Six studies met the inclusion criteria (1,832 participants, range 35 to 1,187). There were two prospective and four retrospective studies. All studies included a representative patient spectrum, used an acceptable reference standard, avoided partial verification bias, provided relevant clinical information and reported uninterpretable results. Incorporation bias was avoided in two studies, differential verification bias was avoided in two studies, selection bias was avoided in four studies and five studies explained withdrawals. None of the studies blinded interpreters of the tests.
The overall rate of postoperative infectious complications was 23.7% (95% CI 16.2 to 32.1; six studies). The highest area under the curve (0.81), specificity (83%, 95% CI 77 to 90 with corresponding sensitivity 68%, 95% CI 57 to 79) and diagnostic odds ratio (11.7 95% CI 6.1 to 22.3) were achieved on day four (five studies). On day four the positive predictive value was 55.7% (95% CI 50.3 to 61.1) and negative predictive value was 89.3% (95% CI 87.1 to 91.5). The mean cutoff value of C-reactive protein on day four was 135mg/L. The highest sensitivity was achieved on day two (75%, 95% CI 58 to 91 with corresponding specificity 58%, CI 46 to 69; four studies). Considerable heterogeneity was observed for all analyses.
Summary estimates of sensitivity and specificity from the bivariate random effects model were reported as being similar.