Can LLMs reliably perform epidemiological extraction in elite football?

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Anterior cruciate ligament (ACL) injuries remain a critical burden in professional football, resulting in prolonged absence and high economic costs. While physical demands have increased, traditional epidemiological surveillance relies on manual data extraction, which is resource-intensive and subjective. Large Language Models (LLMs) offer a potential solution for automated screening, but their reliability in sports medicine remains largely unexplored. This study aimed to: (1) update the epidemiology of ACL injuries in the Spanish First Division (LaLiga) over six consecutive seasons (2019/2020–2024/2025) and (2) validate the accuracy of an AI-assisted screening approach using ChatGPT compared to manual data extraction. A retrospective observational study was conducted using publicly available sources. Manual screening by external investigators (gold standard) was compared against an AI-assisted workflow based on a structured prompt-engineering framework. Analyses focused on injury frequency, mechanism, return-to-play (RTP) time, recurrence rates, and AI case-detection sensitivity. A total of 56 ACL injuries were identified in 52 players. Findings revealed a predominance in defenders (41.1%), a high prevalence of non-contact mechanisms (67.9%), and a median RTP of 280 days. Notably, the match-to-training injury ratio was 1.8–1, a significantly narrower gap than historical records. The recurrence rate was 17.9%, with 70% of cases occurring in the contralateral knee. Methodologically, the AI model exhibited a learning curve: 44.4% sensitivity in the initial refinement season (2019/2020), which progressed to a sustained 100% sensitivity across the following five seasons (2020–2025). ACL injury incidence in LaLiga remains stable, but the narrowing match-to-training ratio suggests that modern training intensities are reaching competitive levels of biomechanical stress. LLMs, once optimized through prompt engineering, provide a highly reliable and scalable tool for near real-time epidemiological surveillance, representing a methodological milestone in sports medicine research.