Logo Logo
Help
Contact
Switch Language to German
Ingrisch, Michael; Schöppe, Franziska; Paprottka, Karolin; Fabritius, Matthias; Strobl, Frederik F.; De Toni, Enrico N.; Ilhan, Harun; Todica, Andrei; Michl, Marlies; Paprottka, Philipp Marius (2018): Prediction of 90Y Radioembolization Outcome from Pretherapeutic Factors with Random Survival Forests. In: Journal of Nuclear Medicine, Vol. 59, No. 5: pp. 769-773
Full text not available from 'Open Access LMU'.

Abstract

Our objective was to predict the outcome of Y-90 radioembolization in patients with intrahepatic tumors from pretherapeutic baseline parameters and to identify predictive variables using a machine-learning approach based on random survival forests. Methods: In this retrospective study, 366 patients with primary (n = 92) or secondary (n = 274) liver tumors who had received Y-90 radioembolization were analyzed. A random survival forest was trained to predict individual risk from baseline values of cholinesterase, bilirubin, type of primary tumor, age at radioembolization, hepatic tumor burden, presence of extrahepatic disease, and sex. The predictive importance of each baseline parameter was determined using the minimal-depth concept, and the partial dependency of predicted risk on the continuous variables bilirubin level and cholinesterase level was determined. Results: Median overall survival was 11.4 mo (95% confidence interval, 9.7-14.2 mo), with 228 deaths occurring during the observation period. The random-survival-forest analysis identified baseline cholinesterase and bilirubin as the most important variables (forest-averaged lowest minimal depth, 1.2 and 1.5, respectively), followed by the type of primary tumor (1.7), age (2.4), tumor burden (2.8), and presence of extrahepatic disease (3.5). Sex had the highest forest-averaged minimal depth (5.5), indicating little predictive value. Baseline bilirubin levels above 1.5 mg/dL were associated with a steep increase in predicted mortality. Similarly, cholinesterase levels below 7.5 U predicted a strong increase in mortality. The trained random survival forest achieved a concordance index of 0.657, with an SE of 0.02, comparable to the concordance index of 0.652 and SE of 0.02 for a previously published Cox proportional hazards model. Conclusion: Random survival forests are a simple and straightforward machine-learning approach for prediction of overall survival. The predictive performance of the trained model was similar to a previously published Cox regression model. The model has revealed a strong predictive value for baseline cholinesterase and bilirubin levels with a highly nonlinear influence for each parameter.