2991 - Predicting Benefit of Elective Nodal Irradiation in Pancreatic Adenocarcinoma: A Supervised Machine Learning Approach

G. K. Harada¹, J. Park¹, N. V. Peterson², A. N. Munjal³, D. Imagawa⁴, Z. Jutric⁴, F. Dayyani⁵, J. Valerin⁵, D. S. Hong Jr⁶, and S. N. Seyedin¹; ¹Department of Radiation Oncology, University of California - Irvine, Orange, CA, ²Department of Radiation Oncology, University of California Irvine, Irvine, CA, ³UC Irvine, Orange, CA, ⁴Division of Hepatobiliary, Pancreas Surgery & Islet Cell Transplantation, Department of Surgery, University of California, Irvine, Orange, CA, ⁵Division of Hematology/Oncology, Department of Medicine, University of California - Irvine, Orange, CA, ⁶Department of Radiation Oncology, Chao Family Comprehensive Cancer Center, University of California, Irvine, Orange, CA

Purpose/Objective(s): The role of neoadjuvant elective nodal irradiation (ENI) remains controversial for patients with clinically node-negative (cN0) disease after induction chemotherapy. The purpose of this study was to create a model to predict the benefit of ENI prior to treatment initiation. Materials/

Methods: Using the National Cancer Database (NCDB), we identified patients with cT1-4N0M0 pancreatic adenocarcinoma diagnosed between 2006 and 2020. All patients had multiagent neoadjuvant chemotherapy (NAC), radiation with ENI, and definitive resection with nodal dissection. Radiation dose was assessed using a biologically effective dose (a/b = 10; BED10). Benefit from ENI was defined as an absence of pathologic nodal disease (ypN0). Following a data split, a cross-validated LASSO logistic regression model was used to select variables to predict benefit from ENI, with assessments of incidence of ypN+ and associated overall survival (OS) in both training and testing cohorts. Out of sample predictions were then made on an NCDB cohort receiving multiagent NAC and radiation without ENI, with similar assessments of ypN+ and OS. The threshold for statistical significance was set to p<0.05.
Results: A total of 258 patients were identified with a median age of 64.0 years (IQR=57–70 years) with 31.0% having cT3-4 disease (N=80). Median dose with ENI was 59.5 Gy (IQR=53.1–59.5 Gy). The final model is displayed in [Table 1]. Notably, a dose-dependent response was seen in patients with cT3-4 disease and BED10 delivered to lymph nodes (Coeff.=0.003 per Gy if cT3-4). Within testing data, model accuracy was 79.2%, with 84.6% sensitivity and 75.4% specificity (AUC=0.780). Only 23.4% of patients identified to benefit from ENI (N=11) developed ypN+ versus 75.5% deemed unlikely to benefit (N=37; p<0.001). Patients unlikely to benefit had a median OS of 22.3 months (95% CI=18.3–NR) while median OS was not reached in those predicted to benefit (95% CI=19.5–NR; p=0.109). Of 1,161 patients without ENI, 31.7% (N=196) predicted to benefit had ypN+ disease, while those unlikely to benefit had 45.1% (N=245; p<0.001). Median OS was similar within the ENI-naïve cohort at 31.6 months (95% CI=27.3–NR) and 28.6 months (95% CI=25.8–33.3; p=0.637), respectively. The absolute risk reduction for ypN+ disease with ENI was 8.3% (NNT=12).
Conclusion: Using a machine learning approach, there was a dose-dependent response for predicted benefit of ENI in those with cT3-4 disease, where those predicted to benefit had a trend toward improved OS and lower rates of ypN+. When compared to an ENI-naïve cohort, the benefit of ENI was estimated as a risk reduction of 8.3% for ypN+.

Abstract 2991 – Table 1. LASSO Logistic Regression Model for Prediction of ENI Benefit

Covariate	Coefficient
Age	0.011 (per year)
Female Sex	0.083
Charlson-Deyo Score = 3	-0.474
Time from Diagnosis to Surgery	0.005 (per day)
Time from Diagnosis to Radiation Therapy	-0.008 (per day)
BED10 to Regional Lymph Nodes if cT3-4 Disease	0.003 (per Gy)