Machine Learning Models Predict Early Postoperative Relapse in Pancreatic Cancer
Received Date : 06 May 2021
Accepted Date : 16 Aug 2021
Available Online : 07 Sep 2021
Cem ŞİMŞEKa, Deniz Can GÜVENb, Furkan CEYLANc, İbrahim Yahya ÇAKIRc,
Taha Koray ŞAHİNc, Ömer DİZDARb, Yasemin BALABANa, Şuayib YALÇINb
aDepartment of Gastroenterology, Hacettepe University Faculty of Medicine, Ankara, TURKEY
bDepartment of Medical Oncology, Hacettepe University Faculty of Medicine, Ankara, TURKEY
cDepartment of Internal Medicine, Hacettepe University Faculty of Medicine, Ankara, TURKEY
Doi: 10.37047/jos.2021-84326 - Article's Language: EN
J Oncol Sci. 2021;7(3):115-24
ABSTRACT
Objective: A risk stratification system for tailoring treatment selection is absent for patients with pancreatic ductal adenocarcinoma
(PDAC). Machine learning models can outperform traditional survival models in predicting outcomes and guiding treatment. Therefore,
the current study aimed to test the performance of machine learning models in predicting disease-free survival (DFS) and overall survival
(OS) in operated PDAC cases. Material and Methods: The demographic, clinical, histopathological, radiologic, and laboratory data for the
resected PDAC samples were retrospectively reviewed. Univariate and multivariate conventional survival analyses were conducted for the 6-
month DFS and 12-month OS. Two machine learning methods were adopted: a machine learning model, DeepHit, and a gradient boosting decision
tree model, LightGBM (Light Gradient Boosting Machine). The performance of these models was compared using the area under the
receiver operator characteristic curves (AUROC). Results: For the study, 121 PDAC cases that underwent resection surgery with curative intent
were included. The median OS of the study population was 21.9 (11.5-44.4) months, and the median DFS was 11.8 (6-25.6) months. The
constructed deep learning model AUROC values were as follows: Relapse at 6 months 0.58 (±0.177) and 0.73 (±0.098); survival over 12
months 0.56 (±0.14) and 0.78 (±0.078); survival over 24 months 0.53 (±0.13), and 0.63 (±0.083). Conclusion: Machine learning models performed
similarly to the Cox regression-based dichotomous models. However, further validation of the model in a different and larger dataset
is required.
Keywords: Pancreatic cancer; prognosis; survival; artificial intelligence; machine learning