Biases of traditional term prediction models: results from different sample-based models evaluated on 41 343 ultrasound examinations

Abstract

Objective: To evaluate two Norwegian traditional, sample-based term prediction models as applied to the data from a large population-based registry. The two models were also compared with an established German model.

Methods: Our database included information from 41 343 non-selected ultrasound scans registered over the years 1987-2005. The prediction models were applied to measurements from the ultrasound examinations, and the resulting term predictions were compared with the actual times of the deliveries. The median bias (the difference between the true and the predicted date of delivery) was calculated for each model, both for the study population as a whole and for subgroups of measurements of biparietal diameter (BPD) and femur length (FL). Secondary measures, i.e. proportion of births within ± 14 days and the rates of preterm and post-term deliveries, were also assessed.

Results: The analyses showed that the models had significant biases, predicting delivery date either too late or too early. For each model the size of the bias varied, depending on the fetal size at the time of the examination; the extremes were minus 4 and plus 4 days for the BPD-based predictions. There were similar results with the FL-based predictions.

Conclusion: Term predictions made with traditional sample-based models had significant biases that varied over each method's measurement range. These models have important shortcomings, probably because of strict selection criteria in the process of constructing the models, and because the methods primarily aim at estimating the last menstrual period-based day of conception, not the day of birth.