This review aims to determine the accuracy with which published risk scores predict spontaneous preterm birth in pregnant women. Studies were identified without language restrictions through nine different databases (up to June 2002), and manual searching of bibliographies of known primary and review articles. Two reviewers selected studies independently and extracted data on their characteristics, quality and accuracy. Accuracy data were used to form 2 x 2 contingency tables of the results of risk scoring with spontaneous preterm birth as the reference standard. Heterogeneity was assessed and its reasons were explored. Summary estimates of accuracy were produced within clinically appropriate subgroups. There were 19 primary accuracy articles that met the selection criteria, including a total of 67390 women. There are 12 different risk-scoring systems, the one developed by Creasy being the most commonly evaluated. Quality features of an ideal study, such as blinding and consecutive enrolment, were frequently missing from the included studies, no study fulfilled all criteria for high quality study, and there was heterogeneity between their accuracy estimates. The reference standard most often used was birth before 37 weeks' gestation. The point estimates for the likelihood ratios (LRs) varied widely among the studies. LRs for an abnormal score (LR+) ranged from 1.0 (95% confidence interval (CI) 0.6-1.4) to 38.8 (95% CI 23.5-63.9) while that for a normal score (LR-) ranged from 0.1 (95% CI 0.02-0.6) to 1.2 (95% CI 0.9-1.6). In otherwise asymptomatic women, risk scoring in early pregnancy has a wide range of accuracy in predicting spontaneous preterm birth before 37 weeks' gestation. The evidence is of a relatively poor quality and lacks clinically important reference standards.