Bayesian Learning to Reduce Cardiac Risk of Patients With Locally Advanced Non-Small Cell Lung Cancer Based on Personalized Radiation Therapy Prescription.

PURPOSE: Radiation-induced heart damage is a significant concern in the treatment of non-small cell lung cancer (NSCLC) that can have debilitating or life-threatening consequences. Current strategies focus on minimizing heart exposure, but individual susceptibility varies. Existing evidence also suggests that a uniform "one-size-fits-all" dosimetric constraint for the heart may not be optimal for all patients.

METHODS AND MATERIALS: We developed a prospective study using Bayesian continuous learning and adaptation to develop a framework for personalized adaptive radiation treatment (PART) to reduce cardiovascular adverse events among patients with locally advanced NSCLC. The trial includes a Bayesian personalized risk prediction model to guide heart dose constraints; sequential learning to refine the model and the PART; continuous adaptation of the target risk level; and go/no-go monitoring of PART effectiveness in clinical implementation. Elevation of high-sensitivity cardiac troponin T (hs-cTnT) after radiation was used as a surrogate biomarker for grade ≥2 cardiovascular adverse events to allow real-time decision-making.

RESULTS

As of July 31, 2025, 100 patients have been enrolled and completed radiation treatment. Standard radiation plans were implemented for cohort 1 (50 patients), and PART for cohort 2 (50 patients). The first model incorporated patient- and disease-related factors and mean heart dose (MHD) as risk factors. The average treated MHDs were 7.84 ± 6.30 Gy in cohort 1 and 6.36 ± 6.01 Gy in cohort 2 (P = .20). The incidence of hs-cTnT elevation was 20.5% in cohort 2 compared with 31.9% in cohort 1. Within cohort 2, patients who satisfied the PART dose constraint had a markedly lower incidence of hs-cTnT elevation (9.7%) compared with those who exceeded the PART dose constraint (46.2%, P = .012).

CONCLUSIONS

Clinical implementation of PART model to guide treatment decision within a prospective trial is feasible. The recommended MHD constraints generated by the first version of PART appear reasonable and clinically relevant. PART was associated with lower incidence of hs-cTnT elevation.