Neonatal outcomes after prenatal exposure to selective serotonin reuptake inhibitors: extraction dataset and analysis code

Abstract

Version 1.20, 12 September 2026. No estimate, table 2 value or certainty rating changes. Three presentation and build corrections. Supplementary table S5b, the analyses whose interval excludes no effect only while every data source is present, was a single eight-column table that wrapped author names mid-word when typeset and left a reader unable to see which omission returned an interval to the null; it is now one compact block per analysis, each with its own three-column table naming every omission, its estimate with interval, and whether that omission returns the interval to no effect. Supplementary file 2 printed all twenty-one analysis scripts verbatim, which put about a hundred pages of source in front of anyone opening the submission proof; it is now a one-page manifest giving each script, what it does and the order to run them in, with the code itself supplied runnable here and in supplementary file 3 exactly as before. The sentence under the leave-one-out table for the after-week-20 pulmonary hypertension stratum printed raw record identifiers rather than author and year labels, because that list is keyed on the bare study name while the table it sits under is keyed on the full record identifier. In the document build, build_bmj.py cleared the entire package folder rather than the manuscript folder alone, so rebuilding the manuscript silently deleted the tables, figures and supplementary items already assembled beside it. Version 1.19, 12 September 2026. Corrects the PRISMA checklist builder in the document build: locations were matched anywhere in the rendered page text, which sent item 5 to page 1 because the abstract carries the phrase "Eligibility criteria for selecting studies". Headings are now matched only where they stand alone on their own line. No analysis output, estimate or study set changed from version 1.18. Version 1.18, 12 September 2026. No study was added or removed; the analysis set is unchanged at 127 studies on 82 independent data sources. Changes: leave-one-out is now computed for every analysis with three or more data sources rather than four or more, so that it matches the fragility test the GRADE imprecision rule already used; leave-one-out confidence intervals are now stored alongside the point estimates and printed, omission by omission, in the new supplementary table S5b, and for the unspecified-timing pulmonary hypertension analysis in supplementary table S4a; the supplementary table S1 column formerly headed "Leave-one-out range" is now "Leave-one-out point-estimate range", because it is a range of point estimates and was open to being read as a confidence interval. The archive now also carries the document build and the manuscript sources, so the submission can be rebuilt from this deposit alone. The complete data extraction and analysis code for the systematic review and meta-analysis "Neonatal outcomes after prenatal exposure to selective serotonin reuptake inhibitors" (PROSPERO CRD420251243990). The registered searches were run on 14 August 2026 from database inception, and all four registered sources were searched again on 7 September 2026 on a broadened exposure string adding antidepressant terms, also from inception. 127 observational studies drawing on 82 independent data sources were included. SSRI_extraction_dataset.xlsx holds the extraction in full. The Estimates sheet has one row per extracted estimate, 844 in all, of which 514 are the adjusted estimates that enter the analyses. Every one carries the verbatim sentence or table row it was taken from and the table or figure number in the source paper, so any extraction can be checked against its source without contacting the authors. Further sheets hold study characteristics, Newcastle-Ottawa domain scores with the adjustment-set assessment, and the reports excluded at full text with their reasons. SSRI_zenodo_deposit.zip holds the same dataset, the native JSON file the analysis reads, the three analysis output files (pooled estimates, subgroup and sensitivity results, GRADE judgements), and the Python scripts that produce every estimate, table and figure in the paper, with a README giving the order to run them in. SSRI_zenodo_README.md is that README, provided separately so it can be read without downloading the archive. Python 3.11 with numpy, scipy, pandas and matplotlib. No proprietary software and no random seeds: every result is deterministic and reproducible from the deposited files alone. Version 1.17, 12 September 2026: a second round of external pre-submission review. Two changes to the analysis. The risk-of-bias downgrade threshold in grade_v1.py was implemented as a proportion below 0.34 while the protocol and the manuscript both say fewer than a third, so a body meeting the minimum adjustment set in exactly one third of its studies was downgraded against the stated rule; the implementation is corrected to the stated rule rather than the wording to the implementation. Exactly one body sits at that boundary, small for gestational age against unexposed women at 4 of 12, and it carries an imprecision downgrade as well, so no certainty rating changes: the table still reads 24 bodies, 22 very low and 2 low. Supplementary table S2 gains a panel listing every row held out of a pooled analysis with its reason, eleven in all, read from the dataset so the table and the analysis cannot disagree; three of those rows changed a printed estimate at version 1.16 and that is now stated rather than silent. The remaining changes are to the manuscript and not to any number: the abstract no longer presents the any-timing pulmonary hypertension estimate as a third finding whose prediction interval excludes no effect, because three of its six sources cannot be removed without the interval returning to the null, which is also why it carries an imprecision downgrade; several sentences that still described the pre-correction analysis were corrected; and the Table 2 row formerly labelled all exposure windows is relabelled unspecified exposure timing. No study was added or removed and no pooled estimate changed. Version 1.16, 12 September 2026: corrects the persistent pulmonary hypertension analysis and two pooling-rule violations found by external pre-submission review. The comparator A pulmonary hypertension pool, reported as covering all exposure windows, was drawing three of its nine inputs from those studies' after-week-20 rows, which the protocol forbids in an any-timing pool; the pool_eligible flag existed only on the wave 4 records, so the selector fell back to the narrowest interval and chose the late-window row for Chambers 2006, Bérard 2017 and Kieler 2012. Those three report no any-timing estimate and leave that pool, and Munk-Olsen 2021's any-timing row is now explicitly distinguished from its after-week-20 row. The estimate moves from 1.77 (1.18 to 2.66) on nine data sources to 1.30 (1.06 to 1.60) on six, with I-squared falling from 63.9 to 0.0 per cent and a prediction interval that now excludes the null. Separately, three studies reporting an Apgar threshold other than below 7 (Cornet 2024 and Ulbrich 2021 at 5 or lower, Lund 2009 at 7 or lower) are removed from the Apgar pools, moving the five-minute estimate from 2.22 (1.78 to 2.77) to 2.06 (1.69 to 2.51) and reducing the graded bodies of evidence from 25 to 24. A leave-one-out of the late-exposure stratum and a Danish participant-overlap sensitivity are added. The eligibility decisions are recorded row by row in the dataset itself, and fix_review_rows.py and screen_ad_research.py are added to the code, the latter so that the screening triage described in the methods can be inspected. No study was added or removed: the review still rests on 127 studies, 82 data sources and 844 extracted rows. Version 1.15, 11 September 2026: keeps the deposited archive identical to the copy submitted with the manuscript. Between version 1.14 and submission, subgroups_v2.py was changed to record every substitution that crosses a significance threshold rather than only the six largest by percentage change, so that both of the flipping substitutions the Results names can be checked against the analysis output instead of only one; three build-check scripts were updated to match the current study set and the record-id stripping pattern was extended to the wave 4 identifiers, which had been printing in the reference list of the submission PDF. No data, no pooled estimate, no subgroup or sensitivity result and no figure changes: the only new content in the analysis output is the flips key in subgroups_v2.json. This version of the archive is the one supplied with the manuscript as supplementary file 3. Version 1.14, 11 September 2026: incorporates the broadened re-search. The registered exposure block named the individual SSRIs and the phrase itself but not the word antidepressant, while the review’s prespecified rules admit reports defining exposure more broadly and assign them to a mixed-exposure stratum, so the search was narrower than the eligibility criteria it fed. All four registered sources were searched again on 7 September 2026, identifying 18448 records of which 4456 were screened in duplicate, together with 491 records carrying no abstract whose title alone names both an antidepressant and pregnancy. Of the 898 unique reports sought, 630 were removed on record type before retrieval, 199 of the remaining 268 were retrieved and assessed at full text in duplicate, 69 could not be retrieved through any institutional or open-access route and are listed in supplementary table S7d rather than assumed ineligible, and 16 studies were added. The review moves from 111 studies on 74 data sources to 127 on 82. Every newly included study was assessed for participant overlap against all 111 studies already in the review and against the other fifteen. Seven cluster decisions follow, all recorded in supplementary table S8: Laine 2019 joins the Finnish national cluster; Rommel 202