Introduction:
The EU Injury Database (EU IDB) was designed as a prevention-oriented data system that successfully helps to identify injury hotspots. A more difficult task, however, is to assess how many injuries of a certain hotspot could potentially be prevented by a specific measure. The broader availability of Artificial Intelligence (AI) tools offers new options for this task.

Objectives:
This project aimed at automatically evaluating reports of accidents that occurred during sport activities. The evaluation should lead to a summary and quantification of the effect of certain preventive measures applied to a given set of injuries. This information should eventually be available through a web application.

Methods:
Current methods of Natural Language Processing (language model "GPT3") were used for the automatic analysis of IDB accident reports, including an accident description and coded information according to the IDB Full Dataset standard (collected by the IDB Austria). This analysis automatically generated up to four prevention measures for each accident. Subsequently, for the essential measures identified, the proportion of accidents that could have been prevented or positively influenced by the measure was quantified using GPT3 question answering.

Results:
Three thousand recent accident reports were used for this evaluation. Overall, the analysis indicates that up to 30% of cases could have been prevented or mitigated by a set of six known behavioural measures: ● 10%, if the person had warmed up beforehand (10% “yes”, 2% “unclear”, 88% “no”) ● 8% by wearing better footwear (15% “unclear”) ● 7% by wearing gloves (2% “unclear”) ● 4% by wearing a helmet (4% “unclear”) ● 3% by wearing elbow and knee pads (2% “unclear”) ● 2% by wearing safety glasses (3% “unclear”) In the presentation, further results will be presented for specific types of sport.

Conclusion:
The “automated evaluation tool” provides a quick insight into the preventive potential of a given set of accidents, exemplified in this study for sport accident reports as available in the IDB full data set. As the tool is open for any question, i.e., any behavioural injury prevention measure, it could prove an interesting new approach for the evaluation of prevention measures. In a next step, these results will be evaluated against empirical studies of the efficacy of the prevention measures used in this study.

Keywords: Sport injury prevention, GPT3, Automated text analysis

Authors:
Robert Bauer*, Anna Maschek**, Andreas Stoeckl**, Johanna Trauner-Karner*. * Austrian Road Safety Board (KFV), ** University of Applied Sciences Upper Austria