Home // International Journal On Advances in Life Sciences, volume 11, numbers 1 and 2, 2019 // View article

Knowledge Distillation from Machine Learning Models for Prediction of Hemodialysis Outcomes

Authors:
Harry Freitas da Cruz
Siegfried Horschig
Christian Nusshag
Matthieu-P. Schapranow

Keywords: clinical prediction model; renal replacement therapy; machine learning; supervised learning; knowledge distillation.

Abstract:
In order to compensate severe impairments of renal function, artificial, extracorporeal devices, so called dialyzers, have been developed to enable renal replacement therapy. The parameters utilized in this form of therapy and the specific patient characteristics substantially affect individual patient outcomes and overall disease progression. In this paper, we present a clinical prediction model for outcomes of critically ill patients that underwent a specific form of renal replacement, hemodialysis. For this purpose, we employed two categories of machine learning models: interpretable (Bayesian rule lists and logistic regression) and non-interpretable (multilayer perceptron and random forest). To provide more transparency to the latter category, we applied mimic learning and feature importance metrics. Results show that non-interpretable models outperform the rule-based classifier (c-statistic ≥ 0.9). Despite this result, the use of interpretability methods enables more thorough model scrutiny by a medical experts, revealing possible model biases, which might have been otherwise disregarded.

Pages: 33 to 43

Copyright: Copyright (c) to authors, 2019. Used with permission.

Publication date: June 30, 2019

Published in: journal

ISSN: 1942-2660