The bone turnover quartet (PINP, β-CTX, osteocalcin, 25-hydroxyvitamin D) and healing of lower-limb fractures: protocol of a systematic review with exploratory meta-analysis (retrospective registration)

RETROSPECTIVE REGISTRATION. This record documents the methods of a systematic review that had already been completed when it was deposited. The searches were run on 4 November 2026; screening, data extraction, quality assessment and quantitative synthesis were completed on 5 November 2026; this registration was created afterwards. It is therefore NOT a prospective preregistration and must not be cited as evidence that the analyses were prespecified. Its purpose is limited and specific: to place the protocol, the decision rules, the extraction dataset and the analysis code on the public record so that a reader can distinguish the decisions taken before the data were seen from those taken after. WHY NOT PROSPERO. PROSPERO does not accept a review once data extraction has begun. Submitting a record there at this stage would have required a false declaration of review stage. A truthful statement of non-registration in the paper, combined with this dated deposit, was preferred to a registration obtained after the fact. REVIEW QUESTION. In patients with lower-limb fractures (or long-bone fractures including the lower limb), treated operatively or conservatively, do serum PINP, β-CTX, osteocalcin (OC) or 25-hydroxyvitamin D (25(OH)D), measured before or after surgery, discriminate between normal healing and delayed union or nonunion; and does the level of any of these markers differ between the two outcome groups? ELIGIBILITY. Included: human clinical studies with at least 10 patients; lower-limb or long-bone fractures including the lower limb; at least one of the four markers measured in serum at one or more time points; and either a group comparison by healing outcome or reporting of predictive parameters (sensitivity, specificity, AUC, odds ratio). Excluded: animal, ex vivo and cell studies; osteoporosis drug trials in which healing outcome was not an endpoint; studies restricted to vertebral or upper-limb fractures; reviews, commentaries, letters and case reports; and studies from which quantitative data could not be extracted and whose authors could not be contacted. SOURCES AND STRINGS. Five primary sources — PubMed, Europe PMC, Web of Science Core Collection, CNKI and Wanfang Data — and two supplementary sources — BIOSIS Previews (Ovid, 1995–2026) and CENTRAL (Cochrane Library, 2026 Issue 11). All searches were run on 4 November 2026, limited to English and Chinese, with no lower year limit. Verbatim strings and yields are given in the deposited protocol (Appendix, section C3); the primary sources returned 1 330 records before deduplication and 506 duplicates were removed. Two coverage constraints are recorded rather than concealed: Embase is not subscribed at this institution and CENTRAL was used as a partial substitute; and Ovid here covers BIOSIS Previews only, with MEDLINE reached through the Web of Science platform and EBSCOhost. SELECTION AND EXTRACTION. Records were deduplicated and screened on title and abstract, then on full text, by two reviewers independently, with disagreements resolved by discussion; reasons for exclusion at full text were recorded. Two reviewers independently extracted first author, year, country, design, fracture site, sample size, number of delayed unions or nonunions, outcome definition and time of assessment, markers and assay methods, sampling time point, and effect sizes with 95% confidence intervals, together with 2x2 counts where available or reconstructible. DIGITISATION RULE. Where group means were published only as charts (Emami 1999, Figures 2B and 3B), values were read by script: the vertical axis was calibrated by least squares on tick-label centres to obtain a linear pixel-to-unit mapping, and marker centres were then converted. The error bars in the source figures are ONE-SIDED, so each half-length equals one standard error; reading them as upper-and-lower caps would double the apparent dispersion. SD was recovered as SEM x sqrt(n). Calibration residuals were 1.7 and 0.7 pixels. All readings are scripted and reproducible. RISK OF BIAS. QUADAS-2 across patient selection, index test, reference standard, and flow and timing, with applicability concerns for the first three domains. SYNTHESIS. Sensitivity and specificity were pooled separately on the logit scale by DerSimonian-Laird random effects with inverse-logit back-transformation; the diagnostic odds ratio on the natural log scale; Hedges' g (small-sample-corrected standardised mean difference) for continuous outcomes, variance g^2/[2(n1+n2-2)] + (n1+n2)/(n1n2). The bivariate and HSROC models were not used, because at k=5 with a narrow range of S the parameters are not identifiable. The area under the SROC curve was not reported: the Moses-Littenberg regression returned a negative slope (b = -0.042), outside the admissible range, and an area computed from it is uninterpretable. That failure is reported in the paper rather than concealed. Markers with a single study were not pooled. RESULTS, FOR THE RECORD. Twenty studies entered the qualitative synthesis. Twelve provide sensitivity and specificity sufficient to rebuild a 2x2 table (PINP 5, β-CTX 4, OC 2, BSAP 1), and eight provide group means suitable for pooling (PINP 2, β-CTX 2, BSAP 2, OC 2). Pooled: PINP sensitivity 85.9% and specificity 76.5%, DOR 22.1; β-CTX 67.9% and 68.9%, DOR 6.3; OC 77.9% and 69.1%. Standardised mean differences: PINP g = -1.12, β-CTX g = -0.52, BSAP g = -0.94, OC g = -0.86. POST HOC DECISIONS, LISTED IN FULL. Six methodological choices were made after part of the data had been seen, and each is recorded here with its stage and its effect: (1) the CENTRAL string was widened from five marker terms to the full list after the first screening round, raising the yield from 89 to 135 records and retrieving two eligible studies the first run had missed; (2) the Chinese Medical Journal Network was searched, returned no record not already retrieved from CNKI or Wanfang, and was removed, with the flow denominators recomputed from 1 617 to 1 330; (3) Bakhsh 2024 moved from narrative-only to quantitative on obtaining the full text, raising PINP from four studies to five and making β-CTX poolable for the first time; (4) Figures 2B and 3B of Emami 1999 were digitised after confirming that the text carried no numerical values, raising BSAP and OC from one study each to two; (5) the SROC area was abandoned when the fitted slope came out negative; (6) the time point entering the main analysis was fixed only after inspecting the time-point data, with one time point per study for β-CTX to avoid counting the Hao cohort twice, and week 4 primary with week 7 as sensitivity for OC. None of these reversed the direction of any pooled estimate. MATERIALS. Deposited with this registration: the protocol (which follows the PRISMA-P structure item by item), the extraction dataset with per-item source annotation, the digitisation outputs including the pixel readings and calibration parameters, the original figure crops, the analysis and plotting scripts, and the complete numerical outputs. The statistical code uses only the Python 3 standard library and Pillow, so every number in the review can be recomputed without external statistical packages.