NIPT timing method explainer
Formulas and results from our paperIllustrative data where marked
Non-invasive prenatal testing (NIPT) reads fetal DNA in the mother's blood. For a male fetus the result is reliable once the Y-chromosome fraction reaches 4 %, which happens later for some women than others, but every week of waiting shortens the time to act on a bad result. Our team had to recommend a testing week for each BMI group that balances the two.
The risk function, the objective and the procedures are the ones in our working paper, and sections 4 and 5 show the paper's results. The contest's patient data is not published here, so sections 2 and 3 run the method on a synthetic group of women instead. Their numbers show how the method behaves; they are not our results.
The cost of waiting
The problem says only that finding a problem before 12 weeks carries low risk, 13–27 weeks high risk and 28 weeks or later very high risk. We turned that into a piecewise linear penalty f(t).
f(t) = 0 up to 12 weeks, 0.1 (t − 12) up to 28 weeks, then 1.6 + 0.3 (t − 28).
Choosing the week
Illustrative dataFor each week t from 10 to 25 (step 0.1), take every woman in the group tested by then. Estimate the share whose Y fraction was at least 4 % (P̂) and the share whose result matched the baby's outcome (Âcc), then score W(t) = a P̂ + y Âcc − b f(t). The best week maximises W, provided at least 10 women have been tested.
250 synthetic women, more of them tested early. Draw a few new samples: the best week moves by weeks, because W is nearly flat around its peak.
Why the estimates are Bayesian
Is the recommendation robust?
Problem 3's check: add measurement noise and diagnosis errors to the data, find the best week again, repeat 1,000 times and look at the spread. Here it runs on the synthetic group from section 2.
Our result for male fetuses (problem 2)
First a best week for each woman, from women with a BMI within ±0.5 of hers; then K-means on (BMI, best week) with whole-number cut points gave five BMI groups, and the objective picked each group's week.
| Group | BMI | Women | Centre (BMI, week) | CV | Best week |
|---|---|---|---|---|---|
| 1 | [20, 29) | 191 | 28.97, 11.7 | 0.087 | 13 |
| 2 | [29, 32) | 300 | 30.77, 11.4 | 0.033 | 12.5 |
| 3 | [32, 35) | 244 | 32.71, 11.2 | 0.0136 | 12 |
| 4 | [35, 38) | 174 | 34.65, 11.3 | 0.0257 | 15.5 |
| 5 | [38, ∞) | 81 | 37.49, 12.3 | 0.0693 | 12.5 |
From the working paper. Group 4's later week and group 5's small size were flagged there as needing more data.
Female fetuses: spotting trisomies (problem 4)
With no Y chromosome to measure, the question becomes classification: 554 female-fetus samples, 63 of them with a T13, T18 or T21 finding. A random forest ranked the features, the strongest went into a logistic regression, and a decision tree gave a readable version.
Random-forest feature importance
Accuracy and AUC
Only 11.4 % of samples are abnormal, so calling every sample normal already scores 88.6 % accuracy. The paper's recall and precision are weighted averages over both classes; the share of abnormal cases caught is not reported.
Logistic regression, odds ratio per standard deviation (log scale)