Purpose:
More than 90% of intraocular melanomas involve the choroid and about 50% cause fatal metastatic disease. Univariate analysis provides only approximate survival estimates, relevant to large groups of patients but not individuals. Our aim was to create a prognostic model that combined pathological, clinical and genetic data, using imputation techniques to compensate for missing information.
Methods:
Survival was first modelled using the Cox proportional hazards (PH) model. However, evaluation of the Grambsch-Therneau residuals showed that the null hypothesis of the proportionality of hazards was rejected (p<0.001). The variable that violated the PH assumptions most strongly was the basal tumour diameter. An alternative approach was therefore explored by formulating the problem in terms of an Accelerated Failure Time model (AFT) which specifies that predictors act additively on the log failure time.Non-linear transformations were carried out on the input variables to model time-dependence. A Bayesian Regularization method was used to reduce overfitting and n-fold cross-validation was deployed to assess the model’s performance. Using data from 3653 patients, we generated two predictive models; the first using clinical data only and the second using clinical and laboratory data. The model’s outcome was all-cause mortality.A web-based user interface system was developed for the model. The interface was built using bespoke tool called MatSOAP which relies on Matlab (The Mathworks) and Simple Object Access Protocol (SOAP). The system provides the outcome in tabular and graphical forms. It also provides pictograms for ease of communication with patients.
Results:
Results showed that the c-index of discrimination was 0.75 (95% CI 0.74-0.76) for the clinical model and 0.79 (95% CI, 0.76-0.82) for the laboratory model. Calibration assessed by calculating the Cox-Snell residuals showed good correlation between predicted and observed mortality. The Kolmogorov-Smirnov statistic was 0.8774204 (p = 0.699) and 0.7980748 (p = 0.8005) for the clinical and laboratory models respectively.
Conclusions:
We have developed methods for multivariate analysis of clinical, pathological and genetic findings, compensating for missing data and taking age and sex into account. These provide survival curves that are relevant to individual patients with choroidal melanoma.
Keywords: melanoma • oncology • pathology techniques