Fifty-two studies were included in the review (2,943 participants). All of the studies fulfilled QUADAS criteria on inclusion of an appropriate patient spectrum, avoidance of differential verification bias and reporting of withdrawals. There was potential for partial verification bias in around 25% of studies. Around 73% of studies did not report on whether the index test results were interpreted blind to the reference standard. Around 85% of studies did not report whether assessment of the reference standard was blinded to the index test.
There was substantial heterogeneity in estimates of sensitivity (I2=79.5%), specificity (I2=84%) and diagnostic odds ratios (I2=67%). Type of tumour was the only variable to show a significant association with accuracy in meta-regression analysis (relative DOR 5.36, 95% CI 2.42 to 11.86). A second meta-regression was restricted to studies that used semi-quantitative scoring for immunohistochemistry and included the threshold as a covariate also found an association with tumour types. This analysis suggested that immunohistochemistry for other tumours was significantly more accurate than for brain tumours (relative DOR 4.38, 95% CI 1.82 to 10.54).
Based on pooling data in homogeneous subgroups of patients, summary estimates of sensitivity ranged from 53% (95% CI 33% to 73%) to 79% (95% CI 73% to 84%) and summary estimates of specificity ranged from 60% (95% CI 52% to 67%) to 93% (95% CI 91% to 95%).
The Egger test suggested a risk of publication bias (p=0.002).