Hello,
It was mentioned briefly in the webinar that there may be some inaccuracy inherent to the training data labels (gestational age or each sample). Can we estimate a confidence measure or RMSE for these values? Through what method were they obtained in the published and unpublished data sets (ultrasound + LMP estimate or more precise methods)? I looked for this info in the documentation for GSE113966 and couldn't find anything.
I've read that the only way to be certain about the date of conception is through artificial insemination or monitoring basal body temperature and serial luteinizing hormone levels (https://www.ncbi.nlm.nih.gov/pubmed/12501080). The Science 2018 (Ngo et al.) paper cites that "in previous studies, ultrasound and last menstrual period estimates of gestational age, which assume delivery at 40 weeks gestation, fell within 14 days of the observed gestational age at delivery with 57.8% and 48.1% accuracy, respectively."
Can we expect a similar level of confidence in the training data?
I'm just trying to contextualize the need met by subchallenge 1: is the purpose to simply develop a way of estimating gestational age that is cheaper than ultrasound, or more accurate as well?
Thank you for your help, and for all of your work organizing this challenge.
Nicola
Created by Nicola Lawford nicola Hi Nicola,
With this challenge we will only be able to assess if gene expression predicts the LMP + Ultrasound based gestational age estimate, irrespective of possible onset of obstetrical disease. This could be useful towards developing a cheaper and more convenient method than ultrasound. We won't be able to tell if prediction models are more accurate than LMP+ Ultrasound since a true gold standard is not availble.
The estimate from Savitz et al 2002 quoted in Ngo et al 2018 would apply here as well. Based on the set of all patients with normal pregnancy that delivered by spontaneous labor, 52% have delivered within 1 week from expected due date, yet keep in mind, those are only a fraction of all patients involved.
Adi