Program Information
A Multi-Objective Dynamic Bayesian Network Approach for Adaptive Personalized Radiotherapy in Non-Small-Cell Lung Cancer (NSCLC)
Y Luo*, D McShan , R Ten Haken , I El Naqa , University of Michigan, Ann Arbor, MI
Presentations
TU-FG-605-1 (Tuesday, August 1, 2017) 1:45 PM - 3:45 PM Room: 605
Purpose: To identify hierarchical biophysical relationships that could simultaneously influence both tumor local control (LC) and radiation-induced toxicities longitudinally over the course of radiotherapy for personalized, adaptive radiotherapy using a multi-objective dynamic Bayesian network (MO-DBN) approach.
Methods: 68 NSCLC patients with 48 cases of LC and 17 events of radiation pneumonitis grade 2 or above (RP2) were considered retrospectively. Each had 360 associated features including: tumor location, 13 clinical factors, 62 microRNAs, 59 single-nucleotide polymorphisms (SNPs) together with tumor and lung gEUDs, 90 cytokines and 129 PET radiomics features measured at 3 different phases of radiotherapy. Relevant biophysical predictors (nodes in the MO-DBN) were identified using a constraint-based local discovery algorithm and the corresponding structure (nodes plus edges) was constructed using a score-based learning algorithm. Cross-validation was employed to guard against overfitting and the area under the free-response receiver operating characteristics (AU-FROC) curve evaluated performance.
Results: The MO-DBN developed predicts LC and RP2 after treatment from information available at 3 radiotherapy time points, before treatment and after delivery of ~1/3 and ~2/3 of the dose. In addition to one SNP, one microRNA, and tumor location, the MO-DBN includes tumor and lung gEUDs, two cytokines, and two PET radiomics features from each phase of the radiotherapy. The AU-FROC of the MO-DBN for the joint prediction of patients’ LC and RP2 across the whole radiotherapy process is 0.80 (95% CI: 0.69-0.86).
Conclusion: A novel approach for adaptive personalized radiotherapy based on a MO-DBN developed from a retrospective dataset with large-scale biophysical features has been demonstrated. Here, a MO-DBN not only predicts a NSCLC patient’s personalized LC and RP2 simultaneously before and during radiotherapy, but also can estimate appropriate plans for adapting treatment based on individual patients’ characteristics and desired radiation outcomes. Testing with external independent datasets would validate the exact MO-DBN structure.
Funding Support, Disclosures, and Conflict of Interest: NIH grant numbers: P01 CA059827
Contact Email: