Background The development of running-related injury (RRI) prevention programmes is based on aetiology described in longitudinal studies. Such studies have been conducted assuming that risk factors may influence the occurrence of RRIs under a linear system approach. Such approach has not been successful in explaining and/or predicting RRIs satisfactorily.
Objectives To develop and validate an artificial intelligence (AI) algorithm in order to identify RRI risk profiles in recreational runners.
Design Mathematical model.
Settings São Paulo, Brazil.
Participants 191 recreational runners.
Assessment of Risk Factors This was a 3-step AI study using data from a prospective cohort study. In step 1, variable selection and exploratory analyses were conducted in the original (n=191) and simulated data (n=5000). In step 2, the AI algorithm was developed using self-organising maps, k-means and probabilistic neural networks. The algorithm was trained in 80% (n=4000) of the simulated data, and validated using the remaining 20% (n=1000). Characterisation of RRI risk profiles was performed in step 3.
Main Outcome Measures RRI risk profiles were established based on the groups created by the developed algorithm. Descriptive analyses were performed to summarise the risk profiles.
Results The variables with greatest influence in the algorithm were: sex; running intensity; history of RRIs; and current musculoskeletal discomfort related to running. Five groups were suggested by the algorithm. Male runners reporting previous RRIs and running in low-to-moderate intensities (>6 min/km) were at the highest risk of RRIs. Male runners reporting previous RRIs and running in high intensities (3 to 5 min/km) in about 32.1% of the time were at the lowest risk of RRIs. The classification accuracy of the algorithm presented a median of 99.6% (interquartile range: 99.5% to 99.8%).
Conclusions A non-linear system approach using AI and machine learning techniques were successful in developing an RRI risk profile algorithm for recreational runners.