Table 2. PROBAST-informed appraisal of primary prediction model studies contributing to Table 1.

Items are recorded as “Yes/No/NR/Unclear” based on study reporting; when methodological details (e.g., missing-data handling, class imbalance handling, calibration/utility) were not reported, they are conservatively coded as “NR/Unclear.” This appraisal is intended to contextualize AUC-based performance comparisons with attention to validation design, reporting completeness, and clinical readiness rather than to assign formal risk-of-bias scores.

THA: total hip arthroplasty; TKA: total knee arthroplasty; DCA: decision curve analysis; NR: not reported; AUC: area under the curve; EMR: electronic medical record; CV: cross-validation; NSQIP: National Surgical Quality Improvement Program; RF: Random Forest; SMOTE: Synthetic Minority Over-sampling Technique; PROBAST: Prediction model Risk Of Bias ASsessment Tool

Reference	Study (Author, Year)	Outcome	Data source	Validation	Missing data handling	Class imbalance	Calibration/Utility	PROBAST domains flagged	Notes
[13]	Huang et al., 2023	Surgical site infection	Ortho patients (train + external validation cohort)	External validation (separate external set)	NR	NR	Calibration + DCA reported	Analysis	Classic regression nomogram w/ calibration + DCA
[40]	Huang et al., 2021	Blood transfusion after THA/TKA	Multi-hospital EMR cohort	Random subsampling + 10-fold CV	Excluded missing/incorrect (0.73%)	NR	NR	Analysis	Reports AUC-focused comparison; calibration not highlighted
[47]	Zang et al., 2024	Perioperative blood transfusion (hip surgery)	Retrospective hip surgery cohort	Temporal split (first 70% train / last 30% test)	Excluded missing data	NR	Calibration + Brier + DCA reported	Analysis	Time split is stronger than random split; still single-system retrospective
[59]	Jiang et al., 2024	Pedicle screw loosening	Lumbar fixation cohort	Random split (8:2) + 10-fold CV	Imputed (<20%) w/ RF regression; otherwise excluded	NR	Calibration plots + Brier	Analysis	Strong reporting on calibration/Brier; still internal validation only
[60]	Xiong et al., 2023	Spine outcome model; per paper	Spine surgery imaging/clinical cohort	Train/validation split (0.75/0.25)	Excluded incomplete imaging	NR	Calibration + DCA reported	Participants / Analysis	Imaging exclusion can introduce selection bias; no external validation shown
[61]	Chen et al., 2023	Surgical site infection prediction	Retrospective cohort	Train/test split (reported)	Excluded incomplete clinical data	NR	Calibration curves reported	Analysis	Calibration reported, but imbalance handling not clearly described
[62]	Zhang et al., 2024	SSI following spine surgery	986 pts “complete data” cohort	5-fold CV (4 folds train / 1 validate)	Complete-case (“complete data” only)	NR	NR	Participants / Analysis	Internal CV only; AUC-heavy evaluation; missing handling = exclusion
[71]	Gupta et al., 2025	30-day mortality (femoral shaft fracture surgery)	NSQIP dataset	Stratified 80/20 split + 10-fold CV	Dropped vars >5% missing; kNN imputation for remaining	SMOTE + Tomek	Calibration slope/intercept + Brier	Analysis	Best reported “methods hygiene” among these (imbalance + calibration + imputation)
[72]	Han et al., 2025	Postoperative infection (nosocomial infection paper)	Retrospective cohort (2011–2024)	Random 70/30 split + 10-fold CV	Imputation reported (details in supplement)	NR	Brier reported	Analysis	Strong model-development description; still internal validation only