<p>Spontaneous preterm birth (sPTB) remains a leading cause of neonatal mortality. While current prediction models perform well in the third trimester, identifying risk in the early-to-mid second trimester is critical for enabling effective interventions. This review evaluates the performance and utility of predictive models for sPTB developed for the second trimester, categorizing them into clinical, biomarker, and ultrasound approaches and suggests a way forward. Clinical models leveraging electronic health records and natural language processing achieve high local accuracy (AUC 0.77–0.85) but often fail to generalize and suffer from low positive predictive values in diverse populations (1–5%). Biomarker-based models, including multi-omics and cell-free RNA, demonstrate superior discrimination in high-risk cohorts (0.80) but struggle with reproducibility and cost. They often have smaller sample sizes, raising concerns over generalizability. Ultrasound based features, such as cervical consistency and texture analysis, significantly outperform standard cervical length measurement (0.68 vs. 0.84) but lack standardization and external validation. Integrated models combining clinical history with biological layers consistently yield the most robust predictions (AUC &gt; 0.82) but face implementation barriers regarding complexity and data harmonization. sPTB is a heterogeneous syndrome that cannot be adequately captured by a single modality. Reliance on isolated markers restricts predictive capability. Future research must prioritize multi-modal integration and phenotype-specific modeling to address distinct etiologies. Furthermore, to bridge the gap between research and practice, model evaluation should shift beyond the Area Under the Receiver Operating Characteristic (AUROC) curve to include rigorous calibration analysis and reporting of detection rates at fixed false-positive cutoffs.</p>