When a positive procalcitonin is not diagnostic of early-onset neonatal infection: a systematic diagnostic-accuracy review and Bayesian decision analysis

Supporting materials and manuscript for a systematic review of the diagnostic accuracy of procalcitonin (PCT) for early-onset neonatal bacterial infection (EONI), coupled with a Bayesian decision re-analysis. CONTENTS. (1) The review protocol, fixed before any screening was performed. (2) The exact search strings for PubMed/MEDLINE and Europe PMC, with the date each was run and its yield. (3) The complete screening log, with a reason recorded for every excluded record at both the title-and-abstract and full-text stages. (4) The per-study extraction table and the QUADAS-2 appraisal, with the supporting quotation for each judgement. (5) The list of arithmetic inconsistencies identified when published predictive values were recomputed from the reported sensitivity, specificity and sample size. (6) The Python code used for the searches, the screening workflow, the decision re-analysis and the figures. PRINCIPAL FINDINGS. Of 941 records screened, 203 were sought for full text, 69 assessed and 26 included. Three studies were at low risk of bias for the index test; 23 selected their threshold after inspecting the data, the values reported as optimal spanning 0.142 to 27 ng/mL; at least seven counted PCT among the criteria defining the infection. In the largest study with a prespecified threshold and blinded adjudication (3051 consecutive at-risk newborns), cord PCT had a sensitivity of 42 percent and a specificity of 98.6 percent. At its own prevalence of 1.2 percent, negative predictive value was 99.3 percent while 58 percent of infections were missed; a test identifying no one would have scored 98.8 percent. ACCESS. Files are under restricted access while the associated manuscript is under peer review. Access will be granted on request to editors and reviewers, and the deposit will be opened on publication. REGISTRATION. The review is not registered. PROSPERO accepts registration only for reviews that have not progressed beyond completion of data extraction. The protocol deposited here was fixed before screening and is provided so that what was planned can be compared with what was done. USE OF ARTIFICIAL INTELLIGENCE. A large language model was used to assist with code generation, a first pass of screening and extraction against pre-specified criteria, and language editing. All output was reviewed by the author, who verified the accuracy estimates on which the conclusions rest directly against the full texts of the corresponding studies. Single reviewer; no independent duplication.