A comparative study on machine learning based algorithms for prediction of motorcycle crash severity.

Area of Science:

• Road safety research
• Transportation engineering
• Data science and machine learning applications

Background:

• Motorcycle crash severity in Ghana is under-researched, with unknown risk factors and outcomes.
• Traditional statistical models may yield inaccurate results due to inherent assumptions.
• Machine learning offers non-parametric alternatives for analyzing complex crash data.

Purpose of the Study:

• To predict and classify motorcycle crash severity using machine learning algorithms.
• To evaluate and compare different modeling approaches for motorcycle crash severity.
• To investigate the impact of various risk factors on motorcycle crash injury outcomes in Ghana.

Main Methods:

• Utilized a motorcycle crash dataset from Ghana (2011-2015) classified into four injury severity levels.
• Developed and compared three machine learning models: J48 Decision Tree, Random Forest (RF), and Instance-Based learning (IBk).
• Validated models using 10-fold cross-validation and compared performance against the multinomial logit model (MNLM).

Main Results:

• Machine learning algorithms demonstrated superior accuracy and effectiveness compared to the MNLM.
• The Random Forest (RF) algorithm exhibited the best performance, attributed to its optimization and extrapolation capabilities.
• Identified critical determinants of motorcycle crash injury severity, including location type, time, settlement type, collision details, and road conditions.

Conclusions:

• Machine learning, particularly Random Forest, provides a robust and accurate approach to modeling motorcycle crash severity in Ghana.
• The study successfully identified key risk factors influencing crash severity, crucial for targeted interventions.
• Findings underscore the need for data-driven safety strategies to mitigate motorcycle crash injuries.