Click here to


Are you sure ?

Yes, do it No, cancel

Improved Estimation of Biological Parameters in TCP/NTCP Modeling by Machine Learning

I El Naqa*, D Owen , M Matuszak , K Cuneo , T Lawrence , R Ten Haken , University of Michigan, Ann Arbor, MI


(Sunday, 7/29/2018) 2:05 PM - 3:00 PM

Room: Karl Dean Ballroom A1

Purpose: Estimation of isoeffective biological dose (EQD) for outcomes prediction, i.e., TCP and NTCP, requires knowledge of multiple parameters including the radiosensitivity parameter (α/β). This can be challenging if the number of events (toxicity/recurrence) is limited. Here, we explore the use of imbalance correction by Synthetic Minority Over-sampling Technique (SMOTE) and show its application to maximum likelihood estimates (MLE) of TCP/NTCP with unknown α/β ratio and generalized equivalent uniform dose (gEUD) parameters.

Methods: 192 hepatocellular cancer patients treated with radiotherapy (146 SBRT and 46 IMRT) to a physical dose of 50 Gy were evaluated. NTCP was modeled using the LKB model and TCP by Poisson model. Toxicity was assessed by Child-Pugh changes (17% rate) and failure as recurrence within the field (7% rate) with a follow-up of 1 year. Data imbalance was corrected using SMOTE, which employs a k-nearest neighbor algorithm with perturbation in the feature space of the minority group. Features were selected as mean EQD2 dose, number of fractions, and dose per fraction. Modeling parameters (D50, γ50, α/β, a_gEUD) were estimated using MLE and confidence intervals (CI) were calculated by likelihood profiling.

Results: Three-parameters LKB (D50, γ50, α/β) estimates yielded an α/β=2.63 (95%CI: 0.914-7.49). The use of SMOTE resulted in a tighter CI with α/β=2.70 (95%CI: 1.49-4.95). Similar trend was noted in the case of 4-parameters LKB (added a_gEUD) yielded α/β= 1.50 (95%CI: 0.282-5.99) and a_gEUD=0.714 (95%CI:0.467-0.944), which improved with SMOTE into α/β= 3.02 (95%CI: 1.78-5.37) and a_gEUD=1.05 (95%CI:0.897-1.23). In case of TCP with 3-parameters Poisson model, SMOTE slightly improved α/β=17.1 (95%CI: -78-181) to α/β= 12.7 (95%CI: -104-153). In case of 4-parameters Poisson model, estimates with SMOTE had wide CIs and an average α/β above 10.

Conclusion: Machine learning techniques can aid in the estimation of traditional radiosensitivity parameters and improve accuracy for EQD conversion and TCP/NTCP predictions.


Not Applicable / None Entered.


TH- Radiobiology(RBio)/Biology(Bio): RBio- LQ/TCP/NTCP/outcome modeling

Contact Email