Abstract
Purpose :
Algorithms utilizing artificial intelligence (AI) have demonstrated promise in detection and survival prediction for ophthalmic cancers. To the authors’ knowledge, this investigation is the first to utilize AI to predict all-cause mortality in patients with choroidal melanoma (CM) among a nationwide US cohort.
Methods :
The Surveillance, Epidemiology, and End Results (SEER) Program, a national repository of standardized cancer statistics, was executed for patient de-identified cases of CM. We included all cases of CM defined by ICD-O-3 diagnosed between January 1, 2000, and December 31, 2018. We excluded all patients with arbitrary or unknown causes of death, cases with multiple cancer diagnoses, or incomplete follow-up dates. We selected 21 variables known to influence CM survival. Binary logistic and Cox regression determined variable selection for multilayer perceptron (MLP) analysis. Two hidden layers and sigmoid activation functions were utilized for 5-year and 10-year survival predictions. Performance was evaluated using area under receiver operating characteristic curves (AUC). Statistical analyses were performed with IBM SPSS Premium Version 29.
Results :
4600 CM cases met inclusionary criteria. Of 21 variables, 9 variables qualified for entry into MLP algorithm after achieving statistical significance during logistic regression and cox regression analyses. For 5-year survival prediction, our 9-feature MLP model demonstrated an overall accuracy of 77.5% and 75.6% in training and testing sets, respectively, and a mean AUC of 85.1% (Figure 1). This model highlighted tumor stage, depth of tumor, age, histology subtype, and sequence of radiation/surgery as the 5 greatest contributors to survival prediction. For 10-year survival prediction, our 9-feature MLP model demonstrated an overall accuracy of 85.8% and 87.1% for training and testing sets, respectively, and a mean AUC of 92.1%. Marital status and median household income demonstrated a normalized importance of at least 20% (Figure 2).
Conclusions :
AI can forecast 10-year CM mortality with high accuracy and 5-year CM mortality with reasonable accuracy. AI continues to validate previously known survival prognosticators while highlighting promise in identifying novel psychosocial influences.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.