Pith. sign in

REVIEW 3 major objections 9 minor 23 references

This paper claims that replacing CNN-based fusion with a Vision Transformer encoder improves emergency triage zone classification on an 11,102-record emergency department dataset, lifting accuracy from 78.66% to 80.19% and F1-score from 68.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 20:17 UTC pith:K4V2KVJV

load-bearing objection A plausible but under-specified ViT-fusion model for triage, undermined by an internally inconsistent evaluation count. the 3 major comments →

arxiv 2607.16662 v1 pith:K4V2KVJV submitted 2026-07-18 cs.LG

Multimodal Attention-based Deep Learning for Emergency Triage with Electronic Health Records

Hazqeel Afyq Athaillah Kamarul Aryffin , Kamarul Aryffin Baharuddin , Mohd Halim Mohd Noor This is my paper
classification cs.LG
keywords emergency triagemultimodal deep learningVision TransformerTabNetBERTelectronic health recordsself-attentiontriage zone classification
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to show that a multimodal model fusing TabNet embeddings of vital signs with BERT embeddings of chief complaints through a Vision Transformer encoder predicts emergency triage zones more accurately than a convolutional fusion baseline. On 11,102 records from a Malaysian emergency department, it reports 80.19% accuracy versus 78.66%, 70.09% F1-score versus 68.39%, and 90.59% macro ROC AUC versus 89.33%, with paired t-test p<0.05 across 10 runs. If correct, the result suggests that global self-attention captures cross-modal dependencies, such as linking an abnormal heart rate to a specific complaint, better than localized convolutional windows. The practical stake is reducing under-triage in emergency departments, especially in the yellow zone, without over-triaging green patients.

Core claim

The central claim is that replacing the 2D convolutional fusion of the CNN-based baseline with a Vision Transformer encoder yields a more accurate triage-zone classifier. The proposed model concatenates the 1x8x8 TabNet embedding and the 12x8x8 BERT embedding along the channel dimension to form a 13x8x8 feature map, flattens it into an 832-element sequence, adds sinusoidal positional encodings, and passes it through an 8-layer ViT encoder with 128 attention heads. The paper reports statistically significant improvements in accuracy, F1-score, and macro ROC AUC, with a per-run error analysis showing reduced red-to-green and yellow-zone under-triage compared to the baseline.

What carries the argument

The load-bearing object is the 13x8x8 multimodal feature map: TabNet's 64-dimensional embedding reshaped to 1x8x8 and BERT's 768-dimensional embedding reshaped to 12x8x8, concatenated along channels, flattened to 832 scalars, and fed after sinusoidal positional encoding to a Vision Transformer encoder (8 layers, 128 heads, feedforward dimension 2048). The ViT is intended to replace convolution's localized sliding-window feature extraction with global self-attention over all positions, allowing a vital-sign anomaly to be linked to a distant textual symptom.

Load-bearing premise

The whole comparison rests on the unstated implementation choice of how the flattened 832-element feature map is fed to the transformer; if the tokenization or projection is arbitrary or mismatched, the reported 1.95% accuracy gain over CNN fusion is not tied to any identifiable mechanism.

What would settle it

Download the released code and verify the input pipeline: count total test predictions across the 10 runs (about 11,100 is expected from an 80/10/10 split of 11,102 records, not 13,876), and inspect whether the 832 scalars are projected into the transformer's working dimension by a single linear layer or treated as individual tokens. Either discrepancy would call the reported mechanism and error counts into question.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the reported gains hold, self-attention fusion can serve as the backbone for emergency triage decision-support tools that combine vital signs with free-text chief complaints.
  • Yellow-zone management improves substantially: across all 10 runs, under-triage to green drops from 1,375 to 1,157 and over-triage to red drops from 402 to 333.
  • Severe red-to-green under-triage decreases from 74 to 66 errors across all runs, a clinically meaningful reduction for the highest-risk patients.
  • The same channel-concatenate-then-ViT pattern applies to other tabular-plus-text EHR tasks, such as admission prediction, disposition, or critical-care risk scoring.
  • Because the ViT attends over all feature positions, the model could in principle yield attention-based explanations of which vital-sign/text pairs drive a triage zone, though the paper does not provide such visualizations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper never specifies how the flattened 832 scalars are tokenized or projected into the transformer's working dimension; if the released code uses a single learned linear projection, the channel structure is likely discarded and the claimed cross-modal attention may be a generic flexible fusion instead of a true modality-aware mechanism.
  • The reported 13,876 test cases across 10 runs is inconsistent with an 80/10/10 split of 11,102 records, which would give roughly 1,110 test cases per run and about 11,100 total; this suggests the error-analysis denominators do not match the stated split.
  • A direct testable extension would hold the TabNet and BERT embeddings fixed and compare CNN fusion, cross-attention fusion, and ViT fusion under identical hyperparameter tuning; if ViT's advantage only appears with the unstated tokenization choice, the mechanistic claim weakens.
  • The performance gains are modest (about 1.95% accuracy), and without an exact reproduction of the preprocessing, class weighting, and paired splits, the practical significance for clinical deployment remains an open question.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 9 minor

Summary. The paper proposes a multimodal deep-learning architecture for emergency triage zone classification on 11,102 records from Hospital Universiti Sains Malaysia. TabNet encodes structured vital signs into a 64-dimensional embedding, reshaped to 1×8×8; BERT (via SentenceTransformer) encodes the chief complaint into a 768-dimensional embedding, reshaped to 12×8×8. These are concatenated into a 13×8×8 feature map, flattened to 832 scalars, and processed by a Vision Transformer (ViT) encoder, replacing the CNN fusion of the baseline model from Leung et al. The authors report an average accuracy of 80.19% vs 78.66%, F1-score 70.09% vs 68.39%, and macro ROC AUC 90.59% vs 89.33% across 10 runs, with paired t-tests p<0.05, and conclude that attention-based fusion improves triage prediction over convolutional fusion.

Significance. If the numerical results withstand scrutiny, the contribution is a modest but clinically relevant improvement in triage zone classification on a real single-center dataset, with a clear comparison to an existing CNN-based fusion baseline. The paper offers a concrete motivation for replacing convolutional fusion with self-attention in this multimodal EHR setting, and it provides source code and a data link, which are strengths. However, the coherence of the evaluation protocol is currently questionable, and the architecture description is too underspecified to be reproduced or independently assessed. The central comparison rests on these two pillars, so the significance of the claimed result cannot be fully evaluated until they are resolved.

major comments (3)
  1. [§4.3 / §3.2.2] The reported evaluation protocol is internally inconsistent. The paper states an 80/10/10 stratified split of 11,102 records, repeated 10 times, which implies about 1,110 test cases per run and a total of roughly 11,100 test predictions across all runs. However, §4.3 explicitly says 'An error analysis across all 10 runs (13,876 test cases)'. This discrepancy is not a rounding error: the aggregate confusion-matrix errors listed in §4.3 (Red 385 + Yellow 1,490 + Green 924 = 2,799) yield an accuracy of 1 − 2,799/13,876 = 79.8%, consistent with the reported 80.19%, while under the stated 10% test split the denominator would be ≈11,102, giving 1 − 2,799/11,102 = 74.8%. The reported metrics and the paired t-tests are therefore computed under an unspecified evaluation protocol. Please state exactly how many test cases were used per run, whether the test sets overlapped across runs, and reconcil
  2. [§3.2.1] The proposed architecture is not reproducible as described. The flattened sequence x_flat ∈ R^832 is said to be combined with sinusoidal 1D positional embeddings and passed to a ViT Encoder with '8 layers, 128 multihead self-attention heads, and a feedforward network dimension of 2048', producing a '1024-dimensional transformer output'. It is not specified how 832 scalars are tokenized or projected into a transformer working dimension (d_model), whether each scalar is treated as one token, how 128 heads are arranged (e.g., head dimension), or how the MLP head maps a sequence of token outputs to three classes. The phrase 'forgoing the image patching step' does not resolve this: a ViT typically requires patch embeddings and a token dimension. Because the paper's central claim is that ViT-based fusion outperforms CNN fusion, this omission must be fixed with a precise specification of the to
  3. [§3.1 / Table 3] The baseline model identity is unclear. Section 3.1 says the baseline is 'the baseline model proposed reported in [11] for triage zone classification', and the paper compares against it. However, Table 3 describes Leung et al. [11] as 'Utilize TabNet and Bert with convolutional layers for hospitalization prediction' with accuracy 0.805 and AUC 0.836. If the reimplemented baseline differs from the published [11] system (e.g., target task, data split, or hyperparameters), the comparison to the published state of the art is ambiguous. Please clarify the exact relationship between the reimplemented baseline and the original [11] model, and state whether the baseline was retrained on the HUSM data under the same protocol.
minor comments (9)
  1. [Abstract] Duplicate word: 'exploits self-attention to to capture' should be 'to capture'.
  2. [§3.2.1] Typo: 'the model architecture integrates integrates vital signs' should be 'integrates'.
  3. [§5] Typo: 'a significant immprovement' should be 'improvement'.
  4. [§3.2.2 / §4.3] Two unresolved 'Appendix??' references (training-loss plots and per-run classification results). These appendices appear to be missing from the manuscript.
  5. [Figures] Figure numbering is inconsistent. The text refers to 'Fig. 1: Proposed model architecture' but a later figure is also captioned 'Fig. 2: Proposed model architecture'; the results section says 'Figure 2 compares the average class precision and recall', which does not match either caption.
  6. [§4.3] Typo: 'in give experiments' should be 'in five experiments'.
  7. [§3.2.1] The phrase '128 multihead self-attention heads' is likely a typo; 128 heads on an 832-length sequence is unusual and the hidden dimension is unspecified. Please correct to the actual number of heads and provide the head dimension.
  8. [§4.2 / Data availability] The data availability statement says 'Submitted pending approval' but also provides a UCI link. Please clarify the current availability status of the data.
  9. [§3.2.2] The sentence 'This superior generalization is primarily driven by the inclusion of attention mechanisms and layer normalization' is not supported by the experimental design, since the models differ in architecture, optimizer, learning rate, and effective training duration. This attribution should be removed or backed by an ablation.

Circularity Check

0 steps flagged

No circularity: the triage comparison is an empirical benchmark against a held-out test set, with no prediction reduced to a fitted parameter or to a self-citation.

full rationale

The paper's central claim is an empirical model comparison: the proposed TabNet+BERT+ViT fusion outperforms the Leung et al. CNN-based baseline on a held-out test portion of the HUSM triage dataset. The architecture and evaluation do not derive the reported accuracy, F1, or ROC AUC from the same quantities used to define or fit the model. The comparison is performed on held-out splits with paired t-tests across 10 runs, so the outcome is not forced by construction. The baseline and proposed models are external design choices, not renamings of the target metric. No load-bearing step reduces to a self-citation: references [11], [12], [21], [23], etc. are independent prior works with no author overlap with the present paper. The post-hoc statement that attention and layer normalization 'primarily driv[e]' the improvement is an unsupported attribution, but it is not circular reasoning. The reported inconsistency between 13,876 aggregate test cases and the stated 80/10/10 split of 11,102 records is a serious experimental-reporting concern, but it concerns the validity of the evaluation protocol, not a definitional or self-referential reduction. Accordingly, no circularity is found; the appropriate score is 0.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The paper introduces no new theoretical entities. It relies on standard pretrained components (BERT, TabNet), clinician-supplied thresholds, and hand-chosen architecture hyperparameters. The main unstated dependency is the exact tokenization/projection of the flattened multimodal feature map into the ViT, which is not documented. The empirical evaluation depends on the nurse-assigned triage labels and on temporal data exclusion.

free parameters (7)
  • Learning rate (proposed model) = 0.01 (Adam)
    Selected by Optuna over 50 trials (Section 3.2.2).
  • Learning rate (baseline model) = 0.003 (AdamW)
    Selected by Optuna over 50 trials (Section 3.2.2).
  • Optimizer choice (Adam vs AdamW) = Adam for proposed, AdamW for baseline
    Optimized via Optuna; different optimizers across the two models complicate comparison (Section 3.2.2).
  • TabNet pre-training ratio = 0.8
    Hand-chosen; obscures 80% of features during unsupervised pretraining (Section 3.2.1).
  • TabNet decision steps = 3
    Hand-chosen; number of decision steps summed to form the 64-dim tabular embedding (Section 3.2.1).
  • ViT depth, heads, FFN dimension = 8 layers, 128 heads, 2048 FFN
    Fixed architecture choices not justified by tuning or ablation (Section 3.2.1).
  • Class weights for loss = Inversely proportional to class frequencies
    Computed from training class distribution to address imbalance (Section 4.1).
axioms (6)
  • domain assumption Vital sign thresholds provided by HUSM doctors correctly binarize physiological abnormality
    Section 4.2 converts raw vitals into binary threshold features using doctor-supplied cutoffs; incorrect thresholds would distort all model inputs.
  • domain assumption BERT (bert-base-uncased) sentence embeddings preserve enough clinical semantics from expanded chief complaints
    Section 3.2.1 uses a generic SentenceTransformer BERT without medical-domain fine-tuning; no evidence that the embeddings capture triage-relevant meaning.
  • domain assumption Ground-truth triage zones (red/yellow/green) assigned by nurses are a valid and reliable prediction target
    Section 4.1 uses nurse-assigned zones as labels; these are themselves subjective, so the model learns to reproduce manual triage, not necessarily true acuity.
  • domain assumption Excluding 2020+ data to avoid COVID effects does not introduce selection bias
    Section 4.1 post hoc removes recent years; the model is only validated on 2017-2019 data, narrowing generalizability.
  • domain assumption The 10 repeated runs on the same train/validation/test split provide a valid basis for paired t-tests
    Section 4.3 applies paired t-tests to repeated runs of the same split; these runs are not independent samples, so the p-values reflect training stochasticity rather than variability across data partitions.
  • domain assumption MICE with Bayesian Ridge imputation is appropriate for the missing clinical data
    Section 4.2 uses MICE without reporting missingness analysis; assumes missing-at-random holds.

pith-pipeline@v1.3.0-alltime-deepseek · 9714 in / 14016 out tokens · 130723 ms · 2026-08-01T20:17:26.303083+00:00 · methodology

0 comments
read the original abstract

Accurate emergency triage decision is critical to avoid clinical deterioration, morbidity, and mortality. Machine learning-based triage system involves acquiring the main presenting complaint in text form and assessing vital signs in numerical data, enabling an automated and efficient analysis of patient information for timely and accurate prioritization of medical attention. However, modelling the intricacies of both data types requires a comprehensive understanding of the temporal structure and dependencies within the data. Thus, the aim of this study is to propose a multimodal deep learning architecture that can effectively handle both tabular and textual data. Furthermore, the proposed model exploits self-attention to to capture both local and global relationships between the features. A dataset consisting of 11,102 triage data collected from emergency department of Hospital Universiti Sains Malaysia is used for model development and validation. The proposed model demonstrated an increase of 1.95% in accuracy, 2.49% in F1-score, and 1.41% in ROC AUC compared to the baseline model. The experimental results demonstrated the potential of the proposed model in predicting triage decisions.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

23 extracted references · 8 canonical work pages · 1 internal anchor

  1. [1]

    Fry, M., Burr, G.: Review of the triage liter- ature: Past, present, future? Australian Emer- gency Nursing Journal5(2), 33–38 (2002) https: //doi.org/10.1016/S1328-2743(02)80018-9

    work page doi:10.1016/s1328-2743(02)80018-9 2002

  2. [2]

    Krylova et al

    Joseph, M.J., Summerscales, M., Yogesan, S., Bell, A., Genevieve, M., Kanagasingam, Y.: 8 T able 3: Comparison of relevant works and the proposed model Relevant W ork Description Performance S. Krylova et al. [17] Utilize TabTransformer to pre- dict Triage24 data of unfinished questionnaire-based digital medical triage interviews. 80% accuracy D. Gligorij...

    2023

  3. [3]

    BMC Emer- gency Medicine18(1), 58 (2018) https://doi.org/ 10.1186/s12873-018-0215-0

    Tam, H.L., Chung, S.F., Lou, C.K.: A review of triage accuracy and future direction. BMC Emer- gency Medicine18(1), 58 (2018) https://doi.org/ 10.1186/s12873-018-0215-0

    work page doi:10.1186/s12873-018-0215-0 2018

  4. [4]

    Journal of Ambient Intelligence and Humanized Comput- ing14(7), 8459–8486 (2023) https://doi.org/10

    Kumar, Y., Koul, A., Singla, R., Ijaz, M.F.: Artificial intelligence in disease diagnosis: a sys- tematic literature review, synthesizing frame- work and future research agenda. Journal of Ambient Intelligence and Humanized Comput- ing14(7), 8459–8486 (2023) https://doi.org/10. 1007/s12652-021-03612-z

    2023

  5. [5]

    International Journal of Intelligent Networks3, 58–73 (2022) https://doi.org/10.1016/j.ijin.2022

    Javaid, M., Haleem, A., Pratap Singh, R., Suman, R., Rab, S.: Significance of machine learning in healthcare: Features, pillars and applications. International Journal of Intelligent Networks3, 58–73 (2022) https://doi.org/10.1016/j.ijin.2022. 05.002

    work page doi:10.1016/j.ijin.2022 2022

  6. [6]

    Future Healthcare Journal6(2), 94–98 (2019) https:// doi.org/10.7861/futurehosp.6-2-94

    Davenport, T., Kalakota, R.: The potential for artificial intelligence in healthcare. Future Healthcare Journal6(2), 94–98 (2019) https:// doi.org/10.7861/futurehosp.6-2-94

    work page doi:10.7861/futurehosp.6-2-94 2019

  7. [7]

    Crit- ical Care (London, England)23(1), 64 (2019) https://doi.org/10.1186/s13054-019-2351-7

    Raita, Y., Goto, T., Faridi, M.K., Brown, D.F.M., Camargo, C.A., Hasegawa, K.: Emer- gency department triage prediction of clinical outcomes using machine learning models. Crit- ical Care (London, England)23(1), 64 (2019) https://doi.org/10.1186/s13054-019-2351-7

    work page doi:10.1186/s13054-019-2351-7 2019

  8. [8]

    Electronics12(5), 1102 (2023) https://doi.org/10.3390/electronics12051102

    P˘ av˘ aloaia, V.-D., Necula, S.-C.: Artificial Intelli- gence as a Disruptive Technology—A Systematic Literature Review. Electronics12(5), 1102 (2023) https://doi.org/10.3390/electronics12051102

    work page doi:10.3390/electronics12051102 2023

  9. [9]

    Acute Medicine & Surgery9(1), 740 (2022) https://doi.org/10.1002/ams2.740

    Mueller, B., Kinoshita, T., Peebles, A., Graber, M.A., Lee, S.: Artificial intelligence and machine learning in emergency medicine: a narrative review. Acute Medicine & Surgery9(1), 740 (2022) https://doi.org/10.1002/ams2.740

    work page doi:10.1002/ams2.740 2022

  10. [10]

    Applied Sciences10(23), 8631 (2020) https://doi.org/10

    Maslej-Kreˇ sˇ n´ akov´ a, V., Sarnovsk´ y, M., Butka, P., 9 Machov´ a, K.: Comparison of Deep Learning Mod- els and Various Text Pre-Processing Techniques for the Toxic Comments Classification. Applied Sciences10(23), 8631 (2020) https://doi.org/10. 3390/app10238631 . Number: 23

    2020

  11. [11]

    In: 2021 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pp

    Leung, K.-C., Lin, Y.-T., Hong, D.-Y., Tsai, C.-L., Huang, C.-H., Fu, L.-C.: A Novel Inter- pretable Deep-Learning-Based System for Triage Prediction in the Emergency Department: A Prospective Study. In: 2021 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pp. 2979–2985 (2021). https://doi.org/ 10.1109/SMC52423.2021.9658729

    arXiv 2021

  12. [12]

    JMIR Medical Informatics12, 48862 (2024)

    Lin, Y.-T., Deng, Y.-X., Tsai, C.-L., Huang, C.-H., Fu, L.-C.: Interpretable deep learning sys- tem for identifying critical patients through the prediction of triage level, hospitalization, and length of stay: Prospective study. JMIR Medical Informatics12, 48862 (2024)

    2024

  13. [13]

    arXiv (2020)

    Arik, S.O., Pfister, T.: TabNet: Attentive Inter- pretable Tabular Learning. arXiv (2020). https: //doi.org/10.48550/arXiv.1908.07442 . http:// arxiv.org/abs/1908.07442 Accessed 2023-12-11

    Pith/arXiv arXiv doi:10.48550/arxiv.1908.07442 2020

  14. [14]

    Surgical Innovation31(6), 583–597 (2024)

    Chai, C., Peng, S.-z., Zhang, R., Li, C.-w., Zhao, Y.: Advancing emergency department triage pre- diction with machine learning to optimize triage for abdominal pain surgery patients. Surgical Innovation31(6), 583–597 (2024)

    2024

  15. [15]

    In: 2020 IEEE International Conference on Big Data (Big Data), pp

    Arnaud, ´E., Elbattah, M., Gignon, M., Dequen, G.: Deep Learning to Predict Hospitaliza- tion at Triage: Integration of Structured Data and Unstructured Text. In: 2020 IEEE International Conference on Big Data (Big Data), pp. 4836–4841 (2020). https: //doi.org/10.1109/BigData50022.2020.9378073 .https://ieeexplore.ieee.org/document/9378073 Accessed 2023-12-11

    arXiv 2020

  16. [16]

    IEEE Access 9, 45552–45561 (2021) https://doi.org/10.1109/ ACCESS.2021.3066164

    Sharafat, A.R., Bayati, M.: PatientFlowNet: A Deep Learning Approach to Patient Flow Pre- diction in Emergency Departments. IEEE Access 9, 45552–45561 (2021) https://doi.org/10.1109/ ACCESS.2021.3066164 . Accessed 2023-12-11

    arXiv 2021

  17. [17]

    In: 2024 International Conference on Machine Learning and Applications (ICMLA), pp

    Krylova, S., Schmidt, F., Vlassov, V.: Leveraging machine learning models to predict the outcome of digital medical triage interviews. In: 2024 International Conference on Machine Learning and Applications (ICMLA), pp. 160–167 (2024). IEEE

    2024

  18. [18]

    Deep Attention Model for Triage of Emergency Department Patients

    Gligorijevic, D., Stojanovic, J., Satz, W., Sto- jkovic, I., Schreyer, K., Del Portal, D., Obradovic, Z.: Deep Attention Model for Triage of Emer- gency Department Patients. arXiv (2018). https: //doi.org/10.48550/arXiv.1804.03240 . http:// arxiv.org/abs/1804.03240 Accessed 2023-12-11

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1804.03240 2018

  19. [19]

    Emergency Medicine Australasia 33(3), 480–484 (2021) https://doi.org/10.1111/ 1742-6723.13656

    Tahayori, B., Chini-Foroush, N., Akhlaghi, H.: Advanced natural language processing technique to predict patient disposition based on emergency triage notes. Emergency Medicine Australasia 33(3), 480–484 (2021) https://doi.org/10.1111/ 1742-6723.13656 . Accessed 2023-12-11

    arXiv 2021

  20. [20]

    Journal of Biomedical Informatics133, 104171 (2022)

    Chen, T.-L., Chen, J.C., Chang, W.-H., Tsai, W., Shih, M.-C., Nabila, A.W.: Imbalanced prediction of emergency department admission using natu- ral language processing and deep neural network. Journal of Biomedical Informatics133, 104171 (2022)

    2022

  21. [21]

    In: 2019 IEEE Interna- tional Conference on Bioinformatics and Biomedicine (BIBM), pp

    Wang, G., Liu, X., Xie, K., Chen, N., Chen, T.: DeepTriager: A Neural Attention Model for Emergency Triage with Electronic Health Records. In: 2019 IEEE Interna- tional Conference on Bioinformatics and Biomedicine (BIBM), pp. 978–982 (2019). https: //doi.org/10.1109/BIBM47256.2019.8983093 .https://ieeexplore.ieee.org/document/8983093 Accessed 2023-12-11

    arXiv 2019

  22. [22]

    ¨O., Pfister, T.: Tabnet: Attentive inter- pretable tabular learning

    Arik, S. ¨O., Pfister, T.: Tabnet: Attentive inter- pretable tabular learning. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, pp. 6679–6687 (2021)

    2021

  23. [23]

    Journal of Biomedical Informatics 133, 104171 (2022) https://doi.org/10.1016/j.jbi

    Chen, T.-L., Chen, J.C., Chang, W.-H., Tsai, W., Shih, M.-C., Wildan Nabila, A.: Imbalanced prediction of emergency department admission using natural language processing and deep neu- ral network. Journal of Biomedical Informatics 133, 104171 (2022) https://doi.org/10.1016/j.jbi. 2022.104171 . Accessed 2023-12-11 10

    arXiv 2022