Diagnostic Accuracy of Artificial Intelligence in Laryngeal Disorders: An Integrative Review
Open accessThe DOI matches its Crossref record; 83 of the 84 references it cites were found in a registry or an index; 50 of the 184 sources its claims cite were checked for support.
Abstract
Background/Objectives: Laryngeal disorders are among the most prevalent conditions in otolaryngology, yet they remain challenging to diagnose without specialized expertise. Artificial intelligence (AI) systems leveraging machine learning (ML) and deep learning (DL) have demonstrated promising performance for the automatic detection and classification of voice disorders and laryngeal lesions. Methods: This review synthesizes findings from 88 studies published between 2015 and 2025 on AI-based laryngeal disorder detection, considering physioacoustic mechanisms, databases and acquisition protocols, AI architectures and validation strategies, and diagnostic performance. Results: The current literature supports high internal accuracies for binary healthy versus pathological detection (88–99%); meanwhile, performance decreases for higher-level tasks such as pathophysiological category classification and identification, particularly under external validation. From a clinical perspective, clinicians do not infer specific diagnoses from isolated acoustic parameters such as percent jitter or shimmer. Instead, they rely on how these perturbation patterns dynamically evolve during connected speech, where alterations guide perceptual differentiation between underlying disorders. Recurrent sources of bias include dependence on a limited number of historical vowel-based databases, class and demographic imbalance, and limited ecological validity of recording protocols. Additional concerns involve the predominant use of internal cross-validation and insufficient reproducibility or code sharing. Conclusions: Drawing on the literature, an integrative three-level clinical recognition framework is proposed, delineating realistic use cases for AI as a decision-support tool rather than an autonomous diagnostic system. Key priorities for future personalized medicine and research are also identified, including diversified multi-center datasets, standardized methodological reporting, rigorous external validation, and compliance with regulatory and ethical requirements for medical AI deployment.
Overview
CoChat Summary
TL;DR
This integrative review of 88 studies concludes that AI for laryngeal disorders performs well for binary healthy-vs-pathological screening under controlled conditions, but performance drops substantially for finer-grained diagnosis and external validation. The paper argues that current systems are best used as decision-support tools, not replacements for laryngoscopy, imaging, or histopathology.
Key findings
- Across the 88 included studies, internal accuracies for binary healthy versus pathological detection were commonly 88–99%; the paper also notes many internal accuracies exceed 90%.
- For more difficult tasks, reported performance was lower: 70–90% for broad pathophysiological category recognition and rarely above 75% for specific pathology identification.
- External validation exposed substantial drops: in some studies, accuracy and sensitivity decreased by about 10–30 percentage points; the review also states a 12–20% performance gap in external testing.
- Only 7 of 88 studies performed both internal and external validation, and fewer than 15% provided open-source code or full model architecture.
Methods
- Critical narrative/integrative review, not a systematic review; search was structured with predefined eligibility criteria and multi-step screening.
- Searched PubMed, Cochrane Library, and Scopus for studies from 2015–2025; 88 papers were included.
- Included original human studies using AI/ML/DL for recognition, detection, classification, or severity assessment of laryngeal disorders; methodological quality was assessed qualitatively rather than with a standardized risk-of-bias tool.
Limitations
- The paper states that the literature is highly heterogeneous in populations, disorders, databases, voice-quality assessment methods, AI architectures, and outcome measures, which is why it used a narrative rather than systematic review approach.
- The review itself notes recurrent evidence problems in the underlying literature: reliance on restricted historical databases, class imbalance, sustained-vowel tasks, internal-only validation, and limited reporting beyond global accuracy.
- Evident from the text: because the search was limited to 2015–2025 and three databases, it may miss older or non-indexed studies; this is an inference from the methods, not explicitly stated.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Figures



Verification
Claims the paper makes
29 Valid · 11 Not valid · 32 Not checked · 31 Own results and interpretations
At least one cited source doesn’t support this claim or points to a different paper, and none confirmed it — see the verdicts below.
In the paper
“In their comparative analysis, ResNet-based encoders offered only incremental gains in multiclass scenarios [48,62].”
§ Diagnostic Accuracy of Artificial Intelligence in Laryngeal Disorders: An Integrative Review > 3. Results and Discussion > 3.2. Panorama of AI Approaches
What the sources say
- Not checked — no full text of the sourcecites [48] Lee J.B. Lee H.G. Quantitative analysis of automatic voice disorder detection studies for hybrid feature and classifier selection Biomed. Signal Process. Control 2024 91 106014 10.1016/j.bspc.2024.106014 source
- Disputedcites [62] Bur A.M. Zhang T. Chen X. Kavookjian H. Kraft S. Karadaghy O. Farrokhian N. Mussatto C. Penn J. Wang G. Interpretable Computer Vision to Detect and Classify Structural Laryngeal Lesions in Digital Flexible Laryngoscopic Images Otolaryngol. Head Neck Surg. 2023 169 1564 1572 10.1002/ohn.411 37350279 PMC12374229 sourceExperimental — accuracy not yet measured
The Experimental Settings state that every detection model used the same ResNet-50 backbone, so the study did not compare ResNet-based encoders to assess incremental gains. The Conclusions describe classification as benign or suspicious for malignancy, not a multiclass scenario. No page numbers are available in our copy. (LLM, conf=0.83)
Checked by a reasoning model — may be wrong.
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Naranjo et al. confirmed that a compact set mixing perturbation, cepstral peak prominence, glottal noise excitation and non-linear dynamics reached 89% cross-validated accuracy for Reinke’s oedema [41].”
§ Diagnostic Accuracy of Artificial Intelligence in Laryngeal Disorders: An Integrative Review > 3. Results and Discussion > 3.1. Physioacoustic Foundations of Automated Detection
What the sources say
- Partly supportedcites [41] Naranjo L. Pérez C.J. Campos-Roca Y. Madruga M. Replication-based regularization approaches to diagnose Reinke’s edema by using voice recordings Artif. Intell. Med. 2021 120 102162 10.1016/j.artmed.2021.102162 34629154 sourceExperimental — accuracy not yet measured
Section 4.2.1 reports 0.88933 cross-validated accuracy for the selected-feature Ridge model. But the selected-feature list (discussion following Table 7) contains CPP, four MFCCs, SHANNON and RPDE—not jitter, shimmer or glottal-to-noise excitation (GNE). Those measures were considered in the initial feature set (Section 2.4), but the claimed mix did not achieve the reported 89% result. (LLM, conf=0.93)
boostedstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“Diagnostic performance ranging from 90% to 99% has been reported across acoustic and laryngeal image-based tasks [26,30,31,32,33,34].”
§ Diagnostic Accuracy of Artificial Intelligence in Laryngeal Disorders: An Integrative Review > 1. Introduction Nearly one-third of the adult population will experience a voice disorder during their lifetime [1], with dysphonia prevalence reaching 7% in the general population [2] and up to 50% among voice professionals [3], resulting in significant quality of life impairment and socioeconomic burden [4,5]. The clinical spectrum includes benign structural lesions (e.g., vocal fold nodules, polyps, and cysts [6,7]), neuromuscular disorders (e.g., vocal fold paralysis, spasmodic dysphonia [8,9]), inflammatory conditions, and malignant lesions [10,11]. While some conditions present with distinctive morphological patterns, most are associated with non-specific voice quality (VQ) modifications that remain difficult to attribute to a specific diagnosis through perceptual evaluation alone [10,12,13,14]. Indeed, perceptual VQ assessment is subjective, with significant inter- and intra-rater variability and disorder identification rates around 60–70% [15], further influenced by auditory fatigue, clinical experience, and cognitive biases [13]. Laryngeal imaging—including videolaryngoscopy, stroboscopy, and high-speed videoendoscopy—represents the reference standard for laryngeal evaluation [9,16,17,18]; however, access depends on practitioner expertise, equipment availability, and cost, while neuromuscular disorders may present without visible structural lesions, complicating diagnosis [19]. The European Consensus guidelines recommend multidimensional VQ assessment combining laryngostroboscopy with perceptual, patient-reported, and acoustic measures [5]. In this context, artificial intelligence (AI)-based acoustic analysis offers a non-invasive adjunctive approach for VQ diagnosis, triage, and follow-up [20,21,22,23,24]. Current AI architectures include classical machine learning (ML) models—such as support vector machines (SVMs), random forests (RFs), and Gaussian mixture models (GMMs) operating on hand-crafted acoustic features [6,7,25]—deep learning (DL) networks processing spectrograms or raw waveforms [3,26,27], and self-supervised encoders such as wav2vec 2.0 or HuBERT pre-trained on large unlabeled speech corpora [23,28,29]. Diagnostic performance ranging from 90% to 99% has been reported across acoustic and laryngeal image-based tasks [26,30,31,32,33,34]. This integrative review aimed to summarize AI system performance for laryngeal disorder detection and classification, while examining methodological biases limiting clinical translatability, including physioacoustic foundations, acquisition protocols, validation strategies, and regulatory and ethical implications.
What the sources say
- Not checked — no full text of the sourcecites [26] Fang S.H. Tsao Y. Hsiao M.J. Chen J.Y. Lai Y.H. Lin F.C. Wang C.T. Detection of Pathological Voice Using Cepstrum Vectors: A Deep Learning Approach J. Voice 2018 33 634 641 10.1016/j.jvoice.2018.02.003 29567049 source
- Not checked — no full text of the sourcecites [30] Mahmood S.A. Multi-Dimensional Features Extraction for Voice Pathology Detection Based on Deep Learning Methods J. Voice 2024, in press 10.1016/j.jvoice.2024.12.048 39894721 source
- Not matchedcites [31] Achtuth Rao M.V. Yamini B.K. Ketan J. Preetie Shetty A. Pal P.K. Shivashankar N. Ghosh P.K. Automatic Classification of Healthy Subjects and Patients With Essential Vocal Tremor Using Probabilistic Source-Filter Model Based Noise Robust Pitch Estimation J. Voice 2021 37 314 321 10.1016/j.jvoice.2021.01.009 33579623
We searched by its title and found no matching paper.
- Not checked — no full text of the sourcecites [32] Islam R. Abdel-Raheem E. Tarique M. Voice pathology detection using convolutional neural networks with electroglottographic (EGG) and speech signals Comput. Methods Programs Biomed. Update 2022 2 100074 10.1016/j.cmpbup.2022.100074 PMC9404230 36039092 source
- Supportedcites [33] Barlow J. Sragi Z. Rivera-Rivera G. Al-Awady A. Daşdöğen Ü. Courey M.S. Kirke D.N. The Use of Deep Learning Software in the Detection of Voice Disorders: A Systematic Review Otolaryngol. Head Neck Surg. 2024 170 1531 1543 10.1002/ohn.636 38168017 source
Table 3 reports acoustic-model accuracies within the stated range, including 92.0%, 98.6%, and 99.0%. The Video Group reports a 91.3% macroaverage accuracy for videolaryngoscopy image models, supporting the claim that performance in this range has been reported across both task types. The paper also reports results outside that range; the cited sentence does not say the range covers every model. (LLM, conf=0.89)
- Not checked — no full text of the sourcecites [34] Tessler I. Primov-Fever A. Soffer S. Anteby R. Gecel N.A. Livneh N. Alon E.E. Zimlichman E. Klang E. Deep learning in voice analysis for diagnosing vocal cord disorders: A systematic review Eur. Arch. Oto-Rhino-Laryngol. 2024 281 863 871 10.1007/s00405-023-08362-6 38091100 source
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“Several pathophysiologically distinct disorders produce overlapping acoustic signatures, leading to inherent diagnostic ambiguity within the vocal signal itself [6,7,19,42].”
§ Diagnostic Accuracy of Artificial Intelligence in Laryngeal Disorders: An Integrative Review > 3. Results and Discussion > 3.1. Physioacoustic Foundations of Automated Detection
What the sources say
- Not checked — no full text of the sourcecites [6] Akbari A. Arjmandi M.K. Employing linear prediction residual signal of wavelet sub-bands in automatic detection of laryngeal pathology Biomed. Signal Process. Control 2015 18 293 302 10.1016/j.bspc.2015.02.008 source
- Not matchedcites [7] Al-Nasheri A. Muhammad G. Alsulaiman M. Ali Z. Investigation of Voice Pathology Detection and Classification on Different Frequency Regions Using Correlation Functions J. Voice 2016 31 3 15 10.1016/j.jvoice.2016.01.014 26992554
We searched by its title and found no matching paper.
- Not checked — no full text of the sourcecites [19] Aichinger P. Schoentgen J. Detection of Diplophonation in Audio Recordings of German Standard Text Readings J. Voice 2018 33 949.e1–949.e10 10.1016/j.jvoice.2018.06.009 30089537 source
- Supportedcites [42] Hu H.C. Chang S.Y. Wang C.H. Li K.J. Cho H.Y. Chen Y.T. Lu C.J. Tsai T.P. Lee O.K.S. Deep Learning Application for Vocal Fold Disease Prediction Through Voice Recognition: Preliminary Development Study J. Med. Internet Res. 2021 23 e25247 10.2196/25247 34100770 PMC8241431 source
The Principal Findings section says organic lesions, unilateral vocal paralysis, and vocal atrophy can produce the same weak, breathy sound and diplophonia despite different vibration patterns. It also reports difficulty distinguishing these conditions by voice, supporting the claim of diagnostic ambiguity in the vocal signal. (LLM, conf=0.90)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“these signatures remain physiologically non-specific, being shared across nodules, polyps, Reinke’s oedema, reflux-related changes and hyperfunctional dysphonia [39,40,41].”
§ Diagnostic Accuracy of Artificial Intelligence in Laryngeal Disorders: An Integrative Review > 3. Results and Discussion > 3.1. Physioacoustic Foundations of Automated Detection
What the sources say
- Not checked — no full text of the sourcecites [39] Leite D.R.A. de Moraes R.M. Wanderley Lopes L. Different performances of machine learning models to classify dysphonic and non-dysphonic voices J. Voice 2025 39 577 590 10.1016/j.jvoice.2022.11.001 36513560 source
- Not checked — no full text of the sourcecites [40] Li Z. Zhou L. Liu M. Huang Z. Exploring novel objective voice assessment parameters: A pilot study J. Voice 2025, in press 10.1016/j.jvoice.2025.05.020 40484739 source
- Partly supportedcites [41] Naranjo L. Pérez C.J. Campos-Roca Y. Madruga M. Replication-based regularization approaches to diagnose Reinke’s edema by using voice recordings Artif. Intell. Med. 2021 120 102162 10.1016/j.artmed.2021.102162 34629154 sourceExperimental — accuracy not yet measured
The Introduction (page unavailable) discusses acoustic features used for nodules, polyps, and Reinke’s oedema, and notes that different pathologies have been grouped into a single class for acoustic diagnosis. That supports the general point that such features are not unique to Reinke’s oedema, but the paper does not establish that the specified signatures are shared across all five conditions; it does not address reflux-related changes or hyperfunctional dysphonia. (LLM, conf=0.83)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
60 more claims · 20 with their cited sources checked
Create a free account to see all 103 claims and to verify the 6 references still open.
- It may state two separate claims as one. 29
- Probably not a substantial claim. 11
Citations
Cites
Clinical Practice Guideline: Hoarseness (Dysphonia)
Schwartz, S. R., Cohen, S. M., Dailey, S. H., Rosenfeld, R. M., Deutsch, E. S., Gillespie, M. B., Granieri, E., Hapner, E. R., Kimball, C. E., Krouse, H. J., et al. (2009) · Otolaryngology–Head and Neck Surgery · doi:10.1016/j.otohns.2009.06.744 · Cited by 453 · cited 1×
“Nearly one-third of the adult population will experience a voice disorder during their lifetime [1], with dysphonia prevalence reaching 7% in the general population [2] and up to 50% among voice professionals [3], resulting in significant quality of life impairment and socioeconomic burden [4,5].”
Clinical Practice Guideline: Hoarseness (Dysphonia) (Update)
Stachler, R. J., Francis, D. O., Schwartz, S. R., Damask, C. C., Digoy, G. P., Krouse, H. J., McCoy, S. J., Ouellette, D. R., Patel, R. R., Reavis, C. (. W., et al. (2018) · Otolaryngology–Head and Neck Surgery · doi:10.1177/0194599817751030 · Cited by 368 · cited 1×
“Current dysphonia guidelines consider laryngoscopic visualization and multidimensional voice assessment as the reference standard [20,21].”
Detection of Pathological Voice Using Cepstrum Vectors: A Deep Learning Approach
Fang, S.-H., Tsao, Y., Hsiao, M.-J., Chen, J.-Y., Lai, Y.-H., Lin, F.-C., & Wang, C.-T. (2019) · Journal of Voice · doi:10.1016/j.jvoice.2018.02.003 · Cited by 278 · cited 8×
“…develop self-supervised representation learning to improve generalization in data-sparse regimes, alongside multimodal fusion strategies combining acoustic signals with laryngeal imaging and clinical metadata, while preserving computational tractability for real-time deployment [23,24,26,57].”
InfluentialSelf‐reported impact of dysphonia in a primary care population: An epidemiological study
Cohen, S. M. (2010) · The Laryngoscope · doi:10.1002/lary.21058 · Cited by 198 · cited 2×
“Nearly one-third of the adult population will experience a voice disorder during their lifetime [1], with dysphonia prevalence reaching 7% in the general population [2] and up to 50% among voice professionals [3], resulting in significant quality of life impairment and socioeconomic burden [4,5].”
Using Ambulatory Voice Monitoring to Investigate Common Voice Disorders: Research Update
Mehta, D. D., Van Stan, J. H., Zañartu, M., Ghassemi, M., Guttag, J. V., Espinoza, V. M., Cortés, J. P., Cheyne, H. A., & Hillman, R. E. (2015) · Frontiers in Bioengineering and Biotechnology · doi:10.3389/fbioe.2015.00155 · Cited by 131 · cited 3×
“…missed diagnoses [34,48,65], whereas longitudinal monitoring demands measurement stability, calibration across repeated assessments, and explicit quantification of within-patient variability—with clinically meaningful change defined as exceeding typical short-term fluctuation [8,13,14,76–78].”
Lechien, J. R., Geneid, A., Bohlender, J. E., Cantarella, G., Avellaneda, J. C., Desuter, G., Sjogren, E. V., Finck, C., Hans, S., Hess, M., et al. (2023) · European Archives of Oto-Rhino-Laryngology · doi:10.1007/s00405-023-08211-6 · Cited by 107 · cited 2×
“Current dysphonia guidelines consider laryngoscopic visualization and multidimensional voice assessment as the reference standard [20,21].”
Self-citationKim, H., Jeon, J., Han, Y. J., Joo, Y., Lee, J., Lee, S., & Im, S. (2020) · Journal of Clinical Medicine · doi:10.3390/jcm9113415 · Cited by 97 · cited 3×
“The clinical spectrum includes benign structural lesions (e.g., vocal fold nodules, polyps, and cysts [6,7]), neuromuscular disorders (e.g., vocal fold paralysis, spasmodic dysphonia [8,9]), inflammatory conditions, and malignant lesions [10,11].”
BAGLS, a multihospital Benchmark for Automatic Glottis Segmentation
Gómez, P., Kist, A. M., Schlegel, P., Berry, D. A., Chhetri, D. K., Dürr, S., Echternach, M., Johnson, A. M., Kniesburges, S., Kunduk, M., et al. (2020) · Scientific Data · doi:10.1038/s41597-020-0526-3 · Cited by 95 · cited 2×
“…forests (RFs), and Gaussian mixture models (GMMs) operating on hand-crafted acoustic features [6,7,25]—deep learning (DL) networks processing spectrograms or raw waveforms [3,26,27], and self-supervised encoders such as wav2vec 2.0 or HuBERT pre-trained on large unlabeled speech corpora [23,28,29].”
InfluentialAl-nasheri, A., Muhammad, G., Alsulaiman, M., & Ali, Z. (2017) · Journal of Voice · doi:10.1016/j.jvoice.2016.01.014 · Cited by 91
Hu, H.-C., Chang, S.-Y., Wang, C.-H., Li, K.-J., Cho, H.-Y., Chen, Y.-T., Lu, C.-J., Tsai, T.-P., & Lee, O. K.-S. (2021) · Journal of Medical Internet Research · doi:10.2196/25247 · Cited by 82 · cited 10×
“Recent research demonstrates that some approaches even maintain high performance under unfavorable signal-to-noise ratios [31,42].”
Influential
Cited by
Chauhan, J. P. S., Chaudhary, K., & Verma, J. K. (2026) · Student's Journal of Health Research Africa · doi:10.51168/sjhrafrica.v7i2.2798
“(7).”
Metadata
- Authors
- Mairesse, S., Maniaci, A., Briganti, G., & Lechien, J. R.
- Affiliations
- University of Mons (BE)
- Corresponding author
- Lechien, Jerome R.
- Venue
- Journal of Personalized Medicine
- Publisher
- MDPI AG
- Published
- 2026-06-01
- Volume
- 16
- Issue
- 6
- Pages
- 301
- ISSN
- 2075-4426 (ISSN-L)
- PMC
- PMC13301393
- OpenAlex
- W7163067316
- Cited by
- 1
- References
- 76
- Citations by year
- 2026: 1
- FWCI
- 1.04
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- Locations
- Journal of Personalized Medicine (publishedVersion, cc-by)
- PubMed (publishedVersion)
- PubMed Central (publishedVersion, cc-by)
- ORBi UMONS (submittedVersion, cc-by)
- Topics
- Health Sciences › Health Professions › Speech and Hearing › Voice and Speech Disorders
- Health Sciences › Health Professions › Health Information Management › Artificial Intelligence in Healthcare
- Physical Sciences › Computer Science › Artificial Intelligence › AI in cancer detection
- Keywords
- laryngeal diseases, voice disorders, artificial intelligence, machine learning, deep learning, laryngeal lesions, acoustic analysis, jitter and shimmer, external validation, clinical decision support, data bias, recording protocols, medical AI regulation
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-01, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-01 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.