Abstract
Abstract: :
Purpose:To evaluate the ability of an artificial neural network (ANN) to forecast the 5–year survival of uveal melanoma. Methods:153 eyes with uveal melanoma, (mean age 61 years) treated with Ru–106 brachytherapy between 1988–1998 were followed clinically and ultrasonically every 6.7 ± 0.3 months (mean follow–up 9.6±3.7 years). 3–4 layers backpropagation ANNs (2–16 neurons each) were constructed and trained on 75 patients. The information fed to the ANN included patient’s age, sex and country of birth; tumor’s base size, height, internal reflectivity, regularity, vascularity, extra–scleral extension, location in the eye and the initial post–brachytherapy regression rate. The ANN’s ability to forecast the 5–year survival was tested on a separate group of 78 patients using ROC curves and likelihood ratios (LR). Results:57 patients had small tumors (2–4mm), 66 medium tumors (4–8mm) and 24 large ones (>8 mm). Multivariate Cox regression of all input parameters showed that the most significant prognostic parameter was tumor height followed by tumor regression rate. Large tumors had a 23% 5–year mortality compared with 10% for small tumors. Fast regressing tumors (>0.7mm/month) had 37% 5–year mortality compared to 16% for the slow regressing tumors. The best 3–layers ANN had a diagnostic precision of 84% (LR=31.5) while the best four layers ANN had a diagnostic precision of 82% (LR=15.7). Conclusions:Three layers ANNs have a very good ability to forecast the 5–year mortality from uveal melanoma. Additional layers do not add forecasting accuracy.