Cross-site pooling with feature harmonization for small-for-gestational-age (SGA) prediction across heterogeneous feature spaces

Small-for-gestational-age (SGA) neonates face health risks when undetected, yet machine learning (ML) models are difficult to apply across sites with heterogeneous feature availability. We developed a cross-site pooling framework, incorporating feature-specific imputation and leakage-safe five-fold partitioning, for SGA prediction using Malaysian ( n = 18,582) and Indian ( n = 1,760) antenatal data. The framework is evaluated on a held-out partition of 853 complete records including 184 SGA events. Three training strategies were evaluated: (i) country-specific training using common features (baseline), (ii) pooled training with feature harmonization, and (iii) cross-country transfer using common features. In the Malaysia test set ( n = 807, 169 SGA), the AUROCs were 0.825, 0.823, and 0.606, whereas in India test set ( n = 46, 15 SGA), they were 0.553, 0.673, and 0.695. The feature-harmonized approach showed results similar to the country-specific baseline in Malaysia (0.823 vs. 0.825) and numerically higher than the India-only baseline in India (0.673 vs. 0.553) with wide overlapping CIs on 46 scans (15 SGA). In a training size-matched analysis, pooling Malaysian data raised India AUROC from 0.573 to 0.653 (95% CI 0.604–0.706) while adding the harmonized feature gave 0.655 (95% CI 0.608–0.706), no measurable difference. Compared to clinical diagnosis using EFW<10th centile, the model achieved slightly lower AUROC (Malaysia: 0.823 vs. 0.838; India: 0.673 vs. 0.695). We showed that cross-site pooling was feasible despite heterogeneous feature availability. Feature harmonization enabled integration of datasets with differing feature spaces rather than providing discriminative gain. Independent validation on unseen sites, clinically justified operating thresholds, and prospective evaluation are required before clinical deployment.