MODL-27. INTEGRATING OVERALL SURVIVAL AND TUMOR CONTROL PROBABILITY MODELS TO PREDICT LOCAL PROGRESSION AFTER BRAIN METASTASIS RADIOSURGERY

Abstract

Abstract PURPOSE Stereotactic radiosurgery (SRS) for brain metastases has competing goals of maximizing local control and minimizing toxicity. Tumor control probability (TCP) models predict decreasing risk of local progression (LP) with dose escalation. However, many patients die from extracranial disease prior to local progression and may not require maximally aggressive treatment, which increases risk of toxicity. Better tools to assess patients’ risk of LP after SRS could help practitioners personalize SRS treatment to maximize therapeutic ratio. Here we assess whether simultaneous modeling of TCP and overall survival (OS) can improve prediction of LP after SRS. METHODS Records of patients undergoing SRS at a single institution were reviewed retrospectively. Using established OS and TCP models, the probability of one-year survival [p(OS)] and probability of one-year LP [p(LP)]) were calculated for each patient and each metastasis, respectively. Joint-probability was used to combine the models [p(LP,OS)=p(LP)*p(OS)]. Analyses were conducted at the individual-metastasis and whole-patient levels. Fine-Gray regression was used to model the effect of p(LP) or p(LP, OS) on risk of local progression after SRS, with death as a competing risk. RESULTS Ninety-two patients and 391 lesions were identified. At the patient level, one-year LP was 0.08 (95%CI, 0.03-0.15), median p(LP,OS) was 0.13 (95%, CI 0.07-0.2), and median p(LP) was 0.29 (95% CI, 0.22-0.38). At the metastasis level, one-year LP was 0.02 (95%CI, 0.01-0.04), median p(LP,OS) was 0.05 (95%, CI 0.02-0.07), and median p(LP) was 0.10 (95% CI, 0.07-0.13). p(LP,OS) was significantly associated with the risk of local progression at the patient level (p=0.048) and metastasis level (p=0.007); however, p(LP) was not (p=0.16 and p=0.28, respectively). CONCLUSION This study suggests simultaneous OS and TCP modeling more accurately predicts LP than TCP modeling alone. Better understanding which patients are at risk of LP after SRS may help personalize treatment to minimize risk without sacrificing efficacy.