Limited accuracy of the Mitchell static-dynamic sport classification for predicting CMR-defined cardiac remodelling patterns in elite athletes

Abstract Background In athletes, the static-dynamic Mitchell classification is recommended to interpret sport-specific cardiac remodelling. Derived mainly from echocardiography and presumed haemodynamic load, it has not been rigorously validated against cardiac MRI (CMR) indices and individual CMR phenotypes in elite athletes. Purpose To validate the Mitchell sport classification by assessing how closely CMR-derived remodelling patterns follow static-dynamic load, and how accurately sport classes predict their expected CMR phenotypes. Methods Cross-sectional CMR study in ELITE+, comprising two harmonised elite athlete cohorts. Static and dynamic load ranks (low/moderate/high) were assigned per primary sport according to guidelines. Indexed LV mass (LVM), end-diastolic volumes (EDV), maximal LV wall thickness (LVWT), ejection fractions, and concentricity indices were converted to Z-scores using sex-matched, non-competitive controls. Associations between each CMR metric and static/dynamic ranks were evaluated using linear regression. The two metrics with the strongest associations defined four high- versus low-load CMR phenotypes (using Z-score≥1.645 vs Z1.645), to test accuracy of the four corresponding static-dynamic sport classes in predicting CMR phenotypes. We also assessed how closely the pattern of sport positions in the CMR metric pair plane matched their theoretical positions in the static-dynamic grid using a shape-similarity test (Procrustes). Results We included 1,009 individuals: 760 elite athletes (24 IQR 21-28 years, 40% women, 43 sports) and 249 controls (30 23-43 years, 49% women). Sports were classified as isolated high-dynamic (41%), isolated high-static (11%), combined high-static-high-dynamic (44%), or low/moderate load (4%). Static load was most, though weakly, correlated with LVM (R²=0.08), followed by concentricity (R²=0.07) and LVWT (R²=0.06), whereas dynamic load was most correlated with RVEDV (R²=0.06) and LVEDV (R²=0.05), and also with LVM (R²=0.04) (all P0.001). On the LVM-RVEDV plane that best followed the static-dynamic axes, sport class discriminated CMR phenotypes with a balanced accuracy of 37% (95% CI 34-40%; vs a 25%-chance level). By class, sport classes predicted the corresponding CMR phenotypes in 41 and 44% of the low-load and high-static-high-dynamic groups, but only in 18% of isolated high-static and high-dynamic groups (Fig. 1). Overall, the pattern of sport positions in the LVM-RVEDV plane weakly matched their pattern in the static-dynamic sport grid (Procrustes correlation 0.24; P0.001; R²=0.12; Fig. 2). Conclusion In elite athletes, sport classes show limited accuracy (37%) for predicting CMR-defined remodelling patterns, even when phenotypes are defined using CMR metrics that best reflect static/dynamic load differences between sports. CMR in athletes should preferably be interpreted alongside training and performance data, within a data-driven framework, rather than relying on sport frameworks.Figure 1:Scatter plotFor image description, please refer to the figure legend and surrounding text. Figure 2:Procrustes analysisFor image description, please refer to the figure legend and surrounding text.