Xueling Zhong, Yutao Wang, Wenting Chai, Yanni Lin, Nengtong Zheng, Jinhua Huang
The prealbumin-containing TyG-FPR-AISI composite was associated with clinically significant early SA-AKI and short-term AKI persistence, but its incremental performance beyond a clinical baseline model was modest. These findings support its exploratory role as a routinely available nutrition-relevant host-response risk-enrichment signal rather than a stand-alone prediction tool; external validation and recalibration are required before clinical application.
BACKGROUND: Postpartum stress urinary incontinence (SUI) is a highly prevalent condition that imposes substantial physical, psychological, and economic burdens, underscoring the necessity of early identification of high-risk populations to improve clinical outcomes. However, existing machine learning (ML) prediction models yield inconsistent results, and their performance and reliability remain uncertain. This review aimed to synthesize the available evidence on ML-based prediction models for postpartum SUI.
OBJECTIVE: This study aimed to systematically evaluate the methodological quality, risk of bias, and predictive performance of ML-based prediction models for postpartum SUI, and to quantitatively synthesize their discrimination metrics.
METHODS: A systematic search of nine databases was conducted from inception to 23 March 2026. Studies developing and validating ML-based risk prediction models for postpartum SUI were included. Methodological quality and risk of bias were assessed using the PROBAST+AI tool, and reporting quality was evaluated with the TRIPOD+AI statement. A meta-analysis of the area under the receiver operating characteristic curve (AUC) was performed, employing robust variance estimation (RVE) to account for dependent effect sizes. The study was registered with PROSPERO (CRD420261369137).
RESULTS: Seven studies encompassing a total of 4,072 patients were included. All studies were rated as having a high risk of bias. The pooled AUC across 20 training models was 0.931 (95% CI: 0.880-0.962) and across 25 validation models was 0.890 (95% CI: 0.832-0.930). After correcting for dependent effect sizes applying RVE, the pooled AUCs were 0.915 (95% CI: 0.796-0.968) and 0.825 (95% CI: 0.637-0.927), respectively. Substantial heterogeneity was observed in both sets (I 2 = 98.2 and 93.6%, respectively). Subgroup analyses revealed that models using only clinical features achieved the highest pooled AUC (0.974), whereas multimodal models integrating clinical features, pelvic floor ultrasound, and pelvic floor electromyography parameters showed lower AUC (0.821) but more stable performance (I 2 = 50.3% vs. 95.2%). The most frequently included predictors were age, body mass index, parity, neonatal birth weight, and bladder neck descent.
CONCLUSION: Current ML prediction models demonstrate acceptable discrimination for postpartum SUI, but they exhibit a high risk of bias, poor reporting standards, and a lack of external validation, rendering them not yet suitable for direct clinical application.
SYSTEMATIC REVIEW REGISTRATION: https://www.crd.york.ac.uk/PROSPERO/view/CRD420261369137, CRD420261369137.