Abstract: Since 20 10, the application of artificial intelligence (A I) in medicine has made substantial progress. The application of artificial intelligence in gastroenterology includes endoscopic lesion analysis, cancer detection and analysis of inflammatory lesions or gastrointestinal bleeding in wireless capsule endoscope. Artificial intelligence is also used to evaluate liver fibrosis and distinguish pancreatic cancer patients from pancreatitis patients. Artificial intelligence can also determine the prognosis of patients or predict their response to treatment based on multiple sets of data. This paper summarizes the methods of artificial intelligence to help doctors diagnose or determine prognosis, and discusses its limitations. It is understood that further randomized controlled studies are needed before the health authorities approve artificial intelligence technology.
Keywords: deep learning; Machine learning; Neural network; digestive system
There is no single definition of artificial intelligence. The concept of artificial intelligence includes programs that perform functions related to our human intelligence, such as learning and exploring and solving problems. Artificial intelligence, machine learning and deep learning are conceptually overlapping disciplines (see figure 1). Machine learning is a broad subject, including computer science and statistics. Machine learning programs iterate repeatedly to improve the performance of specific tasks, thus generating algorithms for analyzing data and learning description and prediction models. Most data used for training are organized in tabular form, in which objects or individuals are rows and variables, whether numbers or classifications, are columns. Machine learning can be roughly divided into supervised methods and unsupervised methods. The purpose of unsupervised learning is to identify groups according to the * * * nature of data without prior knowledge of the number or characteristics of groups. Supervised learning involves using the representation of input-output pairs of each object in training data. Input the characteristic description containing individuals, and output the interesting results that need to be predicted, which can be the category of classification tasks or the numerical value of regression tasks. Supervised machine learning algorithm learns the mapping relationship between input and output pairs, and automatically predicts the corresponding output when a new output appears. Two studies used electronic medical records to determine the risk of abdominal diseases, and 1 study used genetic factors to determine the risk of IBD. Two-thirds of the studies (2 1 2 14) used K-fold cross-validation to avoid over-fitting, and the accuracy rate of patients in 2 1 2 was about 90%.
Many studies have verified the ability of artificial intelligence to predict the treatment response of patients with IBD. Waljee et al. developed a machine learning method based on age and laboratory data, which is cheaper and can predict the clinical response of patients to thiopurine more accurately than measuring the metabolite of 6- thioguanine nucleotide (6-TGN) (AUC 0.86 vs 0.60) [94]. Then, according to biomarkers, imaging data and endoscopic results, they improved the previous ML model to predict the objective remission of patients treated with thiopurine. ML model is superior to the measurement of 6-TGN level (AUC 0.79 vs 0.49)[95]. ML model was used to analyze the data of phase III clinical trial of Vedolizumab in the treatment of ulcerative colitis, and compared with the fecal calcium protection level with AUC of 0.7 1 in the sixth week. AI can predict which patients will achieve endoscopic remission in the 52nd week without corticosteroids, and the AUC value of the predicted performance is 0.73. Therefore, when the benefits of Vedolizumab are not obvious in the first 6 weeks, this algorithm can be used to select patients who continue to use Vedolizumab [96]. In addition, there is an artificial intelligence algorithm, which combines the data of microbial population with clinical data to determine the clinical response of patients with IBD and predict the AUC of patients for integrated treatment to be 0.78[97]. The sensitivity and specificity of neural network in identifying patients with ulcerative colitis who need further surgery after cell replacement therapy are 0.96 and 0.87 respectively [98].
An artificial intelligence system to predict the onset or progress of IBD is also being developed. A neural network that analyzes the early biopsy images of patients with Crohn's disease has an accuracy rate of 83.3% in identifying the disease progress and 86.0% in predicting whether surgery is needed [99]. Waljee et al. established ML method to analyze the data of electronic medical records, and predicted that the AUC value of steroid use in IBD-related inpatients and outpatients would reach 0.87[ 100] within 6 months. Artificial neural network can accurately predict the frequency of clinical recurrence in patients with IBD [10 1].
There have been 12 studies to verify the ability of AI to detect small intestinal bleeding in infinite capsule endoscopic images (Table 3) [55,102-12]. Eight projects in 12 adopted special verification technology, mainly K-fold cross-verification. Among these studies, 9 studies have an accuracy rate of more than 90%.
For patients with acute upper gastrointestinal bleeding or lower gastrointestinal bleeding, the cause of bleeding can be easily determined by endoscopy. However, a large number of patients with recurrent bleeding need repeated endoscopic examination and treatment. Therefore, the ML model was developed to determine the patients at risk of recurrent bleeding and those most likely to need treatment, and to estimate the mortality rate. These models use clinical and/or biological data and identify these patients with about 90% accuracy [113-117]. Based on the retrospective analysis of 22,854 patients with gastric ulcer and 65,438+0,265 patients, the ML model can identify patients with recurrent ulcer bleeding according to their age, hemoglobin level, gastric ulcer, gastrointestinal diseases, malignant tumor and infection. The AUC of patients with recurrent ulcer bleeding within 1 year was 0.78, and the accuracy was 84.3%.
Twenty-two studies have tested the ability of AI to assist in the diagnosis and treatment of pancreatic diseases or liver diseases (Table 4). Among them, there are 6 AI systems for pancreatic cancer, of which 5 studies are based on endoscopic ultrasound [1 18- 122] and 1 study is based on serum markers [123]. The AUC of pancreatic cancer patients determined by these studies is about 90%. Of the 16 liver studies, 7 were aimed at detecting fibrosis related to viral hepatitis [124- 130] and 6 were aimed at detecting nonalcoholic fatty liver [13 1- 136]. Two studies identified esophageal varices [137, 138]. 1 item to evaluate patients with unexplained chronic liver disease [139]. Among them, 13 studies used electronic medical records and/or biometric data to establish algorithms, and 3 studies used elastic imaging data. Except for two projects, all the studies used specific verification techniques, mainly K-fold cross-verification. The accuracy of these models is about 80%.
In addition to improving the accuracy of diagnosis, AI method is also needed to determine the prognosis of patients and predict the progress of diseases. Pearce et al. established a ML model to predict the severity of acute pancreatitis according to APACHE II score and C-reactive protein level. The AUC value of their model is 0.82, the sensitivity is 87%, and the specificity is 7 1%[ 140]. According to the age, hematocrit, serum glucose and calcium levels and urea nitrogen levels of patients with acute pancreatitis, Hong et al. created an artificial neural network to evaluate patients with persistent organ failure, with an accuracy rate of 96.2%[ 14 1]. Jovanovic et al. developed an ANN model to identify the needs of patients with common bile duct stones for therapeutic endoscopic retrograde cholangiopancreatography according to the results of clinical, laboratory and percutaneous ultrasound examination, and its AUC was 0.88[ 142].
Banerjee et al. developed an artificial neural network based on clinical and laboratory data to determine the possibility of death of patients with liver cirrhosis within 1 year with 90% accuracy. This model can be used to determine the best candidate for liver transplantation [143]. Konerman et al. established a machine learning model based on clinical, laboratory and histopathological data to identify the highest risk of disease progression and liver-related outcomes (liver-related death, liver decompensation, hepatocellular carcinoma, liver transplantation or Child-Pugh score increased to 7) in patients with chronic hepatitis C virus infection. In the validation set of 65,438+0,007 patients, the AUC value of the model reached 0.708. Khosravi et al. established a neural network to predict the survival time of 1 168 liver transplant patients. The model can estimate the survival probability of 1-5 years, with AUC of 86.4% and Cox proportional risk regression model of 80.7%[ 146]. Researchers also use artificial neural networks to match liver donors and recipients, thus providing powerful decision-making techniques [147]. In addition, ML model can help predict the response to treatment. Takayama et al. established an artificial neural network to predict the response of patients with chronic hepatitis C virus infection to pegylated interferon a-2b combined with ribavirin. The prediction sensitivity and specificity reached 82% and 88%, respectively.
Artificial intelligence will become an important means for gastroenterologists and hepatologists to diagnose patients, choose treatment methods and predict prognosis. Many methods have been developed under these objectives and show different performance levels. Because of the differences in performance indicators, it is difficult to compare the results of these studies. Artificial intelligence seems to be particularly valuable under endoscope, which can increase the detection of malignant and precancerous lesions, inflammatory lesions, small intestinal bleeding and pancreaticobiliary duct diseases. In hepatology, artificial intelligence technology can be used to determine the risk of liver fibrosis in patients and allow some patients to avoid liver biopsy.
Our review only covers the articles listed in PubMed, and some publications in computer science and medical image analysis journals may be omitted. Nevertheless, in the past 20 years, artificial intelligence has become an important part of gastroenterology and hepatology research. Although the focus of this review is on auxiliary diagnosis and prognosis, artificial intelligence in other research directions is also being explored, such as the quality control evaluation of endoscopy based on machine learning (cecal marking, the follow-up suggestion of machine learning evaluation to detect colonoscopy), and the application of AI in gastrointestinal tract is also expanding.
It is worth noting that the current AI technology is limited by the lack of high-quality data sets. Most of the evidence used to develop ML algorithm comes from preclinical research, but it has not been applied to clinic at present. In addition, DL algorithm is considered as a black box model, which is difficult to understand the decision-making process and prevents doctors from discovering potential confounding factors. It is also important to consider moral challenges. Artificial intelligence does not know the patient's preferences or legal responsibilities. Who should be responsible for the misdiagnosis of endoscopy-endoscopist, programmer or manufacturer? In addition, when determining the risk of liver fibrosis related to viral hepatitis, inherent prejudices such as racial discrimination can easily be incorporated into artificial intelligence algorithms, especially in the field of hepatology. When developing an artificial intelligence model, it is important to consider these factors and verify the model in a series of people. There is always inherent uncertainty in medicine, so it is impossible to make a perfect prediction. Some research gaps related to artificial intelligence still need to be studied in the fields of gastroenterology and hepatology (Table 5).
In the field of gastroenterology and hepatology, the development of artificial intelligence has no turning back, and its future influence is enormous. Using artificial intelligence can increase people's access to care in developing areas, especially in assessing the risk of patients suffering from viral hepatitis or intestinal parasitic diseases. Smart phones can remotely monitor patients' health by using artificial intelligence technology, and a method for measuring fecal calcium protectant for IBD patients at home has been established [149]. Artificial intelligence can also identify new therapeutic targets by integrating molecular, genetic and clinical data of large patient data sets. However, artificial intelligence will not completely replace doctors, and artificial intelligence will still assist doctors in their work. Although the machine can make accurate predictions, in the end, the medical staff still have to make decisions for patients according to their preferences, environment and morality.