Anal sphincter injury at childbirth: a machine-learning impedance test validated on 70 women, 81 without the reference exam

A team from Warsaw University of Technology and OASIS Diagnostics, the company that sells the device, evaluates a machine-learning-assisted impedance spectroscopy tool meant to flag an anal sphincter injury in the hours after childbirth. Among 70 analysable women, the locked model finds 86.2% of injuries confirmed by endoanal ultrasound and wrongly flags 7.3% of women without injury. The result is encouraging for an adjunct tool, but 81 of the 151 women enrolled did not get the reference exam, the study comes from a single centre, several authors are tied to the company, and neither code nor data are public.

Context

Obstetric anal sphincter injuries (OASI) are third- and fourth-degree tears that occur during vaginal delivery. They are not always visible on the immediate clinical exam, yet an unrecognised injury exposes the woman to lasting anal incontinence. The reference exam is three-dimensional endoanal ultrasound (EAUS), done with a probe placed in the anal canal by an experienced operator. It is not available everywhere, nor right after birth.

The device studied, ONIRY (CE-marked, class IIa), offers an alternative: electrical impedance spectroscopy. A single-use 12 mm probe with eight electrodes sends a weak current (under 1 mA, between 1 and 100 kHz) and measures how the tissue resists it. Intact and torn tissue do not conduct the same way. A machine-learning model interprets the measurements and returns a binary verdict, PASS or REFER (send for further assessment). This paper is a preprint posted to medRxiv on 11 October 2026; it has not been peer reviewed.

Method

The work proceeds in three phases, reported against the STARD-AI and TRIPOD+AI checklists (reporting guidelines for diagnostic accuracy studies and prediction models involving AI).

Phase 1, development of version V1. One hundred and fifty-two women, examined up to eight weeks after vaginal delivery, had two measurements and a reference EAUS: 30 without injury, 61 with a first- or second-degree tear, 61 with a third- or fourth-degree tear. After excluding 19 runs in which the two measurements differed by more than 1 kΩ at 1 kHz, 279 measurements remain (117 with OASI, 162 without). The model is a neural network evaluated by 10×10-fold cross-validation (the dataset is split into ten parts, ten times over, with each participant never appearing in both training and test).

Phase 2, real-world adaptation (V2). Post-market surveillance of V1 flagged 15 discordant results (6 false positives, 9 false negatives), mostly within the first 24 hours and linked to marked oedema, biological fluids or recent sutures. The authors therefore retrained the model on 144 early recordings from seven centres, with a reduced feature set (impedance magnitude, resistance, reactance, phase shift, local variance), a 90/10 split between training and internal validation, and an ensemble of two multilayer perceptrons (small stacked neural networks) combined with a logical AND rule. The threshold was chosen to maximise balanced accuracy while keeping sensitivity at or above 80%. The exact architecture is proprietary. V2 was then tested on 125 recordings from two centres, with the "totality of clinical evidence" as reference, since EAUS was not systematically available.

Phase 3, prospective external validation. In a Warsaw hospital, women aged 18 or over with at least one risk factor (first birth, instrumental delivery, birthweight of 4,000 g or more, gestation of 41 weeks or more) or clinical doubt were examined with the locked V2 model. Of 151 consecutive enrolments, 81 had no EAUS; 70 are analysed. Midwives performed 70% of the exams. The reference is 3D EAUS, done by a single expert with more than 25 years of experience, a median of 28 days after delivery, not blinded to clinical or impedance findings.

Results

ML metrics. Sensitivity is the share of real injuries the test detects; specificity, the share of uninjured women it correctly leaves alone. In phase 1, V1 reaches 90.0% accuracy, sensitivity and specificity (internal cross-validation). In phase 2, V2 on 125 recordings: 28 true positives, 4 false negatives, 84 true negatives, 9 false positives, i.e. 89.6% accuracy, 87.5% sensitivity, 90.3% specificity. In phase 3, among 70 women of whom 29 (41.4%) have confirmed OASI: accuracy 90.0% (63/70; 95% CI 80.8–95.1), sensitivity 86.2% (25/29; CI 69.4–94.5), specificity 92.7% (38/41; CI 80.6–97.5). The 4 false negatives and 3 false positives follow from these counts; from them we derive a positive predictive value of about 89% and a negative one of about 90% (not read from the preprint's table; check against the PDF).

Clinical translation. Applied to a population like the study's (41% injuries), these rates give, per 1,000 women examined, about 357 injuries found, 57 missed, 543 uninjured women correctly reassured and 43 false alarms. But this prevalence belongs to a sample enriched in women at risk. To illustrate the trap (our calculation, not the study's), at 5% prevalence the same rates would give, per 1,000 women, about 43 injuries detected, 7 missed and 69 false alarms: a REFER result would then be right only 38% of the time. A test's value depends on the population in which it is used.

What is good

A model locked before validation. V2 was frozen before phase 3, and recruitment is described as consecutive: this avoids tuning the model to the test data and picking easy cases.

An honest analysis of V1's errors. The authors do not hide that the first model mostly erred within 24 hours of birth, and they link these failures to oedema, fluids and sutures. They draw a concrete adaptation from it: retraining on early data, a low-impedance alert, automated quality control of measurements.

A demanding reporting framework. Reporting against STARD-AI and TRIPOD+AI, giving Wilson confidence intervals, and stating that phase 3 had ethics approval and phase 1 is registered (NCT04903977): the paper is more transparent than many preprints in the field.

What is less good

Partial verification bias. Of 151 women enrolled, 81 (54%) did not get the EAUS and are excluded from the analysis. If those who came back for the exam differed from the rest (for example because an abnormality worried them), the measured performance does not transfer as is. The authors acknowledge this without being able to quantify it.

An imperfect reference and a narrow sample. EAUS took place a median of 28 days after birth (range 1 to 40), whereas impedance testing took place within 4 hours: anatomy may have changed in between (healing, oedema resolving). A single expert, not blinded, read the images. And with 29 injuries, the confidence interval for sensitivity runs from 69.4% to 94.5%: it cannot be ruled out that it is closer to 70% than to 90%. In phase 2C, the reference is an overall clinical judgement, not EAUS, which makes those figures more fragile.

Conflicts of interest and no reproducibility. Four authors are or were employees of OASIS Diagnostics, two are founders, four are shareholders, and one is an inventor on the related patent. The adaptation and external validation phases received no outside funding, and neither data, code nor model are public, for commercial confidentiality reasons. Nobody can therefore check the work, and the part of the paper we read reports no numerical comparison with clinical examination alone, which is the alternative to beat. Among the classic AI-health failure modes, this leaves a missing comparator, a prevalence-sensitive metric, and a risk of population bias (one maternity unit, a cohort enriched in at-risk women).

What it changes

For the research community. The paper shows a rare good practice: re-adapting a commercial model after real-world errors, then freezing it and validating it at a site not used for training. The logical next step is a multicentre study with systematic EAUS follow-up (or random sampling of the unverified), several blinded readers, and a direct comparison with clinical examination alone.

For clinicians. On current evidence, this test looks like an add-on to draw attention to women who should be referred for imaging, not a way to rule an injury out. It does not change practice until the test is validated in less selected populations, and it does not replace the clinical exam or EAUS.

For patients and the public. Earlier detection of deep tears could shorten the time to repair or appropriate care. That promise is not yet demonstrated: the paper measures no benefit on continence or quality of life.

Going further

Preprint: Młyńczak M., Korzeniewski K., Rosoł M., Uchman-Musielak M., Barcz E., Grzybowska M. E., Sudoł-Szopińska I., Kropshofer S., Spinelli A., Borycka K., "Real-World Adaptation and External Validation of Machine-Learning Impedance Spectroscopy for Immediate Postpartum Detection of Obstetric Anal Sphincter Injuries", medRxiv, doi: 10.64898/2026.10.08.26365032. Affiliations: Warsaw University of Technology, Holy Family Specialist Hospital (Warsaw), Medical University of Gdańsk, Humanitas University, Innsbruck Medical University, among others. Phase 1 funding: Polish National Centre for Research and Development (POIR.01.01.01-00-0726/18). Phase 1 trial: NCT04903977. Code, data and model: not public.

Editorial transparency note: this article was produced by Tatakoto's automated editorial pipeline from the full preprint; Table 1 was not readable in the extraction, so PPV, NPV and likelihood ratios are our own calculations from the published counts. The English, Spanish and Chinese versions are AI-generated. Spotted an error? Contact us.