Abstract
Introduction: Though cyst fluid glucose levels have been shown to be a reliable marker for differentiating Mucinous and Non-Mucinous pancreatic cysts, a number of cysts have a more complex association with CEA and amylase enzyme values. Our project aimed to utilize machine learning to develop a decision tree that will provide clinicians with a practical tool to optimally determine the mucinous character of the lesion from lab results alone. Methods: This is a retrospective study conducted at a high-volume advanced endoscopy center. Clinical and lab data was abstracted from the electronic medical records of all patients who underwent pancreatic cyst aspiration between June 2015 and November 2017. Statistical analysis was done using STATA v 15.1. A decision tree was built using the J48 classifier algorithm in Weka 3.8.0. Factors used for data modelling were age, gender, diabetic status, and levels of glucose, CEA and amylase. The decision tree thus generated was validated using 10-fold cross validation. Results: A total of 57 lesions were included in the analysis. The average age of the cohort was 62.6 years (SD ±16.4). The most common diagnosis was pseudocyst (n=21, 36.8%), followed by 20 IPMNs (14 branch ducts (24.6%), 3 main duct (5.3%), and 3 mixed duct cysts (5.3%)), 11 mucinous Cystadenomas (19.3%), 4 lymphoepithelial cysts (7.0%), and 1 ciliated foregut cyst (1.6%). Thus, there were 31 Mucinous and 26 Non-Mucinous pancreatic cysts in our cohort. Mucinous cysts had higher CEA levels compared to Non-Mucinous cysts (Median 56.4 ng/mL vs 3.1 ng/mL, Mann-Whitney-U, pvalue=0.0018). However, Glucose and Amylase levels were lower in Mucinous cysts compared to Non-mucinous cysts (Both Mann-Whitney-U, p-value= 0.0002 and 0.0107 respectively). The decision tree algorithm revealed that pancreatic cysts will be non-mucinous if the glucose level is greater than 16 mg/dL. Additionally, cysts will non-mucinous if their CEA < 3.1 ng/mL and undetectable glucose levels. This model thus correctly classified 77.2% of all cysts, and, had a kappa statistic of 0.5473. Its ROC was 0.757 and r2-statistic was 0.85. Conclusion: Our decision tree analysis reveals that in addition to the expected non-mucinous cysts with high glucose values, there exists a small group of zero-glucose low-CEA cysts with CEA < 3.1 ng/mL and undetectable glucose levels. Our decision tree can be used by endoscopists in clinical practice as a simple and easy tool to predict the mucinous/non-mucinous character of pancreatic cysts.73_A Figure 1. Pancreatic Cysts Descriptive Statistics73_B Figure 2. Pancreatic Cyst Fluid Decision Tree Analysis - Results73_C Figure 3. Pancreatic Cyst Decision Tree Model
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call