Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

E-bike agents: Large Language Model-Driven E-Bike Accident Analysis and Severity Prediction

Zhichao Yang, Jiashu He, Mohammad B. Al-Khasawneh, Darshan Pandit, Cirillo Cinzia

TL;DR

Analysis of injury incidents involving e-bikes and traditional bicycles using the CPSRMS and NEISS datasets reveals key differences in mechanical failure modes, injury severity patterns, and affected user groups.

Abstract

E-bikes have rapidly gained popularity as a sustainable form of urban mobility, yet their safety implications remain underexplored. This paper analyzes injury incidents involving e-bikes and traditional bicycles using two sources of data, the CPSRMS (Consumer Product Safety Risk Management System Information Security Review Report) and NEISS (National Electronic Injury Surveillance System) datasets. We propose a standardized classification framework to identify and quantify injury causes and severity. By integrating incident narratives with demographic attributes, we reveal key differences in mechanical failure modes, injury severity patterns, and affected user groups. While both modes share common causes, such as loss of control and pedal malfunctions, e-bikes present distinct risks, including battery-related fires and brake failures. These findings highlight the need for tailored safety interventions and infrastructure design to support the safe integration of micromobility devices into urban transportation networks.

E-bike agents: Large Language Model-Driven E-Bike Accident Analysis and Severity Prediction

TL;DR

Analysis of injury incidents involving e-bikes and traditional bicycles using the CPSRMS and NEISS datasets reveals key differences in mechanical failure modes, injury severity patterns, and affected user groups.

Abstract

E-bikes have rapidly gained popularity as a sustainable form of urban mobility, yet their safety implications remain underexplored. This paper analyzes injury incidents involving e-bikes and traditional bicycles using two sources of data, the CPSRMS (Consumer Product Safety Risk Management System Information Security Review Report) and NEISS (National Electronic Injury Surveillance System) datasets. We propose a standardized classification framework to identify and quantify injury causes and severity. By integrating incident narratives with demographic attributes, we reveal key differences in mechanical failure modes, injury severity patterns, and affected user groups. While both modes share common causes, such as loss of control and pedal malfunctions, e-bikes present distinct risks, including battery-related fires and brake failures. These findings highlight the need for tailored safety interventions and infrastructure design to support the safe integration of micromobility devices into urban transportation networks.

Paper Structure

This paper contains 22 sections, 4 equations, 16 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Methods
  4. Analyzing Unstructured Incident Narratives with LLMs
  5. Data Classifier
  6. Information Extractor
  7. Injury Causes Determiner
  8. Incident-Component Link Detector
  9. Incident Severity Analysis-Ordered Logit Model
  10. Experiments
  11. Experiment Settings
  12. Datasets
  13. LLM Model details
  14. Evaluation metrics
  15. Classification Results
  16. ...and 7 more sections

Figures (16)

  • Figure 1: The overall workflow of Bike and E-bike agents. The dataset classifier agent filters out the accidents related to bikes and e-bikes. Such incidents are then passed to the bike and e-bike data analysis team for further processing, which includes: information extractor, injury causes determiner, and incident-component link detector. The information extractor identifies incident factors, which are then analyzed using an ordered logit model to explore their relationship with injury severity levels. The injury causes determiner identifies the cause of each injury and categorizes them as human-related, equipment-related, or both. The incident-component link detector examines whether each incident is associated with specific e-bike components and identifies the components most frequently linked to incidents. Finally, the analysis results are visualized.
  • Figure 2: Example workflow of the four agents to analyze e-bike accidents
  • Figure 3: Distribution of CPSRMS bicycle- and e-bike–related incidents by location, land use, year, gender, and age category.
  • Figure 4: Distribution of NEISS bicycle- and e-bike–related incidents by location, land use, year, gender, and age category.
  • Figure 5: Top three most frequent injury causes of bicycle and e-bike incidents by age group and gender in the CPSRMS dataset.
  • ...and 11 more figures