Click here to

Session: Personalized Treatment via Outcome Modeling [Return to Session]

A Novel Probabilistic Approach for Personalized Radiotherapy: Integrating Machine Learning and Optimization

A Ajdari*, T Bortfeld, Massachusetts General Hospital & Harvard Medical School, Department of Radiation Oncology, Division of Radiation BioPhysics, Boston, MA.


TH-D-BRC-1 (Thursday, 7/14/2022) 11:00 AM - 12:00 PM [Eastern Time (GMT-4)]

Ballroom C

Purpose: Traditional radiotherapy (RT) treatment planning relies on population-wide estimates of organ tolerance to minimize excess toxicity. Though conventional machine learning (ML) models can predict patient-specific toxicity, they rely on binary classifications which are not amenable to nuanced RT adaptations. We propose a (Bayesian) probabilistic ML-based approach for predicting patient-specific risk and personalizing the RT workflow.

Methods: Sixty-nine non-small cell lung cancer patients with baseline and mid-treatment [18]F-flurodeoxyglucose (FDG)-PET images were retrospectively analyzed. A probabilistic Bayesian Networks (BN) model was developed to update the prior (baseline) risk of radiation pneumonitis (RP) and estimate a posterior risk upon obtaining mid-treatment FDG information. Subsequently, a patient-specific dose modifying factor, DMF, as a surrogate for lung radiosensitivity, was estimated to personalize the LKB normal tissue toxicity probability (NTCP) model. This personalized NTCP was then integrated into a NTCP-based optimization model for RT adaptation, ensuring tumor coverage and respecting patient-specific lung radiosensitivity.

Results: The magnitude of the predicted risks corresponded with the RP severity. Average predicted risk for grade 1-4 RP were 0.18, 0.42, 0.63, and 0.76, respectively (p<0.001). The proposed model yielded an average AUC of 0.87, outperforming the AUCs of LKB-NTCP (0.77), pre-treatment BN (0.81), and random forest (0.83). Average for the radio-tolerant (RP grade=1) and radiosensitive (grade≥2) groups were 0.8 and 1.63, p<0.01, with DMF=1 denoting “average” (population-wide) radiosensitivity. RT personalization in 15 cases resulted in 5 dose escalation strategies (average mean tumor dose increase=6.47 Gy), and 10 dose de-escalation with average mean lung dose reduction of 2.98 Gy, corresponding to average NTCP reduction of 15%. The intensity of RT adaptation was based on the magnitude of the predicted risk.

Conclusion: Moving from binary classifications to more nuanced probabilistic predictions is key in realizing RT personalization. Our proposal shows such probabilistic frameworks could yield significant NTCP-related benefits without compromising tumor control.

Funding Support, Disclosures, and Conflict of Interest: This study was partially funded by the National Cancer Institute under grant #R01CA266275 and the Therapy Imaging Program (TIP) funded by the Federal Share of program income earned by Massachusetts General Hospital on C06CA059267, Proton Therapy Research and Treatment Center.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email