Pith. sign in

REVIEW 3 major objections 6 minor 40 references

Uncertainty-Aware Missing-Data Multimodal Latent for Fetal-Growth Analysis

T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper shows that the four blocks of routine third-trimester fetal measurements are close to mutually uninformative, and that this one property determines what a latent representation can impute, audit, and reveal beyond fetal size.

desk verdict Useful empirical result on block independence and a validated error screen, but the central claim rests on an untested ignorability assumption that needs sensitivity analysis. read the letter →

arxiv 2608.07590 v1 pith:RSUUW5OB submitted 2026-08-05 stat.AP q-bio.QM

Tiago Cortinhal , César Díaz-Parga , Gabriel Bernardino , Marta Nuñez-Garcia This is my paper · ORCID
classification stat.APq-bio.QM MSC 62H2562P10
keywords multimodalfusionmissingmodalitieslatent-variablemodeluncertaintyquantificationfetalgrowthrestrictiondataqualityfactorproductofexperts
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that the four blocks of routine third-trimester fetal measurements — biometry, maternal characteristics, Doppler, and fetal cardiac — are near-perpendicular in information terms: predicting any one block from the other three explains just 2.3% of its variance. It shows that this near-independence is the organising fact: it sets an upper bound on what imputation of a missing panel can supply, defines what a reconstruction-based audit can catch, and explains why fetal size alone cannot separate small fetuses with adverse outcomes from constitutionally small ones. A linear-Gaussian factor model with eight factors, fitted to 25 measurements from 977 fetuses, produces this picture by marginalizing missing panels rather than imputing them, and the same reconstruction residual flags 36 confirmed transcription errors among 38 flagged records. If the claim is right, routine incomplete data can be used not only to place a fetus on a continuous growth spectrum but also to say how much a missing panel leaves unknown.

What carries the argument

The carrying mechanism is a product-of-experts linear-Gaussian factor model: each fetus's 25 measurements are generated by $K$ Gaussian latent factors, and the posterior precision for the latent scores is the sum of one term per observed measurement, so an absent block contributes nothing and is marginalized rather than imputed (Eq. 2). An orthogonal rotation gives the axes clinical names — size, head, haemodynamic redistribution, maternal body mass, stature, and three cardiac axes. The same model supplies the standardized reconstruction residual (Eq. 3), whose large values flag records inconsistent with the fitted measurement structure.

What would settle it

Take a cohort with near-complete Doppler acquisition and refit the same model; if the cross-block $R^2_{cv}$ and the factor loadings change materially, or if fetuses with absent panels show systematically different outcomes conditional on observed measurements, the ignorability assumption fails and the near-independence and coverage claims are artifacts of selective acquisition. A simpler version: audit the records with incomplete Doppler panels and test whether their observed measurements or outcomes differ from fetuses with complete panels.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that biometry, maternal, Doppler, and cardiac measurements are close to mutually uninformative (out-of-fold $R^2_{cv} = 0.023$), and that this one property determines the capacities of any representation built from them. The fitted latent is a continuous growth spectrum with no cluster structure (dip $p = 0.99$, gap statistic $k = 1$, silhouette $0.07$), ordered by birthweight centile (Spearman $\rho = 0.55$), with SGA and LGA at opposite ends. Among SGA fetuses, a Doppler-dominated haemodynamic redistribution axis separates the 25 adverse outcomes (AUC 0.70, 0.585–0.808) where measured EFW does not (AUC 0.60, 0.451–0.738). With Doppler censored, marginalized intervals cover 97% and 93% of held-out values at nominal 95% and 90%, while the reconstruction residual doubles as a data-quality screen that identifies within-block transcription errors; the synthetic benchmark locates appreciable cross-block detection only above $R^2_{cv} \approx 0.13$.

Load-bearing premise

The whole analysis stands on the assumption that why a Doppler or cardiac panel was not acquired can be left out of the model — that the missingness carries no information about the fetus's condition — even though the paper notes acquisition follows clinical discretion, and this assumption is never tested with a sensitivity analysis.

Editorial extensions

If this is right

  • Doppler and cardiac panels that were never acquired cannot be reliably reconstructed: with cross-block $R^2_{cv} = 0.023$, imputation would place a fetus on the basis of almost no information, so any downstream analysis should treat absent panels as absent, not fill them in.
  • The marginalized representation gives an honest per-fetus uncertainty: interval coverage stayed at 97% and 93% against nominal 95% and 90% when Doppler was censored, while the posterior widened only along axes the missing block would have constrained.
  • Among small fetuses, the unsupervised latent recovers the clinically meaningful signal: the redistribution axis separated adverse outcomes where measured size did not, both overall and in term deliveries.
  • The reconstruction residual is a practical registry-audit tool: 36 of 38 flagged records were confirmed transcription errors, all within a measurement block.
  • Cross-block error detection is currently out of reach in this cohort; the synthetic benchmark sets the threshold at $R^2_{cv} \approx 0.13$ before such errors become detectable.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the near-independence result is a property of this cohort's acquisition protocol, not of fetal physiology; a protocol that acquires Doppler routinely could raise cross-block coupling and make imputation and cross-block error detection viable where this paper shows they are not.
  • Editorial inference: the residual screen's false-negative rate was not measured on real data (unflagged records were not audited); a follow-up with full audit of a random sample would quantify sensitivity in practice and is a direct test of the screening claim.
  • Editorial inference: since posterior width depends only on which blocks were seen, the same formalism could be used prospectively to decide which additional panel would most reduce a fetus's uncertainty, turning missingness itself into a triage signal.
  • Editorial inference: the continuous-spectrum finding does not rule out discrete FGR phenotypes; it only says routine third-trimester tabular measurements do not resolve them, and a cohort with earlier-onset disease, maternal and cord-blood metabolomics, or imaging might.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper fits a linear-Gaussian factor model (K=8 via parallel analysis) to 25 measurements in four blocks (biometry, maternal, Doppler, cardiac) from 977 third-trimester fetuses, marginalizing missing measurements rather than imputing them. It reports that the blocks are nearly mutually uninformative (out-of-fold R2_cv = 0.023), that the latent space is a continuous growth spectrum with no cluster structure, that a Doppler-dominated axis separates SGA fetuses with adverse outcomes (AUC 0.70), that censoring the Doppler block yields 97%/93% held-out coverage against nominal 95%/90%, and that a reconstruction residual screen flagged 38 records, of which 36 were confirmed transcription errors. A synthetic benchmark with injected cross-block errors maps when cross-block detection becomes feasible.

Significance. If the central claim holds, the empirical finding that the four measurement blocks are nearly mutually uninformative has concrete consequences for imputation and for data-quality auditing in fetal-growth records. The paper's strengths include a genuinely external held-out coverage validation of the posterior uncertainty, a striking residual-screen result (36/38 confirmed against an independent registry), reproducible code for the synthetic benchmark, and a generally transparent limitations section. The main weakness is that the block-independence estimate and the R2_cv = 0.023 quantity rest on an untested ignorability assumption that the paper itself calls into question in Sect. 3.1; the synthetic threshold R2_cv ≈ 0.13 is also conditional on the generative model class. These issues are load-bearing for the paper's central thesis, but they appear addressable with additional sensitivity analyses.

major comments (3)
  1. [Sect. 3.1 and Eq. (2)] The paper explicitly states in Sect. 3.1 that missingness 'follows a clinical decision rather than at random,' yet Eq. (2) marginalizes missing measurements using a full-information maximum likelihood estimator whose validity requires ignorability (MAR with distinct parameters). Because Doppler and cardiac panels are acquired at clinician discretion, plausibly in response to suspected pathology, the observed cells may not be representative of the complete-data distribution. This could bias the fitted loadings, the block-independence estimate (R2_cv = 0.023), and the coverage intervals. The manuscript needs a sensitivity analysis—for example, a pattern-mixture model, a selection model, or a comparison with complete-case and inverse-probability-weighted estimates—to show that the block-independence conclusion is not an artifact of MNAR. Until this is provided, the load-bearing premise of the paper remains unidentified.
  2. [Sect. 4.5] The report of 'out-of-fold R2_cv = 0.023 (0.020–0.026)' does not specify the predictor. The reader needs to know whether this is the factor-model posterior predictive, a linear regression, or another model; how folds are constructed; whether the model is refit per fold; and whether the R2 is computed per block or pooled. Without this specification, the number cannot be reproduced or its fairness as a measure of cross-block information assessed.
  3. [Sect. 4.6 and Table 1] The synthetic benchmark generates data from the same linear-Gaussian model class that is used for detection, so the 'appreciable above R2_cv ≈ 0.13' threshold is conditional on that model class and on the specific scheme for injecting errors. The manuscript should state this limitation directly in the main text and temper the conclusion that this gives the coupling 'required' for cross-block detection, because real data may have nonlinear dependences and error distributions different from the synthetic panel.
minor comments (6)
  1. [Sect. 3.2] In Eq. (2), the notation E[z_n] is used for the posterior mean but is never defined; please define it explicitly as the posterior mean under the product-of-experts model.
  2. [Sect. 3.2] The sentence 'the model is as following which we fit by expectation-maximization:' is ungrammatical and should be revised.
  3. [Sect. 4.5] Please clarify the order of operations: was the model used for the residual screen fitted on the original (uncorrected) records or on the corrected set? The statement 'The residual analysis itself is computed on the original records' does not say which fitted model is used.
  4. [Table 1 caption] The caption says thresholds are set at a nominal 5% false-positive rate, but the achieved false-positive rates range from 0.048 to 0.066; please clarify whether thresholds are calibrated separately for each number of observed measurements and report the achieved FPR for each cell.
  5. [Abstract and Sect. 4.3] The abstract reports the AUC 0.70 for adverse outcome among SGA fetuses with 25 events; please state in the abstract that this is based on only 25 events to avoid overstating precision.
  6. [Sect. 3.1] The statement about clinical-discretion missingness cites Little and Rubin [7]; consider also citing a clinical source documenting Doppler and cardiac ultrasound acquisition patterns in routine third-trimester care.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are validated against held-out data and external registry audits.

full rationale

The paper's derivation chain is self-contained. The factor model is fitted to the cohort, and its key outputs are checked independently: out-of-fold R2_cv = 0.023 is a cross-validated prediction score; the coverage claims are tested on Doppler blocks censored in five folds and compared against multiple imputation; the reconstruction-residual screen is validated against an external source registry (36 of 38 flagged records confirmed as transcription errors); and the SGA adverse-outcome separation is evaluated with permutation tests and confidence intervals. The posterior-width property (Eq. 2) is a definitional consequence of product-of-experts marginalization, but the paper does not present it as an empirical discovery; it separately validates the resulting intervals against held-out measurements, which is independent support. The synthetic benchmark generates data from the same linear-Gaussian model class as the detector, which makes the R2_cv ≈ 0.13 threshold a simulation-based operating characteristic rather than an externally established constant; this is a mild self-referentiality in the audit-capability extrapolation, but it is not a reduction of a prediction to its inputs by construction, and the paper explicitly positions it as a controlled sensitivity analysis. There are no load-bearing self-citations and no fitted parameter is renamed as a prediction. The untested ignorability assumption is a correctness/validity concern, not a circularity, and does not affect this score.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central results rest on the factor model's fit parameters (K, W, mu, psi), on the ignorability of missingness, and on the correctness of the model class for both the cohort and the synthetic benchmark. No new physical or clinical entity is postulated; the latent factors are statistical constructs.

free parameters (4)
  • Latent dimensionality K = 8
    Selected by Horn's parallel analysis on this cohort; every downstream result uses K = 8, and the choice itself is data-driven.
  • Factor loadings W, means mu, noise variances psi = EM estimates on the cohort
    These fitted model parameters define posterior means, posterior covariances, and reconstruction residuals; every reported result depends on them.
  • False-discovery-rate threshold for residual screen = 0.05
    Used to flag 38 records; changing it changes the screen's sensitivity and the 36/38 confirmation count.
  • Cross-block coupling in synthetic benchmark = R2_cv from 0.00 to 0.66
    Varied as a design parameter in the simulation, not fitted to the cohort; it defines the x-axis of the benchmark.
assumptions (6)
  • domain assumption Missingness is ignorable for likelihood-based estimation (MAR/MCAR); acquisition decisions do not depend on unobserved fetal state beyond observed measurements.
    FIML/EM in Eq. 2 conditions only on observed measurements and treats missingness as ignorable; Sect. 3.1 states acquisition is at clinical discretion, which can violate MAR.
  • domain assumption The linear-Gaussian factor model with diagonal noise correctly captures the joint distribution of the 25 measurements.
    Eq. 1 assumes Gaussian factors and diagonal noise; if the true joint is non-Gaussian or the noise correlated, loadings and intervals shift.
  • domain assumption The 25 measurements are selected a priori to represent the four clinical domains and are not outcome-informed.
    Sect. 3.1 says variables were chosen to represent domains, not outcomes, to avoid label leakage; this cannot be verified from the paper.
  • standard math Horn's parallel analysis provides a valid K for this missing-data setting.
    Horn's method is applied to incomplete data; the paper does not describe how missingness was handled in the eigenvalue comparison.
  • ad hoc to paper The synthetic benchmark's generating model matches the fitted model class, so its coupling thresholds are conditional on that model class.
    Synthetic data are generated from the same linear-Gaussian model as the fitted model, so the R2_cv approximately 0.13 threshold may not transfer if real data deviate.
  • standard math Standard asymptotic or permutation validity of the dip test, gap statistic, and AUC confidence intervals.
    Used to conclude no cluster structure and to bound AUCs; the paper does not discuss power or assumptions of these tests at n = 977.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Uncertainty-Aware Missing-Data Multimodal Latent for Fetal-Growth Analysis." pith.science (2026). https://pith.science/paper/RSUUW5OB

@misc{pith2026260807590,
  author       = {Pith},
  title        = {Pith review of: Uncertainty-Aware Missing-Data Multimodal Latent for Fetal-Growth Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RSUUW5OB}},
  note         = {Machine review of arXiv:2608.07590}
}
read the original abstract

Objective: Routine third-trimester examination yields fetal biometry, maternal, Doppler and fetal-cardiac measurements, acquired at clinical discretion and therefore often incomplete. We show that these four measurement blocks are close to mutually uninformative, and that this single property determines what a representation of them can impute, what it can audit, and what fetal size alone cannot indicate. Methods: A linear-Gaussian factor model (K = 8 by parallel analysis, VARIMAX-rotated) was fitted to 25 measurements in four blocks from 977 fetuses (169 SGA, 61 severe; 77 LGA). Posterior precision sums contributions from observed measurements only, so missing values are marginalized rather than imputed. Data quality was screened using the standardized residual between each measurement and its reconstruction. Results: Predicting any one block from the other three gives an out-of-fold R2 of 0.023. The representation is a continuous growth spectrum with no cluster structure (Hartigan dip p = 0.99, gap statistic k = 1, three-cluster silhouette 0.07) ordering fetuses by birthweight centile (Spearman rho = 0.55). Among 169 SGA fetuses the haemodynamic redistribution axis separated the 25 adverse outcomes (AUC 0.70, 0.585-0.808) where measured size did not (0.60, 0.451-0.738). With Doppler censored, the marginalized interval covered held-out measurements in 97% and 93% of cases against nominal 95% and 90%. The reconstruction residual flagged 38 of 977 records, 36 confirmed transcription errors in the registry. Conclusion: Marginalizing missing measurements yields a representation whose uncertainty reflects the available data, and whose reconstruction residual doubles as a data-quality screen. Because the blocks are nearly independent, confirmed flags are within-block errors, and a synthetic benchmark gives the coupling needed before cross-block detection becomes available.

Figures

Figures reproduced from arXiv: 2608.07590 by the authors.

Figure 1
Figure 1. Method overview, in three stages. (a) Probabilistic generative model. Two fetuses enter the factor model with different blocks available; the dashed empty boxes are blocks that were never acquired (cardiac for the first fetus, Doppler for the second). Each observed block contributes one precision term to the product-of-experts posterior (Eq. 2); an absent block contributes no term and drops out of the sum, and nothi… view at source ↗
Figure 2
Figure 2. VARIMAX loadings of the K = 8 factor latent. Rows are the factors L1–L8, each named by its dominant block: fetal size (L1), head size (L2), haemodynamic redistribution from Doppler (L3), maternal body mass (L4), maternal stature (L5), and three cardiac axes (L6– L8). Color denotes the VARIMAX-rotated loading, with intensity reflecting the strength of the association between each measurement and latent factor. 4 Expe… view at source ↗
Figure 3
Figure 3. Birthweight-defined growth groups in the latent space, on the axis carrying the largest [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Modality-conditional uncertainty on the fetal size (L1) and redistribution (L3) axes. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 40 canonical work pages

  1. [1]

    Ultrasound Obstet

    Gordijn, S.J., Beune, I.M., Thilaganathan, B., et al.: Consensus definition of fetal growth restriction: a Delphi procedure. Ultrasound Obstet. Gynecol.48(3), 333–339 (2016)

    work page 2016

  2. [2]

    Psychometrika47(1), 69–76 (1982)

    Rubin, D.B., Thayer, D.T.: EM algorithms for ML factor analysis. Psychometrika47(1), 69–76 (1982)

    work page 1982

  3. [3]

    Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the EM algorithm. J. R. Stat. Soc. B39(1), 1–38 (1977)

    work page 1977

  4. [4]

    Argelaguet, R., Velten, B., Arnol, D., et al.: Multi-Omics Factor Analysis—a framework for unsupervised integration of multi-omics data sets. Mol. Syst. Biol.14(6), e8124 (2018)

    work page 2018

  5. [5]

    IEEE Trans

    Klami, A., Virtanen, S., Leppäaho, E., Kaski, S.: Group factor analysis. IEEE Trans. Neural Netw. Learn. Syst.26(9), 2136–2147 (2015)

    work page 2015

  6. [6]

    van Buuren, S., Groothuis-Oudshoorn, K.: mice: Multivariate Imputation by Chained Equations in R. J. Stat. Softw.45(3), 1–67 (2011)

    work page 2011

  7. [7]

    Little, R.J.A., Rubin, D.B.: Statistical Analysis with Missing Data. 3rd edn. Wiley (2019) 14

    work page 2019

  8. [8]

    In: MICCAI 2025

    Gu, Y., Saito, K., Ma, J.: Learning contrastive multimodal fusion with improved modality dropout for disease detection and prediction. In: MICCAI 2025. LNCS, vol. 15974, pp. 280–290. Springer (2025)

    work page 2025

Show all 40 references
  1. [9]

    In: MICCAI 2024 Workshops

    Chaptoukaev, H., Marcianó, V., Galati, F., Zuluaga, M.A.: HyperMM: robust multimodal learning with varying-sized inputs. In: MICCAI 2024 Workshops. LNCS, vol. 15401, pp. 170–183. Springer (2024)

    2024

  2. [10]

    In: MICCAI 2025

    Scholz, D., Erdur, A.C., Ehm, V., et al.: MM-DINOv2: adapting foundation models for multi-modal medical image analysis. In: MICCAI 2025. LNCS, vol. 15967, pp. 320–330. Springer (2025)

    2025

  3. [11]

    In: Machine Learning for Healthcare (MLHC)

    Al Jorf, B., Shamout, F.E.: MedPatch: confidence-guided multi-stage fusion for multimodal clinical data. In: Machine Learning for Healthcare (MLHC). PMLR, vol. 298 (2025)

    2025

  4. [12]

    iScience26(9), 107620 (2023)

    Miranda, J., Paules, C., Noell, G., et al.: Similarity network fusion to identify phenotypes of small-for-gestational-age fetuses. iScience26(9), 107620 (2023)

    2023

  5. [13]

    Schouten, D., Nicoletti, G., Dille, B., et al.: Navigating the landscape of multimodal AI in medicine: a scoping review. Med. Image Anal.105, 103621 (2025)

    2025

  6. [14]

    npj Digit

    Mikolaj, K.W., Christensen, A.N., Taksoe-Vester, C.A., et al.: Predicting abnormal fetal growth using deep learning. npj Digit. Med.8, 318 (2025)

    2025

  7. [15]

    Lancet384(9946), 869–879 (2014)

    Papageorghiou, A.T., Ohuma, E.O., Altman, D.G., et al.: International standards for fe- tal growth based on serial ultrasound measurements: the INTERGROWTH-21st Project. Lancet384(9946), 869–879 (2014)

    2014

  8. [16]

    Jiao, J., Zhou, J., Li, X., et al.: USFM: a universal ultrasound foundation model. Med. Image Anal.96, 103202 (2024)

    2024

  9. [17]

    npj Digit

    Maani, F., Saeed, N., Saleem, T., et al.: FetalCLIP: a visual-language foundation model for fetal ultrasound image analysis. npj Digit. Med. (2026)

    2026

  10. [18]

    In: NeurIPS 31, pp

    Wu, M., Goodman, N.: Multimodal generative models for scalable weakly-supervised learn- ing. In: NeurIPS 31, pp. 5575–5585 (2018)

    2018

  11. [19]

    In: MICCAI 2025

    Ambsdorf, J., Munk, A., Llambias, S., et al.: General methods make great domain-specific foundation models: a case-study on fetal ultrasound. In: MICCAI 2025. LNCS. Springer (2025)

    2025

  12. [20]

    Silva, P.I.P., Perez, M.: Prenatal ultrasound diagnosis of biometric changes in the brain of growth restricted fetuses: a systematic review of literature. Rev. Bras. Ginecol. Obstet. 43(7), 545–559 (2021)

    2021

  13. [21]

    Biometrics65(4), 1233–1242 (2009)

    Slaughter, J.C., Herring, A.H., Thorp, J.M.: A Bayesian latent variable mixture model for longitudinal fetal growth. Biometrics65(4), 1233–1242 (2009)

    2009

  14. [22]

    In: AAAI 2024, vol

    Yao, W., Yin, K., Cheung, W.K., Liu, J., Qin, J.: DrFuse: learning disentangled represen- tation for clinical multi-modal fusion with missing modality and modal inconsistency. In: AAAI 2024, vol. 38, pp. 16416–16424 (2024)

    2024

  15. [23]

    DeVore, G.R.: The importance of the cerebroplacental ratio in the evaluation of fetal well- being in SGA and AGA fetuses. Am. J. Obstet. Gynecol.213(1), 5–15 (2015)

    2015

  16. [24]

    Circulation121(22), 2427–2436 (2010) 15

    Crispi, F., Bijnens, B., Figueras, F., et al.: Fetal growth restriction results in remodeled and less efficient hearts in children. Circulation121(22), 2427–2436 (2010) 15

    2010

  17. [25]

    Psychometrika 30(2), 179–185 (1965)

    Horn, J.L.: A rationale and test for the number of factors in factor analysis. Psychometrika 30(2), 179–185 (1965)

    1965

  18. [26]

    Psychometrika 23(3), 187–200 (1958)

    Kaiser, H.F.: The varimax criterion for analytic rotation in factor analysis. Psychometrika 23(3), 187–200 (1958)

    1958

  19. [27]

    Hartigan, J.A., Hartigan, P.M.: The dip test of unimodality. Ann. Statist.13(1), 70–84 (1985)

    1985

  20. [28]

    Tibshirani, R., Walther, G., Hastie, T.: Estimating the number of clusters in a data set via the gap statistic. J. R. Stat. Soc. B63(2), 411–423 (2001)

    2001

  21. [29]

    In: ICLR (2021)

    Han, Z., Zhang, C., Fu, H., Zhou, J.T.: Trusted multi-view classification. In: ICLR (2021)

    2021

  22. [30]

    In: NeurIPS 31 (2018)

    Sensoy, M., Kaplan, L., Kandemir, M.: Evidential deep learning to quantify classification uncertainty. In: NeurIPS 31 (2018)

    2018

  23. [31]

    Angelopoulos, A.N., Bates, S.: A gentle introduction to conformal prediction and distribution-free uncertainty quantification. Found. Trends Mach. Learn.16(4), 494–591 (2023)

    2023

  24. [32]

    In: NeurIPS 30 (2017)

    Geifman, Y., El-Yaniv, R.: Selective classification for deep neural networks. In: NeurIPS 30 (2017)

    2017

  25. [33]

    Sonography2(2), 27–31 (2015)

    Quinton, A.E., Cook, C.M., Peek, M.J.: The prediction of the small for gestational age fetus with the head circumference to abdominal circumference (HC/AC) ratio: a new look at an old measurement. Sonography2(2), 27–31 (2015)

    2015

  26. [34]

    Lancet339(8788), 283–287 (1992)

    Gardosi, J., Chang, A., Kalyan, B., Sahota, D., Symonds, E.M.: Customised antenatal growth charts. Lancet339(8788), 283–287 (1992)

    1992

  27. [35]

    Early Hum

    Wills, A.K., Chinchwadkar, M.C., Joglekar, C.V., et al.: Maternal and paternal height and BMI and patterns of fetal growth: the Pune Maternal Nutrition Study. Early Hum. Dev. 86(9), 535–540 (2010)

    2010

  28. [36]

    Ultrasound Obstet

    Cruz-Lemini, M., Crispi, F., Valenzuela-Alcaraz, B., et al.: Value of annular M-mode dis- placement vs tissue Doppler velocities to assess cardiac function in intrauterine growth restriction. Ultrasound Obstet. Gynecol.42(2), 175–181 (2013)

    2013

  29. [37]

    Diagnostics14(5), 548 (2024)

    Domínguez-Gallardo, C., Ginjaume-García, N., Ullmo, J., et al.: Fetal left ventricle function evaluated by two-dimensional speckle-tracking echocardiography across clinical stages of severity in growth-restricted fetuses. Diagnostics14(5), 548 (2024)

    2024

  30. [38]

    Transactions on Machine Learning Research (2026)

    Wu, R., Wang, H., Chen, H.-T., Carneiro, G.: Deep multimodal learning with missing modality: a survey. Transactions on Machine Learning Research (2026)

    2026

  31. [39]

    Lancet401(10389), 1692–1706 (2023)

    Ashorn, P., Ashorn, U., Muthiani, Y., et al.: Small vulnerable newborns—big potential for impact. Lancet401(10389), 1692–1706 (2023)

    2023

  32. [40]

    Fetal Diagn

    Figueras, F., Gratacós, E.: Update on the diagnosis and classification of fetal growth re- striction and proposal of a stage-based management protocol. Fetal Diagn. Ther.36(2), 86–98 (2014) 16

    2014

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.