G. Kantidakis, H. Putter, C. Lancia, J.D. de Boer, A.E. Braat, §§ Fiocco
Thursday 5 march 2020
14:10 - 14:20h at Leo Franssen zaal
Parallel session: Parallel sessie XIV– Klinische abstracts
Background: Predicting survival of patients after liver transplantation is regarded as one of the most challenging areas in modern medicine. Hence, selecting the best prediction model is of great importance. Nowadays, there is a strong discussion in the medical field about machine learning (ML) and whether it has greater potential than Cox models when it comes to complex data. Criticism to ML is related to unsuitable performance measures and lack of interpretability which is important for the medical personnel to take decisions. Here, the potential of ML is investigated for large data of 62294 patients in USA with 97 prognostic factors (52 donor, 45 recipient characteristics) selected from over 600. It is also of particular interest the identification of potential risk factors for liver transplantation.
Methods: A comparison is performed between 3 different Cox models (Cox with all variables, Cox backward and Cox LASSO) and 3 machine learning techniques: a random survival forest (RSF) and 2 partial logistic artificial neural networks (PLANNs). Emphasis is given on the advantages and pitfalls of each method and on extracting interpretation from the ML techniques. The clinical endpoint is overall graft-survival (defined as the time between transplantation and the date of graft-failure or death).
Results: The most prognostic variables are re-transplantation (1st for the Cox models and the PLANN 1) and donor age (1st for RSF and PLANN 2). Other prognostic variables for all models are life support, HCV serology status, donor type, diabetes, race, and pre-treatment status. RSF shows slightly better predictive performance than the Cox models. In addition, neural networks show in general a good performance. However, instability is present due to the lack of a global measure for performance evaluation in survival setting.
Conclusions: It is shown that machine learning techniques can be a useful tool for both prediction and interpretation. Cox models have a straightforward interpretation but make the proportional hazards assumption. Machine learning techniques are very flexible since they do not make any assumptions about the underlying data. Nevertheless, they have limitations with respect to variable interpretation.