PREDICTION OF EARLY RECURRENCE OF RESECTED PANCREATIC DUCTAL ADENOCARCINOMA: A MACHINE LEARNING METHOD
Toshitaka Sugawara*, Daisuke Ban, Jo Nishino, Tomotaka Kato, Atsushi Kudo, Minoru Tanabe
Tokyo Medical and Dental University, Tokyo, Japan
Background/Aim: Pancreatic ductal adenocarcinoma (PDAC) is a devastating disease; up to 70% of patients is diagnosed as unresectable at initial diagnosis. Even after undergoing potential curative resection, early systemic recurrence occurs in 50% of patients. These patients might already have occult metastatic disease at the point of operation, and can’t achieve long survival after curative-intent surgery. Although some previous studies investigated early recurrence prediction after resection of PDAC, these studies used only one factor (e.g. CA19-9) or scoring method containing several factors. It is easy to assume that early recurrence and occult metastasis will not be predicted in a high accuracy by a simple scoring method, much less one factor. We used elastic net, a sparse modelling method, to create a model with preoperative parameters to predict early recurrence of PDAC. Methods: Data of 184 patients with resectable PDAC who underwent pancreatectomy between January 2005 and September 2018 were retrospectively reviewed. Patients who underwent neoadjuvant therapy and total pancreatectomy for recurrence in the remnant pancreas were excluded. Patients whose follow-up interval was incomplete were also excluded. The cut-off value of early recurrence was defined as 6 months, considering the time lag that occult metastasis becomes detectable. Preoperative data such as demographic data, tumor markers, immune-inflammation factors calculated from blood examination, and clinicopathological data were collected. All missing values were complemented by a machine learning method named factor analysis of mixed data in R. Since the size of the dataset was small, we didn’t have external validation; therefore, we repeated internal 10-fold cross validation 1000 times. Area under the curve (AUC) was used to show the usefulness of the models. Furthermore, a decision curve analysis was performed to evaluate the utility of the models. Results: A total of 136 patients were included in the final analysis. At the point of last follow-up, 103 patients recurred (75.8%), and 35 (34.0%) recurred within 6 months after surgery. In the model formulation, we used 10 parameters: age, gender, CEA, CA19-9, DUPAN2, total lymphocyte count, modified glasgow prognostic score, tumor size (category), lymph node metastasis, and peripancreatic invasion. The mean AUC of all models to the internal repeated validation was better than that of CA19-9 (0.702 vs 0.657). The decision curve analysis showed that the models can detect patients who may suffer from early recurrence and need some other treatment (Fig. 1). Incidentally, the AUC of the model obtained from all patients without validation was 0.758; Fig. 2 showed the model. Conclusion: Elastic net can be a good tool to predict early recurrence of PDAC.
Fig. 1. Decision curve analysis with 1000 times of repeated internal 10-fold cross validation for models obtained from Elastic net and CA19-9.
Fig. 2. p, probability of early recurrence; TLC, total lymphocyte count; TS, tumor size (category); LNM, lymph node metastasis (0 or 1); PI, peripancreatic invasion (0 or 1).
Back to 2020 Abstracts