Detailed Abstract
[Plenary Session]
[PL 4] Development of Risk Prediction Platform for Pancreatic Fistula after Pancreatoduodenectomy using Artificial Intelligence
In Woong HAN1, Naru KIM1, Youngju RYU1, Dae Joon PARK1, Sang Hyun SHIN1, Jin Seok HEO1, Dong Wook CHOI1, Baek Hwan CHO2
1Departments of Surgery, Samsung Medical Center, Sungkyunkwan University College of Medicine, Korea
2Smart Healthcare & Device Research Center, Samsung Medical Center, Sungkyunkwan University College of Medicine, Korea
Introduction : Postoperative pancreatic fistula (POPF) is a frequent and life-threatening complication following pancreatoduodenctomy (PD). Artificial intelligence technology is receiving a lot of attention and is actively being distributed in medical field. However, no studies have been reported on the application of this technique to outcomes after pancreatic surgery. As a result, this study aimed to develop risk prediction platform for POPF using artificial intelligence model.
Methods : From 2007 to 2016, medical records of 1771 patients at Samsung Medical Center who underwent PD were reviewed. A total of 53 variables were inserted into training algorism. Algorithms which made risk prediction platform were ‘the random forest method (RFM)’, ‘Neural Network (NN)’, and ‘Recursive Feature Elimination (RFE)’.
Results : The number of Clinically-Relevant POPF was 222 (12.5%) according to ISGPS definition 2016. Initial AUCs using RFM and NN using 6 variables (body mass index, preoperative albumin level, pancreatic duct size, sex, ASA score, and location of tumor) which were identified with independent risk factors after multivariate analysis were 0.669 and 0.696, which were higher than AUC 0.652 using conventional methods. After exclusion of high rate of missing values, AUCs using 48 variables were increased as 0.674 and 0.708. The maximal AUC using RFE was 0.756.
Conclusions : Up to now, we think this study is the first report to predict POPF using artificial Intelligence. The performance of new risk prediction platform is reliable as AUC 0.756. After external validation, this new platform could be used for selecting patients who need more intensified therapy, and establishing preoperatively effective treatment strategy.
e res
Methods : From 2007 to 2016, medical records of 1771 patients at Samsung Medical Center who underwent PD were reviewed. A total of 53 variables were inserted into training algorism. Algorithms which made risk prediction platform were ‘the random forest method (RFM)’, ‘Neural Network (NN)’, and ‘Recursive Feature Elimination (RFE)’.
Results : The number of Clinically-Relevant POPF was 222 (12.5%) according to ISGPS definition 2016. Initial AUCs using RFM and NN using 6 variables (body mass index, preoperative albumin level, pancreatic duct size, sex, ASA score, and location of tumor) which were identified with independent risk factors after multivariate analysis were 0.669 and 0.696, which were higher than AUC 0.652 using conventional methods. After exclusion of high rate of missing values, AUCs using 48 variables were increased as 0.674 and 0.708. The maximal AUC using RFE was 0.756.
Conclusions : Up to now, we think this study is the first report to predict POPF using artificial Intelligence. The performance of new risk prediction platform is reliable as AUC 0.756. After external validation, this new platform could be used for selecting patients who need more intensified therapy, and establishing preoperatively effective treatment strategy.
e res
SESSION
Plenary Session
Room A 4/6/2019 11:45 AM - 12:00 PM