Machine learnINg for the rElapse Risk eValuation in Acute Biliary Pancreatitis.

Overview

The MINERVA (Machine learnINg for the rElapse Risk eValuation in Acute biliary pancreatitis) project stems from the need in the clinical practice of taking an operational decision in patients that are admitted to the hospital with a diagnosis of acute biliary pancreatitis. In particular, the MINERVA prospective cohort study aims to develop a predictive score that allows to assess the risk of hospital readmission for patients diagnosed with mild biliary acute pancreatitis using Machine Learning and artificial intelligence. The objectives of the MINERVA study are to: 1. Propose a novel methodology for the assessment of the risk of relapse in patients with mild biliary acute pancreatitis who did not undergo early cholecystectomy (within 3 to 7 days from hospital admission); 2. Propose a Machine Learning predictive model using a Deep Learning architecture applied to easily collectable data; 3. Validate the MINERVA score on an extensive, multicentric, prospective cohort; 4. Allow national and international clinicians, medical staff, researchers and the general audience to freely and easily access the MINERVA score computation and use it in their daily clinical practice. The MINERVA score model will be developed on a retrospective cohort of patients (MANCTRA-1, already registered in ClinicalTrials.gov) and will be validated on a novel prospective multicentric cohort. After validation, the MINERVA score will be free and easy to compute instantly for all medical staff; it will be accessible at any time on the MINERVA website and web app, and will provide an immediate and reliable result that can be a clear indication for the best treatment pathway for the clinician and for the patient.

Acute pancreatitis is the most common pancreatic disease, with a global incidence of 34 cases per 100,000 individuals. This disease counts more than 1.5 million new patients per year worldwide, with a mortality that approaches 1%. Mild biliary acute pancreatitis patients, when admitted to the hospital, can be treated with index, early cholecystectomy (within 3 to 7 days from the acute episode) or conservatively. While conservative treatment can be resolutive, up to 35% of these patients have a relapse within 30 days, and require emergency surgery in a significantly worse overall patient condition, reducing the chances of success. Other than that, relapse dramatically increases the chances of chronic pancreatitis, pancreatic cancer, postoperative complications and overall mortality. Relapse episodes have also an economic impact on healthcare facilities, as a second and longer hospital admission per patient increases the overall medical cost per patient by at least 100%. So far, however, there are no standardised methods to predict relapse of biliary acute pancreatitis in patients who did not undergo early cholecystectomy after the first episode of mild biliary acute pancreatitis. The MINERVA (Machine learnINg for the rElapse Risk eValuation in Acute biliary pancreatitis) project stems from the need in the clinical practice of taking an operational decision in patients that are admitted to the hospital with a diagnosis of mild acute biliary pancreatitis. The MINERVA project aims to reach the following objectives and results: 1. Propose a novel methodology for the assessment of the risk of relapse in patients with mild biliary acute pancreatitis who did not undergo early cholecystectomy after the first episode of mild biliary acute pancreatitis; 2. Propose a Machine Learning predictive model using a Deep Learning architecture applied to data easy to collect from patients; 3. Validate the MINERVA score on an extensive, multicentric, prospective cohort; 4. Allow national and international clinicians, medical staff, researchers and the general audience to freely and easily access the MINERVA score computation and use it in their daily clinical practice. The MINERVA score will provide the clinicians with a validated and standardized assessment of relapse risk that takes into account the personal history, demographic data and laboratory characteristics of each patient. The MINERVA score will be free and easy to compute instantly for all medical staff; it will be accessible at any time on the MINERVA website and web app, and will provide an immediate and reliable result that can be a clear indication for the best treatment pathway for the clinician and for the patient. The MINERVA score model will be developed on a retrospective cohort of patients (MANCTRA-1, already registered in ClinicalTrials.gov) and will be validated on a novel prospective multicentric cohort. Retrospective cohort The model development and initial training will be performed on a retrospective cohort of patients (n=692) collected during a preliminary multicentric study, the MANCTRA-1 study (approved by the Ethics Committee of the University of Cagliari Hospital, MANCTRA-1 - NCT04747990, Prot. PG/2021/7108) conducted by the PI (Dr. Mauro Podda) and the University of Cagliari local responsible of the MINERVA project. Prospective cohort A total of 430 patients will be recruited in the prospective cohort of the MINERVA study. Methods The MINERVA score will be grounded on a Machine Learning model that will be developed and trained on a retrospective cohort and validated on a prospective cohort of patients. All model variables will be processed with kernel Principal Component Analysis (kPCA). The Convex Hull of the scatterplot of the main components will be computed and the smallest rectangle will be extracted. The rectangle will be transformed into a 2d image with a fixed resolution using feature averaging and normalization. The model will be developed at the University of Naples Federico II by Dr. Daniela Pacella with the Machine Learning expertise and supervision. To prevent overfitting, the dataset will be split into training set, test set and validation set. Additionally, k-fold cross-validation will be used. The performance of the MINERVA model will be evaluated using the most adopted measures of accuracy, such as precision, recall and AUC (Area Under the ROC Curve). Additionally, its performance will be compared with that achieved using traditional machine learning methods (SVM, ANN). Missing data will be handled with imputation methods. Variables Age (Years) Sex (Male:Female) Previous episodes of biliary pancreatitis (Yes; No) Admitting speciality (HepatoPancreatoBiliary surgery, General surgery, Internal medicine, Gastroenterology) Body mass index -BMI- (Kg/m2) Clinical history of diabetes (No diabetes; Yes with organ dysfunction; Yes without organ dysfunction) Clinical history of chronic pulmonary disease (Yes; No) Clinical history of hypertension (Yes; No) Clinical history of atrial fibrillation (Yes; No) Clinical history of ischemic heart disease (Yes; No) Clinical history of chronic kidney disease (No; Yes under medications; Yes in permanent renal replacement therapy or in preparation for it) Clinical history of of diseases of the hematopoietic system (Yes; No) Patient on immunosuppressive medications on hospital admission (Yes; No) White Blood Cells -WBC- (cells/mm3) Neutrophils (cells/mm3) Platelets (Plt/mm3) INR (International Normalized Ratio) C-reactive protein -CRP- (mg/L) Aspartate aminotransferase -AST- (U/L) Alanine aminotransferase -ALT- (U/L) Total bilirubin (mg/dL) Conjugated bilirubin (mg/dL) Gamma-glutamil-transpeptidase -GGT- (U/L) Serum amylase (U/L) Serum lipase (U/L) Lactate DeHydrogenase -LDH- (U/L) Choledocholithiasis (Yes; Yes with common bile duct obstruction; No) Cholangitis (Yes; No) ERCP with sphincterotomy (Yes within 24h from hospital admission; Yes within 24-48h from hospital admission; Yes within 48-72h from hospital admission; No) Acute biliary pancreatitis relapse at 30-day, 60-day, 90-day, 1 year.