Pith. sign in

REVIEW 3 major objections 5 minor 71 references

A unified tokenization framework lets one shared transformer process facial video and fNIRS brain signals for pain recognition.

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 11:53 UTC pith:V5SBWTBP

load-bearing objection Solid Perceiver-style framework for video+fNIRS pain recognition, but the SOTA claim is a single 12-subject test-set point with no error bars and is not established. the 3 major comments →

arxiv 2607.19716 v2 pith:V5SBWTBP submitted 2026-07-22 cs.CV

A Unified Tokenization Framework for Pain Recognition using Heterogeneous 3D Modalities

Stefanos Gkikas , Christian Arzate Cruz , Valentina Becchetti , Muhammad Umar Khan , Alessandro Giuseppi , Raul Fernandez Rojas This is my paper
classification cs.CV MSC 68T0768T45
keywords pain recognitionmultimodal fusiontokenizationfNIRSfacial videodeep learningAI4Pain
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.

This paper claims that facial video and functional near-infrared spectroscopy (fNIRS) brain recordings can be processed by a single shared transformer model, without separate encoders or handcrafted features. The approach works by tokenizing each modality into a common token space using axis folding and Fourier position encodings, then letting learned latent vectors attend to the tokens. On the AI4Pain benchmark, the stacked video–fNIRS model reaches 57.33% test accuracy, the highest the authors report, while running in real time on CPU and GPU. A sympathetic reader would care because pain assessment today relies heavily on self-report or single-modality models; a unified pipeline could make continuous objective monitoring more practical.

Core claim

The paper's central claim is that heterogeneous 3D inputs — spatiotemporal facial video and multichannel fNIRS waveforms or spectrograms — can be flattened into one token sequence and processed by a single attention backbone with no modality-specific encoders and no handcrafted features. Using a segment-latent variant, the authors reach 57.33% accuracy on the AI4Pain test set by simple channel-wise stacking of 20 video frames with the fNIRS waveform, exceeding the previous best reported 55.69% while keeping inference real-time on GPU and CPU. The same backbone also works unimodally: 57.00% with video only and 53.67% with fNIRS spectrograms.

What carries the argument

The carrier is the tokenization pipeline: axis folding packs modality-specific dimensions into channels (video frames become a 2D grid; 1D waveforms are interpolated onto a 2D grid), then spatial axes are flattened into tokens with Fourier-feature positional encodings. A compact set of learned latent vectors acts as queries in cross-attention over the tokens and refines itself through self-attention, in either a global-latent or segment-latent configuration. Segment-latent splits the token sequence into S contiguous segments, each with its own latent state that attends locally and then exchanges information across segments, giving higher accuracy at lower cost than the global variant.

Load-bearing premise

The claim that 57.33% is state of the art assumes the 12-subject AI4Pain test set gives stable accuracy differences and that all compared methods used the same preprocessing and train/validation split; a 1.64-point gap over the prior best is within normal sampling noise.

What would settle it

Retrain the model and the strongest prior method on the same train/validation split with several random seeds, and also run leave-one-subject-out; if the proposed model does not consistently exceed the prior best by more than chance, or its accuracy drops below roughly 53% across seeds, the state-of-the-art claim is refuted. Re-running with a random re-split of the 65 subjects is the quickest check.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • A single model can serve both behavioral and brain-activity pain assessment, so a hospital or wearable system no longer needs separate architectures for video and physiological signals.
  • The consistent edge of fNIRS spectrograms over raw waveforms points to time–frequency structure as the informative part of the brain signal, guiding future signal-representation choices.
  • Simple channel-stack fusion outperforms more complex fusion modules on this data, suggesting modality integration can be trivial once tokenization is shared.
  • The model runs in real time on CPU, which would make continuous bedside or remote pain monitoring practical without specialized hardware.
  • Performance stays roughly flat as video frame count drops from 300 to 10, so cheap low-frame-rate capture suffices.

Where Pith is reading between the lines

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

  • Beyond the paper, the same axis-folding tokenization should transfer to other 3D physiological arrays (EEG, ECG, EDA) and to other behavioural states, because nothing in the method is pain-specific beyond the three-class output head.
  • Beyond the paper, the 57.33% figure rests on a 12-subject test set; a multi-seed replication or leave-one-subject-out evaluation would tell whether the 1.64-point margin over 55.69% is signal or noise.
  • Beyond the paper, the video-only 57.00% result hints that facial behaviour alone carries most of the pain signal in this dataset; a useful test is whether adding fNIRS still helps on populations where facial expression is suppressed or masked.

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 / 5 minor

Summary. The paper proposes a unified tokenization framework for heterogeneous 3D pain-recognition modalities. Inputs such as facial video volumes, raw fNIRS waveforms, and fNIRS PSD spectrograms are mapped to a common token space by folding modality-specific axes into channels, flattening spatial/temporal positions, and adding Fourier positional features. Two latent-transformer variants are evaluated: a global-latent Perceiver-style model and a segment-latent model that partitions the token sequence into contiguous segments. On the AI4Pain dataset, the authors report extensive validation ablations over 13 video frame counts, 6 segment counts, and fNIRS representations, and claim a state-of-the-art test accuracy of 57.33% for a simple channel-stacking fusion of video and fNIRS waveforms (Table 11), while also emphasizing real-time CPU/GPU inference.

Significance. If the main claim is substantiated, the contribution is useful: a single shared backbone that processes behavioral (video) and neurophysiological (fNIRS) data with no modality-specific encoders, a simple stack fusion, and broad ablation coverage. The validation study is unusually systematic in exploring tokenization resolution and computational cost, and the inference measurements are a helpful practical addition. However, the headline result—surpassing prior state of the art on AI4Pain—rests entirely on one test-set point estimate without uncertainty quantification, and the test set has only 12 subjects. The reported 1.64-point margin over the previous best is within plausible sampling noise, so the significance of the empirical claim is currently not established to a conventional standard.

major comments (3)
  1. [§5, Table 11] The central SOTA claim is not supported by uncertainty quantification. The AI4Pain test set contains 12 subjects, and the reported 57.33% versus PainFormer's 55.69% is a 1.64-point difference. With 12 independent subjects, the standard error of a proportion near 0.56 is about 14 points; even if trials per subject are pooled, subject-level correlations make the effective sample small. No confidence interval, bootstrap, per-subject breakdown, or repeated-seed variation is provided. The manuscript should add these; without them the claim that the framework 'achieves the highest reported performance' is fragile and could simply be sampling noise.
  2. [§4.3/Table 10 and §5] The model-selection procedure is inconsistent with the test report. On validation, the video–fNIRS-PSD stack achieves 50.93% accuracy, whereas the video–fNIRS-Wave stack achieves 47.96%; yet only the latter is reported on the test set (57.33%), with no test-set accuracy for the PSD fusion. The stated reason—a favorable accuracy–efficiency trade-off—does not follow from the validation accuracy, since PSD fusion is 2.97 points more accurate on validation. If the choice was made after seeing the test set, or if multiple configurations were tried and the best test number selected, the claim is post-hoc. Please report both test results, or specify a pre-registered selection rule and adhere to it.
  3. [§5, Table 11] The comparability of the baselines in Table 11 is not established. The table cites accuracy numbers from different papers ([12], [27], [57], [66], [38], [53], [22]) without showing that they all used the same test split (the 12-subject AI4Pain test set), the same metric (macro-averaged versus per-sample accuracy), and the same preprocessing (face detection, resizing, frame sampling, fNIRS channel selection). Because the SOTA claim rests on comparing a single number across methods, the authors should state for each baseline whether it was re-evaluated under identical conditions or cite the original protocol in enough detail to verify comparability. Also, the exact configuration for 'Our' rows (e.g., 57.00% video-only with '20 channels' and 53.67% fNIRS-PSD) is not traceable from Tables 3–10, so the test-set numbers need explicit configuration identifiers.
minor comments (5)
  1. [§3.2, Eq. (2)] The Fourier feature encoding depends on hyperparameters K and f_max, but these are not listed in Table 1 or anywhere else. Please state their values or provide a reference for the default.
  2. [§3.2/Table 10 and §5] The notation '#Chan.' is ambiguous for video. In Table 3 the column reports frame count, while the text says the effective channel count is frame count × 3 (RGB). In Table 10 and the footnote of Table 11, 'Video 20' and '20 channels' are unclear; use '20 frames' or report the actual channel dimension consistently.
  3. [§5, Table 11] The layout of the 'Our' rows in Table 11 is difficult to parse. A separate row per configuration with explicit modality checkmarks and a footnote column would improve readability.
  4. [Abstract] Typo: 'theAI4Painbenchmark' should be 'the AI4Pain benchmark'.
  5. [References] Reference [58] is incomplete: 'Wenli Qu et al. 2026' has no full author list, venue, or page/article number. Please complete the entry.

Circularity Check

0 steps flagged

No significant circularity: the central SOTA claim is a held-out test-set measurement, not a derivation reduced to its inputs; self-citations in Table 11 are not load-bearing.

full rationale

The paper's central claim — that the stacked video–fNIRS fusion reaches 57.33% on the AI4Pain test set — is an empirical measurement on a held-out test split after validation-based model selection. Section 5 states: "For the final test-set evaluation, the video–fNIRS waveform stack fusion was selected over the video–PSD combination, as it achieves a more favorable accuracy–efficiency trade-off on the validation set." This is standard model selection, not fitting the test labels. The architecture equations (1)–(13) define a generic tokenization and attention mechanism; no equation is fitted to the benchmark outcome, and the framework does not encode test labels in its construction. The comparison table includes externally authored baselines ([57], [53], [66]), so the state-of-the-art claim does not reduce to a self-citation chain even though [22] and [27] have overlapping authors. Concerns that the reported 57.33% is a single test-set point estimate with no confidence intervals on only 12 test subjects, or that validation-preferred video–fNIRS-PSD (50.93%) is not reported on the test set, are statistical robustness issues rather than circularity. Likewise, the Safe and Responsible Innovation Statement's caveat that the controlled dataset limits generalization is an acknowledged limitation, not a circular step. No pattern of self-definition, fitted-input-called-prediction, or imported uniqueness is present.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

This is an empirical deep-learning study, so its predictive power comes from trained weights and hand-set hyperparameters rather than a closed-form derivation. The only free constants in the analytic description are the Fourier encoding parameters K and f_max (not reported), the segment count S, the frame count, and the fusion choice. The main domain assumptions concern the dataset labels and the split. No new physical entities are introduced; segment-latent states are learned architectural components.

free parameters (5)
  • Segment count S = 32 (final config)
    Swept over {2,4,8,16,32,64} on validation; S=32 selected for fusion because it maximizes average accuracy (46.61%) at low cost.
  • Video frame count = 20 frames (final fusion)
    Validation sweep over 10–300 frames; the video+fNIRS-wave stack at 20 frames chosen for accuracy/efficiency trade-off on validation.
  • Fusion selection = video + fNIRS-waveform stack
    Chosen on validation over video+PSD and triple-stack for best accuracy–efficiency; affects the reported test SOTA.
  • Architecture hyperparameters = L=4, M=32, d=128, heads=1/8, R=8
    Table 1; fixed across all experiments and presumably tuned on validation; not justified in text.
  • Fourier feature bands K and max frequency f_max = not reported
    Eq. 2 introduces K and f_max but the paper never states their values; they control positional encoding and are part of the method.
axioms (5)
  • domain assumption AI4Pain labels (no/low/high pain from electrical-stimulation threshold/tolerance) are valid ground truth
    Section 4: pain threshold/tolerance protocol from [9,12]; if labels are noisy or subjective, all accuracies are affected.
  • domain assumption The 41/12/12 subject split is representative and the test set is a fair estimator of generalization
    Section 4; final model chosen on validation; assumes no distribution shift between validation and test subjects.
  • ad hoc to paper Packing modality-specific axes into the channel dimension and flattening preserves task-relevant structure
    Section 3.2; video frames are stacked as channels and spatial axes flattened; no proof that temporal/spectral dynamics needed for pain are retained.
  • ad hoc to paper Fourier features with unspecified K and f_max suffice for positional encoding
    Eq. 2; K and f_max are not reported; the choice is a hyperparameter, not derived.
  • domain assumption Deep-learning training on 41 subjects generalizes to held-out subjects
    Standard supervised-learning assumption; small training set makes this fragile.
invented entities (1)
  • Segment-latent states (e_s^(l)) no independent evidence
    purpose: Local latent bottlenecks that aggregate tokens segment-wise before cross-segment attention; the novel component of the architecture.
    Section 3.2.2. These are learned latent vectors; there is no external falsifiable prediction attached to them. Their only support is validation/test accuracy of the full model.

pith-pipeline@v1.3.0-alltime-deepseek · 19175 in / 15614 out tokens · 165033 ms · 2026-08-01T11:53:44.090282+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of A Unified Tokenization Framework for Pain Recognition using Heterogeneous 3D Modalities." pith.science (2026). https://pith.science/paper/V5SBWTBP

@misc{pith2026260719716,
  author       = {Pith},
  title        = {Pith review of: A Unified Tokenization Framework for Pain Recognition using Heterogeneous 3D Modalities},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V5SBWTBP}},
  note         = {Machine review of arXiv:2607.19716}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Pain is a complex and pervasive phenomenon affecting a large percentage of the population, and accurate assessment is essential for effective clinical management and intervention. Computational pain recognition systems enable continuous monitoring, support clinical decision-making, and help mitigate pain-related distress and functional decline. This study introduces a unified tokenization framework for heterogeneous 3D modalities in pain recognition that provides a single processing pipeline across behavioral and brain-activity 3D data, without requiring separate architectures for each modality or handcrafted inductive biases. The framework preserves spatial, temporal, and time--frequency structure while mapping diverse inputs into a shared token space. Extensive experiments show that the proposed approach effectively processes facial videos and fNIRS data in both raw-signal and spectrogram-based representations. On the AI4Pain benchmark dataset, the proposed framework achieves state-of-the-art performance while maintaining high computational efficiency and enabling real-time assessment on both GPU and CPU hardware.

Figures

Figures reproduced from arXiv: 2607.19716 by Alessandro Giuseppi, Christian Arzate Cruz, Muhammad Umar Khan, Raul Fernandez Rojas, Stefanos Gkikas, Valentina Becchetti.

Figure 1
Figure 1. Figure 1: Overview of the proposed tokenization framework, using videos and fNIRS. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of the Global-Latent and Segment￾Latent variants in terms of accuracy, GFLOPs, and GPU la￾tency for the video and fNIRS modalities. from 45.82% to 46.61%. Finally, we found that the maximum aver￾age value was attained at 32 segments (46.61%). Compared to the Global-Latent variant, the Segment-Latent tokenization consistently achieves higher performance while requiring lower computational and inf… view at source ↗

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

71 extracted references · 13 canonical work pages

  1. [1]

    Kamran Ali and Charles E. Hughes. 2025. A Unified Biosensor–Vision Multi- Modal Transformer network for emotion recognition.Biomedical Signal Process- ing and Control102 (2025), 107232. doi:10.1016/j.bspc.2024.107232

    arXiv 2025

  2. [2]

    Sumair Aziz, Calvin Joseph, Niraj Hirachan, Luke Murtagh, Girija Chetty, Roland Goecke, and Raul Fernandez-Rojas. 2025. A two-stage architecture for identifying and locating the source of pain using novel multi-domain binary patterns of EDA. Biomedical Signal Processing and Control104 (2025), 107454. doi:10.1016/j.bspc. 2024.107454

    arXiv 2025

  3. [3]

    Aleksandra Badura, Aleksandra Masłowska, Andrzej Myśliwiec, and Ewa Pietka

  4. [4]

    Ghazal Bargshady, Sumair Aziz, Stefanos Gkikas, Manolis Tsiknakis, Roland Goecke, and Raul Fernandez Rojas. 2025. Pain Assessment Using Multi-Kernel- FCN-LSTM and Haemoglobin Difference in fNIRS.ACM Trans. Comput. Health- care(2025). doi:10.1145/3757931

    work page doi:10.1145/3757931 2025

  5. [5]

    Ghazal Bargshady, Calvin Joseph, Niraj Hirachan, Roland Goecke, and Raul Fer- nandez Rojas. 2024. Acute Pain Recognition from Facial Expression Videos using Vision Transformers. In2024 46th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). 1–4. doi:10.1109/EMBC53108. 2024.10781616

    arXiv 2024

  6. [6]

    Ramsin Benyamin, Andrea M Trescot, Sukdeb Datta, Ricardo M Buenaventura, Rajive Adlaka, Nalini Sehgal, Scott E Glaser, and Ricardo Vallejo. 2008. Opioid complications and side effects.Pain physician11, 2S (2008), S105

    2008

  7. [7]

    Yanlong Chen, Mattia Orlandi, Pierangelo Maria Rapa, Simone Benatti, Luca Benini, and Yawei Li. 2025. PhysioWave: A Multi-Scale Wavelet-Transformer for Physiological Signal Representation. arXiv:2506.10351 [cs.LG]

    arXiv 2025

  8. [8]

    Jaleh Farmani, Ghazal Bargshady, Stefanos Gkikas, Manolis Tsiknakis, and Raul Fernandez Rojas. 2025. A CrossMod-Transformer deep learning framework for multi-modal pain detection through EDA and ECG fusion.Scientific Reports15, 1 (2025), 29467. doi:10.1038/s41598-025-14238-y

    doi:10.1038/s41598-025-14238-y 2025

  9. [9]

    Raul Fernandez Rojas, Niraj Hirachan, Nicholas Brown, Gordon Waddington, Luke Murtagh, Ben Seymour, and Roland Goecke. 2023. Multimodal physiological sensing for the assessment of acute pain.Frontiers in Pain Research4 (2023). doi:10.3389/fpain.2023.1150264

    arXiv 2023

  10. [10]

    Raul Fernandez Rojas, Calvin Joseph, Ghazal Bargshady, and Keng-Liang Ou

  11. [11]

    Raul Fernandez Rojas, Mingyu Liao, Julio Romero, Xu Huang, and Keng-Liang Ou. 2019. Cortical Network Response to Acupuncture and the Effect of the Hegu Point: An fNIRS Study.Sensors19, 2 (2019). doi:10.3390/s19020394

    work page doi:10.3390/s19020394 2019

  12. [13]

    Stefanos Gkikas. 2025. A Pain Assessment Framework based on multimodal data and Deep Machine Learning methods. arXiv:2505.05396 [cs.AI] https: //arxiv.org/abs/2505.05396 arXiv preprint arXiv:2505.05396

    Pith/arXiv arXiv 2025

  13. [14]

    Stefanos Gkikas., Chariklia Chatzaki., Elisavet Pavlidou., Foteini Verigou., Kyri- akos Kalkanis., and Manolis Tsiknakis. 2022. Automatic Pain Intensity Estimation based on Electrocardiogram and Demographic Factors.Proceedings of the 8th In- ternational Conference on Information and Communication Technologies for Ageing Well and e-Health - ICT4A WE,, 155–...

    work page doi:10.5220/0010971700003188 2022

  14. [15]

    Stefanos Gkikas, Chariklia Chatzaki, and Manolis Tsiknakis. 2023. Multi-task Neural Networks for Pain Intensity Estimation Using Electrocardiogram and Demographic Factors. InInformation and Communication Technologies for Ageing Well and e-Health. Springer Nature Switzerland, 324–337. doi:10.1007/978-3-031- 37496-8_17

    doi:10.1007/978-3-031- 2023

  15. [16]

    Stefanos Gkikas, Christian Arzate Cruz, Yu Fang, Lu Cao, Muhammad Umar Khan, Thomas Kassiotis, Giorgos Giannakakis, Raul Fernandez Rojas, and Randy Gomez. 2026. A Lightweight Transformer for Pain Recognition from Brain Activity. arXiv:2604.16491 [cs.CV]

    Pith/arXiv arXiv 2026

  16. [17]

    Stefanos Gkikas, Christian Arzate Cruz, Calvin Joseph, Giorgos Giannakakis, and Raul Fernandez Rojas. 2026. Towards a Unified Modality-Agnostic Multimodal Framework for Cognitive Workload Assessment. In2026 14th International Con- ference on Affective Computing and Intelligent Interaction (ACII)(Puebla, Mexico). IEEE

    2026

  17. [18]

    Stefanos Gkikas, Yu Fang, Christian Arzate Cruz, Muhammad Umar Khan, and Raul Fernandez Rojas. 2026. ReFace: Reorganizing Facial Spatiotemporal Repre- sentations for Improved Pain Assessment. In2026 14th International Conference on Affective Computing and Intelligent Interaction (ACII)(Puebla, Mexico). IEEE

    2026

  18. [21]

    Stefanos Gkikas, Ioannis Kyprakis, and Manolis Tsiknakis. 2025. Tiny-BioMoE: a Lightweight Embedding Model for Biosignal Analysis. InCompanion Proceedings of the 27th International Conference on Multimodal Interaction (ICMI Companion ’25). Association for Computing Machinery, New York, NY, USA, 117–126. doi:10. 1145/3747327.3764788

    arXiv 2025

  19. [22]

    Stefanos Gkikas, Raul Fernandez Rojas, and Manolis Tsiknakis. 2025. PainFormer: A Vision Foundation Model for Automatic Pain Assessment.IEEE Transactions on Affective Computing16, 4 (2025), 3369–3386. doi:10.1109/TAFFC.2025.3605475

    arXiv 2025

  20. [23]

    Tachos, Stelios Andreadis, Vasileios C

    Stefanos Gkikas, Nikolaos S. Tachos, Stelios Andreadis, Vasileios C. Pezoulas, Dimitrios Zaridis, George Gkois, Anastasia Matonaki, Thanos G. Stavropou- los, and Dimitrios I. Fotiadis. 2024. Multimodal automatic assessment of acute pain through facial videos and heart rate signals utilizing transformer-based architectures.Frontiers in Pain Research5 (2024...

    arXiv 2024

  21. [24]

    Stefanos Gkikas and Manolis Tsiknakis. 2023. Automatic assessment of pain based on deep learning methods: A systematic review.Computer Methods and Programs in Biomedicine231 (2023), 107365. doi:10.1016/j.cmpb.2023.107365

    arXiv 2023

  22. [25]

    Stefanos Gkikas and Manolis Tsiknakis. 2023. A Full Transformer-based Frame- work for Automatic Pain Estimation using Videos. In2023 45th Annual Interna- tional Conference of the IEEE Engineering in Medicine & Biology Society (EMBC). 1–6. doi:10.1109/EMBC40787.2023.10340872

    arXiv 2023

  23. [28]

    Dan Hendrycks, Norman Mu, Ekin D Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. 2019. Augmix: A simple data processing method to improve robustness and uncertainty.arXiv preprint arXiv:1912.02781(2019)

    Pith/arXiv arXiv 2019

  24. [29]

    Coyne, Elizabeth Ely, Céline Gélinas, and Renee C.B

    Keela Herr, Patrick J. Coyne, Elizabeth Ely, Céline Gélinas, and Renee C.B. Man- worren. 2019. Pain Assessment in the Patient Unable to Self-Report: Clinical Practice Recommendations in Support of the ASPMN 2019 Position Statement. Pain Management Nursing20, 5 (2019), 404–417. doi:10.1016/j.pmn.2019.07.005

    work page doi:10.1016/j.pmn.2019.07.005 2019

  25. [30]

    Dong Huang, Xiaoyi Feng, Haixi Zhang, Zitong Yu, Jinye Peng, Guoying Zhao, and Zhaoqiang Xia. 2022. Spatio-Temporal Pain Estimation Network With Mea- suring Pseudo Heart Rate Gain.IEEE Transactions on Multimedia24 (2022), 3300–3313. doi:10.1109/TMM.2021.3096080

    arXiv 2022

  26. [31]

    Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. 2021. Perceiver: General Perception with Iterative Attention. In Proceedings of the 38th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 139), Marina Meila and Tong Zhang (Eds.). PMLR, 4651–4664

    2021

  27. [32]

    Murugappan, D Manasa, and R Gayathri

    S Jerritta, M. Murugappan, D Manasa, and R Gayathri. 2022. Discrete Wavelet Transform based Pain Assessment using Multiple Physiological Signals. In2022 First International Conference on Electrical, Electronics, Information and Commu- nication Technologies (ICEEICT). 1–7. doi:10.1109/ICEEICT53079.2022.9768545

    arXiv 2022

  28. [33]

    Xinwei Ji, Tianming Zhao, Wei Li, and Albert Zomaya. 2023. Automatic Pain Assessment with Ultra-short Electrodermal Activity Signal. InProceedings of the 38th ACM/SIGAPP Symposium on Applied Computing(Tallinn, Estonia)(SAC ’23). Association for Computing Machinery, New York, NY, USA, 618–625. doi:10. 1145/3555776.3577721

    arXiv 2023

  29. [34]

    Lucille A Joel. 1999. The fifth vital sign: pain.AJN The American Journal of Nursing99, 2 (1999), 9

    1999

  30. [35]

    Thomas Kassiotis, Stefanos Gkikas, Nikolaos Smyrnis, and Giorgos Giannakakis

  31. [36]

    Joanna G Katzman and Rollin Mac Gallagher. 2024. Pain: The Silent Public Health Epidemic.Journal of Primary Care & Community Health15 (2024), 21501319241253547

    2024

  32. [37]

    Alan David Kaye, Mark R Jones, Adam M Kaye, Juan G Ripoll, Vincent Galan, Burton D Beakley, Francisco Calixto, Jamie L Bolden, Richard D Urman, and Laxmaiah Manchikanti. 2017. Prescription opioid abuse in chronic pain: an updated review of opioid abuse predictors and strategies to curb opioid abuse: part 1.Pain physician20, 2 (2017), S93

    2017

  33. [38]

    Muhammad Umar Khan, Sumair Aziz, Luke Murtagh, Girija Chetty, Roland Goecke, and Raul Fernandez Rojas. 2025. Empirically Transformed Energy Pat- terns: A novel approach for capturing fNIRS signal dynamics in pain assessment. Computers in Biology and Medicine192 (2025), 110300. doi:10.1016/j.compbiomed. 2025.110300

    arXiv 2025

  34. [39]

    Muhammad Umar Khan, Girija Chetty, Stefanos Gkikas, Manolis Tsiknakis, Roland Goecke, and Raul Fernandez-Rojas. 2026. GIAFormer: A Gradient-Infused Attention and Transformer for Pain Assessment with EDA-fNIRS Fusion.Infor- mation Fusion131 (2026), 104173. doi:10.1016/j.inffus.2026.104173

    arXiv 2026

  35. [40]

    Muhammad Umar Khan, Girija Chetty, Roland Goecke, and Raul Fernandez- Rojas. 2025. A Systematic Review of Multimodal Signal Fusion for Acute Pain Assessment Systems.ACM Comput. Surv.(2025). doi:10.1145/3737281

    work page doi:10.1145/3737281 2025

  36. [41]

    Muhammad Umar Khan, Maryam Sousani, Niraj Hirachan, Calvin Joseph, Maryam Ghahramani, Girija Chetty, Roland Goecke, and Raul Fernandez-Rojas

  37. [42]

    Tanvi Khera and Valluvan Rangasamy. 2021. Cognition and Pain: A Review. Frontiers in PsychologyVolume 12 - 2021 (2021). doi:10.3389/fpsyg.2021.673962

    arXiv 2021

  38. [43]

    Youngsun Kong and Ki H. Chon. 2024. Electrodermal activity in pain assessment and its clinical applications.Applied Physics Reviews11, 3 (08 2024), 031316. doi:10.1063/5.0200395

    doi:10.1063/5.0200395 2024

  39. [44]

    Ioannis Kyprakis, Stefanos Gkikas, Eric Nichols, Yu Fang, and Manolis Tsiknakis

  40. [45]

    doi:10.3390/s24020458

    Multilevel Pain Assessment with Functional Near-Infrared Spectroscopy: Evaluating ΔHBO2 and ΔHHB Measures for Comprehensive Analysis.Sensors 24, 2 (2024). doi:10.3390/s24020458

    work page doi:10.3390/s24020458 2024

  41. [46]

    Jia Li, Xiaohan Lyu, Xiaoyun Li, Xilin Yang, Lingling Weng, Yi Wang, and Weiwei Peng. 2025. Pain in focus: How persistent pain disrupts the attentional bias towards pain-related information.NeuroImage321 (2025), 121539. doi:10.1016/j. neuroimage.2025.121539

    arXiv 2025

  42. [47]

    Zequan Liang, Hao Luo, Xi Chen, Zhipeng Zhong, Cheng Fan, Xingrong Song, Bilian Li, and Jianming Lv. 2025. A Multi-Modal Multi-Expert Framework for Pain Assessment in Postoperative Children.IEEE Transactions on Affective Computing 16, 4 (2025), 2828–2841. doi:10.1109/TAFFC.2025.3567307

    arXiv 2025

  43. [48]

    Zhenyuan Lu, Burcu Ozek, and Sagar Kamarthi. 2023. Transformer encoder with multiscale deep learning for pain classification using physiological signals. Frontiers in Physiology14 (2023). doi:10.3389/fphys.2023.1294577

    arXiv 2023

  44. [49]

    arXiv:2601.00000 [cs.LG]

    An Exploratory Analysis of Pain Localization via Explainable Computa- tional Modeling. arXiv:2601.00000 [cs.LG]

    arXiv

  45. [50]

    JiaHao Li, JinCheng Luo, YanSheng Wang, YunXiang Jiang, Xu Chen, and YuJuan Quan. 2025. Automatic Pain Assessment Based on Physiological Signals: Applica- tion of Multi-Scale Networks and Cross-Attention Cross-Attention. InProceedings of the 2024 13th International Conference on Bioinformatics and Biomedical Science (ICBBS ’24). Association for Computing ...

    arXiv 2025

  46. [51]

    Moore, Samantha M

    David J. Moore, Samantha M. Meints, Asimina Lazaridou, Devin Johnson, Olivia Franceschelli, Marise Cornelius, Kristin Schreiber, and Robert. R. Edwards. 2019. The Effect of Induced and Chronic Pain on Attention.The Journal of Pain20, 11 (2019), 1353–1361. doi:10.1016/j.jpain.2019.05.004

    work page doi:10.1016/j.jpain.2019.05.004 2019

  47. [52]

    Müller and Frank Hutter

    Samuel G. Müller and Frank Hutter. 2021. TrivialAugment: Tuning-free Yet State-of-the-Art Data Augmentation. In2021 IEEE/CVF International Conference on Computer Vision (ICCV). 754–762. doi:10.1109/ICCV48922.2021.00081

    arXiv 2021

  48. [53]

    Minh-Duc Nguyen, Hyung-Jeong Yang, Soo-Hyung Kim, Ji-Eun Shin, and Seung- Won Kim. 2024. Transformer with Leveraged Masked Autoencoder for video- based Pain Assessment. arXiv:2409.05088 [cs.CV]

    Pith/arXiv arXiv 2024

  49. [54]

    2024.The pain phenomenon

    Serge Marchand. 2024.The pain phenomenon. Vol. 13. Springer Nature

    2024

  50. [55]

    DA Meehan, ME McRae, DA Rourke, C Eisenring, and FA Imperial. 1995. Anal- gesic administration, pain intensity, and patient satisfaction in cardiac surgical patients.American Journal of Critical Care4, 6 (1995), 435–442

    1995

  51. [56]

    Kim Ngan Phan, Ngumimi Karen Iyortsuun, Sudarshan Pant, Hyung-Jeong Yang, and Soo-Hyung Kim. 2023. Pain Recognition With Physiological Signals Using Multi-Level Context Information.IEEE Access11 (2023), 20114–20127. doi:10. 1109/ACCESS.2023.3248654

    arXiv 2023

  52. [57]

    Pooja Prajod, Dominik Schiller, Daksitha Withanage Don, and Elisabeth André

  53. [58]

    Wenli Qu et al. 2026. Symmetry-Aware Structured Representation Learning for Unified Multi-Modal Physiological Modeling in Affective State and Preference Inference. (2026)

    2026

  54. [59]

    Patil and Hitendra D

    Manisha S. Patil and Hitendra D. Patil. 2024. Ensemble Neural Networks for Mul- timodal Acute Pain Intensity Evaluation using Video and Physiological Signals. Journal of Computational Analysis and Applications (JoCAAA)33, 05 (Sep. 2024), 779–791

    2024

  55. [60]

    Elisavet Pavlidou and Manolis Tsiknakis. 2025. Multimodal Pain Assessment Based on Physiological Biosignals: The Impact of Demographic Factors on Per- ception and Sensitivity. InProceedings of the 11th International Conference on Information and Communication Technologies for Ageing Well and e-Health - ICT4A WE. INSTICC, SciTePress, 320–329. doi:10.5220/0...

    work page doi:10.5220/0013426800003938 2025

  56. [61]

    Heng Shi, Belkacem Chikhaoui, and Shengrui Wang. 2022. Tree-Based Models for Pain Detection from Biomedical Signals. InParticipative Urban Health and Healthy Aging in the Age of AI. Springer International Publishing, 183–195

    2022

  57. [62]

    Rolf A. H. Snijders, Linda Brom, Maurice Theunissen, and Marieke H. J. van den Beuken-van Everdingen. 2023. Update on Prevalence of Pain in Patients with Cancer 2022: A Systematic Literature Review and Meta-Analysis.Cancers15, 3 (2023). doi:10.3390/cancers15030591

    work page doi:10.3390/cancers15030591 2023

  58. [63]

    In2024 12th International Conference on Affective Computing and Intelligent Interaction Workshops and Demos (ACIIW)

    Faces of Experimental Pain: Transferability of Deep-Learned Heat Pain Features to Electrical Pain*. In2024 12th International Conference on Affective Computing and Intelligent Interaction Workshops and Demos (ACIIW). 31–38. doi:10.1109/ACIIW63320.2024.00009

    arXiv 2024

  59. [64]

    Mingze Sun, Haoxiang Wang, Bilian Li, Jianming Lv, Guancheng Yao, and Yuhao Han. 2025. Reconstruction and fusion: Using pseudo physiological modality for pain recognition.Engineering Applications of Artificial Intelligence142 (2025), 109905. doi:10.1016/j.engappai.2024.109905

    arXiv 2025

  60. [65]

    Raul Fernandez Rojas, Julio Romero, Jehu Lopez-Aparicio, and Keng-Liang Ou

  61. [66]

    In2021 10th International IEEE/EMBS Conference on Neural Engineering (NER)

    Pain Assessment based on fNIRS using Bi-LSTM RNNs. In2021 10th International IEEE/EMBS Conference on Neural Engineering (NER). 399–402. doi:10. 1109/NER49283.2021.9441384

    arXiv 2021

  62. [67]

    Vivian Santiago. 2022. Painful Truth: The Need to Re-Center Chronic Pain on the Functional Role of Pain.Journal of Pain Research15 (2022), 497–512. doi:10.2147/JPR.S347780

    work page doi:10.2147/jpr.s347780 2022

  63. [68]

    Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. 2016. Joint face detection and alignment using multitask cascaded convolutional networks.IEEE signal processing letters23, 10 (2016), 1499–1503

    2016

  64. [69]

    Ruicong Zhi and Junwei Yu. 2019. Multi-modal Fusion Based Automatic Pain Assessment. In2019 IEEE 8th Joint International Information Technology and Artificial Intelligence Conference (ITAIC). 1378–1382. doi:10.1109/ITAIC.2019. 8785727

    work page doi:10.1109/itaic.2019 2019

  65. [70]

    Argyrios Stampas, Claudia Pedroza, Jennifer N Bush, Adam R Ferguson, John L Kipling Kramer, and Michelle Hook. 2020. The first 24 h: opioid administration in people with spinal cord injury and neurologic recovery.Spinal Cord58, 10 (2020), 1080–1089

    2020

  66. [72]

    Kestler, and Friedhelm Schwenker

    Patrick Thiam, Peter Bellmann, Hans A. Kestler, and Friedhelm Schwenker. 2019. Exploring deep physiological models for nociceptive pain recognition.Sensors 19 (10 2019), 4503. Issue 20. doi:10.3390/s19204503

    work page doi:10.3390/s19204503 2019

  67. [73]

    Jo Vianto, Anjitha Divakaran, Hyungjeong Yang, Soonja Yeom, Seungwon Kim, Soohyung Kim, and Jieun Shin. 2025. Multimodal Model for Automated Pain Assessment: Leveraging Video and fNIRS.Applied Sciences15, 9 (2025). doi:10. 3390/app15095151

    2025

  68. [74]

    Chaoqi Yang, M Westover, and Jimeng Sun. 2023. BIOT: Biosignal Transformer for Cross-data Learning in the Wild. InAdvances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine (Eds.), Vol. 36. Curran Associates, Inc., 78240–78260

    2023

  69. [2021]

    doi:10.3390/s21041311

    Multimodal Signal Analysis for Pain Recognition in Physiotherapy Using Wavelet Scattering Transform.Sensors21, 4 (2021). doi:10.3390/s21041311

    work page doi:10.3390/s21041311 2021

  70. [2024]

    Frontiers in Neuroinformatics(2024)

    Empirical comparison of deep learning models for fNIRS pain decoding. Frontiers in Neuroinformatics(2024). doi:10.3389/fninf.2024.1320189

    arXiv 2024

  71. [2026]

    In2026 14th International Conference on Affective Computing and Intelligent Interaction (ACII)(Puebla, Mexico)

    Explainable Graph Attention Network for Stress Recognition (StressGAT) via Differential Action Units. In2026 14th International Conference on Affective Computing and Intelligent Interaction (ACII)(Puebla, Mexico). IEEE