Speech as a Multimodal Digital Phenotype for Multi-Task LLM-based Mental Health Prediction

Christopher Lucasius; Deepa Kundur; Jacob Vorstman; Madison Aitken; Mai Ali; Marco Battaglia; Peter Szatmari; Tanmay P. Patel

REVIEW 4 major objections 5 minor 30 references

Speech as a Multimodal Digital Phenotype for Multi-Task LLM-based Mental Health Prediction

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper proposes treating patient speech as a trimodal data source—transcripts, acoustic landmarks, and vocal biomarkers—and uses a longitudinal, multi-task LLM architecture to predict depression, suicidal ideation, and sleep…

desk verdict Competent incremental pipeline whose 70.8% headline isn't yet trustworthy until the patient-level split and lambda_aux selection are clarified. read the letter →

arxiv 2505.23822 v3 pith:MMUV3BN7 submitted 2025-05-28 cs.CL cs.MM

Mai Ali , Christopher Lucasius , Tanmay P. Patel , Madison Aitken , Jacob Vorstman , Peter Szatmari , Marco Battaglia , Deepa Kundur This is my paper

classification cs.CLcs.MM

keywords multimodalspeechanalysismulti-tasklearninglargelanguagemodelsmentalhealthpredictionadolescentdepressionlongitudinalacousticlandmarksvocalbiomarkers

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

This paper tries to show that a patient's speech, treated as three parallel data streams rather than one, can support automated mental-health screening in adolescents. The three streams are the transcript of what was said, a discretized sequence of acoustic landmarks capturing how sounds are produced, and quantitative vocal biomarkers such as pitch and jitter. The authors build an LLM-based architecture that fuses these streams, carries a memory across successive clinical visits, and simultaneously predicts depression, suicidal ideation, and sleep disturbance. On the Depression Early Warning dataset, the full configuration reaches a balanced accuracy of 70.8% for the primary depression task, higher than the unimodal, single-task, and non-longitudinal alternatives tested. If the result generalizes, routine speech recordings could become a noninvasive, multi-condition screening signal for adolescent mental health.

What carries the argument

The load-bearing mechanism is the fusion of three speech-derived modalities into one LLM pipeline. Transcripts come from Whisper speech recognition; acoustic landmarks are discretized symbolic sequences representing spectrogram energy events such as vocal-fold vibration and nasal release; vocal biomarkers are windowed statistics like MFCC, fundamental frequency, jitter, shimmer, and harmonic-to-noise ratio. A P-tuned LLM (Mental-Alpaca or LLaMA-2-7B) with LoRA encodes the text and landmarks, a transformer encoder handles the biomarker time series, and a GRU propagates a hidden state across visits for longitudinal modeling. Three task heads with weighted binary cross-entropy produce the predictions, and the total loss is $\mathcal{L} = \mathcal{L}_M + \lambda_{\text{aux}}(\mathcal{L}_{A_0}+\mathcal{L}_{A_1})$, where $\lambda_{\text{aux}}$ controls how strongly the auxiliary suicidal-ideation and sleep tasks shape the shared representation.

What would settle it

Take the same trimodal longitudinal multi-task model and select $\lambda_{\text{aux}}$ on the validation set alone, then report the test balanced accuracy for depression. If the number drops from 70.8% toward the 58.9% observed at $\lambda_{\text{aux}}=0$, the reported multi-task gain is an artifact of test-set-based hyperparameter selection rather than a property of the model.

Watch

Extended reading notes

Core claim

The paper's central claim is that multimodal, longitudinal, multi-task learning over speech-derived signals outperforms each of its components. Concretely, the authors report that adding vocal biomarkers to text-plus-landmarks improves balanced accuracy for depression under the Mental-Alpaca base model, and that the longitudinal GRU memory plus auxiliary tasks on suicidal ideation and sleep disturbances pushes the primary depression balanced accuracy to 70.8% at auxiliary-task weight $\lambda_{\text{aux}}=1.00$. The authors interpret this as evidence that the three conditions share underlying speech signatures, so modeling them jointly extracts more signal than predicting depression alone.

Load-bearing premise

The headline 70.8% result holds only if the auxiliary-task weight $\lambda_{\text{aux}}$ was chosen by a rule that did not inspect the test set; the paper never states that rule.

Editorial extensions

If this is right

If the result holds, speech recordings collected during routine clinical visits could screen for multiple comorbid mental-health conditions at once, using a single noninvasive signal.
Longitudinal multi-task modeling would make each new visit's prediction informed by the patient's trajectory, potentially catching deterioration earlier than cross-sectional screening.
Adding vocal biomarkers to transcript-based LLM analysis would be worth adopting in other speech-health applications beyond depression.
The finding that a mental-health-tuned base model outperforms a larger general LLM supports domain-specific fine-tuning for clinical tasks.
The reported synergy between depression, suicidal ideation, and sleep tasks suggests that auxiliary labels can strengthen primary-task prediction when labeled data are limited.

Reading between the lines

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

Because the paper does not state how $\lambda_{\text{aux}}$ was chosen, the 70.8% figure should be treated as an upper estimate until the selection rule is reported; a validation-set-only selection could lower the headline number.
A natural next test is to transfer the same trimodal MTL recipe to public longitudinal speech datasets; the current evaluation rests entirely on the Depression Early Warning dataset, so dataset-specific effects cannot yet be ruled out.
The GRU-based longitudinal memory makes an explicit, testable prediction: ablating the hidden state should eliminate the longitudinal gain, and that ablation is straightforward to run.
The same architecture could be extended to non-speech modalities such as facial expressions or movement data, where longitudinal comorbidity patterns are also clinically relevant.

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

Short version: this is a competent, incremental empirical paper. The new thing is the specific combination—Whisper transcripts, acoustic landmarks, vocal biomarkers, multi-task learning, and a GRU over visits—applied to a private adolescent depression dataset. The authors are honest that the longitudinal idea is not new; the contribution is the application. The experimental design has real strengths: clear label definitions from validated scales, ROC-threshold selection on validation then applied blindly to test, and balanced accuracy as the primary metric.

The soft spots are the usual ones and they matter here. No error bars or repeated runs; a private dataset with no code; and no description of how the train/validation/test split was made. That last point is load-bearing because DEW has multiple visits per patient and the longitudinal model conditions on previous visits. If the split is by arm rather than by patient, the model can memorize a patient's earlier sessions in training and leak into test. The paper never says. Similarly, Table III sweeps lambda_aux and the abstract quotes the best value (1.00) without stating whether that hyperparameter was locked before seeing the test set. So the 70.8% headline could be a selection artifact, a leakage artifact, or both. Those are fixable reporting gaps, not evidence of fraud, but they mean the central claim is not yet supported as written.

The auxiliary tasks show unstable results—sleep disturbance recall jumps around, suicidal ideation precision is often zero in baselines—which reinforces the need for error bars or at least more runs. The self-citation to [11] is fine.

Bottom line: for a paper reporting a clinical screening tool, this needs a serious referee, and the authors should be asked to clarify the split protocol, report patient-level split results, and show a lambda_aux selection procedure that doesn't touch the test set. If they can do that, the paper becomes a useful contribution. As it stands, it's a plausible method with an unverified headline. I'd send it to review rather than desk reject, with the expectation of major revision.

Referee Report

4 major / 5 minor

Summary. The paper proposes a multimodal LLM-based framework for predicting depression, suicidal ideation, and sleep disturbances from adolescent speech, using three speech-derived modalities (transcripts, acoustic landmarks, and vocal biomarkers), multi-task learning (MTL), and a longitudinal extension that propagates a latent representation across a patient's visits. The method is evaluated on the private Depression Early Warning (DEW) dataset, with the central claim being a balanced accuracy of 70.8% for depression, higher than each unimodal, single-task, and non-longitudinal baseline. The architecture and loss formulations appear clearly described, but the experimental reporting lacks critical protocol details.

Significance. If the reported result holds under rigorous evaluation, the paper would provide a meaningful demonstration that combining linguistic, acoustic-landmark, and vocal-biomarker modalities within an LLM-based multi-task, longitudinal framework can improve mental-health prediction from speech. The idea of treating speech as a trimodal digital phenotype is sensible and the multi-task synergy direction is interesting. The paper also follows some good practices, such as choosing classification thresholds on a validation set and reporting balanced accuracy to address class imbalance. However, the significance is currently undercut by the absence of a clear model-selection protocol, patient-level split information, and any uncertainty quantification. The private dataset and lack of released code further limit the ability of other researchers to verify or build on the results.

major comments (4)

[Abstract and Section V-C, Table III] The headline 70.8% balanced accuracy is the maximum value of the auxiliary-task weight lambda_aux in the sweep reported in Table III, while Table II fixes lambda_aux = 0.25. The paper does not state which data split was used to select lambda_aux. If the test set was used for this selection, the headline number is an optimistic estimate; if a validation set was used, the selection protocol must be described so the comparison with the fixed-lambda_aux baselines in Table II is fair. Please specify the exact selection protocol, report validation and test performance across the full lambda_aux sweep, and indicate which value was locked before test evaluation.
[Section III-A and Section IV-D] The DEW dataset contains multiple follow-up arms per patient_id, and the proposed longitudinal model uses a patient's previous arms as context. The paper never states whether the train/validation/test split was performed at the patient level, i.e., whether all arms of every test patient were excluded from training. If the split is instead done at the arm level, the model could memorize a patient's earlier arms during training and use that patient-identity information at test time, inflating all reported balanced accuracies. Please describe the split explicitly, report the numbers of patients and arms in each split, and confirm that no patient appears in more than one split.
[Section V, Tables II and III] No error bars, confidence intervals, or significance tests are reported, and there is no statement about the number of random seeds or repeated fine-tuning runs. Several of the key comparisons, such as depression balanced accuracy 0.666 versus 0.644 in Table II and 0.672 versus 0.708 in Table III, are small and could easily fall within the stochastic variation of LLM fine-tuning. Please report means and standard deviations over at least three seeds, or provide bootstrapped confidence intervals, to support the claim that the proposed method outperforms the baselines.
[Section III-B and Table I] The labeling criteria for sleep disturbances are internally inconsistent: the text says the label uses question 3 of PHQ-9 and questions 32 and 33 of the MFQ, but Table I lists 'PHQ-9/ Q9 + MFQ/ Q32 + MFQ/ Q32' for the threshold, and the summary column does not match the items. This ambiguity affects the ground truth for one of the three tasks and must be corrected. Please also specify the exact numeric cutoffs and clarify whether the combination is a sum of item scores or a logical condition.

minor comments (5)

[Section I-A] The word 'suprisingly' appears in the introduction; it should be 'surprisingly'.
[Section IV-A1] The manual correction of Whisper transcripts is not quantified. Please report how many utterances were manually corrected and whether any inter-annotator reliability was assessed.
[Section IV-A3] The selection of a 500-millisecond window and the specific summary statistics for vocal biomarkers are stated but not justified; please provide a reference or an explanation for these choices.
[Section IV-A2] The paper does not mention how long transcript and landmark sequences are truncated to fit the LLM's context window; please state the tokenization and truncation strategy.
[Section VI] The conclusion claims improvement over 'baseline models [7]', but the baselines in this paper are architectural ablations, not the Mental-LLM system from reference [7]. Please rephrase to avoid implying a direct comparison with that system.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; minor non-load-bearing self-citation and unstated hyperparameter/split protocols prevent a clean 0, but no reported result reduces to its inputs by construction.

full rationale

The paper's derivation chain is a standard supervised pipeline: extract three speech-derived modalities (Whisper text, acoustic landmarks per [9], vocal biomarkers per [14]), embed them, optionally pass through a GRU for longitudinal context, and train three classification heads with the weighted BCE loss of Eq. (1) and the total loss of Eq. (2). The headlined 70.8% balanced accuracy is measured on a test set with thresholds selected on the validation set and then blindly applied to the test set, so the primary metric is not defined in terms of the fitted quantities. The only author-overlapping citation is reference [11], used in Section II as a general background pointer to prior work on speech/video modalities; it is not invoked to justify the architecture, the longitudinal strategy, the multi-task formulation, or the reported accuracy, so it is not load-bearing. The base models LLaMA-2-7B and Mental-Alpaca are external, and Mental-Alpaca's mental-health fine-tuning does not make the DEW evaluation circular because DEW is a separate dataset. Two reporting gaps are noted as correctness risks rather than circularity: (a) the train/validation/test split for the repeated-measures DEW arms is never described, so patient-level leakage cannot be excluded; and (b) the protocol for selecting the auxiliary-task weight lambda_aux is unstated, and the abstract highlights the Table III maximum (lambda_aux = 1.00) while Table II fixes lambda_aux = 0.25 for all compared architectures. These omissions undermine the strength of the superiority claim but do not, on the evidence in the paper, demonstrate that any prediction is equivalent to an input or fit by construction.

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

No new physical or conceptual entities are introduced. The central claim rests on dataset-label validity, split hygiene, and the incremental value of acoustic features, none of which are independently verified in this paper.

free parameters (3)

lambda_aux (auxiliary task weight) = 0.25 in Table II; 1.00 yields best depression BA 0.708 in Table III
Manually chosen to balance auxiliary losses; the paper reports a sweep and highlights the maximum, making the headline result dependent on this hyperparameter.
ROC threshold per task (3 tasks) = Selected on validation set (point closest to top-left corner)
Threshold selection is standard but contributes to evaluation; no analysis of sensitivity is provided.
Vocal biomarker window size = 500 ms
Fixed without an ablation; changes feature statistics and could affect results.

assumptions (4)

domain assumption The DEW dataset labels and the equipercentile unification of CDRS and HAM-D depression scores are valid.
Labels are treated as ground truth in Section III-B; the linking method is cited but not locally validated.
domain assumption The train/validation/test split prevents participant-level information leakage.
The split is never described; the longitudinal model uses data from earlier visits of the same participant, so leakage is a real risk if splits are not by participant.
domain assumption Non-text speech features (acoustic landmarks and vocal biomarkers) carry predictive signal beyond the text transcript.
The central multimodal claim depends on incremental signal; improvements in Table II are small and untested for significance.
domain assumption Whisper transcription with manual corrections is sufficiently accurate for clinical prediction.
Section IV-A1 describes manual scanning of transcripts, but transcription errors remain possible and are not quantified.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Speech as a Multimodal Digital Phenotype for Multi-Task LLM-based Mental Health Prediction." pith.science (2026). https://pith.science/paper/MMUV3BN7

@misc{pith2026250523822,
  author       = {Pith},
  title        = {Pith review of: Speech as a Multimodal Digital Phenotype for Multi-Task LLM-based Mental Health Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MMUV3BN7}},
  note         = {Machine review of arXiv:2505.23822}
}

read the original abstract

Speech is a noninvasive digital phenotype that can offer valuable insights into mental health conditions, but it is often treated as a single modality. In contrast, we propose the treatment of patient speech data as a trimodal multimedia data source for depression detection. This study explores the potential of large language model-based architectures for speech-based depression prediction in a multimodal regime that integrates speech-derived text, acoustic landmarks, and vocal biomarkers. Adolescent depression presents a significant challenge and is often comorbid with multiple disorders, such as suicidal ideation and sleep disturbances. This presents an additional opportunity to integrate multi-task learning (MTL) into our study by simultaneously predicting depression, suicidal ideation, and sleep disturbances using the multimodal formulation. We also propose a longitudinal analysis strategy that models temporal changes across multiple clinical interactions, allowing for a comprehensive understanding of the conditions' progression. Our proposed approach, featuring trimodal, longitudinal MTL is evaluated on the Depression Early Warning dataset. It achieves a balanced accuracy of 70.8%, which is higher than each of the unimodal, single-task, and non-longitudinal methods.

Figures

Figures reproduced from arXiv: 2505.23822 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 21 canonical work pages

[11]

Prediction of relapse in adolescent depression using fusion of video and speech data

C. Lucasius, M. Ali, M. Battaglia, J. Strauss, P. Szatmari, and D. Kundur, “Prediction of relapse in adolescent depression using fusion of video and speech data.” [Online]. Available: https://ceur- ws.org/V ol-3649/Paper22.pdf

work page
[1]

Adolescent depression: National trends, risk factors, and healthcare disparities,

W. Lu, “Adolescent depression: National trends, risk factors, and healthcare disparities,” American Journal of Health Behavior , vol. 43, pp. 181–194, 01 2019

work page 2019
[2]

Sleep’s role in the development and resolution of adolescent depression,

M. Gradisar, M. Kahn, G. Micic, M. Short, C. Reynolds, F. Orchard, S. Bauducco, K. Bartel, and C. Richardson, “Sleep’s role in the development and resolution of adolescent depression,” Nature Reviews Psychology , vol. 1, p. 1–12, 06 2022. [Online]. Available: https://www.nature.com/articles/s44159-022-00074-8

work page 2022
[3]

Elevated risk of depression among adolescents presenting with sleep disorders,

S. M. Inkelis, S. Ancoli-Israel, J. D. Thomas, and R. Bhattacharjee, “Elevated risk of depression among adolescents presenting with sleep disorders,” Journal of Clinical Sleep Medicine , 11 2020

work page 2020
[4]

Language as a biomarker for psychosis: A natural language processing approach,

C. M. Corcoran, V . A. Mittal, C. E. Bearden, R. E. Gur, K. Hitczenko, Z. Bilgrami, A. Savic, G. A. Cecchi, and P. Wolff, “Language as a biomarker for psychosis: A natural language processing approach,” Schizophrenia Research, vol. 226, pp. 158–166, 12 2020

work page 2020
[5]

Speech as a promising biosignal in precision psychiatry,

M. Kappen, M.-A. Vanderhasselt, and G. M. Slavich, “Speech as a promising biosignal in precision psychiatry,” Neuroscience & Biobehavioral Reviews, vol. 148, p. 105121, 2023. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0149763423000908

work page 2023
[6]

The relationship between linguistic expression in blog content and symptoms of depression, anxiety, and suicidal thoughts: A longitudinal study,

B. O’Dea, T. W. Boonstra, M. E. Larsen, T. Nguyen, S. Venkatesh, and H. Christensen, “The relationship between linguistic expression in blog content and symptoms of depression, anxiety, and suicidal thoughts: A longitudinal study,” PLOS ONE , vol. 16, p. e0251787, 05 2021

work page 2021
[7]

Mental-llm: Leveraging large language models for mental health prediction via online text data,

X. Xu, B. Yao, Y . Dong, S. Gabriel, H. Yu, J. Hendler, M. Ghassemi, A. K. Dey, and D. Wang, “Mental-llm: Leveraging large language models for mental health prediction via online text data,” Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 8, no. 1, p. 1–32, Mar. 2024. [Online]. Available: http://dx.doi.org/10.1145/3643540

doi:10.1145/3643540 2024

Show all 30 references

[8]

Multitask learning for mental health: Depression, anxiety, stress (das) using wearables,

B. Saylam and ¨Ozlem Durmaz ˙Incel, “Multitask learning for mental health: Depression, anxiety, stress (das) using wearables,” Diagnostics, vol. 14, pp. 501–501, 02 2024

2024
[9]

When llms meets acoustic landmarks: An efficient approach to integrate speech into large language models for depression detection,

X. Zhang, H. Liu, K. Xu, Q. Zhang, D. Liu, B. Ahmed, and J. Epps, “When llms meets acoustic landmarks: An efficient approach to integrate speech into large language models for depression detection,”
[10]

Opportunities and risks of large language models in psychiatry,

N. Obradovich, S. S. Khalsa, W. U. Khan, J. Suh, R. H. Perlis, O. Ajilore, and M. P. Paulus, “Opportunities and risks of large language models in psychiatry,” NPP—Digital Psychiatry and Neuroscience , vol. 2, p. 1–8, 05 2024. [Online]. Available: https://www.nature.com/article...

2024
[12]

Large language models could change the future of behavioral healthcare: a proposal for responsible development and evaluation,

E. C. Stade, S. W. Stirman, L. H. Ungar, C. L. Boland, H. A. Schwartz, D. B. Yaden, J. Sedoc, R. J. DeRubeis, R. Willer, and J. C. Eichstaedt, “Large language models could change the future of behavioral healthcare: a proposal for responsible development and evaluation,” npj m...

2024
[13]

Read, diagnose and chat: Towards explainable and interactive llms-augmented depression detection in social media,

W. Qin, Z. Chen, L. Wang, Y . Lan, W. Ren, and R. Hong, “Read, diagnose and chat: Towards explainable and interactive llms-augmented depression detection in social media,” 2023. [Online]. Available: https://arxiv.org/abs/2305.05138

2023 arXiv
[14]

Depac: a corpus for depression and anxiety detection from speech,

M. Tasnim, M. Ehghaghi, B. Diep, and J. Novikova, “Depac: a corpus for depression and anxiety detection from speech,” 2023. [Online]. Available: https://arxiv.org/abs/2306.12443

2023 arXiv
[15]

Lora: Low-rank adaptation of large language models,

E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “Lora: Low-rank adaptation of large language models,”
[16]

Multitask learning,

R. Caruana, “Multitask learning,” Machine Learning, vol. 28, pp. 41– 75, 1997

1997
[17]

Multitask learning for mental health conditions with limited social media data,

A. Benton, M. Mitchell, and D. Hovy, “Multitask learning for mental health conditions with limited social media data,” ACL Anthology, pp. 152–162, 04 2017. [Online]. Available: https://aclanthology.org/E17- 1015/

2017
[18]

Detecting moments of change and suicidal risks in longitudinal user texts using multi-task learning,

T. Azim, L. Singh, and S. Middleton, “Detecting moments of change and suicidal risks in longitudinal user texts using multi-task learning,” ePrints Soton (University of Southampton) , pp. 213–218, 01 2022. [Online]. Available: https://aclanthology.org/2022.clpsych-1.19/

2022
[19]

Translating the BDI and BDI-II into the HAMD and vice versa with equipercentile linking,

T. A. Furukawa, M. Reijnders, S. Kishimoto, M. Sakata, R. J. DeRubeis, S. Dimidjian, D. J. A. Dozois, U. Hegerl, S. D. Hollon, R. B. Jarrett, F. Lesp ´erance, Z. V . Segal, D. C. Mohr, A. D. Simons, L. C. Quilty, C. F. Reynolds, C. Gentili, S. Leucht, R. R. Engel, and P. Cuijp...

2019
[20]

Does response on the phq-9 depression questionnaire predict subsequent suicide attempt or suicide death?

G. E. Simon, C. M. Rutter, D. Peterson, M. Oliver, U. Whiteside, B. Operskalski, and E. J. Ludman, “Does response on the phq-9 depression questionnaire predict subsequent suicide attempt or suicide death?” Psychiatric Services , vol. 64, pp. 1195–1202, 12 2013

2013
[21]

Validation of a composite of suicide items from the mood and feelings questionnaire (mfq) in offspring of recurrently depressed parents,

G. Hammerton, S. Zammit, R. Potter, A. Thapar, and S. Collishaw, “Validation of a composite of suicide items from the mood and feelings questionnaire (mfq) in offspring of recurrently depressed parents,” Psychiatry Research, vol. 216, pp. 82–88, 04 2014

2014
[22]

Evaluation of the phq-9 item 3 as a screen for sleep disturbance in primary care,

K. L. MacGregor, J. S. Funderburk, W. Pigeon, and S. A. Maisto, “Evaluation of the phq-9 item 3 as a screen for sleep disturbance in primary care,” Journal of General Internal Medicine, vol. 27, p. 339–344, 03 2012. [Online]. Available: https://www.ncbi.nlm.nih.gov/pmc/article...

2012
[23]

Llama: Open and efficient foundation language models,

H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi `ere, N. Goyal, E. Hambro, F. Azhar, A. Rodriguez, A. Joulin, E. Grave, and G. Lample, “Llama: Open and efficient foundation language models,” 2023. [Online]. Available: https://arxiv.org/abs/2302.13971

2023 arXiv
[24]

Robust speech recognition via large-scale weak super- vision,

A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak super- vision,” 2022. [Online]. Available: https://arxiv.org/abs/2212.04356

2022 arXiv
[25]

Speechmark: landmark detection tool for speech analysis,

S. Boyce, H. Fell, and J. Macauslan, “Speechmark: landmark detection tool for speech analysis,” in Interspeech, 09 2012, pp. 1894–1897

2012
[26]

Avec 2019 workshop and challenge: State-of-mind, detecting depression with ai, and cross-cultural affect recognition,

F. Ringeval, B. Schuller, M. Valstar, N. Cummins, R. Cowie, L. Tavabi, M. Schmitt, S. Alisamir, S. Amiriparian, E.-M. Messner, S. Song, S. Liu, Z. Zhao, A. Mallol-Ragolta, Z. Ren, M. Soleymani, and M. Pantic, “Avec 2019 workshop and challenge: State-of-mind, detecting depressi...

2019
[27]

Gpt understands, too,

X. Liu, Y . Zheng, Z. Du, M. Ding, Y . Qian, Z. Yang, and J. Tang, “Gpt understands, too,” 2023. [Online]. Available: https://arxiv.org/abs/2103.10385

2023 arXiv
[2019]

Available: https://arxiv.org/abs/1907.11510

[Online]. Available: https://arxiv.org/abs/1907.11510

1907 arXiv
[2021]

Available: https://arxiv.org/abs/2106.09685

[Online]. Available: https://arxiv.org/abs/2106.09685

arXiv
[2024]

Available: https://arxiv.org/abs/2402.13276

[Online]. Available: https://arxiv.org/abs/2402.13276

arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [11]

Prediction of relapse in adolescent depression using fusion of video and speech data

C. Lucasius, M. Ali, M. Battaglia, J. Strauss, P. Szatmari, and D. Kundur, “Prediction of relapse in adolescent depression using fusion of video and speech data.” [Online]. Available: https://ceur- ws.org/V ol-3649/Paper22.pdf

work page

[2] [1]

Adolescent depression: National trends, risk factors, and healthcare disparities,

W. Lu, “Adolescent depression: National trends, risk factors, and healthcare disparities,” American Journal of Health Behavior , vol. 43, pp. 181–194, 01 2019

work page 2019

[3] [2]

Sleep’s role in the development and resolution of adolescent depression,

M. Gradisar, M. Kahn, G. Micic, M. Short, C. Reynolds, F. Orchard, S. Bauducco, K. Bartel, and C. Richardson, “Sleep’s role in the development and resolution of adolescent depression,” Nature Reviews Psychology , vol. 1, p. 1–12, 06 2022. [Online]. Available: https://www.nature.com/articles/s44159-022-00074-8

work page 2022

[4] [3]

Elevated risk of depression among adolescents presenting with sleep disorders,

S. M. Inkelis, S. Ancoli-Israel, J. D. Thomas, and R. Bhattacharjee, “Elevated risk of depression among adolescents presenting with sleep disorders,” Journal of Clinical Sleep Medicine , 11 2020

work page 2020

[5] [4]

Language as a biomarker for psychosis: A natural language processing approach,

C. M. Corcoran, V . A. Mittal, C. E. Bearden, R. E. Gur, K. Hitczenko, Z. Bilgrami, A. Savic, G. A. Cecchi, and P. Wolff, “Language as a biomarker for psychosis: A natural language processing approach,” Schizophrenia Research, vol. 226, pp. 158–166, 12 2020

work page 2020

[6] [5]

Speech as a promising biosignal in precision psychiatry,

M. Kappen, M.-A. Vanderhasselt, and G. M. Slavich, “Speech as a promising biosignal in precision psychiatry,” Neuroscience & Biobehavioral Reviews, vol. 148, p. 105121, 2023. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0149763423000908

work page 2023

[7] [6]

The relationship between linguistic expression in blog content and symptoms of depression, anxiety, and suicidal thoughts: A longitudinal study,

B. O’Dea, T. W. Boonstra, M. E. Larsen, T. Nguyen, S. Venkatesh, and H. Christensen, “The relationship between linguistic expression in blog content and symptoms of depression, anxiety, and suicidal thoughts: A longitudinal study,” PLOS ONE , vol. 16, p. e0251787, 05 2021

work page 2021

[8] [7]

Mental-llm: Leveraging large language models for mental health prediction via online text data,

X. Xu, B. Yao, Y . Dong, S. Gabriel, H. Yu, J. Hendler, M. Ghassemi, A. K. Dey, and D. Wang, “Mental-llm: Leveraging large language models for mental health prediction via online text data,” Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 8, no. 1, p. 1–32, Mar. 2024. [Online]. Available: http://dx.doi.org/10.1145/3643540

doi:10.1145/3643540 2024