Pith. sign in

REVIEW 3 major objections 4 minor 36 references

DynaBridge: Dynamic Summary-Guided Cross-Task Multimodal Fusion for DASS-Structured Mental Health Assessment

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A multimodal model that exploits the fixed DASS-21 item-to-subscale structure beats generic fusion baselines on adolescent depression, anxiety, and stress prediction, reaching 0.5012 mean F1 for risk and 0.3216 mean QWK for item responses o

desk verdict Worth a look for the DASS-structured cross-task idea, but the validation-only evaluation means the headline gains shouldn't be taken at face value. read the letter →

arxiv 2607.25679 v1 pith:K2WEVPIP submitted 2026-07-28 cs.LG cs.AIcs.MM

Shiyu Teng , Haichen Yu , Jiaqing Liu , Hao Sun , Yu Song , Shurong Chai , Ruibo Hou , Lanfen Lin
show 1 more author
Yen-Wei Chen
This is my paper
classification cs.LGcs.AIcs.MM
keywords multimodalmentalhealthassessmentDASS-21ordinalitempredictioncross-tasklearningLLM-generatedsummariesitem-to-riskreconstructiondepressionanxietystressfusion
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

DynaBridge argues that multimodal mental health assessment should not treat risk labels and questionnaire items as independent outputs. Instead, it reconstructs depression, anxiety, and stress risk from ordinal item-level predictions using the fixed DASS-21 item-to-subscale mapping, then fuses that structured evidence with a direct multimodal risk prediction. The framework also injects frozen-LLM-generated DASS-aware summaries as participant-level semantic evidence and refines item predictions only when the summary is confident. On the official AdoDAS validation split, this yields 0.5012 mean F1 for D/A/S risk prediction and 0.3216 mean QWK for DASS-21 item prediction, improving on the official baseline and representative multimodal methods. The paper's central claim is that exploiting the psychometric structure of questionnaire-derived labels is a useful inductive bias for scalable, behavior-based screening.

What carries the argument

The load-bearing object is the DASS-21 item-to-subscale mapping—fixed item groups D={3,5,10,13,16,17,21}, A={2,4,7,9,15,19,20}, S={1,6,8,11,12,14,18}—which DynaBridge uses to reconstruct risk evidence from predicted item distributions. For each item, a softmax head predicts a distribution over four ordered levels; expected scores are summed per subscale, doubled, and sigmoid-calibrated (with dimension-specific a_r, b_r) to form a reconstructed risk probability. That reconstruction is interpolated with a direct risk prediction via fusion weight lambda_r, and both branches are supervised jointly with a consistency loss. At inference, a confidence-aware refinement gate (requiring summary confid

What would settle it

Evaluate DynaBridge on the hidden test split or an independent cohort and compare against the official baseline and the summary-enhanced fusion variant; if the mean F1 or QWK advantage does not persist, the central claim fails. A more direct structural test: retrain with the DASS-21 item groups randomly shuffled (breaking the psychometric mapping); if performance does not drop below the intact-mapping model, the claimed benefit of the item-to-risk reconstruction is not real.

Watch

Extended reading notes

Core claim

DynaBridge establishes that a structured cross-task formulation of DASS assessment—predicting ordinal item distributions, reconstructing subscale risk via the DASS-21 scoring rule (subscale score = 2 × sum of item scores), and combining that reconstruction with direct multimodal risk prediction—improves both risk-level and item-level accuracy over generic fusion. It further shows that frozen LLMs can serve as safe auxiliary evidence extractors when prompted to summarize only transcript-supported cues and never access ground-truth labels. On the official subject-disjoint validation split, the method achieves 0.5012 mean F1 for D/A/S risk prediction and 0.3216 mean QWK for DASS-21 item predict

Load-bearing premise

The reported comparisons are measured on the same validation split used to select calibration parameters (a_r, b_r), fusion weights (lambda_r), refinement thresholds (tau_s, tau_m, alpha_max), loss weights, and decision thresholds; the paper assumes these validation-selected hyperparameters yield an unbiased performance estimate, so if they overfit this split, the claimed consistent improvement over baselines may not transfer to new participants.

Editorial extensions

If this is right

  • The DASS-21 item-to-subscale structure is a transferable inductive bias: other questionnaire-based assessments with fixed item-to-scale mappings could benefit from the same reconstruction logic.
  • Frozen LLMs can contribute semantic evidence to multimodal screening without leaking labels, provided they are constrained to evidence extraction and their outputs are gated conservatively.
  • Ordinal item prediction improves risk-level accuracy downstream, so joint training of item and risk tasks is better than training them independently.
  • The confidence-gated refinement mechanism prevents hallucinated semantic cues from overriding confident behavioral predictions, a practical safeguard for LLM-assisted assessment.
  • The validation gains suggest potential for the hidden test set, though the paper stops short of claiming generalization beyond the development split.

Reading between the lines

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

  • The reconstruction branch makes risk predictions decomposable into item-level evidence, which a clinician could inspect to understand why a participant is flagged as elevated risk.
  • If the item-to-subscale mapping is a genuine source of signal, then randomly perturbing that mapping during training should hurt performance; this is a direct, testable falsification.
  • The large auxiliary-metadata oracle gap suggests contextual factors outside the behavioral stream carry substantial signal; test-time-available proxies (school/class context, summaries) are only a partial substitute.
  • Because calibration parameters, fusion weights, refinement thresholds, and decision thresholds are all selected on the validation split, the reported margins may shrink on new cohorts; independent-cohort evaluation would clarify the method's robustness.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper presents DynaBridge, a multimodal framework for DASS-21-based mental health assessment. It integrates acoustic, visual, textual, and frozen-LLM-generated semantic summaries, and enforces psychometric structure through ordinal item prediction and item-to-risk reconstruction. On the official AdoDAS validation split, the authors report mean F1 of 0.5012 for D/A/S risk prediction and mean QWK of 0.3216 for item prediction, claiming improvements over the official baseline, CubeMLP, and a summary-enhanced fusion method.

Significance. The core idea—exploiting the DASS-21 item-to-subscale mapping to couple item-level and risk-level predictions—is timely and plausible. The design includes careful leakage controls: the LLM summaries are generated without access to labels or model predictions, and the auxiliary metadata oracle is explicitly isolated. However, the empirical evidence is compromised by the evaluation protocol: all headline results come from the same validation split used for model selection, with no error bars or significance tests. If the performance holds under a proper out-of-sample evaluation, the contribution would be interesting; as it stands, the central claim of superiority is not robustly supported.

major comments (3)
  1. [Section 4.2, Table 1] The headline results are measured on the official validation split, but Section 4.2 states that calibration parameters a_r,b_r (Eq. 16), fusion weights λ_r (Eq. 18), refinement thresholds τ_s,τ_m,α_max (Eq. 23), loss weights, and decision thresholds are all selected on that same validation split. Hence the reported numbers are in-sample estimates and likely optimistically biased. The manuscript acknowledges this in Section 4.2 ('validation results are reported as development-set evidence'), but the abstract and contributions present these numbers as the main evidence of superiority. Please evaluate on a truly held-out test set (e.g., by obtaining official test labels) or perform model selection on a subset of the training set and then use the validation set as the final test. At minimum, report bootstrap confidence intervals or results across multiple seeds to quantify variability.
  2. [Section 4.3] The comparison with representative methods (CubeMLP, summary-enhanced fusion) may be unfair if those baselines were not tuned on the validation split with the same selection protocol. The paper does not specify how baseline hyperparameters were chosen. If baselines used default hyperparameters while DynaBridge was tuned on the validation set, part of the gain may reflect tuning effort rather than architectural benefit. Please describe the hyperparameter selection procedure for all methods, and if possible, subject the baselines to the same validation-based oracle selection.
  3. [Section 3.6, Eq. 23] The confidence-aware refinement step uses thresholds τ_s and τ_m and the interpolation bound α_max, all selected on the validation split. The refined item distributions are then used to compute the reported A2 QWK on the same split. This double use of the validation set for both threshold selection and final evaluation may overstate the benefit of the refinement module. A separate development split for threshold selection or a sensitivity analysis (e.g., QWK as a function of τ_s, τ_m, α_max) is needed to establish that the gains are not due to overfitting these hyperparameters.
minor comments (4)
  1. [Abstract and Section 4.2] The abstract says 'DynaBridge outperforms the official baseline and representative multimodal methods, achieving 0.5012 mean F1...'. This is stated without qualification, despite the later caveat that these are development-set results. The abstract should explicitly say 'on the validation split' and note that this is not a test-set claim.
  2. [Section 3.2, Figure 1] The term 'dynamic' in DynaBridge is not fully justified; the three-stage summary generation is a fixed progressive protocol, and the summaries are cached and static during training. Clarify what is 'dynamic' beyond the sequential summarization steps.
  3. [Table 2] The auxiliary metadata oracle row is clearly labeled as an upper bound, but it may still mislead readers. Consider moving it to a separate discussion or adding a remark that it uses fields unavailable at test time and is not part of the final system.
  4. [Equation (22)] The consistency loss is written as ||p_direct - p_recon||_2^2. Ensure the notation distinguishes between the squared L2 norm and the square of the L2 norm; the current form is acceptable but could be clearer.

Circularity Check

0 steps flagged · score 0.0 of 10

No definitional or self-citation circularity: the DASS mapping is external, LLM summaries are generated without label access, and self-citations are not load-bearing. The validation-split tuning is a statistical limitation that the paper acknowledges, not a circular derivation.

full rationale

The paper's derivation chain is not circular. The item-to-subscale mapping in Eq. (4) is the externally published DASS-21 scoring structure (Antony et al. 1998; Lovibond & Lovibond 1995), and Eq. (15) applies that same public mapping with the official ×2 factor to the model's own soft item predictions; the model is not defining the mapping, and the DASS risk labels are generated from the same external instrument, so using that mapping is a domain constraint rather than an output of this paper. The reconstruction branch (Eqs. 16–18) contains calibration parameters a_r,b_r and fusion weights λ_r that are learned or selected, not structural identities smuggled in as predictions. The LLM summary branch is explicitly frozen and "never accesses ground-truth DASS-21 item scores, D/A/S risk labels, model predictions, or validation feedback" (Section 3.2), and summaries are cached before neural training, so the semantic evidence does not encode the target labels. Self-citations, including [9,28–31], appear in related work and as a comparison baseline ([31]), but no load-bearing claim is reduced to those citations: the three-stage prompt protocol, the ordinal item prediction head, the item-to-risk reconstruction, and the confidence-aware refinement are specified in this paper and are empirically compared with that baseline. The main methodological caveat is not circularity: Section 4.2 says calibration parameters, fusion weights, refinement thresholds, and decision thresholds are "selected on the validation split," and Table 1 reports on that same split. This makes the validation numbers optimistically biased development-set evidence; the paper itself states that "validation results are reported as development-set evidence rather than as a claim of hidden-test generalization." That is an evaluation-protocol limitation and a reason to be cautious about the absolute numbers, but it is not a case of a prediction being equivalent by construction to its fitted input. No circular step can be exhibited from the paper's equations or citations.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the DASS-21 scoring structure (external standard), the assumption that LLM summaries faithfully encode transcript evidence without label leakage, and the assumption that validation-selected hyperparameters transfer. These are domain assumptions rather than free parameters in the derivation; the fitted quantities are calibration and fusion weights.

free parameters (5)
  • a_r, b_r (D/A/S calibration) = not reported
    Sigmoid calibration parameters in Eq. 16 mapping reconstructed subscale score to risk probability; fitted on validation labels.
  • lambda_r (risk fusion weights) = not reported
    Dimension-specific interpolation weights in Eq. 18; selected on validation.
  • loss weights lambda_risk, lambda_item, lambda_recon, lambda_cons, gamma = not reported
    Weights in Eq. 19 and Eq. 21; chosen by hand/validation.
  • refinement thresholds tau_s, tau_m, alpha_max = not reported
    Confidence gate and interpolation bound in Eq. 23; selected on validation.
  • binary decision thresholds for D/A/S = not reported
    Fixed dimension-specific thresholds used to convert fused probabilities to labels; chosen on validation.
assumptions (4)
  • standard math DASS-21 item-to-subscale grouping (Eq. 4) and the factor-of-2 scoring rule (Eq. 15)
    External psychometric definition; assumed correct for AdoDAS labels.
  • domain assumption Monotone sigmoid mapping from subscale score to binary risk (Eq. 16)
    Assumes a single calibration curve per dimension; if the actual label threshold depends on covariates, reconstruction branch is misspecified.
  • domain assumption Frozen LLM summaries do not leak ground-truth labels or hallucinate item evidence
    Summaries are generated before training with fixed prompts and no label access, but the paper provides no audit of the generated summaries for leakage or hallucination.
  • domain assumption Validation split performance is an unbiased estimator of test performance
    Reported results are on the same split used for hyperparameter selection, so this assumption is questionable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DynaBridge: Dynamic Summary-Guided Cross-Task Multimodal Fusion for DASS-Structured Mental Health Assessment." pith.science (2026). https://pith.science/paper/K2WEVPIP

@misc{pith2026260725679,
  author       = {Pith},
  title        = {Pith review of: DynaBridge: Dynamic Summary-Guided Cross-Task Multimodal Fusion for DASS-Structured Mental Health Assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K2WEVPIP}},
  note         = {Machine review of arXiv:2607.25679}
}
read the original abstract

Multimodal behavioral analysis offers a scalable approach to assessing depression, anxiety, and stress, yet generic fusion models often ignore the psychometric structure of questionnaire labels. In DASS-21, risk labels are derived from ordered symptom items through fixed item-to-subscale mappings. We propose \textbf{DynaBridge}, a dynamic summary-guided cross-task multimodal framework for DASS-structured mental health assessment. DynaBridge encodes acoustic, visual, and textual cues across multiple sessions and augments them with frozen-LLM-generated DASS-aware summaries as participant-level semantic evidence. It predicts ordinal item distributions, reconstructs depression, anxiety, and stress risk evidence from item-level soft scores, and fuses this evidence with direct multimodal risk predictions. A confidence-aware refinement strategy further incorporates high-confidence semantic cues conservatively. On the official AdoDAS validation split, DynaBridge outperforms the official baseline and representative multimodal methods, achieving 0.5012 mean F1 for D/A/S risk prediction and 0.3216 mean QWK for DASS-21 item prediction. These results show the value of bridging multimodal cues, semantic summaries, and DASS-21 psychometric structure.

Figures

Figures reproduced from arXiv: 2607.25679 by the authors.

Figure 1
Figure 1. Overview of the proposed DynaBridge framework. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 3 canonical work pages

  1. [1]

    Antony, Peter J

    Martin M. Antony, Peter J. Bieling, Brian J. Cox, Murray W. Enns, and Richard P. Swinson. 1998. Psychometric Properties of the 42-Item and 21-Item Versions of DynaBridge: Dynamic Summary-Guided Cross-Task Multimodal Fusion for DASS-Structured Mental Health Assessment Conference’26, November 2026, Rio de Janeiro, Brazil the Depression Anxiety Stress Scales...

    doi:10.1037/1040-3590.10.2.176 1998

  2. [2]

    Zico Kolter, and Vladlen Koltun

    Shaojie Bai, J. Zico Kolter, and Vladlen Koltun. 2018. An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling.arXiv preprint arXiv:1803.01271(2018)

    arXiv 2018

  3. [3]

    Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al . 2020. Language Models are Few-Shot Learners. InAdvances in Neural Information Processing Systems, Vol. 33. 1877–1901

    2020

  4. [4]

    Jacob Cohen. 1968. Weighted kappa: Nominal scale agreement with provision for scaled disagreement or partial credit.Psychological Bulletin70, 4 (1968), 213–220

    1968

  5. [5]

    Shubham Dham, Anirudh Sharma, and Abhinav Dhall. 2017. Depression Scale Recognition from Audio, Visual and Text Analysis.arXiv preprint arXiv:1709.05865 (2017). doi:10.48550/arXiv.1709.05865

    work page Pith review arXiv doi:10.48550/arxiv.1709.05865 2017

  6. [6]

    Jonathan Gratch, Ron Artstein, Gale Lucas, Giota Stratou, Stefan Scherer, An- gela Nazarian, Rachel Wood, Jill Boberg, David DeVault, Stacy Marsella, David Traum, Skip Rizzo, and Louis-Philippe Morency. 2014. The Distress Analysis Interview Corpus of Human and Computer Interviews. InProceedings of the Ninth International Conference on Language Resources a...

    2014

  7. [7]

    Devamanyu Hazarika, Roger Zimmermann, and Soujanya Poria. 2020. MISA: Modality-Invariant and -Specific Representations for Multimodal Sentiment Analysis. InProceedings of the 28th ACM International Conference on Multime- dia. Association for Computing Machinery, New York, NY, USA, 1122–1131. doi:10.1145/3394171.3413678

    arXiv 2020

  8. [8]

    Kaiming He, Xinlei Chen, Saining Xie, et al. 2022. Masked Autoencoders Are Scalable Vision Learners. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2022

Show all 36 references
  1. [9]

    Ruibo Hou, Shiyu Teng, Jiaqing Liu, Shurong Chai, Yinhao Li, Lanfen Lin, and Yen-Wei Chen. 2025. Retrieval-Augmented Multimodal Depression Detection. In2025 47th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). IEEE, 1–5. doi:10.1...

    2025

  2. [10]

    Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, et al. 2021. HuBERT: Self- Supervised Speech Representation Learning by Masked Prediction of Hidden Units.IEEE/ACM Transactions on Audio, Speech, and Language Processing29 (2021), 3451–3460

    2021

  3. [11]

    Maximilian Ilse, Jakub Tomczak, and Max Welling. 2018. Attention-based Deep Multiple Instance Learning. InInternational Conference on Machine Learning

    2018

  4. [12]

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of Hallucination in Natural Language Generation.Comput. Surveys55, 12 (2023), 1–38. doi:10.1145/ 3571730

    2023

  5. [13]

    Zhun Liu, Ying Shen, Varun Bharadhwaj Lakshminarasimhan, Paul Pu Liang, AmirAli Bagher Zadeh, and Louis-Philippe Morency. 2018. Efficient Low-rank Multimodal Fusion with Modality-Specific Factors. InProceedings of the 56th Annual Meeting of the Association for Computational Li...

    2018

  6. [14]

    Ilya Loshchilov and Frank Hutter. 2019. Decoupled Weight Decay Regularization. InInternational Conference on Learning Representations

    2019

  7. [15]

    S. H. Lovibond and P. F. Lovibond. 1995.Manual for the Depression Anxiety Stress Scales(2 ed.). Psychology Foundation of Australia, Sydney, Australia

    1995

  8. [16]

    OpenAI. 2023. GPT-4 Technical Report.arXiv preprint arXiv:2303.08774(2023). doi:10.48550/arXiv.2303.08774

    2023 doi arXiv

  9. [17]

    OpenAI. 2024. New Embedding Models and API Updates. https://openai.com/ index/new-embedding-models-and-api-updates/

    2024

  10. [18]

    Adam Paszke, Sam Gross, Francisco Massa, et al . 2019. PyTorch: An Imper- ative Style, High-Performance Deep Learning Library. InAdvances in Neural Information Processing Systems

    2019

  11. [19]

    Syed Arbaaz Qureshi, Mohammed Hasanuzzaman, Sriparna Saha, and Gaël Dias

  12. [20]

    Wasifur Rahman, Md Kamrul Hasan, Sangwu Lee, AmirAli Bagher Zadeh, Chengfeng Mao, Louis-Philippe Morency, and Ehsan Hoque. 2020. Integrat- ing Multimodal Information in Large Pretrained Transformers. InProceedings of the 58th Annual Meeting of the Association for Computational...

    2020 doi

  13. [21]

    Fabien Ringeval, Björn Schuller, Michel Valstar, Nicholas Cummins, Roddy Cowie, Leili Tavabi, Maximilian Schmitt, Sina Alisamir, Shahin Amiriparian, Eva-Maria Messner, Siyang Song, Shuo Liu, Ziping Zhao, Adria Mallol-Ragolta, Zhao Ren, Mohammad Soleymani, and Maja Pantic. 2019...

    2019

  14. [22]

    Stepanov, Stéphane Lathuilière, Shammur Absar Chowdhury, Arindam Ghosh, Radu-Laurentiu Vieriu, Nicu Sebe, and Giuseppe Riccardi

    Evgeny A. Stepanov, Stéphane Lathuilière, Shammur Absar Chowdhury, Arindam Ghosh, Radu-Laurentiu Vieriu, Nicu Sebe, and Giuseppe Riccardi. 2018. Depres- sion Severity Estimation from Multiple Modalities. InProceedings of the 2018 IEEE International Conference on E-health Netwo...

    2018

  15. [23]

    Hao Sun, Yen-Wei Chen, and Lanfen Lin. 2023. TensorFormer: A Tensor-Based Multimodal Transformer for Multimodal Sentiment Analysis and Depression Detection.IEEE Transactions on Affective Computing14, 4 (2023), 2776–2786. doi:10.1109/TAFFC.2022.3233070

    2023

  16. [24]

    Hao Sun, Hongyi Wang, Jiaqing Liu, Yen-Wei Chen, and Lanfen Lin. 2022. Cube- MLP: An MLP-based Model for Multimodal Sentiment Analysis and Depression Estimation. InProceedings of the 30th ACM International Conference on Multi- media. Association for Computing Machinery, New Yo...

    2022

  17. [25]

    Shiyu Teng, Shurong Chai, Jiaqing Liu, Tomoko Tateyama, Lanfen Lin, and Yen-Wei Chen. 2024. Multi-Modal and Multi-Task Depression Detection with Sentiment Assistance. In2024 IEEE International Conference on Consumer Electron- ics (ICCE) (Digest of Technical Papers - IEEE Inter...

    2024

  18. [26]

    Shiyu Teng, Jiaqing Liu, Shurong Chai, Hao Sun, Tomoko Tateyama, Lanfen Lin, and Yen-Wei Chen. 2026. Enhancing Depression Detection Using Pre-trained Multi-modal Sentiment Analysis Models with Deep Prefix Tuning.ACM Trans- actions on Computing for Healthcare(2026). doi:10.1145/3788672

    2026 doi

  19. [27]

    Shiyu Teng, Jiaqing Liu, Yue Huang, Shurong Chai, Tomoko Tateyama, Xinyin Huang, Lanfen Lin, and Yen-Wei Chen. 2024. An Intra- and Inter-Emotion Transformer-Based Fusion Model with Homogeneous and Diverse Constraints Using Multi-Emotional Audiovisual Features for Depression De...

    2024

  20. [28]

    Shiyu Teng, Jiaqing Liu, Rahul Kumar Jain, Shurong Chai, Ruibo Hou, Tomoko Tateyama, Lanfen Lin, and Yen-Wei Chen. 2025. Enhancing Depression Detection with Chain-of-Thought Prompting: From Emotion to Reasoning Using Large Language Models. In2025 47th Annual International Conf...

    2025

  21. [29]

    Shiyu Teng, Jiaqing Liu, Hao Sun, Shurong Chai, Tomoko Tateyama, Lanfen Lin, and Yen-Wei Chen. 2025. Enhanced Multimodal Depression Detection With Emotion Prompts. In2025 IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP). IEEE. doi:10.1109/ICAS...

    2025

  22. [30]

    Shiyu Teng, Jiaqing Liu, Hao Sun, Yue Huang, Rahul Kumar Jain, Shurong Chai, Ruibo Hou, Tomoko Tateyama, Lanfen Lin, Lang He, and Yen-Wei Chen. 2026. DepressionLLM: Emotion- and Causality-Aware Depression Detection with Foun- dation Models.Displays92 (April 2026), 103304. doi:...

    2026

  23. [31]

    Shiyu Teng, Jiaqing Liu, Hao Sun, Yu Li, Shurong Chai, Ruibo Hou, Tomoko Tateyama, Lanfen Lin, and Yen-Wei Chen. 2026. Dynamic Summary Generation for Interpretable Multimodal Depression Detection. In2026 IEEE International Conference on Acoustics, Speech, and Signal Processing...

    2026

  24. [32]

    Zico Kolter, Louis-Philippe Morency, and Ruslan Salakhutdinov

    Yao-Hung Hubert Tsai, Shaojie Bai, Paul Pu Liang, J. Zico Kolter, Louis-Philippe Morency, and Ruslan Salakhutdinov. 2019. Multimodal Transformer for Unaligned Multimodal Language Sequences. InProceedings of the 57th Annual Meeting of the Association for Computational Linguisti...

    2019 doi

  25. [33]

    World Health Organization. 2022. World Mental Health Report: Transforming Mental Health for All

    2022

  26. [34]

    Amir Zadeh, Minghai Chen, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. 2017. Tensor Fusion Network for Multimodal Sentiment Analysis. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguisti...

    2017 doi

  27. [2019]

    doi:10.48550/arXiv.1904.07656

    The Verbal and Non Verbal Signals of Depression: Combining Acoustics, Text and Visuals for Estimating Depression Level.arXiv preprint arXiv:1904.07656 (2019). doi:10.48550/arXiv.1904.07656

    1904 doi arXiv

  28. [2256]

    doi:10.18653/v1/P18-1209

    doi

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.