Optimizing fractionation schedules for de-escalation radiotherapy in head and neck cancers using deep reinforcement learning.

PURPOSE: Patients with locally-advanced head and neck squamous cell carcinomas (HNSCCs), particularly those related to human papillomavirus (HPV), often achieve good locoregional control (LRC), yet they suffer significant toxicities from standard chemoradiotherapy. This study aims to optimize the daily dose fractionation based on individual responses to radiotherapy (RT), minimizing toxicity while maintaining a low risk of LRC failure.

METHOD: A virtual environment was developed to simulate tumor dynamics under RT for optimizing dose schedules. Patients predicted to maintain LRC were selected for de-escalation experiments. The proliferation saturation index (PSI) and linear-quadratic model were used to predict responses. A deep reinforcement learning (DRL) agent optimized fractionation schemes by interacting with the simulation environment, aiming to reduce the OAR's biologically effective dose (BED) while preserving LRC. The impact of model uncertainty was analyzed and a support vector machine (SVM) model was used to segment parameter space and identify patients more robust to noise.

RESULTS

Personalized de-escalation plans were compared to conventional RT in a cohort of 5000 virtual patients. Personalized fractionation reduced the tumor dose and OAR's BED by 29%, with an average OAR BED reduction of 5.61 ± 2.96 Gy. Prognostic outcomes were nearly identical, with 99.80% of patients in the low-risk LRC failure group. Model uncertainty impacted dosimetric indicators and prognosis, but the high-BED benefit group showed greater robustness to noise. SVM decision boundaries defined parameters range for patient selection.

CONCLUSION

Optimizing fractionated doses based on patient responses minimizes toxicity while maintaining LRC in HNSCCs. Stratifying patients can mitigate model uncertainty and reduce treatment risks.