Artificial Intelligence for Prediction and Diagnosis of GI Diseases
Welcome to Health Connect, the podcast for health professionals where we will share the latest news and information on science and technology in the medical field. In this episode, we'll talk about the role of artificial intelligence in the prediction and identification of gastrointestinal diseases.
Did you know that, among the fields of application of artificial intelligence, diagnostics within endoscopy, radiology, and histopathology are the most promising fields reported by gastroenterologists?1
According to a study by van der Zander and colleagues, more than 80 percent of gastroenterology specialists believed that artificial intelligence systems will improve quality of care. Particularly, with a special interest for developments that can help in colorectal polyp detection and in capsule endoscopy.1
While it may seem like a recent trend, the application of artificial intelligence in the field of gastroenterology has been studied for many years to automate the interpretation of diagnostic procedures in multiple pathologies; with special interest in procedures like endoscopy, pill video endoscopy, ultrasound manometry, and microcytoscopy. These procedures produce large amounts of data that require the time and attention of an expert to draw clinical conclusions.2
Artificial intelligence models have been shown to recognize polyps as well as areas and degrees of inflammation accurately; for example, the recent randomized controlled trial led by Glissen Brown reported that the miss rates of a deep learning model for polyp detection were 20.1 percent, which is lower than the 31.2 percent achieved with the standard method.2
So, how accurate could artificial intelligence models be in diagnosing gastrointestinal luminal pathologies? To answer this, Parkash and colleagues conducted a systematic review and meta-analysis of 73 studies to evaluate the diagnostic efficacy of AI models compared to the gold standard and histopathology in various gastrointestinal luminal pathologies.2
According to the findings reported by the team, artificial intelligence models such as deep learning and ensemble methods were the most commonly deployed models to detect gastrointestinal luminal pathologies, with high reported sensitivity and specificity that exceeded 90 percent.2
The most common pathologies investigated were polyps and neoplasms, both identified with high accuracy; while conditions such as Barrett's esophagus and H. Pylori infection reported even higher accuracy than human experts.2
It is worth noting that of all the studies reviewed by Parkash's team, only one investigation came from a low- or middle-income country, that being Pakistan, while the remaining ones came from upper-middle-income or high-income countries.2
This is relevant given that the burden of benign and neoplastic gastrointestinal pathologies in low- or middle-income countries is disproportionately higher, which together with poor access to healthcare facilities, lack of equipment and trained professionals, contributes to poor health outcomes. Therefore, despite its potential, artificial intelligence research in these regions is sparse.2
Continuing with the potential of machine learning and deep learning to identify new disease characteristics, disease severity and clinical outcomes, some of the most interesting achievements in the last 10 years correspond to research in inflammatory bowel disease, or IBD, which offers a promising path to better clinical outcomes for patients with this condition.3
The gold standard for the diagnosis of IBD is the thorough examination of biopsies obtained from the patient during ileocolonoscopy and upper endoscopy. However, current diagnostic methods cannot definitively classify the disease in some patients after initial examination. Establishing the correct diagnosis as early as possible is crucial because treatment for ulcerative colitis and Crohn's disease can vary, especially in severe cases, where surgical interventions need to be considered.3
Since the diagnosis and management of IBD requires the synthesis of endoscopic, radiological and histopathological data, artificial intelligence technologies can leverage this large collection of data to enable the integration of multiple diagnostic modalities into a unified disease assessment model.3
In this context, Zulqarnain and Rhoads authored a review on the use of deep and machine learning models to streamline IBD diagnosis and assessment. First, to address advances in machine learning in endoscopy, they feature the progress achieved by Chierici's team, who developed a deep learning model that could distinguish between healthy and inflamed tissue, between ulcerative colitis and healthy tissue, and between ulcerative colitis and Crohn's disease.3
Although the model’s best performance was on the first two tasks, it could still distinguish ulcerative colitis from Crohn's disease fairly accurately, making it a promising step toward the creation of a useful tool for diagnosing IBD.3
On the other hand, Tekenaka and colleagues developed a deep learning model that can be used in real-time during endoscopy to predict histologic remission and calculate an endoscopic index of ulcerative colitis severity score, known as UCEIS.3
Their model's histologic prediction agreed 97 percent of the time with pathologists' assessments, demonstrating the model's ability to accurately assess endoscopic inflammation.3
Likewise, Maeda and researchers used an artificial intelligence system called EndoBRAIN-UC and endocytoscopy techniques to predict the likelihood of active disease or healing states in ulcerative colitis patients undergoing colonoscopy.3
They found that for patients with a prediction of active disease, the risk of clinical relapse in 12 months was 28.4% versus only 4.9% in patients with predictions of healing.3
Furthermore, Ferreira and team developed a model to automatically detect ulcers in small bowel capsular endoscopy data, while Afonso and colleagues developed a model that automatically quantifies ulcers and predicts bleeding potential. A clinician could take up to 120 minutes to review the same video, but the model was able to analyze the images in an average of 16 minutes.3
In the field of machine learning in radiology, many metrics have been investigated to evaluate inflammation and stricture in the bowel from magnetic resonance enterography data, but machine and deep learning algorithms are under development to identify features related to the condition.3
For example, the model developed by Ziselman's team was able to accurately segment magnetic resonance enterography images of the terminal ileum and classify them as normal or increased wall thickness.3
Even more interestingly, this model was trained on data obtained without intravenously administered contrast, which increases the accessibility and safety of its use.3
Finally, the use of machine learning to automate histopathological scoring could extract meaningful disease features, and help develop new metrics of ulcerative colitis.3
Thus, let’s present the case of histological scoring indices such as The Nancy index, first published in 2017. Using a dataset of 200 images in patients with ulcerative colitis, a novel artificial intelligence system was developed to automate the evaluation of this index, illustrating the potential of using these models to standardize histological scoring.3
Deep learning has previously been used to analyze intestinal histopathological images of surgical specimens in patients with Crohn's disease, showing adipocyte shrinkage and mast cell infiltration to predict postoperative recurrence of Crohn's disease within 2 years after surgery.3
Further, Ohara’s team trained a deep learning model to quantify goblet cell mucus area in whole slide images of patients with ulcerative colitis in endoscopic remission, and found that this method is useful in predicting the clinical relapse in patients with UC in clinical and endoscopic remission.3
On the other hand, there’s the example of the PICaSSO Histologic Remission Index, or PHRI. This was developed using biopsies from 307 patients across 11 centers in Europe and North America under the idea of developing a simple automated histologic index that would correlate with published histologic scores, endoscopic activity scores and clinical outcomes, focusing on a single variable neutrophil infiltration. Since its development, PHRI has been used to develop multiple artificial intelligence algorithms.3
Let us now take a closer look at the development of a computer-aided diagnosis system based on artificial intelligence that was created for the evaluation of ulcerative colitis biopsies and the prediction of prognosis.4
To date, histologic remission is the optimal treatment goal for ulcerative colitis; however, histologic evaluation is limited by low concordance between pathologists.4
According to the results, this validated artificial intelligence model can accurately detect ulcerative colitis activity defined by different histological indices, including the PHRI, the Robarts Histology Index or RHI, and the Nancy Histology Index. It can also estimate the corresponding endoscopic activity and stratify the risk of flares as human clinicians do.4
Note that the PHRI defines disease activity and remission according to the presence or absence of neutrophils in areas of the biopsy: superficial epithelium, lamina propria, cryptal epithelium, and cryptal lumen; so, the absence of neutrophils in all areas is considered remission, while their presence defines disease activity.4
This developed AI model identified histologic activity and/or remission with a sensitivity and specificity of 89 percent and 85 percent, respectively in the PHRI; 94 percent and 76 percent, respectively in the Robarts histology index; and 89 percent and 79 percent, respectively in the Nancy histology index.4
Similarly, this model predicted endoscopic remission/activity with 79 percent and 82 percent accuracy. The hazard ratio for disease recurrence between the histologic activity or remission groups was 3.56 for the pathologist-assessed PHRI and 4.64 for the artificial intelligence-assessed PHRI.4
Of note, the most innovative application of this system is outcome stratification, as it is the first AI tool to stratify the risk of disease flare based on histologic data, considered the most rigorous assessment of ulcerative colitis activity, and observed a strong association between histologic activity and disease flare regardless of whether the classification was performed by pathologists or by the AI system.4
While this system cannot distinguish between different degrees of disease severity, it holds promise for accelerating, simplifying, and standardizing the histologic evaluation of ulcerative colitis in clinical practice, providing accurate diagnostic information to the clinician.4
In addition, the use of PHRI tells us that automated detection of neutrophils may help shed light on their role in mucosal inflammation. Similar systems may be developed for other cell types or tissues, broadening the fields of application.4
Inicio de cápsula
According to a study by van der Zander and colleagues, gastroenterology physicians expect artificial intelligence to change their work. Most of them expect the implementation of AI in healthcare within five years. In exploring the attitudes of GI specialists towards this technology, the most frequently mentioned advantages were improved quality of care and time savings for both patients and physicians, as well as reduced risk of medical errors, more time available for patient-clinician interaction, standardization in interpretation of results, more objective diagnosis, gains in efficiency, and cost reduction.1
It is interesting to compare these positive perspectives with those of gastroenterology patients, who showed greater distrust and a greater lack of knowledge about artificial intelligence, while reporting greater compliance for physicians using artificial intelligence or physicians alone compared with an artificial intelligence system alone.1
This highlights that artificial intelligence should be a tool that enhances or assists human intelligence rather than replacing it, creating a synergy between technology and doctors that allows for more effective health care.1
Fin de cápsula
Welcome back. As we know, gastric cancer is the fourth most common neoplastic disease and the second most common cause of cancer death worldwide. The majority of gastric cancer patients do not present specific symptoms, and some early signs in patients mimic gastritis and indigestion, making it easy for patients to disregard the symptoms.5
Current screening for gastric cancer is heavily reliant on invasive and expensive endoscopic procedures and pathologic biopsies, making prevention and modification of risk factors critical. In their retrospective, single-center study, Afrash and colleagues set out to develop an inexpensive, non-invasive, rapid, and highly accurate diagnostic model using six machine learning algorithms to classify 2,029 patients at high or low risk of developing gastric cancer by analyzing individual lifestyle factors.5
The authors found 11 important influencing factors for gastric cancer risk, mainly Helicobacter pylori infection, which was shown to be three times more effective than the other attributes for early prediction of this cancer.5
High salt intake, chronic atrophic gastritis and low fruit consumption were significantly associated with a high risk of gastric cancer too. In contrast, weight loss and stress ranked as the 10th and 11th most important risk factors in the prediction of gastric cancer, respectively.5
With these risk factors in mind, the authors applied machine learning techniques to stratify gastric cancer risk noninvasively. Based on the results, the eXtreme Gradient Boosting classifier performed best in predicting gastric cancer risk compared to the other machine learning techniques, achieving an average accuracy of 83.4 percent, an average specificity of 85.9 percent, and an average sensitivity of 83.7 percent.5
Patients were first screened by the optimal models developed, and then high-risk cases were referred to specialized centers for additional diagnostic procedures, demonstrating a highly adaptable and low-cost screening approach and increasing the coverage of gastric cancer screening in clinical practice. 5
It is interesting to think that the application of AI models such as this one could help physicians to choose timely interventions that improve the chance of survival and the quality of life of patients. Moreover, it is critical that physicians are familiar with the development of these technologies, as this allows them to implement optimal predictive models and evaluate the most appropriate features for clinical implementation.5
Thanks for joining us on this episode of Health Connect. Don't miss out on our next episode.
Discover more medical news on Viatris Connect.
Referências
- 1. Van der Zander QEW, van der Ende-van Loon MCM, Janssen JMM, Winkens B, van der Sommen F, Masclee AAM, et al. Artificial intelligence in (gastrointestinal) healthcare: patients' and physicians' perspectives. Sci Rep. 2022;12(1):16779. Available at: Artificial intelligence in (gastrointestinal) healthcare: patients' and physicians' perspectives
- 2. Parkash O, Siddiqui ATS, Jiwani U, Rind F, Padhani ZA, Rizvi A, et al. Diagnostic accuracy of artificial intelligence for detecting gastrointestinal luminal pathologies: A systematic review and meta-analysis. Front Med (Lausanne). 2022;9:1018937. Available at: https://www.frontiersin.org/articles/10.3389/fmed.2022.1018937/full
- 3. Zulqarnain F, Rhoads SF, Syed S. Machine and deep learning in inflammatory bowel disease. Curr Opin Gastroenterol. 2023; doi: 10.1097/MOG.0000000000000945. Epub ahead of print. Available at: https://journals.lww.com/co-gastroenterology/Fulltext/9900/Machine_and_deep_learning_in_inflammatory_bowel.71.aspx
- 4. Iacucci M, Parigi TL, Del Amor R, Meseguer P, Mandelli G, Bozzola A, et al. Artificial Intelligence Enabled Histological Prediction of Remission or Activity and Clinical Outcomes in Ulcerative Colitis. Gastroenterology. 2023; S0016-5085(23)00216-0. doi: 10.1053/j.gastro.2023.02.031. Epub ahead of print. Available at: https://www.gastrojournal.org/article/S0016-5085(23)00216-0/fulltext
- 5. Afrash MR, Shafiee M, Kazemi-Arpanahi H. Establishing machine learning models to predict the early risk of gastric cancer based on lifestyle factors. BMC Gastroenterol. 2023;23(1):6. Available at: Establishing machine learning models to predict the early risk of gastric cancer based on lifestyle factors
NON-2023-9597
NON-2023-2512