Clinical Predictors of Regression of Choroidal Melanomas after Brachytherapy: A Growth Curve Model

Abstract

To build multivariate models to assess correctly and efficiently the contribution of tumor characteristics on the rate of regression of choroidal melanomas after brachytherapy in a way that adjusts for confounding and takes into account variation in tumor regression patterns. Modeling of longitudinal observational data. Ultrasound images from 330 of 388 consecutive choroidal melanomas (87%) irradiated from 2000 through 2008 at the Helsinki University Hospital, Helsinki, Finland, a national referral center. Images were obtained with a 10-MHz B-scan during 3 years of follow-up. Change in tumor thickness and cross-sectional area were modeled using a polynomial growth-curve function in a nested mixed linear regression model considering regression pattern and tumor levels. Initial tumor dimensions, tumor-node-metastasis (TNM) stage, shape, ciliary body involvement, pigmentation, isotope, plaque size, detached muscles, and radiation parameters were considered as covariates. Covariates that independently predict tumor regression. Initial tumor thickness, largest basal diameter, ciliary body involvement, TNM stage, tumor shape group, break in Bruch's membrane, having muscles detached, and radiation dose to tumor base predicted faster regression, whether considering all tumors or those that regressed in a pattern compatible with exponential decay. Dark brown pigmentation was associated with slower regression. In multivariate modeling, initial tumor thickness remained the predominant and robust predictor of tumor regression (P < 0.0001). In addition, use of ruthenium isotope as opposed to iodine isotope (P= 0.018) independently contributed to faster regression of tumor thickness. For both isotopes considered alone, initial tumor thickness was the sole clinical predictor of regression (P < 0.0001). Regression of choroidal melanoma after brachytherapy was associated with several clinical tumor and treatment parameters, most of which were shown to reflect initial tumor size. An independent predictor of regression of tumor thickness was the isotope used. These 2 covariates need to be adjusted for when exploring the associations with the rate of regression of histopathologic or genetic features of the tumor. Our model allows such future analyses efficiently without matching.