Premium
Neural networks compared with Cox regression
Author(s) -
DAMATO B,
TAKTAK A,
ELEUTERI A
Publication year - 2008
Publication title -
acta ophthalmologica
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.534
H-Index - 87
eISSN - 1755-3768
pISSN - 1755-375X
DOI - 10.1111/j.1755-3768.2008.3242.x
Subject(s) - proportional hazards model , survival analysis , artificial neural network , medicine , regression analysis , regression , natural history , population , intensive care medicine , artificial intelligence , statistics , computer science , machine learning , mathematics , environmental health
Purpose Survival prediction is useful in patient care and research. Most studies rely on Cox analysis and Kaplan‐Meier curves whereas we have preferred neural networks. The aim of this presentation is to compare these methods and to discuss the advantages and limitations of each. Methods This presentation will be based on our experience with uveal melanoma. A neural network was trained with data from 1780 patients and evaluated with data from another 874 patients. Clinical, histopathological and cytogenetic data were included in the model. All cause mortality was reported, both for patients and for the matched general population. Results Cox analysis assumes linear correlations between variables and proportional hazards throughout the follow‐up period. Kaplan‐Meier analysis requires large patient categories, so that the precision of any prognostication is reduced. Neural networks overcome these limitations. Our model does censor non‐metastatic deaths so that melanoma‐related mortality is not exaggerated in groups of patients with significant competing risks. Conclusion Neural networks allow large numbers of variables to be included in predictive models with relatively small numbers of patients, thereby improving prognostication. Nevertheless, care must be taken when interpreting survival results to avoid serious misconceptions about the natural history of a disease and the impact of treatment.