This is the accepted version of the journal article:
Ibeas, Jose; Macias Toro, Edwar Hernando; Rubiella, Carol; [et al.]. «Re-
nal failure and mortality : from evidence to artificial intelligence, change
of paradigm?».    Nephrology Dialysis Transplantation, Vol.        34, Issue
Supplement1 (June2019), art.gf z103.SP 689.DOI10.1093/ndt/gf z103.SP 689


This version is available at https://ddd.uab.cat/record/273461

under the terms of the            license
RENAL FAILURE AND MORTALITY: FROM EVIDENCE TO ARTIFICIAL INTELLIGENCE, CHANGE
OF PARADIGM?

Jose Ibeas1, Edwar Macias2, Carol Rubiella1, Antoni Morell2, Javier Serrano2, Angel Rodriguez-Jornet1, Jose Vicario2, Dolores Rexachs2
1
 Parc Tauli University Hospital, Barcelona, Spain and 2Autonomous University of Barcelona, Barcelona, Spain


INTRODUCTION: The mortality of the patient with renal insufficiency is high and especially in
dialysis. There are many risk factors involved, although mainly those related to cardiovascular
risk, which in turn are closely linked to those related to uremia, mutually reinforcing.
The approach to identifying these factors is difficult, and those recommended by Guides or
predictive models have not been validated in the renal patient. Mortality risk models implicitly
assume that each risk factor is linearly related to events, simplifying what are really complex
relationships that would include a huge number of factors, with non-linear relationships.
Approaches that incorporate multiple elements that identify real relationships are needed.
Machine-learning can be an alternative. Based on computational methods that detect
complex and non-linear interactions between variables identify latent variables, unlikely to
observe directly.
The aim is to evaluate if machine-learning algorithms based on neural networks can establish
a predictive model of mortality in the patient on hemodialysis.

METHODS: Design: Retrospective analysis of a historical cohort of 250 patients on
hemodialysis in a single centre, from November 2008 to February 2018. Inclusion criteria: >
18 y.o. and death in the period described.
Methodology:
1. Preprocessing the data to be used after collecting the historical information for several
years.
2. Exploration of variables, training models with samples of different time windows to predict
mortality at different time ranges.
3. Variables selection by two methods: a) By an expert; b) By Recursive feature elimination
(RFE), with random forest (RF) as a learning model. In a second phase, train a model using
artificial neural networks (ANN) to compare the two methods.
4. Optimise neural network performance by considering all the variables and architecture
consisted of two layers with 50 and 25 neurons, respectively.
5. Creation of time windows for mortality prediction by models with the training of ANNs of
two hidden layers with 300 and 100 neurons. Samples randomisation and separation with a
ratio of 70-30% for training and testing, using 5-fold for cross- validation.
Variables: a) Age, sex, time in dialysis, date of death; b) Diagnosis of comorbidity by ICD-9; c)
All variables of all hemodialysis sessions; d) All the analytical variables of the entire follow-up
time.

RESULTS: Age: 73.3 6 11.09 years, 64.8% men and 35.2% women, time in HD: 2.4 6 1.9 years.
The best mortality prediction was the one who collected information for six months and
predicted mortality to 18, with a ROC area of 0.982, an accuracy of 0.982, sensitivity of 0.971,
specificity of 0.935, positive predictive value of 0.949 and negative predictive value of 0.962.
CONCLUSIONS: The methodology of machine learning demonstrates a training and validity
with excellent capacity of recognition and prediction of the event of death.
The method of machine learning may involve a change in the approach of algorithms for the
establishment of risk factors for mortality in the renal patient.