Authors:
Harry Freitas da Cruz
;
Frederic Schneider
and
Matthieu-P. Schapranow
Affiliation:
Hasso Plattner Institute, Digital Health Center, Rudolf-Breitscheid-Straße 187, Potsdam and Germany
Keyword(s):
Clinical Prediction Models, Supervised Learning, Interpretability, Nephrology, Acute Kidney Injury.
Related
Ontology
Subjects/Areas/Topics:
Artificial Intelligence
;
Biomedical Engineering
;
Business Analytics
;
Cardiovascular Technologies
;
Computing and Telecommunications in Cardiology
;
Data Engineering
;
Decision Support Systems
;
Decision Support Systems, Remote Data Analysis
;
Health Engineering and Technology Applications
;
Health Information Systems
;
Knowledge-Based Systems
;
Pattern Recognition and Machine Learning
;
Symbolic Systems
Abstract:
Acute kidney injury is a common complication of patients who undergo cardiac surgery and is associated with additional risk of mortality. Being able to predict its post-surgical onset may help clinicians to better target interventions and devise appropriate care plans in advance. Existing predictive models either target general intensive care populations and/or are based on traditional logistic regression approaches. In this paper, we apply decision trees and gradient-boosted decision trees to a cohort of surgical heart patients of the MIMIC-III critical care database and utilize the locally interpretable model agnostic approach to provide interpretability for the otherwise opaque machine learning algorithms employed. We find that while gradient-boosted decision trees performed better than baseline (logistic regression), the interpretability approach used sheds light on potential biases that may hinder adoption in practice. We highlight the importance of providing explanations of the
predictions to allow scrutiny of the models by medical experts.
(More)