# Scientific review: what the available evidence can support

Review date: 2026-10-16. Scope: the free and Pro timing calculations, not medical management. This review supersedes the active-model recommendations in `scientific-model-review.md`; that earlier document describes the retired Gaussian/hybrid implementations.

## Decision

Replace the unsupported Gaussian baseline and the assembled personal hazard model with a reproducible **descriptive reference for the timing of eventual spontaneous labour**. Use the published NICE weekly frequencies. Do not describe the result as an individual's probability of labour today, their baby's birthday, or the probability of remaining pregnant. Personal measurements must no longer change the reference numbers.

This is an improvement in evidential validity and interpretability, **not demonstrated improvement in individual predictive accuracy**. No external validation dataset has been acquired. A large source population does not make an incorrectly defined prediction valid.

## Search and appraisal

This was a targeted critical search, not a registered systematic review, meta-analysis, or exhaustive search of every European language. Searches on 16 October 2026 covered NICE/NHS official statistics, Euro-Peristat, PubMed and primary publisher or institutional-repository records. Search concepts included gestational-age distributions, spontaneous labour onset, induction, competing risks, parity, prior pregnancy duration, BMI, assisted conception, and individual prediction. National registers and large European cohorts were prioritised; smaller prospective prediction studies and randomised trials were retained when their outcome definitions addressed a different question.

For each candidate, the review distinguished geography and recruitment period, selected population, endpoint, denominator, measurement resolution, treatment pathway and availability of a reproducible probability equation. Published abstracts were used where full text was inaccessible; no unavailable table was reconstructed from an abstract. No patient-level registry data were available. Sample sizes below refer to their named study, never to a pooled training dataset.

## Primary evidence and applicability

| Source | Population and endpoint | What it contributes; why it is not a personal daily model |
| --- | --- | --- |
| [NICE NG207, recommendation 1.2.2/table 1](https://www.nice.org.uk/guidance/ng207/chapter/Recommendations#pregnancy-lasting-longer-than-41-weeks), 2021 | Weekly timing among labours that eventually started spontaneously, attributed to NHS HES/MSDS 2019–20. Exact selected denominator, parity mix and individual counts are not supplied in this table. | Direct, reproducible descriptive weekly reference. It selects on the eventual event and does not supply competing-event hazards, individual predictors or day-level counts. |
| [NHS Maternity Statistics, England 2019–20](https://digital.nhs.uk/data-and-information/publications/statistical/nhs-maternity-statistics/2019-20), published 2020 | English NHS hospital activity, 1 April 2019–31 March 2020: 591,759 HES delivery episodes overall; also MSDS data. MSDS changed to version 2.0 in this period. | Establishes source setting and reporting limitations. **591,759 is not the NICE spontaneous-labour table's denominator** and cannot be used to manufacture its standard errors. Delivery method and onset method are distinct variables. |
| [Euro-Peristat 2015–2019 report, fact sheet C5](https://www.europeristat.com/wp-content/uploads/2022/11/Euro-Peristat_Fact_sheets_2022_for_upload.pdf) | National gestational-age-at-birth data, reported in completed weeks; 32 countries supplied gestational-age information, with some UK reporting overlap. Rates are among live births, with plurality-specific summaries. | Strong population context. These are birth outcomes including care interventions, not a spontaneous-onset curve. Country ascertainment and dating can differ. Do not sum overlapping national/UK series or turn ecological differences into personal country offsets. |
| [Gunnarsdóttir et al., BJOG 2025](https://pubmed.ncbi.nlm.nih.gov/40637056/), DOI `10.1111/1471-0528.18292` | Ecological analysis of births at ≥37 weeks in 28 European countries, 2015–2020. National weekly aggregates include live births and stillbirths. | Early-term proportions ranged from 17.8% to 49.1%; substantial heterogeneity remained. The study explicitly uses early-term birth as a proxy for elective birth practices. It supports separating health-system practice from spontaneous biological timing, not biological country multipliers. |
| [Corbett et al., Ireland, 2022](https://pubmed.ncbi.nlm.nih.gov/35640443/), DOI `10.1016/j.ejogrb.2022.05.010` | 12,323 low-risk nulliparous singleton pregnancies with spontaneous onset, one tertiary unit, 2011–2017. Excludes later booking, multiples, deaths and other delivery indications except post-maturity. | Median onset reported as 40.1 weeks; 80.5% of observed spontaneous labours occurred by 41+0. Useful independent shape check, but eventual-event selection and a restricted single-centre population prevent conversion to a general ongoing-pregnancy risk. Do not infer exact daily counts or reinterpret decimal weeks without the full convention. |
| [Ytterberg et al., Sweden, 2023](https://pubmed.ncbi.nlm.nih.gov/37714417/), DOI `10.1016/j.annepidem.2023.09.004` | Approximately 1.1 million spontaneous deliveries in the Swedish Medical Birth Register, 1990–2012; mixed models examine mean and variance of gestational duration. | First pregnancy averaged only 0.29 days shorter than second; parity also changed variance and interacted with prior preterm delivery. A mean difference is not a universal onset-day shift, and changing only the mean discards the reported heterogeneity. |
| [Modzelewska et al., Sweden, 2020](https://pubmed.ncbi.nlm.nih.gov/32706811/), DOI `10.1371/journal.pone.0236494` | Registry 1973–2012; 2,247,062 full-sibling pairs plus other family relationships. The source singleton dataset contained 3,914,795 live births. Endpoint is gestational age at delivery. | Full-sibling correlations varied by age gap and calendar period. Correlation is not a transportable regression coefficient; pair counts are not independent pregnancies. This does not justify the old 0.30 history coefficient, recency weights or ±6-day cap. |
| [Thagaard et al., Denmark, 2018](https://pubmed.ncbi.nlm.nih.gov/28738716/), DOI `10.1080/14767058.2017.1359825` | 9,055 singleton live births at Holbaek, 2006–2011; pregnancy-duration analysis censored medical intervention and compared dating methods. | Ultrasound-based duration was associated with BMI by about 0.20–0.22 days per BMI unit. An adjusted duration association is not a validated coefficient for another model's spontaneous-onset hazard. Censoring interventions does not itself identify real-world competing-event probabilities. |
| [Westvik-Johari et al., Nordic registers, 2021](https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1003683), DOI `10.1371/journal.pmed.1003683` | 4,510,790 live-born singletons in Denmark (1994–2014), Norway and Sweden (1988–2015); 33,056 sibling groups discordant for conception method. Outcomes include gestational age at birth and preterm birth. | Within-sibling mean differences were −1.0 day for fresh transfer and +0.3 day for frozen transfer versus natural conception. Selection, missing BMI/smoking and birth-versus-onset differences matter. Neither estimate licenses an IVF adjustment to an onset forecast. |
| [Migliorelli et al., PREDICT, Switzerland, 2026](https://pubmed.ncbi.nlm.nih.gov/41930568/), DOI `10.1111/aogs.70188` | Prospective cohort of 429 women at two university hospitals. Assessment at 39 weeks; singleton, cephalic, intact membranes, choosing expectant management. Outcome combines spontaneous labour **or prelabour membrane rupture** before 41 weeks. | Competing-risk regression reported 72% outcome occurrence and AUC 0.71–0.72; final predictors include age, BMI, prior vaginal delivery, cervical length and fetal fibronectin. This is much closer to prediction, but it is a different endpoint and selected population. Published associations or one aggregate outcome rate cannot recreate its baseline function, calibration or external validity. |
| [Wennerholm et al., SWEPIS, Sweden, 2019](https://www.bmj.com/content/367/bmj.l6131), DOI `10.1136/bmj.l6131` | Multicentre randomised trial: 2,760 analysed low-risk pregnancies; induction at 41 weeks versus expectant management to 42 weeks. Expectant arm n=1,379. | In that arm, 920/1,379 (66.7%) began labour spontaneously, 457 (33.1%) were induced and 2 had scheduled caesareans. Trial stopped early after mortality imbalance. Demonstrates that an ongoing-pregnancy denominator contains competing outcomes; not a day-by-day natural-history baseline or advice to defer care. |
| [Keulen et al., INDEX, Netherlands, 2019](https://www.bmj.com/content/364/bmj.l344), DOI `10.1136/bmj.l344` | 1,801 women in a multicentre randomised comparison of induction at 41 weeks versus expectant management until 42 weeks, with intervention when indicated. | Useful care-policy and safety context. The expectant strategy is not an intervention-free counterfactual. Randomisation compares strategies, not personal biological due dates; published treatment outcomes must not become ad hoc hazard coefficients. |

These sources are independent checks, not ingredients to combine merely because their associations have compatible-looking units. The Nordic datasets also overlap in populations and periods; their sample sizes must not be added to advertise a larger evidence base.

## Reproducible replacement reference

The input values below reproduce the NICE table's weekly percentages, not invented or digitised daily observations:

| Completed gestational age | Published percentage |
| --- | ---: |
| 31 weeks and under | 2.4 |
| 32+0–36+6 | 5.3 |
| 37+0–37+6 | 5.1 |
| 38+0–38+6 | 12.1 |
| 39+0–39+6 | 25.4 |
| 40+0–40+6 | 32.5 |
| 41+0–41+6 | 16.2 |
| 42+0 and later | 0.9 |

The published percentages total **99.9**, reflecting rounding. For internal arithmetic, each bin's probability mass is its published percentage divided by 99.9. Preserve the original inputs and disclose this normalisation. Do not mix independently rounded source cumulative values with the normalised distribution.

Implementation choices, distinct from observed data:

- Map the entered estimated due date to 40+0 (gestational day 280), using date-only arithmetic. A calendar mapping is not clinical re-dating.
- Display the source weekly shares as the primary evidence. If a daily calendar view is retained, spread a week's mass equally across its seven days **only from 37+0 through 41+6**, clearly labelled interpolation. This is not evidence that each day has the same real-world chance.
- Preserve the earlier and 42+ week masses in totals. Do not renormalise the visible 37–41-week window to 100%, fit a smooth tail, spread the open-ended 42+ bin over an arbitrary duration, or return zero where a daily value is unavailable.
- Any computed cumulative quantity describes the same retrospectively selected cohort. Do not restart its cumulative total at today's date or at the visible chart boundary.
- No adjustment by country, parity, age, BMI, conception method, cervical measures, prior birth or dating method. These can inform context or a future properly trained model, but not this descriptive table.
- Do not provide numerical confidence intervals: selected cell counts and a sampling model have not been established. Normalisation, daily interpolation and transportability uncertainty cannot be represented by an invented “±” band.

### The denominator distinction

Let `J` be the eventual onset pathway and `T` its gestational timing. The reference approximates `P(T in week w | J = spontaneous)` in its source population. A prospective user's question is closer to `P(t ≤ T < t+h, J = spontaneous | pregnant without onset at t, information at t)`. Those are different estimands.

Dividing a day's descriptive mass by the remaining selected-cohort mass is mathematically possible. It conditions on knowing the future pathway, however, so it must not be exposed as a user's daily chance. For illustration, the final two source bins would yield `16.2 / (16.2 + 0.9) ≈ 94.7%` in week 41 among selected eventual spontaneous labours. SWEPIS instead observed 66.7% spontaneous onset in its expectant arm, whose denominator included subsequent induction and caesarean. The cohorts are not interchangeable; this comparison demonstrates the estimand problem rather than a calibration factor.

The NICE reference also does not separately identify membrane rupture, duration of labour, or time of birth. A “labour or waters breaking” endpoint and a delivery-date probability must not be silently substituted.

## Why the old personal assembly is retired

Internal probability bounds and monotonicity are necessary software properties, not clinical validation. The previous implementation combined a hand-shaped logistic hazard, a 72% aggregate compatibility target, mean-duration differences, family correlations and shrunk subdistribution hazard ratios from different studies. It did not estimate those components jointly or retain their covariance and baseline functions.

A subdistribution hazard ratio is not an ordinary daily-hazard multiplier. Treating intervention as ordinary censoring, removing induced pregnancies, or constructing a pure onset survival curve answers different questions. Arbitrary shrinkage and bounded shifts do not establish transportability. An assumption-sensitivity envelope is not a confidence interval or a validated prediction interval. The responsible correction is to withdraw unsupported personal probabilities, not replace their coefficients with new guesses.

## Requirements before restoring individual predictions

1. **Define the target first.** Choose spontaneous contractions, spontaneous labour-or-PROM, or actual delivery as separate endpoints. Specify landmark gestational ages and horizons such as 24 hours, 7 days and 14 days. Define eligible users and clinically important exclusions without assuming missing information is reassuring.
2. **Obtain fit-for-purpose data with permission.** Need event dates/times, gestational dating method, spontaneous onset/PROM, induction, prelabour caesarean, censoring, planned care, and predictors measured before prediction. Include all eligible ongoing pregnancies at each landmark. Completed-week aggregates alone cannot identify day-level predictive hazards.
3. **Model the event process jointly.** Estimate cause-specific hazards with cumulative-incidence conversion or an appropriate competing-risk model. For real-world probability, induction and prelabour caesarean are competing pathways, not disappearances. A hypothetical no-intervention question requires explicit causal assumptions and cannot be recovered by routine censoring alone.
4. **Respect development uncertainty.** Pre-specify transformations and predictor handling; account for correlated pregnancies/mothers, hospitals and countries. Handle missingness explicitly. Use penalisation and bootstrap/internal-external validation rather than arbitrary coefficient shrinkage. Do not let post-outcome information leak into predictors.
5. **Externally validate in the intended population and era.** Use independent hospitals/countries and later calendar periods, including Germany if claiming applicability there. Evaluate calibration-in-the-large, calibration slope and curves at each horizon, Brier score and discrimination with competing-risk methods. Assess important subgroups, missing-data patterns and changing induction policies. AUC alone is insufficient. See [van Geloven et al., BMJ 2022](https://doi.org/10.1136/bmj-2021-069249) for competing-risk validation methodology.
6. **Separate software tests from scientific performance.** Test date boundaries, units, conservation, nondecreasing cumulative incidence, incompatible inputs and model versioning. Separately report external prediction errors and uncertainty against a simple prespecified baseline. A release must not call invariant tests “validated accuracy”.
7. **Review intended use and communications.** Obtain independent obstetric and statistical review before clinical claims. State source population, endpoint, horizon, calibration limits and treatment-policy dependence. Predictions must never recommend postponing assessment, monitoring or advised intervention. Establish drift monitoring and a rollback rule before wider use.

Until these requirements are met, the defensible product is an explicitly labelled population reference with a due-date-aligned calendar, not a personalised medical probability calculator.

## Implementation verification (18 October 2026)

The active Free, Pro, saved-account and email renderers now use the same versioned reference. Regression tests cover independently transcribed weekly totals, preservation of earlier and later mass, every displayed day's Free/Pro equality, invariance to optional personal inputs, invalid civil dates and window boundaries. Email chart rendering now supports all 35 days from 37+0 through 41+6.

All 51 local automated suites passed, including account, backup, privacy and publication-boundary checks. The public build completed. Local browser checks in German and English confirmed matching values, local number formatting and unavailable-day handling; no production visits or email sends were used. These are software checks, not clinical validation or an independent expert review.

The revised explanations are complete in German and English. The main source-and-limitations notice is also translated for the other 11 supported languages; some newly introduced detailed copy still uses the existing English fallback and requires translation review before a multilingual release. No release or production deployment is part of this verification.