KEY POINTS
- Investigators developed a Bayesian hierarchical NTCP model using a 160-patient prospective training cohort and evaluated it in 91 additional patients from an ongoing trial of chemoradiotherapy for NSCLC.
- The endpoint was an increase in high-sensitivity cardiac troponin T of >5 ng/L during RT, used as a biomarker of myocardial injury rather than a clinical cardiovascular event.
- Cardiac dose sensitivity was heterogeneous. The strongest association between mean heart dose and troponin elevation occurred in patients >64 years with left-sided or mediastinal tumors, with an odds ratio of 2.16 per modeled dose effect (95% credible interval 1.07–4.14).
- Rather than deriving one population-wide constraint, the Bayesian model generated risk-group-specific mean heart dose limits while explicitly incorporating uncertainty around the NTCP estimate.
- Across internal validation, same-institution external validation, and prospective implementation, uncertainty-aware constraints achieved higher specificity of 0.57–0.68 vs 0.28–0.51 and higher accuracy of 0.60–0.69 vs 0.43–0.59 than constraints based on point estimates alone.
- During prospective implementation, cardiac troponin elevation occurred in 4.8% of patients who met their personalized constraint versus 34.8% of those who exceeded it (p=0.042).
- Incorporating uncertainty could also relax unnecessarily restrictive constraints in lower-risk patients. In one subgroup, the allowed mean heart dose increased from 9.07 to 16.18 Gy(RBE), reducing simulated plan revisions from 20 to 8 without reducing sensitivity.
CLINICAL TAKEAWAY
A universal mean-heart-dose constraint may not reflect the large differences in baseline cardiac susceptibility among patients receiving thoracic RT. This Bayesian approach is an interesting step toward risk-adapted cardiac sparing, but validation against hard cardiovascular outcomes—not only troponin elevation—is needed before personalized limits replace conventional planning constraints.