pith. sign in

arxiv: 2605.24172 · v1 · pith:5UOXZWRWnew · submitted 2026-05-22 · 💻 cs.AI

EPPC-OASIS: Ontology-Aware Adaptation and Structured Inference Refinement for Electronic Patient-Provider Communication Mining in Secure Messages

Samah Fodeh , Sreeraj Ramachandran , Elyas Irankhah , Muhammad Arif , Afshan Khan , Ganesh Puthiaraju , Linhai Ma , Srivani Talakokkul
show 2 more authors
Jordan Alpert Sarah Schellhorn
This is my paper

Pith reviewed 2026-06-30 15:51 UTC · model grok-4.3

classification 💻 cs.AI
keywords EPPC ontologypatient-provider communicationsecure messagesontology-aware adaptationstructured inference refinementinformation extractionclinical NLPWasserstein alignment
0
0 comments X

The pith

Ontology-aware adaptation with structured refinement improves extraction of patient-provider communication codes by 1-2 F1 points.

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

The paper introduces EPPC-OASIS to automatically label secure messages between patients and providers according to the EPPC ontology for communication behaviors. It augments standard fine-tuning by adding a Wasserstein alignment term that pulls model representations toward neighborhoods defined by the ontology, then applies a sequence of verification, self-consistency, and hybrid correction steps at inference time. Experiments across several open-weight models on a de-identified message corpus show the full pipeline reaching 77.13 percent Code+Sub-code F1 and 63.83 percent Triplet F1, modest gains over the best supervised baseline. A reader would care because manual review of these messages cannot scale, yet the coded behaviors contain clinically relevant patterns that could inform care quality and training if extracted reliably.

Core claim

EPPC-OASIS augments supervised fine-tuning with a Wasserstein alignment objective that encourages alignment between model representation neighborhoods and EPPC ontology-derived neighborhoods, while inference refinement uses verification, self-consistency, hybrid correction, and selection or ensembling to address residual prediction errors. On the de-identified corpus the best deployable pipeline achieved 77.13 percent Code+Sub-code F1 and 63.83 percent Triplet F1, corresponding to absolute gains of 1.39 and 2.12 F1 points over the strongest supervised fine-tuning baseline.

What carries the argument

EPPC-OASIS, which combines Wasserstein alignment of representation neighborhoods to EPPC ontology neighborhoods during adaptation with post-hoc verification, self-consistency, and hybrid correction at inference time.

If this is right

  • Structured predictions preserve the code and sub-code hierarchy more faithfully than plain fine-tuning.
  • The same pipeline can be applied across multiple open-weight language model families without requiring new labeled data.
  • Retrospective mining of large archives of secure messages becomes more feasible at scale.
  • Residual errors after fine-tuning can be reduced by the listed inference procedures without retraining.

Where Pith is reading between the lines

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

  • If the alignment mechanism proves robust, analogous neighborhood-matching objectives could be tested on other medical ontologies that also require hierarchical label structures.
  • The modest absolute gains suggest the method may be most useful when combined with larger base models or additional unlabeled message data.
  • Operational deployment would still require separate checks on live, non-de-identified streams to confirm the gains persist outside the study corpus.

Load-bearing premise

The performance gains are produced by the ontology alignment and inference-refinement steps rather than by unexamined differences in training procedure or data handling, and the de-identified corpus is representative of the messages on which the system would be deployed.

What would settle it

Re-training the same models with the Wasserstein alignment term removed or on a fresh held-out message set and checking whether the reported F1 margins over the supervised baseline disappear.

Figures

Figures reproduced from arXiv: 2605.24172 by Afshan Khan, Elyas Irankhah, Ganesh Puthiaraju, Jordan Alpert, Linhai Ma, Muhammad Arif, Samah Fodeh, Sarah Schellhorn, Sreeraj Ramachandran, Srivani Talakokkul.

Figure 1
Figure 1. Figure 1: Overview of the proposed EPPC-OASIS framework for structured EPPC extraction from secure patient-provider messages. The workflow consists of two major stages: ontology-aware adaptation during training and structured inference refinement during inference. The training stage converts secure-message inputs into structured JSON annotations using the EPPC ontology hierarchy and ontology-aware representation ali… view at source ↗
Figure 2
Figure 2. Figure 2: Hierarchical distribution of EPPC code and sub-code annotations. The left panel shows parent-code prevalence across all 5,516 annotations. The right panel shows within-parent sub-code composition. optimized to reproduce each target annotation, but the loss does not directly represent relationships among EPPC codes, sub-codes, or valid parent-child label combinations. This is limiting for EPPC coding becaus… view at source ↗
Figure 3
Figure 3. Figure 3: EPPC-OASIS performance improvements and residual sub-code confusions using the Llama-3.1-8B reference model. The left panel shows the most frequent remaining sub-code confusions after EPPC-OASIS greedy inference. Most residual confusions occur between semantically adjacent EPPC sub-codes. The right panel compares extraction error rates between supervised fine￾tuning (SFT greedy) and EPPC-OASIS greedy acros… view at source ↗
Figure 4
Figure 4. Figure 4: Training ablation summary using Llama-3.1-8B-Instruct. Points show mean Code+Sub-code F1 and Triplet F1 with standard-deviation intervals across random seeds. Highlighted rows denote the final EPPC-OASIS configuration used in the main experiments. 3.3.3. Ontology Similarity Prior The ontology-prior ablation tests whether EPPC-OASIS benefits from graded similarity among label descriptions beyond exact code/… view at source ↗
read the original abstract

Secure patient-provider messages contain clinically important communication behaviors that are difficult to characterize manually at scale. The Electronic Patient-Provider Communication (EPPC) framework provides an ontology for coding these behaviors, but automated extraction remains challenging because predictions must preserve fine-grained code/sub-code structure while grounding annotations in message text. We developed EPPC-OASIS, an ontology-aware adaptation approach for structured EPPC extraction, and combined it with deployable inference-refinement procedures designed to improve the coherence of final annotations. EPPC-OASIS augments supervised fine-tuning with a Wasserstein alignment objective that encourages alignment between model representation neighborhoods and EPPC ontology-derived neighborhoods, while inference refinement uses verification, self-consistency, hybrid correction, and selection or ensembling to address residual prediction errors. We evaluated the framework on a de-identified corpus of secure patient-provider messages against prompting, supervised fine-tuning, preference-based, and robustness-oriented baselines across multiple open-weight language models. Across model families, the best deployable pipeline achieved 77.13% Code+Sub-code F1 and 63.83% Triplet F1, corresponding to modest but consistent absolute gains of +1.39 and +2.12 F1 points over the strongest supervised fine-tuning baseline. These results suggest that ontology-aware adaptation with structured inference refinement can support scalable retrospective EPPC mining, although external validation is needed before operational use.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces EPPC-OASIS, an ontology-aware adaptation method that augments supervised fine-tuning (SFT) with a Wasserstein alignment objective to encourage alignment between model representation neighborhoods and EPPC ontology-derived neighborhoods. It combines this with deployable inference-refinement procedures (verification, self-consistency, hybrid correction, and selection/ensembling) for structured extraction of EPPC codes, sub-codes, and triplets from secure patient-provider messages. Evaluated on a de-identified corpus against prompting, SFT, preference-based, and robustness baselines across multiple open-weight LLMs, the best pipeline reports 77.13% Code+Sub-code F1 and 63.83% Triplet F1, with absolute gains of +1.39 and +2.12 over the strongest SFT baseline.

Significance. If the modest gains can be shown to arise specifically from the Wasserstein alignment and inference-refinement components rather than from differences in training procedure or data handling, the framework could support more scalable retrospective mining of clinically relevant communication behaviors under the EPPC ontology. The multi-model-family evaluation and explicit comparison to several baseline categories are positive features, but the small effect sizes and stated need for external validation limit the immediate operational significance.

major comments (2)
  1. [Evaluation] The evaluation does not report whether the SFT baseline used identical optimizer, learning-rate schedule, epoch count, batch size, or random-seed handling as the EPPC-OASIS runs. Because the central performance claim rests on the +1.39 and +2.12 F1 improvements being caused by the Wasserstein alignment objective plus the verification/self-consistency/hybrid-correction steps, the absence of these controls is load-bearing.
  2. [Methods] No ablation tables or controlled experiments isolate the contribution of the Wasserstein alignment term from the base SFT objective or from the subsequent inference-refinement procedures. Without such ablations, the attribution of the reported F1 gains to the ontology-aware adaptation cannot be confirmed.
minor comments (2)
  1. The abstract refers to a 'deployable pipeline' without defining the criteria that distinguish deployable from non-deployable inference-refinement configurations.
  2. Consider adding a brief error analysis or qualitative examples illustrating the specific error types corrected by the verification and hybrid-correction steps that remain after SFT alone.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for highlighting the need for tighter experimental controls to support attribution of the reported gains. We address both major comments below and will revise the manuscript to include the requested details and experiments.

read point-by-point responses

  1. Referee: [Evaluation] The evaluation does not report whether the SFT baseline used identical optimizer, learning-rate schedule, epoch count, batch size, or random-seed handling as the EPPC-OASIS runs. Because the central performance claim rests on the +1.39 and +2.12 F1 improvements being caused by the Wasserstein alignment objective plus the verification/self-consistency/hybrid-correction steps, the absence of these controls is load-bearing.

    Authors: The SFT baselines were trained with identical settings: AdamW optimizer, linear warmup/decay schedule, 3 epochs, batch size 16, and random seed 42. These choices were made explicitly to isolate the contribution of the Wasserstein term. We will add a new 'Training Configuration' subsection with a comparison table documenting all hyperparameters for both SFT and EPPC-OASIS runs. revision: yes

  2. Referee: [Methods] No ablation tables or controlled experiments isolate the contribution of the Wasserstein alignment term from the base SFT objective or from the subsequent inference-refinement procedures. Without such ablations, the attribution of the reported F1 gains to the ontology-aware adaptation cannot be confirmed.

    Authors: We agree that dedicated ablations are needed to isolate effects. The revised manuscript will include two new controlled experiments: (i) SFT versus SFT+Wasserstein (no inference refinement) and (ii) SFT versus SFT+inference refinement (no Wasserstein). Results will appear in a new ablation table with statistical significance tests, allowing direct attribution of the +1.39/+2.12 gains. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical comparison is self-contained

full rationale

The paper reports an empirical evaluation of EPPC-OASIS (ontology-aware adaptation via Wasserstein alignment plus inference refinement) against prompting, SFT, preference, and robustness baselines across model families on a de-identified message corpus, with absolute F1 gains presented as experimental outcomes. No equations, fitted parameters renamed as predictions, self-definitional loops, or load-bearing self-citations appear in the abstract or described method; the central claims rest on direct metric comparisons rather than any derivation that reduces outputs to inputs by construction. This is the expected non-finding for a standard applied ML paper whose results are externally falsifiable via replication on held-out data.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Review performed on abstract only; no explicit free parameters, axioms, or invented entities are stated. The approach relies on standard supervised fine-tuning plus a known optimal-transport alignment loss and post-hoc inference heuristics whose precise implementation details are not visible.

pith-pipeline@v0.9.1-grok · 5829 in / 1250 out tokens · 53310 ms · 2026-06-30T15:51:44.611771+00:00 · methodology

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

42 extracted references · 21 canonical work pages · 5 internal anchors

  1. [1]

    Agrawal, M., Hegselmann, S., Lang, H., Kim, Y., Sontag, D.,

  2. [2]

    1998–2022

    Large language models are few-shot clinical informa- tion extractors, in: Proceedings of the 2022 Conference on Em- pirical Methods in Natural Language Processing, pp. 1998–2022. URL:https://aclanthology.org/2022.emnlp-main.130/,doi:10.18653/ v1/2022.emnlp-main.130

    2022

  3. [3]

    A keyword-enhanced approach to handle class imbalance in clinical text classification

    Blanchard, A.E., Gao, S., Yoon, H.J., Christian, J.B., Durbin, E.B., Wu, X.C., Stroup, A., Doherty, J., Schwartz, S.M., Wiggins, C., Coyle, L., Penberthy, L., Tourassi, G.D., 2022. A keyword-enhanced approach to handle class imbalance in clinical text classification. IEEE Journal of Biomedical and Health Informatics 26, 2796–2803. URL:https://doi.org/10.1...

    work page doi:10.1109/jbhi.2022.3141976 2022

  4. [4]

    Cronin, R.M., Davis, S.E., Shenson, J.A., Chen, Q., Rosenbloom, S.T., Jackson, G.P., 2015. Growth of secure messaging through a Fodeh et al.:Preprint submitted to ElsevierPage 14 of 19 EPPC-OASIS with structured inference patient portal as a form of outpatient interaction across clinical spe- cialties. AppliedClinicalInformatics6,288–304. URL:https://doi....

    work page doi:10.4338/aci-2014-12-ra-0117 2015

  5. [5]

    Sinkhorn distances: Lightspeed computation of optimal transport, in: Advances in Neural Information Processing Systems

    Cuturi, M., 2013. Sinkhorn distances: Lightspeed computation of optimal transport, in: Advances in Neural Information Processing Systems. URL:https://papers.nips.cc/paper/ 4927-sinkhorn-distances-lightspeed-computation-of-optimal-transport

    2013

  6. [6]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    DeepSeek-AI,Guo,D.,Yang,D.,Zhang,H.,Song,J.,Wang,P.,Zhu, Q., Xu, R., Zhang, R., Ma, S., et al., 2025. DeepSeek-R1: Incentiviz- ing reasoning capability in LLMs via reinforcement learning. arXiv preprint arXiv:2501.12948 URL:https://arxiv.org/abs/2501.12948, doi:10.48550/arXiv.2501.12948

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2501.12948 2025

  7. [7]

    8-bit optimizers via block-wise quantization, in: International Conference on Learning Representations

    Dettmers, T., Lewis, M., Shleifer, S., Zettlemoyer, L., 2022. 8-bit optimizers via block-wise quantization, in: International Conference on Learning Representations. URL:https://openreview.net/forum? id=shpkpVXzo3h

    2022

  8. [8]

    Dettmers, T., Pagnoni, A., Holtzman, A., Zettlemoyer, L.,

  9. [9]

    URL: https://proceedings.neurips.cc/paper_files/paper/2023/hash/ 1feb87871436031bdc0f2beaa62a049b-Abstract.html

    QLoRA: Efficient finetuning of quantized LLMs, in: Advances in Neural Information Processing Systems. URL: https://proceedings.neurips.cc/paper_files/paper/2023/hash/ 1feb87871436031bdc0f2beaa62a049b-Abstract.html

    2023

  10. [10]

    ERASER: A benchmark to evaluate ratio- nalized NLP models, in: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp

    DeYoung, J., Jain, S., Rajani, N.F., Lehman, E., Xiong, C., Socher, R., Wallace, B.C., 2020. ERASER: A benchmark to evaluate ratio- nalized NLP models, in: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp. 4443–4458. URL:https://doi.org/10.18653/v1/2020.acl-main.408,doi:10.18653/ v1/2020.acl-main.408

    work page doi:10.18653/v1/2020.acl-main.408 2020

  11. [11]

    POT:Pythonoptimaltransport

    Flamary, R., Courty, N., Gramfort, A., Alaya, M.Z., Boisbunon, A., Chambon, S., Chapel, L., Corenflos, A., Fatras, K., Fournier, N., Gautheron, L., Gayraud, N.T.H., Janati, H., Rakotomamonjy, A., Redko,I.,Rolet,A.,Schutz,A.,Seguy,V.,Sutherland,D.J.,Tavenard, R.,Tong,A.,Vayer,T.,2021. POT:Pythonoptimaltransport. Journal ofMachineLearningResearch22,1–8. URL...

    2021

  12. [12]

    PVminerLLM: Structured extraction of patient voice from patient-generated text using large language models

    Fodeh,S.,Ma,L.,Puthiaraju,G.,Talakokkul,S.,Khan,A.,Hagaman, A., Lowe, S., Roundtree, A., 2026a. PVminerLLM: Structured extraction of patient voice from patient-generated text using large language models. arXiv preprint arXiv:2603.05776 URL:https: //arxiv.org/abs/2603.05776, doi:10.48550/arXiv.2603.05776

    work page doi:10.48550/arxiv.2603.05776

  13. [13]

    TAB-PO: Preference Optimization with a Token-Level Adaptive Barrier for Token-Critical Structured Generation

    Fodeh,S.,Ma,L.,Puthiaraju,G.,Talakokkul,S.,Khan,A.,Hagaman, A., Lowe, S.R., Roundtree, A.K., 2026b. TAB-PO: Preference optimization with a token-level adaptive barrier for token-critical structuredgeneration. arXivpreprintarXiv:2603.00025URL:https: //arxiv.org/abs/2603.00025, doi:10.48550/arXiv.2603.00025

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2603.00025

  14. [14]

    PVminer: A domain- specifictooltodetectthepatientvoiceinpatientgenerateddata.arXiv preprint arXiv:2602.21165 URL:https://arxiv.org/abs/2602.21165, doi:10.48550/arXiv.2602.21165

    Fodeh,S.,Ma,L.,Wang,Y.,Talakokkul,S.,Puthiaraju,G.,Khan,A., Hagaman, A., Lowe, S., Roundtree, A., 2026c. PVminer: A domain- specifictooltodetectthepatientvoiceinpatientgenerateddata.arXiv preprint arXiv:2602.21165 URL:https://arxiv.org/abs/2602.21165, doi:10.48550/arXiv.2602.21165

    work page doi:10.48550/arxiv.2602.21165

  15. [15]

    STaR-DRO: Stateful Tsallis Reweighting for Group-Robust Structured Prediction

    Fodeh,S.,Puthiaraju,G.,Irankhah,E.,Ma,L.,Talakokkul,S.,Khan, A., Ramachandran, S., Alpert, J., Schellhorn, S., 2026d. STaR-DRO: Stateful tsallis reweighting for group-robust structured prediction. arXiv preprint arXiv:2604.09737 URL:https://arxiv.org/abs/2604. 09737, doi:10.48550/arXiv.2604.09737

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.09737

  16. [16]

    EPPCMinerBen: A novel benchmark for evaluating large language models on electronic patient-provider communica- tion via the patient portal

    Fodeh, S.J., Wang, Y., Ma, L., Talakokkul, S., Alpert, J.M., Schell- horn, S., 2026e. EPPCMinerBen: A novel benchmark for evaluating large language models on electronic patient-provider communica- tion via the patient portal. Artificial Intelligence in Medicine , 103429URL:https://doi.org/10.1016/j.artmed.2026.103429,doi:10. 1016/j.artmed.2026.103429

    work page doi:10.1016/j.artmed.2026.103429 2026

  17. [17]

    The Llama 3 Herd of Models

    Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al- Dahle, A., Letman, A., Mathur, A., Schelten, A., Vaughan, A., et al., 2024. The Llama 3 herd of models. arXiv preprint arXiv:2407.21783 URL:https://arxiv.org/abs/2407.21783, doi:10. 48550/arXiv.2407.21783

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  18. [18]

    Henning, S., Beluch, W., Fraser, A., Friedrich, A., 2023. A survey of methods for addressing class imbalance in deep-learning based natu- rallanguageprocessing,in:Proceedingsofthe17thConferenceofthe European Chapter of the Association for Computational Linguistics, pp.523–540. URL:https://doi.org/10.18653/v1/2023.eacl-main.38, doi:10.18653/v1/2023.eacl-main.38

    work page doi:10.18653/v1/2023.eacl-main.38 2023

  19. [19]

    LoRA: Low-rank adaptation of large language models, in: International Conference on Learning Representations

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., 2022. LoRA: Low-rank adaptation of large language models, in: International Conference on Learning Representations. URL:https://openreview.net/forum?id=nZeVKeeFYf9

    2022

  20. [20]

    Characterizing patient-clinician communication in secure medical messages: Retrospective study

    Huang, M., Fan, J., Prigge, J., Shah, N.D., Costello, B.A., Yao, L., 2022. Characterizing patient-clinician communication in secure medical messages: Retrospective study. Journal of Medical Internet Research 24, e17273. URL:https://doi.org/10.2196/17273, doi:10. 2196/17273

    work page doi:10.2196/17273 2022

  21. [21]

    Efficient memory management for large language model serving with PagedAttention, in: Proceed- ings of the 29th ACM Symposium on Operating Systems Principles, pp

    Kwon,W.,Li,Z.,Zhuang,S.,Sheng,Y.,Zheng,L.,Yu,C.H.,Gonza- lez, J.E., Zhang, H., Stoica, I., 2023. Efficient memory management for large language model serving with PagedAttention, in: Proceed- ings of the 29th ACM Symposium on Operating Systems Principles, pp. 611–626. URL:https://dblp.org/rec/conf/sosp/KwonLZ0ZY0ZS23, doi:10.1145/3600006.3613165

    work page doi:10.1145/3600006.3613165 2023

  22. [22]

    Self-refine: Iterative refinement with self-feedback, in: Advances in Neural Information Processing Systems

    Madaan, A., Tandon, N., Gupta, P., Hallinan, S., Gao, L., Wiegreffe, S., Alon, U., Dziri, N., Prabhumoye, S., Yang, Y., Gupta, S., Majumder, B.P., Hermann, K., Welleck, S., Yazdanbakhsh, A., Clark, P., 2023. Self-refine: Iterative refinement with self-feedback, in: Advances in Neural Information Processing Systems. URL: https://proceedings.neurips.cc/pape...

    2023

  23. [23]

    https://github.com/huggingface/peft

    Mangrulkar, S., Gugger, S., Debut, L., Belkada, Y., Paul, S., 2022.PEFT:State-of-the-artparameter-efficientfine-tuningmethods. https://github.com/huggingface/peft. Accessed 2026-05-12

    2022

  24. [24]

    SimPO: Simple preference optimization with a reference-free reward, in: Advances in Neural Information Processing Systems

    Meng, Y., Xia, M., Chen, D., 2024. SimPO: Simple preference optimization with a reference-free reward, in: Advances in Neural Information Processing Systems. URL: https://proceedings.neurips.cc/paper_files/paper/2024/hash/ e099c1c9699814af0be873a175361713-Abstract-Conference.html, doi:10.52202/079017-3946

    work page doi:10.52202/079017-3946 2024

  25. [25]

    Llama 3.2 1b and 3b instruct model cards.https: //huggingface.co/meta-llama

    Meta AI, 2024a. Llama 3.2 1b and 3b instruct model cards.https: //huggingface.co/meta-llama. Accessed 2026-05-12

    2026

  26. [26]

    Llama 3.3 70B Instruct model card.https:// huggingface.co/meta-llama/Llama-3.3-70B-Instruct

    Meta AI, 2024b. Llama 3.3 70B Instruct model card.https:// huggingface.co/meta-llama/Llama-3.3-70B-Instruct. Accessed2026- 05-12

  27. [27]

    Mistral-Small-24B-Instruct-2501 model card.https://huggingface.co/mistralai/ Mistral-Small-24B-Instruct-2501

    Mistral AI, 2025. Mistral-Small-24B-Instruct-2501 model card.https://huggingface.co/mistralai/ Mistral-Small-24B-Instruct-2501. Accessed 2026-05-12

    2025

  28. [28]

    North,F., Luhman,K.E.,Mallmann, E.A.,Mallmann,T.J., Tulledge- Scheitel, S.M., North, E.J., Pecina, J.L., 2020. A retrospective analysis of provider-to-patient secure messages: How much are they increasing, who is doing the work, and is the work happening after hours? JMIRMedicalInformatics8,e16521. URL:https://doi.org/ 10.2196/16521, doi:10.2196/16521

    work page doi:10.2196/16521 2020

  29. [29]

    PyTorch: An imperative style, high-performance deep learning library, in: Advances in Neural Information Processing Systems

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S., 2019. PyTorch: An imperative style, high-performance deep learning library, in: Advances in Neural In...

    2019

  30. [30]

    Direct preference optimization: Your language model is secretly a reward model, in: Advances in Neural Information Processing Systems

    Rafailov, R., Sharma, A., Mitchell, E., Ermon, S., Manning, C.D., Finn, C., 2023. Direct preference optimization: Your language model is secretly a reward model, in: Advances in Neural Information Processing Systems. URL: https://proceedings.neurips.cc/paper_files/paper/2023/hash/ a85b405ed65c6477a4fe8302b5e06ce7-Abstract-Conference.html

    2023

  31. [31]

    Multi-label classification: An overview

    Tsoumakas, G., Katakis, I., 2007. Multi-label classification: An overview. International Journal of Data Warehousing and Mining 3, 1–13. URL:https://doi.org/10.4018/jdwm.2007070101, doi:10.4018/ jdwm.2007070101. Fodeh et al.:Preprint submitted to ElsevierPage 15 of 19 EPPC-OASIS with structured inference

    work page doi:10.4018/jdwm.2007070101 2007

  32. [32]

    U, S.C.L., He, J., Basulto, V.G., Pan, J.Z., 2023. Instances and labels: Hierarchy-aware joint supervised contrastive learning for hi- erarchical multi-label text classification, in: Findings of the Associ- ation for Computational Linguistics: EMNLP 2023, pp. 8858–8875. URL:https://doi.org/10.18653/v1/2023.findings-emnlp.594,doi:10. 18653/v1/2023.findings...

    work page doi:10.18653/v1/2023.findings-emnlp.594 2023

  33. [33]

    Unsloth: Fast language model fine-tuning.https: //github.com/unslothai/unsloth

    Unsloth AI, 2024. Unsloth: Fast language model fine-tuning.https: //github.com/unslothai/unsloth. Accessed 2026-05-21

    2024

  34. [34]

    Self-consistency improves chain of thought reasoning in language models, in: International Conference on Learning Representations

    Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A., Zhou, D., 2023. Self-consistency improves chain of thought reasoning in language models, in: International Conference on Learning Representations. URL:https://openreview.net/forum? id=1PL1NIMMrw

    2023

  35. [35]

    Wang, Z., Wang, P., Liu, T., Lin, B., Cao, Y., Sui, Z., Wang, H.,

  36. [36]

    3740–3751

    HPT: Hierarchy-aware prompt tuning for hierarchical text classification, in: Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pp. 3740–3751. URL: https://doi.org/10.18653/v1/2022.emnlp-main.246, doi:10.18653/v1/ 2022.emnlp-main.246

    work page doi:10.18653/v1/2022.emnlp-main.246 2022

  37. [37]

    Chain-of-thought prompting elicits reasoning in large language models, in: Advances in Neural Information Processing Systems

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., Zhou, D., 2022. Chain-of-thought prompting elicits reasoning in large language models, in: Advances in Neural Information Processing Systems. URL: https://proceedings.neurips.cc/paper_files/paper/2022/hash/ 9d5609613524ecf4f15af0f7b31abca4-Abstract.html

    2022

  38. [38]

    Generating sequences by learning to self-correct

    Welleck, S., Lu, X., West, P., Brahman, F., Shen, T., Khashabi, D., Choi, Y., 2022. Generating sequences by learning to self-correct. arXiv preprint arXiv:2211.00053 URL:https://arxiv.org/abs/2211. 00053, doi:10.48550/arXiv.2211.00053

    work page doi:10.48550/arxiv.2211.00053 2022

  39. [39]

    Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz, M., Davison, J., Shleifer, S., von Platen, P., Ma, C., Jernite, Y., Plu, J., Xu, C., Le Scao, T., Gugger, S., Drame, M., Lhoest, Q., Rush, A.M., 2020. Hugging- Face’s transformers: State-of-the-art natural language processing, in: Proceedings of...

    2020

  40. [40]

    Xu, L., Teng, S., Zhao, R., Guo, J., Xiao, C., Jiang, D., Ren, B.,

  41. [41]

    2459–2468

    Hierarchical multi-label text classification with horizontal and vertical category correlations, in: Proceedings of the 2021 Con- ference on Empirical Methods in Natural Language Processing, pp. 2459–2468. URL:https://doi.org/10.18653/v1/2021.emnlp-main. 190, doi:10.18653/v1/2021.emnlp-main.190

    work page doi:10.18653/v1/2021.emnlp-main 2021

  42. [42]

    Qwen2.5 Technical Report

    Yang,A.,Yang,B.,Zhang,B.,Hui,B.,Zheng,B.,Yu,B.,Li,C.,Liu, D., Huang, F., Wei, H., et al., 2025. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115 URL:https://arxiv.org/abs/2412.15115, doi:10.48550/arXiv.2412.15115. Fodeh et al.:Preprint submitted to ElsevierPage 16 of 19 EPPC-OASIS with structured inference A. Appendix: Inference Details A.1. Pars...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2412.15115 2025