Clustering analysis of hamstring injuries—A pilot study to determine phenogroups based on injury characteristics and recovery duration
Bálint Zsidai, György Gulácsi, Tibor Rudisch, Bence Márk Kovács, Pete Friar, Richard Nagy, Robert de Jonge, Eric Hamrin Senorski, Kristian Samuelsson, Gergely PánicsAbstract
Purpose
The objective of this pilot study was to determine clinically relevant phenogroups of professional football players with hamstring injuries using unsupervised cluster analysis based on demographic and injury‐related characteristics, and to assess return to play (RTP) duration with respect to the characteristics of these phenogroups.
Methods
This retrospective cohort study analysed 57 professional football players from Ferencvárosi Torna Club with magnetic resonance imaging (MRI)‐confirmed acute hamstring injuries reported between 2018 and 2024. Three clustering approaches were implemented and compared: k‐means clustering, hierarchical clustering using Ward's linkage method, and mixed‐type clustering using Gower distance with partitioning around medoids (PAMs). The optimal number of clusters was determined using silhouette analysis. Between‐cluster differences in RTP duration and injury characteristics were assessed using Kruskal–Wallis tests for continuous variables and chi‐square tests for categorical variables, with post‐hoc Bonferroni correction.
Results
Mixed‐type clustering identified seven phenogroups based on British Athletics Muscle Injury Classification (BAMIC) grade, lesion length, anatomical tear site, affected muscle and injury extensiveness. The average silhouette width was 0.290, indicating weak to moderate cluster separation. Clusters 1, 3, 4, 5 and 6 were characterised by low‐ to moderate‐grade injury parameters and median RTP durations ranging from 9 to 27 days. Clusters 2 and 7 displayed high‐grade injury characteristics with considerably prolonged RTP (medians of 46.5 and 94.5 days, respectively). Significant between‐cluster differences were observed for days to RTP, lesion length, BAMIC grade, injury severity and anatomical tear site ( p < 0.05).
Conclusions
Unsupervised machine learning may help identify clinically relevant hamstring injury phenogroups based on MRI‐derived tear characteristics, which may facilitate prognosis and inform decision‐making regarding expected RTP duration in professional football players.
Level of Evidence
Level IV, retrospective cohort study.