Room: Exhibit Hall | Forum 4
Purpose: Dynamic Bayesian networks (DBNs) have been developed for personalized radiotherapy in our previous research. A robust radiation treatment planning mechanism is generated in this study to investigate each individual patientâ€™s possible dose escalation based on the developed DBNs by optimizing the trade-off between tumor local control (LC) and radiation induced toxicities (RITs).
Methods: 68 NSCLC patients were employed to learn the structure and conditional probabilities of a DBN during the course of radiotherapy. A linear time-varying discrete state-space (LTV-DSS) system was developed to represent the DBN, where population-based coefficient matrices of the LTV-DSS can be numerically derived from the DBN for optimal control analysis. Based on the LTV-DSS system, an H-infinity method was used to compute an optimal feedback controller to match the targeted outcome of radiation treatment by using a combination of loop shaping and robust stabilization. Similarly, the individual-based LTV-DSS could also be derived from the DBN according to each patientâ€™s responses during a course of radiotherapy. Whether to escalate dose or not during the next treatment course can be obtained by applying the H-infinity loop shaping method to the LTV-DSS. 50 additional patients were used to evaluate the robust control approach.
Results: The mathematical relationship between a DBN and a LTV-DSS system has been derived. The dose escalation of 68 patients has been verified by applying the loop-shaping method to the population-based LTV-DSS resulting in the H-infinity optimal cost equal to 1.36. The H-infinity optimal costs for the 50 additional patients based on an individual-based LTV-DSS indicate that the DBN-based loop-shaping design can support personalized radiation treatment decision making during the course of the radiotherapy.
Conclusion: The LTV-DSS system is developed based on our data-driven DBNs to model the dynamic process of radiotherapy. The H-infinity loop-shaping design is able to optimally find robust personalized treatment plans.
Not Applicable / None Entered.