Applying interpretable machine learning to classify tree and utility pole
related crash injury types
Subasish Das a,⁎, Songjukta Datta a, Hamsa Abbas Zubaidi b, Ihsan Ali Obaidb
a Texas A&M Transportation Institute, 1111 RELLIS Parkway, Bryan, TX 77807, United States of America
b Oregon State University, 233 Owen Hall, Corvallis, OR 97331-3212, United States of America
a r t i c l e i n f o a b s t r a c t
Article history:
Received 3 September 2020
Received in revised form 14 January 2021
Accepted 21 January 2021
Available online 29 January 2021
Keywords:
TUOP crashes
Crash narratives
Text mining
Interpretable machine learning
Misclassification
Explanation
In spite of enormous improvements in vehicle safety, roadway design, and operations, there is still an excessive
amount of traffic crashes resulting in injuries and major productivity losses. Despite the many studies on factors
of crash frequency and injury severity, there is still further research to be conducted. Tree and utility pole/other
pole related (TUOP) crashes present approximately 12 to 15% of all roadway departure (RwD) fatal crashes in the
U.S. The count of TUOP crashes comprise nearly 22% of all fatal crashes in Louisiana. From 2010 to 2016, there
were 55,857 TUOP crashes reported in Louisiana. Individually examining each of these crash reports is not a realistic
option to investigate crash factors. Therefore, this study employed text mining and interpretable machine
learning (IML) techniques to analyze all TUOP crashes (with available crash narratives) that occurred in Louisiana
from 2010 to 2016. This study has two major goals: 1) to develop a framework for applying machine learning
models to classify injury levels from unstructured textual content, and 2) to apply an IML framework that provides
probability measures of keywords and their association with the injury classification. The present study
employed three machine learning algorithms in the classification of injury levels based on the crash narrative
data. Of the used modeling techniques, the eXtreme gradient boosting (XGBoost) model shows better performance,
with accuracy ranging from 0.70 to 24% for the training data and from 0.30% to 16% for the test data.
© 2021 International Association of Traffic and Safety Sciences. Production and hosting by Elsevier Ltd. This is an
open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).