Models trained in Sweden, Chile and Singapore are ahead of current evaluation methods — but none work as well in all three countries.
A study published in the Journal of Medical Internet Research, relayed on September 18, 2026, looked at more than half a million pregnancies followed in Sweden, Chile and Singapore. The machine learning models only used the information available during the first fourteen weeks of gestation to estimate the risk of serious complications, both maternal and neonatal.
The results are clearly in favour of the models in Sweden and Chile, where they far exceed the early assessment methods currently used; in Singapore, the gain remains statistically significant but more modest. The calibration — the agreement between predicted and observed risk — is correct for the Swedish and Singaporean models, and much worse for the Chilean model.
Two lessons emerge. First, social and demographic factors are among the most predictive variables in certain populations, information that traditional evaluation grids often ignore. Second, no model has been equally effective everywhere: the authors conclude that these tools will have to be adjusted and tested on a population-by-population basis, and that they are intended to support clinical decision-making, not replace it.
Read the full article on the original source
Go to the source →