Exploring Machine Learning and Language Models for Multimodal Depression Detection

Javier Si Zhao Hong; Pai Chet Ng; Sherwyn Chan Yin Kit; Timothy Zoe Delaya; Xiaoxiao Miao

REVIEW 4 major objections 5 minor 33 references

Exploring Machine Learning and Language Models for Multimodal Depression Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a 1.06-million-parameter transformer is the most effective model for multimodal depression detection on the MPDD dataset, beating a 7-billion-parameter LLM and a PCA-tuned XGBoost, especially for younger speakers and s

desk verdict A reasonable challenge-paper comparison, but the transformer's claimed superiority rests on a single unrepeated dev split with selection bias and a possible PCA leak; treat as a baseline, not a settled result. read the letter →

arxiv 2508.20805 v1 pith:47ZTYOQI submitted 2025-08-28 cs.CL cs.AIcs.SD

Javier Si Zhao Hong , Timothy Zoe Delaya , Sherwyn Chan Yin Kit , Pai Chet Ng , Xiaoxiao Miao This is my paper

classification cs.CLcs.AIcs.SD

keywords multimodaldepressiondetectionMPDDchallengeXGBoosttransformerlargelanguagemodelsaudio-visual-textfusionclassimbalanceF1score

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

The paper sets out to compare three families of models—gradient-boosted trees, a small multimodal transformer, and a 7-billion-parameter LLM—on a new audio-video-text depression dataset with elderly and young tracks. Its central claim is that a compact 1.06-million-parameter transformer is the most effective overall, particularly for younger speakers and short 1-second windows, while XGBoost wins only the elderly 5-second binary task. If true, this matters because clinical depression datasets are small and labels are costly; a small regularized model that fuses modalities without massive pretraining would be more practical than a huge LLM. The paper also shows that PCA plus class weighting turns XGBoost into a strong classical baseline, and that the 7B-parameter LLM underperforms despite its size.

What carries the argument

The comparison rests on three distinct pipelines: XGBoost with per-modality PCA reduction to 50 dimensions, class weighting, and early stopping; a two-layer transformer with shared linear projections, positional encodings, learned attention pooling, mixup augmentation, and focal loss; and a two-stage fine-tuned LLaMA-2-7B with linear modality projection and LoRA. The decisive mechanisms are PCA plus class weighting for the classical model and mixup plus focal-loss regularization for the small transformer, both aimed at preventing overfitting on a small, imbalanced dataset.

What would settle it

Run the same three systems on the official MPDD test set, or repeat the 90/10 patient split many times with PCA fit only inside each training fold; if the transformer no longer beats XGBoost and the 7B LLM by comparable margins, the claimed ranking is an artifact of the chosen development split.

Watch

Extended reading notes

Core claim

On the reported patient-level development split, the authors find that a deliberately small two-layer multimodal transformer (1.06M parameters) achieves the best weighted F1 on most MPDD-Elderly and MPDD-Young tasks, reaching 95.83% weighted F1 on the young 1-second binary task. XGBoost, with only 0.002M parameters, reaches the highest F1 on the elderly 5-second binary task (94.29% weighted) after per-modality PCA and class weighting. The 6,843M-parameter LLM, adapted from Emotion-LLaMA with two-stage LoRA fine-tuning, consistently underperforms both, especially on ternary classification. The authors conclude that model size does not guarantee performance on small clinical multimodal dataset

Load-bearing premise

The reported rankings rest on the assumption that the single 10% patient-level development split, together with the hyperparameters and feature pairs selected on it, represents the unseen official test set; no test-set results, repeated splits, or PCA-fit-on-training-folds-only checks are reported to back this up.

Editorial extensions

If this is right

A 1.06M-parameter transformer can outperform a 6.8B-parameter LLM on small clinical multimodal datasets, so large scale is not the main driver of depression-detection performance.
For elderly 5-second binary screening, a PCA-reduced XGBoost model is competitive or superior, suggesting simple, interpretable pipelines may be sufficient for coarse screening.
Shorter 1-second windows favor the transformer, especially for younger speakers, implying fine temporal modeling captures useful depression cues.
Personalized features (Big Five traits, demographics, RoBERTa embeddings) help the baseline but do not always help XGBoost, so feature fusion choices are task- and model-dependent.
Class weighting and PCA are effective remedies for class imbalance in small depression datasets, improving XGBoost's unweighted F1 substantially.

Reading between the lines

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

My inference: the claimed ranking would be much stronger if the authors reported repeated patient-level cross-validation with PCA fit strictly inside each training fold; the current single 90/10 split does not rule out development-set overfitting.
My inference: a practical takeaway is that low-resource mental-health screening systems should tune window length and model family to the target age group rather than defaulting to the largest available model.
My inference: the LLM's weak ternary performance may reflect adaptation underfitting (only a few LoRA epochs) or prompt mismatch rather than an inherent ceiling; a different tuning budget could change the relative ranking.
My inference: the same three-model scaffold could be applied to other small clinical datasets with imbalanced labels; if the transformer advantage reproduces across datasets, the conclusion would generalize beyond MPDD.

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.

Desk Editor's Note

read the letter

First thing you should know: this is a straightforward challenge-paper comparison of three model families on the new MPDD dataset, and the authors are honest about the biggest limitation upfront — test labels are unavailable, so all results are on a 10% dev split. That transparency counts for something. But the central claim, that a 1.06M-parameter transformer is the most effective model, is not yet established by the evidence presented.

What's new: it is the first systematic three-way comparison of XGBoost, a compact transformer, and a 7B LLM on this benchmark using the official features. The ablation table is useful, showing the effect of class weighting, PCA, and mixup on one task. The finding that a small transformer beats a 7B LLM on small clinical data is consistent with prior work and worth documenting for practitioners.

The soft spots are real and interlocking. First, the dev sets are tiny — 10% of 337 and 264 samples — so the young binary WF1 of 95.83 is based on roughly 26 samples, with no error bars. Second, PCA in Section III-A is described as simply computing the covariance matrix on each modality without stating that it is fit on training folds only. If the covariance is estimated on the full set before splitting, dev samples leak into the XGBoost features, which could inflate its elderly 5s win (94.29 vs 85.27). Third, hyperparameters and feature pairs were selected on the same dev split used for reporting in Tables IV and V, so those numbers are not independent. The single cross-validated row in Table III (87.08 for the transformer) is notably lower than several dev-set numbers, suggesting selection bias. The quinary task also has classes with 3 and 1 samples, making those results essentially meaningless, yet the paper reports them without caution.

None of these are fatal; they are addressable. The authors should run nested cross-validation, clarify the PCA fitting procedure, and either report official test results when available or explicitly caveat the ranking as dev-only. As it stands, the paper is a useful baseline for the MPDD challenge community, but not a proof that compact transformers are superior.

I'd send this to review if I were the editor — it deserves referee time as a challenge paper — but the verdict should be major revision, not acceptance. The comparison itself is worth having on record, once the evaluation is tightened.

Referee Report

4 major / 5 minor

Summary. The manuscript reports a benchmark of three model families—XGBoost, a compact 1.06M-parameter multimodal Transformer, and a 7B-parameter LLaMA-based model—on the first MPDD depression-detection challenge. Because the official test labels are unavailable, all experiments are performed on a single patient-level 90/10 split of the training set as a development set. The authors evaluate weighted and unweighted F1 for elderly and young tracks at 1s and 5s windows, and ablate preprocessing choices such as PCA, class weighting, Mixup, and personality-aware features. They conclude that the Transformer is the most effective model overall, especially for younger speakers and shorter windows, that XGBoost wins the elderly 5-second binary task, and that LLMs underperform despite their size.

Significance. Depression screening from multimodal behavioral signals is an important application, and the MPDD challenge is new. The paper's systematic comparison of three model classes, its ablation of PCA/class weighting/Mixup, and its reporting of class distributions and model sizes are useful contributions. If the ranking were robust, the finding that a 1.06M-parameter Transformer outperforms a 7B LLM and a very small XGBoost model on a small clinical dataset would be practically valuable. However, the current evidence is limited to a single development split, so the quantitative ranking is not yet established. The paper does not release code, and no official test-set results are available, limiting reproducibility and verification.

major comments (4)

[Section IV, Tables IV–V; Table I footnote] All headline results are on a single patient-level 90/10 split of the training set; the official test set is not used (Table I footnote). This split yields roughly 34 elderly and 26 young development samples, yet Tables IV and V report single-shot F1 scores without error bars. Only one row in Table III uses 10-fold cross-validation. The final claim in Section IV-B.2 that the Transformer is 'most effective' is not supported by a single, unrepeated split; repeated patient-level splits or nested CV with variance estimates are needed.
[Section III-A, PCA projection] The PCA procedure is described as centering, computing covariance Cm, and projecting Zm=(Xm-mean(Xm))V^k_m, but the text nowhere states that the covariance is estimated on training folds only before splitting. If PCA is fit on the full 90+10 set, development samples influence the XGBoost feature representation, leaking information and inflating XGBoost scores—e.g., the elderly 5s binary WF1 of 94.29 vs. the Transformer's 85.27 in Table IV. The authors must clarify, and if necessary correct, this by fitting PCA inside each training fold.
[Section IV-A.1–2, Tables IV–V] Hyperparameters, modality pairs, and early-stopping settings were selected on the same 90/10 development split (Section IV-A.1: 'Hyperparameters were tuned per modality pair... using early stopping on a speaker-level validation split'). Thus the reported scores are model-selection results, not independent test estimates. The Transformer's margins over XGBoost on some tasks are modest (e.g., 93.44 vs. 90.67 WF1 for elderly 1s binary in Table IV), so nested or repeated evaluation is required to show the ranking is not an artifact of selection on this split.
[Section III-A vs. Table V] The XGBoost row in Table V is marked PF=✓, but Section III-A's footnote and Section IV-A.1 state that personality-aware features were excluded from XGBoost. This is a direct contradiction in the reported configuration. Since the PF column is used to compare feature sets, this inconsistency must be corrected in the table or the text before the XGBoost results can be interpreted.

minor comments (5)

[Section III-A, notation] X_a and X_v are both written as n×d matrices, but the sequence lengths for audio and video may differ; using T_a and T_v (as in Section III-B) would avoid ambiguity.
[Section IV, opening paragraph] The text says experiments begin with the 5-second MPDD-Elderly dataset, but results for 1-second windows are also reported. Clarify the order of experimentation and that the 5-second binary ablation drives the configuration choices.
[Table III, header] The header '90( CROSS-VALIDATION )/10' is ambiguous. Specify explicitly which rows use a single 90/10 split and which rows use 10-fold cross-validation.
[Table IV, LLM row] The LLM row reports identical quinary scores for 1s and 5s (WF1 77.89/UF1 30.60). This is surprising and should be explained, or the experimental setup for the LLM windows should be clarified.
[References] Several references are incomplete or contain typos (e.g., [21] 'deprefssion', [22] and [23] use placeholder-style author names). Please verify and complete the reference list.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper reports measured F1 scores from a patient-level dev split; model-selection and PCA-fit caveats are statistical validity concerns, not definitional circularity.

full rationale

This is an empirical benchmark paper, not a derivation. All central claims (e.g., 'the Transformer model (1.06M parameters) is the most effective for multimodal depression detection') are rankings of F1 scores computed on a 90/10 patient-level split of the MPDD training set, with labels provided externally by the challenge (HAMD-24/PHQ-9). No fitted parameter is renamed as a prediction: XGBoost's PCA projection, class weights, and hyperparameters, and the Transformer's architecture/hyperparameters, are selected on the dev split, but the reported F1 numbers are measured outcomes, not identities forced by the fitting procedure. The PCA step in Section III-A is not stated to be fit on training folds only, and hyperparameters are tuned on the same dev split; these are data-leakage/selection-bias concerns that could inflate scores, but they do not make a result equal to its input by construction. There is no load-bearing self-citation: Emotion-LLaMA is cited as an external pretrained model, and the paper's contribution is an empirical comparison, not a theorem resting on that citation. The paper explicitly notes the test labels are unavailable and that results are based on the 90/10 split, which is an honest limitation rather than a circular step. No equation in the paper reduces to an earlier equation by definition, and no 'prediction' is statistically forced by a fitted constant. Therefore the circularity score is 0.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new theoretical entities or constants. Its results rest on the challenge dataset, the provided features, and standard model hyperparameters. The main hidden inputs are the hand-set hyperparameters and the implicit trust in the provided features and labels.

free parameters (4)

PCA dimensionality k = 50 per modality
Hand-set; each audio and visual modality is reduced to 50 components before concatenation (Section III-A). The choice is not justified by a selection criterion.
Transformer hidden size d = 128
Hand-set for small-dataset efficiency (Section IV-A.2).
Mixup alpha = 0.2
Beta(0.2, 0.2) sampled with 50% probability; chosen to smooth decision boundaries (Section IV-A.2).
XGBoost max depth = 3
Shallow trees to reduce overfitting; selected from a small set of options (Section IV-A.1).

assumptions (3)

domain assumption The provided feature embeddings (MFCC, OpenSMILE, Wav2Vec2, DenseNet, ResNet, OpenFace, RoBERTa) are sufficient and well-aligned for depression detection.
The paper uses the challenge-provided pre-extracted embeddings (Table II) without raw feature extraction or explicit alignment verification; this is loaded from the dataset design.
domain assumption The patient-ID 90-10 split prevents subject-level information leakage.
Stated in Section IV ('ensuring subject independence'), but the dev set is also used for hyperparameter tuning and feature-pair selection, which introduces selection bias.
domain assumption HAMD-24 and PHQ-9 scores are reliable depression severity labels.
The labels are derived from clinical questionnaires as described in Section II-A; the paper does not independently validate them.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Exploring Machine Learning and Language Models for Multimodal Depression Detection." pith.science (2026). https://pith.science/paper/47ZTYOQI

@misc{pith2026250820805,
  author       = {Pith},
  title        = {Pith review of: Exploring Machine Learning and Language Models for Multimodal Depression Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/47ZTYOQI}},
  note         = {Machine review of arXiv:2508.20805}
}

read the original abstract

This paper presents our approach to the first Multimodal Personality-Aware Depression Detection Challenge, focusing on multimodal depression detection using machine learning and deep learning models. We explore and compare the performance of XGBoost, transformer-based architectures, and large language models (LLMs) on audio, video, and text features. Our results highlight the strengths and limitations of each type of model in capturing depression-related signals across modalities, offering insights into effective multimodal representation strategies for mental health prediction.

Figures

Figures reproduced from arXiv: 2508.20805 by the authors.

**Figure 1.** Various multimodal depression detection models. The leftmost is the baseline model, while the right three are the models investigated in this paper. [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

33 extracted references · 27 canonical work pages

[1]

Depressive disorder (depression),

World Health Organization, “Depressive disorder (depression),” https: //www.who.int/news-room/fact-sheets/detail/depression, 2023, accessed: 7 July 2025

work page 2023
[2]

Depression underdiagnosis: Prevalence and associated factors. a population-based study,

A. Faisal-Cury, C. Ziebold, D. M. O. Rodrigues, and A. Matijase- vich, “Depression underdiagnosis: Prevalence and associated factors. a population-based study,” Journal of Psychiatric Research , vol. 151, pp. 157–165, July 2022, epub 2022 Apr 23

work page 2022
[3]

The phq-9: validity of a brief depression severity measure,

K. Kroenke, R. L. Spitzer, and J. B. W. Williams, “The phq-9: validity of a brief depression severity measure,” Journal of General Inter- nal Medicine , vol. 16, no. 9, pp. 606–613, September 2001, pMID: 11556941

work page 2001
[4]

A. T. Beck, R. A. Steer, and G. K. Brown, Beck Depression Inven- tory—Second Edition Manual . San Antonio, TX: The Psychological Corporation, 1996, © 1996, 1987 by Aaron T. Beck, Robert A. Steer, Gregory K. Brown

work page 1996
[5]

Dynamic multimodal measurement of depression severity using deep autoencoding,

H. Dibeklio ˘glu, Z. Hammal, and J. F. Cohn, “Dynamic multimodal measurement of depression severity using deep autoencoding,” IEEE Journal of Biomedical and Health Informatics , vol. 22, no. 2, pp. 525– 536, 2017

work page 2017
[6]

D-vlog: Multimodal vlog dataset for depression detection,

J. Yoon, C. Kang, S. Kim, and J. Han, “D-vlog: Multimodal vlog dataset for depression detection,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, 2022, pp. 12 226–12 234

work page 2022
[7]

Automatic depression detection: An emo- tional audio-textual corpus and a gru/bilstm-based model,

Y . Shen, H. Yang, and L. Lin, “Automatic depression detection: An emo- tional audio-textual corpus and a gru/bilstm-based model,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2022, pp. 6247–6251

work page 2022
[8]

Semi-structural interview-based chinese multimodal depression corpus towards automatic preliminary screening of depressive disorders,

B. Zou, J. Han, Y . Wang, R. Liu, S. Zhao, L. Feng, X. Lyu, and H. Ma, “Semi-structural interview-based chinese multimodal depression corpus towards automatic preliminary screening of depressive disorders,” IEEE Transactions on Affective Computing , vol. 14, no. 4, pp. 2823–2838, 2022

work page 2022

Show all 33 references

[9]

A multi-modal open dataset for mental-disorder analysis,

H. Cai, Z. Yuan, Y . Gao, S. Sun, N. Li, F. Tian, H. Xiao, J. Li, Z. Yang, X. Li et al., “A multi-modal open dataset for mental-disorder analysis,” Scientific Data, vol. 9, no. 1, p. 178, 2022

2022
[10]

Amigos: A dataset for affect, personality and mood research on individuals and groups,

J. A. Miranda-Correa, M. K. Abadi, N. Sebe, and I. Patras, “Amigos: A dataset for affect, personality and mood research on individuals and groups,” IEEE Transactions on Affective Computing , vol. 12, no. 2, pp. 479–493, 2018

2018
[11]

Deap: A database for emotion analysis; using physiological signals,

S. Koelstra, C. Muhl, M. Soleymani, J.-S. Lee, A. Yazdani, T. Ebrahimi, T. Pun, A. Nijholt, and I. Patras, “Deap: A database for emotion analysis; using physiological signals,” IEEE Transactions on Affective Computing, vol. 3, no. 1, pp. 18–31, 2011

2011
[12]

Personality and depression: explanatory models and review of the evidence,

D. N. Klein, R. Kotov, and S. J. Bufferd, “Personality and depression: explanatory models and review of the evidence,” Annual Review of Clinical Psychology, vol. 7, no. 1, pp. 269–295, 2011

2011
[13]

Genome- wide analyses for personality traits identify six genomic loci and show correlations with psychiatric disorders,

M.-T. Lo, D. A. Hinds, J. Y . Tung, C. Franz, C.-C. Fan, Y . Wang, O. B. Smeland, A. Schork, D. Holland, K. Kauppi et al. , “Genome- wide analyses for personality traits identify six genomic loci and show correlations with psychiatric disorders,” Nature Genetics, vol. 49, no. ...

2017
[14]

The first mpdd challenge: Multimodal personality- aware depression detection,

C. Fu, Z. Fu, Q. Zhang, X. Kuang, J. Dong, K. Su, Y . Su, W. Shi, J. Yao, Y . Zhao, S. Zhao, J. Wang, S. Song, C. Liu, Y . Yoshikawa, B. Schuller, and H. Ishiguro, “The first mpdd challenge: Multimodal personality- aware depression detection,” arXiv preprint arXiv:2505.10034, ...

2025 arXiv
[15]

Measuring personality in one minute or less: A 10-item short version of the big five inventory in english and german,

B. Rammstedt and O. P. John, “Measuring personality in one minute or less: A 10-item short version of the big five inventory in english and german,” Journal of research in Personality, vol. 41, no. 1, pp. 203–212, 2007

2007
[16]

Recent developments in opens- mile, the munich open-source multimedia feature extractor,

F. Eyben, M. W ¨ollmer, and B. Schuller, “Recent developments in opens- mile, the munich open-source multimedia feature extractor,” Proceedings of the 21st ACM international conference on Multimedia , 2013

2013
[17]

Openface 2.0: Facial behavior analysis toolkit,

T. Baltrusaitis et al., “Openface 2.0: Facial behavior analysis toolkit,” in IEEE FG, 2018, pp. 59–66

2018
[18]

Python fer (facial expression recognition),

J. Shenk, “Python fer (facial expression recognition),” https://github.com/ justinshenk/fer, 2019

2019
[19]

Xgboost: A scalable tree boosting system,

T. Chen and C. Guestrin, “Xgboost: A scalable tree boosting system,” in Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining , 2016, pp. 785–794

2016
[20]

Sup- port vector machines,

M. A. Hearst, S. T. Dumais, E. Osuna, J. Platt, and B. Scholkopf, “Sup- port vector machines,” IEEE Intelligent Systems and their applications , vol. 13, no. 4, pp. 18–28, 1998

1998
[21]

Detecting deprefssion on video logs using audiovisual features,

J. Gideon et al., “Detecting deprefssion on video logs using audiovisual features,” Humanities & Social Sciences Communications, vol. 10, no. 1, pp. 1–12, 2023

2023
[22]

First transformer-based depression detection using multi-head attention,

Y . Zhang et al. , “First transformer-based depression detection using multi-head attention,” Sensors, vol. 21, no. 14, p. 4764, 2021

2021
[23]

Late fusion strategies for multimodal depression classi- fication,

J. Smith et al., “Late fusion strategies for multimodal depression classi- fication,” Psychiatry AI, vol. 12, pp. 45–60, 2023

2023
[24]

Cross-attention multimodal fusion using macbert for depression detection,

L. Chen et al. , “Cross-attention multimodal fusion using macbert for depression detection,” arXiv:2407.12825, 2024

2024 arXiv
[25]

Mentalbert: A clinical language model for mental health assessment,

S. Ji et al. , “Mentalbert: A clinical language model for mental health assessment,” Natural Language Processing Journal , vol. 1, p. 100003, 2022

2022
[26]

Emotion-llama: Multimodal emotion recognition and reasoning with instruction tuning,

Z. Cheng, Z.-Q. Cheng, J.-Y . He, K. Wang, Y . Lin, Z. Lian, X. Peng, and A. Hauptmann, “Emotion-llama: Multimodal emotion recognition and reasoning with instruction tuning,” in Advances in Neural Information Processing Systems , A. Globerson, L. Mackey, D. Belgrave, A. Fan, U...

2024
[27]

The hamilton rating scale for depression,

M. Hamilton, “The hamilton rating scale for depression,” in Assessment of depression. Springer, 1986, pp. 143–152

1986
[28]

wav2vec 2.0: A framework for self-supervised learning of speech representations,

A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” Advances in Neural Information Processing Systems, vol. 33, pp. 12 449– 12 460, 2020

2020
[29]

Densely connected convolutional networks,

G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely connected convolutional networks,” Proceedings of the IEEE conference on computer vision and pattern recognition , pp. 4700–4708, 2017

2017
[30]

Deep residual learning for image recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” Proceedings of the IEEE conference on computer vision and pattern recognition , pp. 770–778, 2016

2016
[31]

Roberta: A robustly optimized bert pretraining approach,

Y . Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, and V . Stoyanov, “Roberta: A robustly optimized bert pretraining approach,” arXiv preprint arXiv:1907.11692 , 2019

1907 arXiv
[32]

Focal loss for dense object detection,

T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar, “Focal loss for dense object detection,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 2980–2988

2017
[33]

mixup: Beyond empirical risk minimization,

H. Zhang, M. Cisse, Y . N. Dauphin, and D. Lopez-Paz, “mixup: Beyond empirical risk minimization,” arXiv preprint arXiv:1710.09412 , 2018. 6

2018 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Depressive disorder (depression),

World Health Organization, “Depressive disorder (depression),” https: //www.who.int/news-room/fact-sheets/detail/depression, 2023, accessed: 7 July 2025

work page 2023

[2] [2]

Depression underdiagnosis: Prevalence and associated factors. a population-based study,

A. Faisal-Cury, C. Ziebold, D. M. O. Rodrigues, and A. Matijase- vich, “Depression underdiagnosis: Prevalence and associated factors. a population-based study,” Journal of Psychiatric Research , vol. 151, pp. 157–165, July 2022, epub 2022 Apr 23

work page 2022

[3] [3]

The phq-9: validity of a brief depression severity measure,

K. Kroenke, R. L. Spitzer, and J. B. W. Williams, “The phq-9: validity of a brief depression severity measure,” Journal of General Inter- nal Medicine , vol. 16, no. 9, pp. 606–613, September 2001, pMID: 11556941

work page 2001

[4] [4]

A. T. Beck, R. A. Steer, and G. K. Brown, Beck Depression Inven- tory—Second Edition Manual . San Antonio, TX: The Psychological Corporation, 1996, © 1996, 1987 by Aaron T. Beck, Robert A. Steer, Gregory K. Brown

work page 1996

[5] [5]

Dynamic multimodal measurement of depression severity using deep autoencoding,

H. Dibeklio ˘glu, Z. Hammal, and J. F. Cohn, “Dynamic multimodal measurement of depression severity using deep autoencoding,” IEEE Journal of Biomedical and Health Informatics , vol. 22, no. 2, pp. 525– 536, 2017

work page 2017

[6] [6]

D-vlog: Multimodal vlog dataset for depression detection,

J. Yoon, C. Kang, S. Kim, and J. Han, “D-vlog: Multimodal vlog dataset for depression detection,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, 2022, pp. 12 226–12 234

work page 2022

[7] [7]

Automatic depression detection: An emo- tional audio-textual corpus and a gru/bilstm-based model,

Y . Shen, H. Yang, and L. Lin, “Automatic depression detection: An emo- tional audio-textual corpus and a gru/bilstm-based model,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2022, pp. 6247–6251

work page 2022

[8] [8]

Semi-structural interview-based chinese multimodal depression corpus towards automatic preliminary screening of depressive disorders,

B. Zou, J. Han, Y . Wang, R. Liu, S. Zhao, L. Feng, X. Lyu, and H. Ma, “Semi-structural interview-based chinese multimodal depression corpus towards automatic preliminary screening of depressive disorders,” IEEE Transactions on Affective Computing , vol. 14, no. 4, pp. 2823–2838, 2022

work page 2022