterns were observed between the groups (Supplementary Table 1, Supplementary Digital Content 5, https://links.lww.com/AJG/C604). Therefore, observations of bowel preparation quality seemed to be consistent and homogenous between the 2 groups (Breslow-Day P = 0.976).

Bowel preparation quality

There were no significant differences in the quality metrics of colonoscopy (e.g., cecal intubation time and withdrawal time) between groups. The proportion of patients who achieved adequate bowel preparation for colonoscopy (total BBPS score ≥ 6) did not differ significantly between the groups (AI-CNN: 90.3% vs control: 91.5%, P = 0.429). Similarly, the mean total BBPS score and mean BBPS score of each segment were comparable between the AI-CNN and control groups (Table 2).

Table 2.: Quality of bowel preparation and outcomes of colonoscopy for AI-CNN and control groups

We further performed subgroup analysis for patients evaluated as “pass” and those as “not pass.” In the analysis of the “pass” subgroups, no significant difference was identified between the 2 groups in adequate bowel preparation rates, except for in BBPS scores: The mean total BBPS score and mean BBPS score of the right colon were significantly higher in the AI-CNN group than in the control group (total: 7.32 ± 1.40 vs 7.16 ± 1.46, respectively, P = 0.044; right colon: 2.33 ± 0.61 vs 2.26 ± 0.60, respectively, P = 0.04; Table 3). Among patients in the “not pass” subgroups, no significant differences were evident between the AI-CNN and control groups in adequate bowel preparation rates and BBPS scores of each segment (Supplementary Table 2, Supplementary Digital Content 5, https://links.lww.com/AJG/C604).

Table 3.: Quality of bowel preparation and outcomes of colonoscopy for AI-CNN and Control groups (subgroup: “pass”)

Outcomes of colonoscopy

The ITT analysis indicated no significant statistical differences in PDR, ADR, and number of detected polyps between the AI-CNN and control groups (Table 2). The PDR and ADR were 40.8% and 20.1%, respectively, in the AI-CNN group and 43.5% and 22.2%, respectively, in the control group. The mean ± SD numbers of identified polyps were 3.0 ± 3.3 and 2.9 ± 3.4 in the AI-CNN and control groups, respectively. Similarly, the subgroup analyses of patients who received “pass” and “not pass” results indicated comparable PDRs, ADRs, and numbers of polyps between 2 groups (Table 3 and Supplementary Table 2, Supplementary Digital Content 5, https://links.lww.com/AJG/C604).

DISCUSSION

The novel AI-CNN model used in this study demonstrated comparable performance to the standard routine practice in evaluating the quality of bowel preparation before colonoscopy. The bowel preparation results and outcomes of colonoscopy, as evaluated by the PDR, ADR, and number of identified polyps, did not differ significantly between patients who used the AI-CNN model and those who adhered to routine practice. These results suggest that the AI-CNN model, in conjunction with regular instructions, may be applied in the routine practice of bowel preparation before colonoscopy.

Although outcomes seemed to be similar between the AI-CNN and control groups, the AI-CNN “pass” subgroup obtained higher BBPS scores in total and for the right colon than did the control “pass” subgroup. We, therefore, can infer that the AI-CNN model may apply an assessment standard stricter than that applied in patient self-evaluation per routine instructions; therefore, the AI-CNN model may help further improve the quality of bowel cleansing in patients who exhibit adequate bowel preparation. This finding demonstrates the importance of patient education: The “pass” subgroup may have consisted of patients who were compliant with the bowel preparation instructions. In an effort to improve bowel preparation quality and avoid repeated colonoscopies, in the future, we may investigate the benefit of enhanced patient education by using the AI-CNN model or incorporating a patient education system into the AI-CNN model.

The comparable outcomes between the AI-CNN and control groups suggest that our model might be a feasible tool that can be operated on a patient's smartphone. This feature may benefit patients, caregivers, and healthcare providers during the COVID-19 pandemic because with this model, caregivers and nurses would not be required to assess the adequacy of bowel cleansing for patients, which in turn would reduce person-to-person contact and the spread of infectious diseases. In addition, the improved bowel preparation quality of the right colon under the aid of the AI-CNN model may be crucial for the prevention of interval CRC. A retrospective analysis of 75,314 adult patients who underwent colonoscopies revealed that interval cancers were more common in the right colon and in the hepatic flexure (¹⁷). Finally, the AI-CNN model was a manpower-saving option that reduced the need for nurses to evaluate the quality of bowel preparation. Therefore, both patients and healthcare providers may experience benefits from using the AI-CNN model, such as the prevention of interval CRC, decreased frequency of colonoscopies, and reduced medical expenditures, especially in the COVID-19 era.

Despite the fact that the AI-CNN model significantly improved the bowel preparation quality in the “pass” subgroup, the ADR of patients who obtained a “pass” result was similar between the groups. Jover et al. demonstrated that the quality of bowel preparation did not influence ADR (¹⁸). However, a 2020 post hoc analysis of 3 randomized trials reported that high-quality cleansing was associated with a higher ADR (¹⁹). A cross-sectional study involving over 4,900 patients also revealed that higher quality bowel cleansing contributed to higher ADRs and BBPS scores (²⁰). In this study, the lack of a significant difference in ADR between the 2 “pass” subgroups may be attributed to the overall high-quality bowel preparation in these subgroups of patients (mean BBPS: 7.32 vs 7.16 in the AI-CNN and control groups, respectively). Their overall BBPS scores were high, and the difference may have been too small to cause a difference in ADR. Furthermore, assistance for patients in the control group was well organized, which could lead to better bowel preparation in the control group than in the routine clinical practice and underestimate the potential benefits of the AI-CNN model. It is possible that the AI-CNN model may perform better compared with routine practice without a strict protocol.

In this study, we used the nonsplit regimen of a 2L PEG solution for bowel preparation. The 2L PEG solution has been a standard protocol based on Chinese guideline recommendations (²¹), despite guidelines recommending split-dose preparation using large volume of 4L solution (^22,23). The quality of bowel preparation for 2L solution was found to be comparable to that of 4L solution in the low-risk population; in addition, 2L significantly reduced the incidence rates of adverse events (²¹). Although there was no significant difference between the control and study groups in the overall bowel preparation score; however, a difference was noted in the right colon. The preparation regimen may still be a confounding factor here. In the future, the 4L regimen may be applied to validate the current findings.

Different interventions for improving patient education and the quality of bowel preparation have been developed and investigated in previous studies (^24–28). Instructions delivered through visual cartoons and videos have been demonstrated to effectively improve bowel cleansing and users' understanding of colonoscopy (^24,25). In addition, instructions delivered through smartphone applications have been reported to be more effective than traditional written or verbal instructions in increasing the adequacy of bowel preparation (^26–28). A text message program (short message service [SMS]) providing timed-release instructions on diet and bowel preparation was reported to improve the quality of CRC screening and increase adequate bowel preparation rates (²⁹). However, unlike these interventions that target patient education, the novel AI-CNN model was formulated to help patients judge the adequacy of their bowel preparation, which can be influenced by the subjective judgments of individuals. This a novel study to assist patients in decision making during bowel preparation, and the early results are promising.

The main strengths of this study are its large sample size and its design as a colonoscopist-blinded, randomized controlled trial, which make the outcome more reliable. However, the AI-CNN model and this study both have limitations. First, the model can only provide a qualitative evaluation of the bowel preparation images. Future studies focusing on quantitative analysis are necessary to explore the relationship between the bowel preparation images and intestinal cleansing grading. Second, the AI-CNN model relies on the use of the internet and smartphones, which limits the unrestricted use of this method for all patients, especially in developing countries and areas. Older adults may also find the need for a smartphone to be a hindrance because of financial limitations, visual impairments, or a lack of interest or skill in using technological devices. In a study on the use of cartoons to educate patients on bowel preparation, young age was determined to be associated with effective bowel preparation (²⁴). Furthermore, the AI-CNN model was trained based on images labeled with “pass” or “not pass” by an experienced nurse, which may cause the bias pertaining to intraobserver and interobserver discordance. In the future, the AI-CNN model should be upgraded according to the evaluation of several experienced healthcare professionals. Regarding this study, the colonoscopies were performed by 24 colonoscopists across 2 hospitals. Each colonoscopist's procedural methods and evaluation of colonoscopy outcomes may have reflected subconscious bias. To diminish interobserver variability, this study involved single-blinded assessment by experienced colonoscopists, and all colonoscopists and nurses were trained in standard BBPS evaluation. In addition, patient-related factors may have influenced bowel preparation. Previous studies have determined that factors such as incomplete compliance with instructions, constipation, having Parkinson disease, being overweight, and having diabetes are associated with poor bowel preparation (^26,30,31). However, to involve a cohort that mostly represented the general population, this study did not exclude patients with these comorbidities. Second, this study did not investigate the user acceptability of the AI-CNN model. To improve the model and broaden its application in routine practice, evaluating its convenience, accessibility, aspects that cause users difficulty, and user satisfaction is crucial. Finally, the results may be difficult to generalize to a more diverse population because this study was only conducted at 2 hospitals.

The AI-CNN model presented in this study is a novel technology to assist patients in evaluating the quality of bowel preparation before colonoscopy. This AI-CNN model, which achieved comparable outcomes to the routine practice, may serve as an alternative approach for evaluating the quality of bowel preparation before colonoscopy. Future studies should measure the user satisfaction and acceptability of this model.

CONFLICTS OF INTEREST

Guarantor of the article: Yang-Bor Lu, MD and Wei Gong, MD.

Specific author contributions: Y.B.L and W.G.: developed the study concept and design. Y.X.H, S.T.C, and H.Z.X.: designed and trained the AI-CNN model. S.T.C, W.T.C, and F.Y.H.: acquired the data. S.C.L, Y.N.H, H.S.H, and G.L.X.: analyzed the data. Y.B.L and H.S.H.: wrote the first draft of the manuscript. Y.B.L, P.H.L, and W.G.: provided critical revision of the manuscript for important intellectual content. Y.B.L, Y.X.H, and W.G.: jointly developed the structure and arguments of the paper. Y.B.L and W.G.: made critical revisions to the paper.

Financial support: The study was supported by the Xiamen Medical Health Science and Technology Project (Project numbers: 3502Z20199172 and 3502Z20209026) and Xiamen Chang Gung Hospital Science Project (CMRPG1E0891).

Potential competing interests: The authors declare that they have no conflicts of interest related to the subject matter or materials discussed in this article. Any disclosures are made in this section.