Reply to the letter to the editor: "Machine learning model for predicting preeclampsia-related adverse outcomes"

We thank Liu et al. for their constructive comments on our article reporting the internal and external validation of a reduced-feature machine learning model for the prediction of preeclampsia-related adverse outcomes. The letter raises four points, which we address in turn. Regarding calibration and clinical utility, we generated calibration curves for the gradient-boosted tree model in both cohorts. Neither cohort shows a systematic, clinically concerning miscalibration pattern; deviations are concentrated at the extremes of the risk distribution, where observations are fewest. Brier scores were 0.010 (German cohort) and 0.068 (North American cohort), indicating good overall probabilistic accuracy. We agree that a formal clinical utility analysis, such as net benefit across threshold probabilities, represents an important next step. Regarding incremental value over the sFlt-1/PlGF ratio, we note that while the ratio offers strong short-term rule-out performance, its positive predictive value for ruling in preeclampsia related adverse outcomes remains limited. Our model, which incorporates the sFlt 1/PlGF ratio alongside ten additional routine clinical features, achieved AUCs of 92% and 87% in the German and North American cohorts, respectively. A formal head-to head comparison remains an important direction for future work. Regarding dataset heterogeneity, we acknowledge the structural differences between cohorts as a relevant limitation, while noting that discriminative performance remained statistically indistinguishable across cohorts despite significant differences in baseline characteristics. Regarding longer-term outcomes, we agree this is a valuable direction and intend to pursue it as follow-up data become available.