Artificial Intelligence Prediction for the Severity of Acute Pancreatitis

Overview

The incidence of acute pancreatitis (AP) is increasing nowadays. The diagnosis of AP is defined according to Atlanta criteria with the presence of two of the following 3 findings; a) characteristic abdominal pain b) amylase and lipase values ≥3 times c) AP diagnosis in ultrasonography (USG), magnetic resonance imaging (MRI), or computerized tomography (CT) imaging. While 80% of the disease has a mild course, 20% is severe and requires intensive care treatment. Mortality varies between 10-25% in severe (severe) AP, while it is 1-3% in mild AP. Scoring systems with clinical, laboratory, and radiological findings are used to evaluate the severity of the disease. Advanced age (\>70yo), obesity (as body mass index (BMI, as kg/m2), cigarette and alcohol usage, blood urea nitrogen (BUN) ≥20 mg/dl, increased creatinine, C reactive protein level (CRP) \>120mg/dl, decreased or increased Hct levels, ≥8 Balthazar score on abdominal CT implies serious AP. According to the revised Atlanta criteria, three types of severity are present in AP. Mild (no organ failure and no local complications), moderate (local complications such as pseudocyst, abscess, necrosis, vascular thrombosis) and/or transient systemic complications (less than 48h) and severe (long-lasting systemic complications (\>48h); organ insufficiencies such as lung, heart, gastrointestinal and renal). Although Atlanta scoring is considered very popular today, it still seems to be in need of revision due to some deficiencies in the subjects of infected necrosis, non-pancreatic infection and non-pancreatic necrosis, and the dynamic nature of organ failure. Even though the presence of 30 severity scoring systems (the most accepted one is the APACHE 2 score among them), none of them can definitely predict which patient will have very severe disease and which patient will have a mild course has not been discovered yet. Today, artificial intelligence (machine learning) applications are used in many subjects in medicine (such as diagnosis, surgeries, drug development, personalized treatments, gene editing skills). Studies on machine learning in determining the violence in AP have started to appear in the literature. The purpose of this study is to investigate whether the artificial intelligence (AI) application has a role in determining the disease severity in AP.

In a retrospective way, 1550 patients who were followed up at the Gastroenterology Clinic of Bezmialem Foundation University between October 2010 and February 2020 period and who were diagnosed with AP according to Atlanta criteria were screened. After the removal of 216 patients with missing data, 1334 patients were included in the study for evaluation. 1. Patient demographic information; \[age (yo), gender (male/female), cigarette/alcohol usage (as yes or no)\], clinical information; \[height (centimeters), weight (kilograms), BMI (as kg/m2), presence of diabetes mellitus and hypertension (yes or no)\], etiology of AP such as gallstones, alcohol, etc., and laboratory tests those taken within the first 24 hours of the admission; \[CRP level (mg/dl, normally: 0-5), BUN level (mg/dl, normally; 9,8 - 20,1), creatinine level (mg/dl, normally; 0,57 - 1,11), number of leukocytes (normally 4.5 to 11.0 ×109/L) and hematocrit level (%, normally: 35,5-48%)\], as well as Balthazar tomographic scoring \[0: normal, 1: an increase in pancreatic size, 2: inflammatory changes in pancreatic tissue and peripancreatic fatty tissue, 3: irregularly bordered, single fluid collection, 4: irregularly bordered 2 or more fluid collections, 5 to 10 different degrees of necrosis)\], will be recorded in the excel file. 2. Revised Atlanta scoring will also be recorded within a week period of hospital admission as mild, moderate, and severe scores. Infected pancreatic necrosis and sepsis that developed during the course of acute pancreatitis will be accepted as severe acute pancreatitis due to the inadequacy of some issues in Atlanta scoring. The severity of the disease will be evaluated according to the Atlanta scores. And the results of the artificial intelligence study will be matched according to the results of Atlanta scoring. 3. Complications are classified as 0; none, 2; local complications: pseudocyst, abscess, necrosis, thrombosis, and mesenteric panniculitis, 3; systemic complications: lung, kidney, gastrointestinal and cardiovascular complications, 4; mixed serious complications/co-morbidity situations, 5: infectious and septic complications. 4. Additionally, invasive procedure requirements such as endoscopic ultrasonography (EUS), endoscopic retrograde cholangiopancreatography (ERCP) (as yes or no), length of hospital stay (less than 10 days or more than 11 days), intensive care unit requirement (present or not), number of future AP attacks (in duration after a month of hospital admission, as of one attack or more than one attack), and survival (death, alive) will also be recorded. Machine Learning Algorithm is used: Gradient Boosted Ensemble Trees Trees. ("Greedy Function Approximation: A Gradient Boosting Machine" by Jerome H. Friedman (1999)). The dataset has been partitioned with a 90%-10% ratio. 10% is for validation and 90% is for AI machine learning. 90% machine learning part has also been divided into two parts as 70% for AI Learning and 30% for testing the learning. For this purpose, 5-fold stratified sampling has been used Artificial Intelligence Methods of the Study Features Used for AI Machine Learning: 1. Gender: M/F 2. Age: Continuous Value 3. Height (cm): Continuous Value 4. Weight (Kg): Continuous Value 5. BMI Groups: Group 1: ≤ 25 kg/m2; Group 2; 25-30 kg/m2; Group 3: \>30,1 kg/m2 6. Cigarette: 0; No, 1; Yes 7. Alcohol: 0; No, 1; Yes 8. Diabetes mellitus: 0; No, 1; Yes 9. Hypertension: 0; No, 1; Yes 10. Etiology: 1; biliary, 2; Alcohol, 3; hypertriglyceridemia, 4; hypercalcemia, 5; drug, 6; congenital, 7; cryptogenic, 8; endoscopic retrograde cholangiography (ERCP), 9; oddy sphincter dysfunction (OSD), 10; malignity, 11; intra papillary mucinous neoplasia (IPMN), 12: primary sclerosing cholangiography (PSC) 13: autoimmune, 14: multiple etiology 11. Leucocyte number (WBC): N; 4,5-11x100 12. Hematocrit (Hct): N; %35,5-48 13. C reactive protein (CRP): N: 0-5 mg/dl 14. Blood urea nitrogen (BUN): N: 9,8-20,1 mg/dl 15. Creatinine (KREA): N: 0,57-1,11 mg/dl 16. Baltazar Scoring (BLTZR): 0; Normal P, 1; Increase in pancreatic size, 2; Inflammatory changes in pancreatic tissue and peripancreatic fatty tissue, 3; Irregularly bordered, single fluid collection, 4, Irregularly bordered 2 or more fluid collections, with various degrees of necrosis (ranging between 5 and 10) In Artificial Intelligence, Decision Tree Models are widely used for supervised machine learning. They may depend on the Gini index, gain ratio/entropy, chi-square, regression, and so on. In AI they are preferred because they generate understandable rules for humans unlike other machine learning algorithms such as Artificial Neural Networks and Support Vector Machines. On the other hand, they are considered to be weak learners. That means they are highly affected by noise and outliers existing in the data set. In order to go around this handicap, models like Random Forest, Ensemble Trees, Gradient Boosting have been developed. Random forest and Ensemble trees generate rules by applying a certain decision tree algorithm to the portions of the data set vertically and horizontally. This technique dramatically reduces the error occurring in learning. After learning processes are completed, they combine weak decision trees into a strong and bigger decision tree model. Ensemble learning models achieve better learning by minimizing the average value of the loss function on the training set via a F ̂(x) approximation. The idea is to apply a steepest descent step to the minimization problem in a greedy fashion. In this study, the gradient boost tree model which was proposed by Friedman has been used for machine learning. This model chooses a separate optimal value for each of the tree's parts rather than a single one for the whole tree. This approach can be used to minimize any differentiable loss L(y, F) in conjunction with forwarding stage-wise additive modeling. It is reported that the gradient boosting tree model outperforms random forest and regular ensemble trees in many cases. The goal of the algorithm is to find an approximation F\_m (x\_i) which minimizes the expected L(y,F(x)) loss function. The algorithm may be summarized as follows: Inputs: A training data set: {(x\_i,y\_i )} i=1 to n with n dimension and a class variable A differentiable loss function: L(y,F(x)) The number of iterations: M. Output: F\_m (x\_i) Algorithm: Initialize the model with a constant value: F\_0 (x)=arg min⁡∑\_(i=1)\^n▒〖L(y\_i,γ)〗 For m = 1 to M: Compute pseudo-residuals rim r\_im=-\[(∂L(y\_(i,) F(x\_i )))/(∂F(x\_i))\] Train a base learner to pseudo-residuals, using the training set: {(x\_i,y\_i )} i=1 to n Compute multiplier γ γ=arg min⁡∑\_(i=1)\^n▒〖L(y\_i,F\_(m-1) (x\_i )+γh\_m (x\_i ))〗 Update the model: 〖F\_m (x\_i)=F〗\_(m-1) (x\_i )+γ\_m h\_m (x\_i ) Output F\_m (x\_i) In the analysis, Synthetic Minority Oversampling Technique (SMOTE) \[5\] has been used in order to avoid the disadvantage of class variable imbalance. SMOTE is a data augmentation technique to increase data. In some cases, the class variable may not have an equal amount of values from all cases. For example, there may be much more survived patients than those who lost their lives. In this kind of situation, data are augmented. There was an imbalance in the class variables in the data set of this study. So, SMOTE has been applied to increase the minority classes for training. The dataset has been partitioned with a 90%-10% ratio. 10% is for validation and 90% is for AI machine learning. 90% machine learning part has also been divided into two parts as 70% for AI Learning and 30% for testing the learning. For this purpose, 5-fold stratified sampling has been used. KNIME analytic platform has been used for the AI machine learning.