Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

A Data-Centric Approach to Detecting and Mitigating Demographic Bias in Pediatric Mental Health Text: A Case Study in Anxiety Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Removing the 20% of sentences with the lowest TF-IDF scores from pediatric mental-health training notes reduces the model's anxiety false-negative gap between female and male patients by 27% while preserving accuracy.

desk verdict The bias-detection result is worth taking seriously, but the mitigation claim overstates what a per-bin read of Table 1 actually shows. read the letter →

arxiv 2501.00129 v1 pith:JI72PNWR submitted 2024-12-30 cs.CL cs.AI

Julia Ive , Paulina Bondaronek , Vishal Yadav , Daniel Santel , Tracy Glauser , Tina Cheng , Jeffrey R. Strawn , Greeshma Agasthya
show 5 more authors
Jordan Tschida Sanghyun Choo Mayanka Chandrashekar Anuj J. Kapadia John Pestian
This is my paper · ORCID
classification cs.CLcs.AI
keywords biasmitigationclinicalNLPpediatricmentalhealthanxietydetectionelectronicrecordsTF-IDFfilteringgenderfairness
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

This paper sets out to show that an AI model screening pediatric electronic health records for anxiety systematically under-diagnoses female adolescents, and that a simple data-side fix can close a large part of that gap. The authors fine-tune a clinical transformer on notes from about 73,000 pediatric patients and find that female patients get a 4% lower accuracy and a 9% higher false-negative rate than male patients, while male notes run about 500 words longer. They trace the disparity to differences in information density and word choice rather than to the clinical content itself, and they show that deleting the 20% of sentences with the lowest average word-level TF-IDF scores from the training data reduces the false-negative gap by 27% on average with no meaningful drop in accuracy. The upshot is a cheap, model-agnostic debiasing recipe for clinical text: cut the low-information filler, keep the salient clinical terms, and the model's hesitation about female patients drops sharply.

What carries the argument

The load-bearing mechanism is information-density filtering (tf-idf_filt): each sentence in a patient's concatenated notes receives an importance score equal to the average word-level TF-IDF of its words, and the 20% of sentences with the lowest scores are deleted from the training data before the model is fine-tuned. The companion mechanism, gender-word substitution (gen_sub), detects proper names with Stanza and replaces them with generic identifiers such as 'person1', and maps gender-specific pronouns to 'they'. Filtering is the active ingredient in reducing the false-negative gap; substitution mainly helps by lowering the model's reliance on gender-linked words once the filler is gone.

What would settle it

A chart review of the sentences removed by tf-idf_filt: if a substantial share of them contain anxiety-specific symptom statements (e.g., 'worries about tests', 'avoids birthday parties'), then the apparent debiasing is partly signal loss. A second concrete check is training the same model on male notes truncated to the average female note length without TF-IDF selection; if the FNR gap closes as much as it does under tf-idf_filt, the operative cause is note length, not sentence informativeness.

Watch

Extended reading notes

Core claim

The central claim is that the anxiety-detection model's sex bias is a textual-bias artifact, not a reflection of the underlying clinical signal. Female patients are under-diagnosed (average false-negative gap of +0.09 over males, up to +0.13 in the youngest and oldest age bins), and this tracks quantifiable differences in the notes: male notes are on average ~540 words longer, and male/female vocabularies show low Jaccard similarity (0.54 overall, 0.34 for medical terms). The paper's proposed remedy, tf-idf_filt, removes the 20% of sentences with the lowest average word-level TF-IDF scores from the training notes; this cuts the average FNR gap by 0.024 points (27%), with the largest drops in bins 5 and 15, while preserving accuracy. Random sentence removal does not replicate the effect, which the authors take as evidence that the bias is carried by low-information filler and gender-linked wording rather than by clinical content. The gender-word substitution method (gen_sub), applied alone, does not reduce the gap, but combined with filtering it further stabilizes the model and reduces reliance on gendered names and pronouns.

Load-bearing premise

The method assumes the 20% of sentences with the lowest average word-level TF-IDF scores are non-clinical filler, so dropping them does not remove anxiety-relevant symptom information, and that the shorter, differently worded notes of female patients reflect documentation style rather than a real difference in their clinical presentation.

Editorial extensions

If this is right

  • Retraining on tf-idf-filtered notes cuts the average false-negative gap between female and male patients by 27% (from 0.09 to about 0.066) while overall accuracy stays essentially unchanged.
  • The filter reduces the model's higher rate of uncertain predictions for female patients by about half; adding gender-word substitution eliminates that uncertainty gap entirely.
  • The combined filter-plus-substitution also lowers the false-negative gap for non-white race subgroups by 0.034 on average, suggesting the method is not limited to sex bias.
  • Debiased models rely less on gendered names and pronouns and more on contextual words such as 'complaint' and 'presents', so the intervention changes the evidence the model uses, not just its error rates.
  • Because random sentence removal does not reproduce the improvement, the effect is tied to dropping low-information sentences specifically, not to shortening the notes.

Reading between the lines

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

  • A targeted chart review of the sentences removed by tf-idf_filt would test whether any anxiety-relevant symptom language is being discarded; if it is, part of the reported 27% 'bias reduction' is actually signal loss.
  • The same thresholding logic could be tried on other under-documented groups where note length differs by documentation style, but the empirically chosen 20% cutoff would need to be re-tuned per setting and per task.
  • The paper uses equality of false-negative rates as its fairness target; under alternative definitions such as calibration or positive predictive value parity, the same filter could produce different apparent effects.
  • A complementary length-matching experiment—truncating male notes to female average length without TF-IDF selection—would isolate whether the operative variable is note length or sentence informativeness.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper studies gender bias in a Clinical-BigBird model fine-tuned to detect pediatric anxiety from electronic health record notes at Cincinnati Children's Hospital Medical Center. The authors report that the model systematically under-diagnoses female patients, with an average false-negative rate (FNR) 0.09 higher for females than males across five age bins, along with lower accuracy and more uncertain predictions. They analyze textual differences between male and female notes, such as note length and vocabulary overlap, and propose two data-centric debiasing methods: filtering sentences by TF-IDF informativeness and replacing gender-biased words with neutral tokens. The central claim is that removing the 20% of sentences with the lowest TF-IDF scores reduces the female-male FNR gap by 27% on average while maintaining overall accuracy.

Significance. The bias-detection result is a meaningful empirical contribution: the finding that female FNR exceeds male FNR consistently across all five age bins is a clear, out-of-sample observation on a large real-world EHR dataset, and the paper's analysis of information-density and vocabulary differences provides a plausible mechanism. The proposed data-centric debiasing approach is interpretable and tailored to clinical text, which is a useful addition to the fairness-in-NLP toolkit. However, the mitigation claim is currently overstated: the reported average gap reduction is not accompanied by a per-group demonstration that the under-diagnosed group actually benefits, no uncertainty quantification is provided, and the key threshold is described only as 'empirically defined' without a validation procedure. The study's code and data are not shared, which limits reproducibility.

major comments (4)
  1. [Results, Objective 3; Table 1; Discussion] The claim that tf-idf_filt is a 'clear winner' that reduces bias 'particularly benefitting the non-privileged subgroup (females)' is not supported by the per-group FNR values. The average gap reduction is dominated by Bin 5, where male FNR worsens from 0.24 to 0.34 while female FNR changes only from 0.37 to 0.36; the gap widens in Bin 10 (from 0.08 to 0.14) and Bin 12 (from 0.08 to 0.09). Please report female and male FNR changes separately across all bins and reframe the conclusion accordingly, or restrict the claim to the bins where female FNR actually improves.
  2. [Methods, Text De-biasing Methods; Results, Objective 3] The 20% TF-IDF retention threshold is described only as 'defined empirically' with no validation procedure. If the threshold was chosen based on the test set, the reported 27% reduction is an in-sample estimate and could result from selection overfitting. Please specify the validation method (e.g., a separate validation split or cross-validation) and show results for a range of thresholds (e.g., 10%, 20%, 30%) to demonstrate that the conclusion is not threshold-specific.
  3. [Discussion, Limitations] The method assumes that the lowest-TF-IDF sentences are non-clinical filler whose removal does not discard anxiety-relevant information. The paper does not test this assumption; the Discussion itself acknowledges that note-length differences may reflect 'symptomatic variations and social circumstances.' If the removed sentences contain clinically meaningful content, the observed 'bias reduction' is partly a loss of signal. Please provide evidence (e.g., manual annotation of removed sentences or an analysis of model performance on a held-out set with and without filtering) that the retained content preserves anxiety-relevant clinical information.
  4. [Tables 1, 2, and 6] All results are from single model runs without error bars, confidence intervals, or significance tests. Given the small average FNR-gap reduction (0.024) and test-set sizes on the order of 850–1,280 patients, the central mitigation finding may be within noise. Please provide uncertainty quantification, such as bootstrap confidence intervals on the FNR gap or multiple fine-tuning runs with different random seeds.
minor comments (5)
  1. [Equation (1)] The balanced error rate formula is typeset ambiguously; it should clearly read BER = 0.5 * (FP/(FP+TN) + FN/(FN+TP)).
  2. [Abstract and Conclusion] The abstract says 'reduced diagnostic bias by up to 27%' while the Results report an average reduction of 27%; the Conclusion states 'up to one-third,' which is inconsistent with the 27% figure and with the per-bin reductions. Please align these numbers.
  3. [Table 5] The row 'Av model length' is not defined in the text; please clarify whether this is the average number of tokens after truncation to the model's input limit or something else.
  4. [Methods, Age binning] The text says 'see the Age Binning subsection below' within the Age Binning subsection itself; correct this self-referential cross-reference.
  5. [Discussion] The paper states that 'notes for males are 500 words longer' on average, but Table 5 shows differences ranging from 27 to 755 words across bins, with an average of about 540; 'roughly 540 words longer' would be more accurate.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the parity and debiasing results are out-of-sample empirical findings, and the only threshold-selection caveat is a reproducibility concern rather than a demonstrated reduction to model inputs.

full rationale

I walked the claimed derivation chain. The bias-detection result (Objective 1) is an empirical parity measurement on held-out test data: Clinical-BigBird models are fine-tuned without any fairness constraint, and the reported 9% higher FNR for female patients is not constructed by the evaluation metric. The text-distribution analysis (Objective 2) is descriptive and does not feed back into the model. For the debiasing method (Objective 3), the paper states: 'Each time we modified the training data, we re-trained our models to obtain new models, which were then tested on the original test data.' The TF-IDF filtering rule removes low-salience sentences; it does not encode the FNR gap or the female/male label, and the Table 1 results show the gap widens in Bins 10 and 12, so the average 27% reduction is an empirical outcome rather than a mathematical identity. The gender-substitution result (reduced reliance on gendered words) is expected, but it is measured via LIME on original test text, so it is not a tautology. There are no load-bearing self-citations: the pre-trained Clinical-BigBird model and LIME are external tools. The one circularity-adjacent concern is the phrase in Objective 3, 'threshold defined empirically' for the 20% sentence-removal cutoff. No validation procedure is described, so one cannot rule out that the cutoff was chosen by inspecting FNR gaps; if so, part of the reported reduction would be a fitted quantity. However, the paper does not state that the threshold was selected on the test FNR gap, and the per-bin behavior is inconsistent with a selection that forces improvement. That is a correctness and reproducibility limitation, not a demonstrated circular reduction. The skeptic's observation that the average reduction is driven by degrading male FNR and that the gap widens in some bins is a substantive fairness critique, but it does not establish circularity. Accordingly, the derivation is self-contained and the score is 0.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claims rest on an empirical cohort definition, a matched-control design, and a debiasing procedure whose main free parameter (the 20% sentence-removal threshold) is set empirically without a validation protocol. There are no invented theoretical entities; the 'person1' placeholders are part of the preprocessing method, not new ontological commitments.

free parameters (4)
  • tf-idf sentence retention threshold = remove 20% of sentences
    Threshold set 'empirically' with no validation split; directly controls how much text is discarded by the main debiasing method.
  • number of most recent notes per patient = 25
    Truncation of patient history to the most recent 25 notes; chosen as the average minimum count across bins.
  • input sequence length = 1000 tokens
    Chosen because longer inputs 'did not result in further improvement'; modulates how much context the model sees.
  • age bin boundaries = 5, 8, 10, 12, 15
    Bins are deliberately non-standard and produce widely varying sex ratios (36% to 69% female); the FNR gaps and debiasing effects differ across bins.
assumptions (3)
  • domain assumption ICD code list in Appendix Table 7 correctly defines anxiety cases and the absence of these codes defines controls.
    The cohort definition relies entirely on billing codes; misclassification would affect both the bias and mitigation results.
  • domain assumption Age and sex matching removes confounding so residual textual differences between male and female notes reflect reporting practices rather than true differences in disease presentation or severity.
    The paper matches cases and controls 1:1 by age and sex and then attributes FNR gaps to non-biological differences; if symptom distribution differs by sex, the gap is partly genuine clinical signal.
  • ad hoc to paper Low-TF-IDF sentences are non-clinical filler whose removal does not discard anxiety-relevant information.
    The tf-idf_filt method assumes sentence importance by TF-IDF equals clinical relevance for anxiety; the paper does not validate that removed sentences lack diagnostic content.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Data-Centric Approach to Detecting and Mitigating Demographic Bias in Pediatric Mental Health Text: A Case Study in Anxiety Detection." pith.science (2026). https://pith.science/paper/JI72PNWR

@misc{pith2026250100129,
  author       = {Pith},
  title        = {Pith review of: A Data-Centric Approach to Detecting and Mitigating Demographic Bias in Pediatric Mental Health Text: A Case Study in Anxiety Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JI72PNWR}},
  note         = {Machine review of arXiv:2501.00129}
}
read the original abstract

Introduction: Healthcare AI models often inherit biases from their training data. While efforts have primarily targeted bias in structured data, mental health heavily depends on unstructured data. This study aims to detect and mitigate linguistic differences related to non-biological differences in the training data of AI models designed to assist in pediatric mental health screening. Our objectives are: (1) to assess the presence of bias by evaluating outcome parity across sex subgroups, (2) to identify bias sources through textual distribution analysis, and (3) to develop a de-biasing method for mental health text data. Methods: We examined classification parity across demographic groups and assessed how gendered language influences model predictions. A data-centric de-biasing method was applied, focusing on neutralizing biased terms while retaining salient clinical information. This methodology was tested on a model for automatic anxiety detection in pediatric patients. Results: Our findings revealed a systematic under-diagnosis of female adolescent patients, with a 4% lower accuracy and a 9% higher False Negative Rate (FNR) compared to male patients, likely due to disparities in information density and linguistic differences in patient notes. Notes for male patients were on average 500 words longer, and linguistic similarity metrics indicated distinct word distributions between genders. Implementing our de-biasing approach reduced diagnostic bias by up to 27%, demonstrating its effectiveness in enhancing equity across demographic groups. Discussion: We developed a data-centric de-biasing framework to address gender-based content disparities within clinical text. By neutralizing biased language and enhancing focus on clinically essential information, our approach demonstrates an effective strategy for mitigating bias in AI healthcare models trained on text.

Figures

Figures reproduced from arXiv: 2501.00129 by the authors.

Figure 1
Figure 1. Types of predictive bias and their origin. We consider biases of four types9: (a) selection bias in the training or testing data that are not representative; (b) label bias originating from biased human annotations; (c) textual bias originating from differences in word distributions; (d) nally, over-amplication bias present in statistical models which amplify the discrepancies from the training data. Our framework… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Mental Health Disorder Detection Beyond Social Media: A Systematic Review of Available Datasets

    cs.CL 2026-07 accept novelty 6.5 of 10

    The first PRISMA review of non-social-media free-text mental-health datasets shows English/depression dominance and large gaps in diversity, labeling standards, and public clinical data.

Reference graph

Works this paper leans on

43 extracted references · 24 canonical work pages · cited by 1 Pith paper

  1. [1]

    Racine, N.et al.Global prevalence of depressive and anxiety symptoms in children and adolescents during covid-19: A meta-analysis.JAMA Pediatr.175, 1142–1150, DOI: 10.1001/JAMAPEDIA TRICS.2021.2482(2021)

    work page arXiv 2021

  2. [2]

    for Psychol

    A call to action on assessing and mitigating bias in articial intelligence applications for mental health.Perspectives on psychological science : a journal Assoc. for Psychol. Sci.18, 1062, DOI: 10.1177/17456916221134490 (2022)

    doi:10.1177/17456916221134490 2022

  3. [3]

    Behrens, B., Swetlitz, C., Pine, D. S. & Pagliaccio, D. The screen for child anxiety related emotional disorders (scared): Informant discrepancy, measurement invariance, and test-retest reliability.Child psychiatry human development50, 473–482, DOI: 10.1007/S10578-018-0854-0 (2019)

    work page doi:10.1007/s10578-018-0854-0 2019

  4. [4]

    Golden, G.et al.Applying articial intelligence to clinical decision support in mental health: What have we learned? (2023)

    work page 2023

  5. [5]

    psychiatry14, DOI: 10.1038/S41398-024-02970-4 (2024)

    Perlman, K.et al.Development of a differential treatment selection model for depression on consolidated and transformed clinical trial datasets.Transl. psychiatry14, DOI: 10.1038/S41398-024-02970-4 (2024)

    work page doi:10.1038/s41398-024-02970-4 2024

  6. [6]

    M., Ji, S

    Zhang, T., Schoene, A. M., Ji, S. & Ananiadou, S. Natural language processing applied to mental illness detection: a narrative review.npj Digit. Medicine 2022 5:15, 1–13, DOI: 10.1038/s41746-022-00589-7 (2022)

    doi:10.1038/s41746-022-00589-7 2022

  7. [7]

    A., Lester, J

    Omiye, J. A., Lester, J. C., Spichak, S., Rotemberg, V . & Daneshjou, R. Large language models propagate race-based medicine.npj Digit. Medicine 2023 6:16, 1–4, DOI: 10.1038/s41746-023-00939-z (2023)

    doi:10.1038/s41746-023-00939-z 2023

  8. [8]

    Y ., Johnson, E., Simon, G

    Coley, R. Y ., Johnson, E., Simon, G. E., Cruz, M. & Shortreed, S. M. Racial/ethnic disparities in the performance of prediction models for death by suicide after mental health visits.JAMA Psychiatry78, 726–734, DOI: 10.1001/ JAMAPSYCHIA TRY .2021.0493(2021)

    work page arXiv 2021

Show all 43 references
  1. [9]

    S., Schwartz, H

    Shah, D. S., Schwartz, H. A. & Hovy, D. Predictive biases in natural language processing models: A conceptual framework and overview. InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 5248–5264, DOI: 10.18653/v1/2020.acl-main.468 (Assoc...

    doi

  2. [10]

    DOI: 10.13026/61xq-mj56 (2021)

    Logé, C.et al.Q-pain: A question answering dataset to measure social bias in pain management. DOI: 10.13026/61xq-mj56 (2021)

    2021 doi

  3. [11]

    & Chang, K.-W

    Zhao, J., Wang, T., Y atskar, M., Ordonez, V . & Chang, K.-W. Men also like shopping: Reducing gender bias amplication using corpus-level constraints. In Palmer, M., Hwa, R. & Riedel, S. (eds.)Proceedings of the 2017 Conference on Empirical Methods in Natural Language Process...

    2017 doi

  4. [12]

    Angwin, L. J. M. S., J. & Kirchner, L. Machine bias: There’s software used across the country to predict future criminals. and it’s biased against blacks. | benton institute for broadband society. (2016). 13.Chouldechova, A. & Roth, A. The frontiers of fairness in machine lear...

    2016

  5. [15]

    & Haas, C

    Caton, S. & Haas, C. Fairness in machine learning: A survey.ACM Comput. Surv.56, 1–38, DOI: 10.1145/3616865/ SUPPL_FILE/3616865-SUPP .PDF(2024)

    2024

  6. [16]

    & Kalai, A

    Bolukbasi, T., Chang, K.-W., Zou, J., Saligrama, V . & Kalai, A. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. InNeurIPS(2016)

    2016

  7. [17]

    K., Delobelle, P ., Berendt, B

    Tokpo, E. K., Delobelle, P ., Berendt, B. & Calders, T. How far can it go? on intrinsic gender bias mitigation for text classication. In Vlachos, A. & Augenstein, I. (eds.)Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguist...

    doi

  8. [18]

    J., Bashir, S

    Raza, S., Garg, M., Reji, D. J., Bashir, S. R. & Ding, C. Nbias: A natural language processing framework for bias identication in text.Expert. Syst. with Appl.237, 121542, DOI: 10.1016/J.ESW A.2023.121542(2024)

    2024

  9. [19]

    & Sugiyama, M

    Fang, T., Lu, N., Niu, G. & Sugiyama, M. Rethinking importance weighting for deep learning under distribution shift. (2020). 12/15

    2020

  10. [20]

    & Zettlemoyer, L

    Clark, C., Y atskar, M. & Zettlemoyer, L. Don’t take the easy way out: Ensemble based methods for avoiding known dataset biases. In Inui, K., Jiang, J., Ng, V . & Wan, X. (eds.)Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th I...

    2019 doi

  11. [21]

    & Lusa, L

    Blagus, R. & Lusa, L. Smote for high-dimensional class-imbalanced data.BMC Bioinforma.14, 1–16, DOI: 10.1186/ 1471-2105-14-106/FIGURES/7 (2013)

    2013

  12. [22]

    He, H., Bai, Y ., Garcia, E. A. & Li, S. Adasyn: Adaptive synthetic sampling approach for imbalanced learning.Proc. Int. Jt. Conf. on Neural Networks1322–1328, DOI: 10.1109/IJCNN.2008.4633969 (2008)

    2008

  13. [23]

    P .et al.Towards debiasing sentence representations

    Liang, P . P .et al.Towards debiasing sentence representations. In Jurafsky, D., Chai, J., Schluter, N. & Tetreault, J. (eds.)Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 5502–5515, DOI: 10.18653/v1/2020.acl-main.488 (Association for...

    doi

  14. [24]

    & Chi, E

    Beutel, A., Chen, J., Zhao, Z. & Chi, E. H. Data decisions and theoretical implications when adversarially learning fair representations. (2017)

    2017

  15. [25]

    & Cohn, T

    Li, Y ., Baldwin, T. & Cohn, T. Towards robust and privacy-preserving text representations. In Gurevych, I. & Miyao, Y . (eds.)Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (V olume 2: Short Papers), 25–30, DOI: 10.18653/v1/P18-2005 (A...

    doi

  16. [26]

    Woodworth, B., Gunasekar, S., Ohannessian, M. I. & Srebro, N. Learning non-discriminatory predictors. In Kale, S. & Shamir, O. (eds.)Proceedings of the 2017 Conference on Learning Theory, vol. 65 ofProceedings of Machine Learning Research, 1920–1953 (PMLR, 2017). 27.Berk, R.et...

    2017

  17. [28]

    Medicine 2024 30:430, 1174–1190, DOI: 10.1038/s41591-024-02885-z (2024)

    V aidya, A.et al.Demographic bias in misdiagnosis by computational pathology models.Nat. Medicine 2024 30:430, 1174–1190, DOI: 10.1038/s41591-024-02885-z (2024). 29.Guo, C., Pleiss, G., Sun, Y . & Weinberger, K. Q. On calibration of modern neural networks (2017)

    2024 doi

  18. [30]

    Kessler, R. C. Epidemiology of women and depression.J. Affect. Disord.74, 5–13, DOI: 10.1016/S0165-0327(02)00426-3 (2003)

    2003 doi

  19. [31]

    K.et al.Antidepressant prescribing by pediatricians: A mixed-methods analysis.Curr

    Tulisiak, A. K.et al.Antidepressant prescribing by pediatricians: A mixed-methods analysis.Curr. Probl. Pediatr. Adolesc. Heal. Care47, 15–24, DOI: 10.1016/J.CPPEDS.2016.11.009 (2017)

    2017 doi

  20. [32]

    R., Lu, L., Peris, T

    Strawn, J. R., Lu, L., Peris, T. S., Levine, A. & Walkup, J. T. Research review: Pediatric anxiety disorders - what have we learnt in the last 10 years?J. child psychology psychiatry, allied disciplines62, 114–139, DOI: 10.1111/JCPP .13262 (2021)

    2021 doi

  21. [33]

    O’Reilly Media, Inc

    Bird, S., Klein, E. & Loper, E.Natural language processing with Python: analyzing text with the natural language toolkit (" O’Reilly Media, Inc.", 2009). 34.Pedregosa, F.et al.Scikit-learn: Machine Learning in Python.J. Mach. Learn. Res.12, 2825–2830 (2011)

    2011

  22. [35]

    M., Ahmad, F

    Li, Y ., Wehbe, R. M., Ahmad, F. S., Wang, H. & Luo, Y . Clinical-longformer and clinical-bigbird: Transformers for long clinical sequences.arXiv preprint arXiv:2201.11838(2022)

    2022 arXiv

  23. [36]

    Wolf, T.et al.Transformers: State-of-the-art natural language processing. InProceedings of the 2020 Conference on Em- pirical Methods in Natural Language Processing: System Demonstrations, 38–45, DOI: 10.18653/v1/2020.emnlp-demos.6 (Association for Computational Linguistics, O...

    2020 doi

  24. [37]

    & Hutter, F

    Loshchilov, I. & Hutter, F. Decoupled weight decay regularization.7th Int. Conf. on Learn. Represent. ICLR 2019(2017)

    2017

  25. [38]

    why should i trust you?

    Ribeiro, M. T., Singh, S. & Guestrin, C. "why should i trust you?": Explaining the predictions of any classier. 1135–1144, DOI: 10.1145/2939672.2939778 (Association for Computing Machinery, 2016)

    2016

  26. [39]

    & Mohammad, S

    Kiritchenko, S. & Mohammad, S. Examining gender and race bias in two hundred sentiment analysis systems. In Nissim, M., Berant, J. & Lenci, A. (eds.)Proceedings of the Seventh Joint Conference on Lexical and Computational Semantics, 43–53, DOI: 10.18653/v1/S18-2005 (Associatio...

    doi

  27. [40]

    & Manning, C

    Qi, P ., Zhang, Y ., Zhang, Y ., Bolton, J. & Manning, C. D. Stanza: A Python natural language processing toolkit for many human languages. InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations(2020)

    2020

  28. [41]

    & V enkatasubramanian, S

    Feldman, M., Friedler, S., Moeller, J., Scheidegger, C. & V enkatasubramanian, S. Certifying and removing disparate impact.Proc. 21st ACM SIGKDD Int. Conf. on Knowl. Discov. Data Min.(2015). 13/15

    2015

  29. [42]

    Fair prediction with disparate impact: A study of bias in recidivism prediction instruments.Big data5, 153–163, DOI: 10.1089/BIG.2016.0047 (2017)

    Chouldechova, A. Fair prediction with disparate impact: A study of bias in recidivism prediction instruments.Big data5, 153–163, DOI: 10.1089/BIG.2016.0047 (2017)

    2017

  30. [43]

    Kurita, K., Vyas, N., Pareek, A., Black, A. W. & Tsvetkov, Y . Measuring bias in contextualized word representations. 166–172, DOI: 10.18653/v1/w19-3823 (2019)

    2019 doi

  31. [44]

    Zhang, Y ., Zhang, Y ., Qi, P ., Manning, C. D. & Langlotz, C. P . Biomedical and clinical English model packages for the Stanza Python NLP library.J. Am. Med. Informatics Assoc.(2021)

    2021

  32. [45]

    & Mullainathan, S

    Obermeyer, Z., Powers, B., V ogeli, C. & Mullainathan, S. Dissecting racial bias in an algorithm used to manage the health of populations.Science366, 447–453, DOI: 10.1126/SCIENCE.AAX2342/SUPPL_FILE/AAX2342_OBERMEYER_SM. PDF (2019)

    2019 doi

  33. [46]

    Medicine 2024 7:1 7, 1–14, DOI: 10.1038/s41746-023-00970-0 (2024)

    Large language models to identify social determinants of health in electronic health records.npj Digit. Medicine 2024 7:1 7, 1–14, DOI: 10.1038/s41746-023-00970-0 (2024)

    2024 doi

  34. [47]

    & Schaar, M

    Seedat, N., Imrie, F. & Schaar, M. V . D. Navigating data-centric articial intelligence with dc-check: Advances, challenges, and opportunities.IEEE Transactions on Artif. Intell.5, 2589–2603, DOI: 10.1109/TAI.2023.3345805 (2024)

    2024

  35. [48]

    & Ash, E

    Li, N., Goel, N. & Ash, E. Data-centric factors in algorithmic fairness.AIES 2022 - Proc. 2022 AAAI/ACM Conf. on AI, Ethics, Soc.396–410, DOI: 10.1145/3514094.3534147/SUPPL_FILE/DA TA_CENTRIC_FACTORS.MP4(2022). 14/15 A Anxiety Diagnosis codes V ocabulary Code Description ICD9C...

    2022

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.