Seventy studies (4,881 patients) were included in the review. Mean study quality ranged from 10.9 to 11.1 points (maximum 15).
Both during treatment (43 studies, 3,235 patients) and after treatment (27 studies, 1,646 patients) exercise was associated with an improvement in fatigue (d=0.32, 95% CI 0.21 to 0.43 and d=0.38, 95% CI 0.21 to 0.54). There was moderate heterogeneity at both time points (Ι²=48.4% and 60.7%).
Univariate moderator analysis showed that for the during treatment model, adherence by baseline fatigue score interaction was related to fatigue outcomes. For the post-treatment model, moderating variables were duration post-treatment, length of exercise programme and type of comparison group. Larger effects were found for studies that used a wait-list comparison (d=0.66, 95% CI 0.42 to 0.90) compared with the effect for other comparison types (d=0.19, 95% CI 0.00 to 0.37).
Funnel plot asymmetry suggested potential small-study bias for all models. Begg and Egger tests both showed significant evidence of publication bias for the overall model (p<0.001), the during treatment model (p=0.003 and p=0.004) and the post-treatment model (p=0.025 and p=0.006).
Sensitivity analyses that investigated the effects of study quality and adherence reinforced the results found in the primary meta-analysis.