Accident prevention with Deep Learning language models

1. INTRODUCTION
There are many measures that can be taken to prevent
accidents or reduce their consequences, including regulations such as mandatory helmet use as well as investments in infrastructure or training activities. Accident
descriptions (narratives), in an explicit way or within
medical anamneses, hold useful information as to which
measures could have prevented a particular case and
should therefore be promoted.
2. OBJECTIVES
The aim of this study is to develop an automated procedure for deriving preventive actions from accident descriptions with the help of machine learning. This could
then be applied to large amounts of data, and quantitative statements could be made about which interventions have which impact on accident prevention.
3. METHODS
In a first step, 1,000 sports accident descriptions from
the IDB Austria injury surveillance system are analysed
by human experts, and possible prevention measures are
derived for each case. Then, existing deep learning language models are evaluated by means of these training
data. The best fitted model is adapted to the task by fine
tuning and applied to the full amount of available data
(50,000 sports accidents).
4. RESULTS
The results are lists of interventions for prevention and
harm reduction for all available cases together with the
number of accidents in which these interventions would
have had an influence. The results will be presented in
the form of a data dashboard that lists the top measures,
broken down by further data items like type of sport, age
and sex.
5. CONCLUSION
Preliminary results indicate that recent deep learning
language models are apt to automatically analyse large
amounts of textual accident data in the view of supporting evidence-based injury prevention strategies and policies.