Pith. sign in

REVIEW 5 major objections 5 minor 64 references

Toward Generalizable Cognitive Impairment Detection with Speech-Based Multimodal Large Language Models

T0 review · 5 major / 5 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 By fusing acoustic embeddings and automatically transcribed text embeddings from open LLMs, this paper reports a state-of-the-art 92.4% accuracy for cognitive-impairment classification and strong cross-dataset transfer using only locally de

desk verdict A clearly written but over-sold evaluation: the 92.4% headline and cross-dataset numbers are not trustworthy until the authors check speaker overlap between ADReSS20 and ADReSSo21 and stop selecting classifiers on the test set. read the letter →

arxiv 2607.21496 v1 pith:3TDKJ6B3 submitted 2026-07-23 eess.SP cs.LG

Yingchao Huang , Xin Wang , Yuhan Su , Shanshan Yao This is my paper
classification eess.SPcs.LG
keywords cognitiveimpairmentdetectionspeechbiomarkersmultimodalfusionlargelanguagemodelsaudioembeddingstextautomaticrecognitionprivacy-preservingscreening
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 tries to establish that a straightforward multimodal fusion—acoustic embeddings from an audio-language model combined with linguistic embeddings from a text-language model applied to automatically transcribed speech—can detect cognitive impairment from spontaneous speech more accurately than prior single-modality systems. On the combined ADReSS20 and ADReSSo21 benchmarks the pipeline reports 92.4% accuracy and an F1 of 92.2%, and in cross-dataset tests it reports 91.5–91.7% accuracy without fine-tuning. The authors argue this shows the fused representations capture complementary disease-related signal rather than dataset-specific artifacts, and that because the components are open-source and locally deployable, the approach is compatible with privacy-sensitive clinical screening. The practical payoff is a non-invasive speech-based screening tool that could be run inside a clinic or even on a local device.

What carries the argument

The load-bearing mechanism is embedding fusion by concatenation. Each subject's audio is segmented, encoded into fixed vectors by an audio-language model (the paper uses Qwen2-Audio), and average-pooled; the same segments are transcribed by a single ASR pipeline and the full transcript is encoded by a text language model (the paper uses Qwen3). The two subject-level vectors are concatenated, Min–Max normalized using training-set statistics, and classified by a neural network. Concatenation plus normalization is the whole fusion strategy, so the paper's claim is that the complementary signal is already present in the off-the-shelf embeddings and only needs scaling and a flexible classifier.

What would settle it

De-duplicate the combined datasets at the speaker level (using anonymized identifiers or voice embeddings) and rerun the combined and cross-dataset protocols; if accuracy drops toward the single-modality baselines, the fusion-specific generalization claim is unsupported.

Watch

Extended reading notes

Core claim

The central claim is that concatenating two frozen embedding vectors—one representing how speech sounds, one representing what is said—and feeding the normalized vector to a small neural network yields state-of-the-art cognitive-impairment detection: 92.4% accuracy, 94.6% precision, 89.8% recall, and 92.2% F1 on the held-out combined test set. The paper further claims that the same pipeline transfers between the two benchmark datasets (91.5% and 91.7% accuracy in the two cross directions) without any adaptation, and that most errors occur in the mild-impairment band near the MMSE boundary, which is where early screening needs to work.

Load-bearing premise

The evaluation assumes the two benchmark datasets have no speaker overlap, because they were drawn from the same recruitment pool; the paper neither checks nor discusses whether a participant's speech appears in both the training and test splits.

Editorial extensions

If this is right

  • A 92.4% result on the combined test set would set a new published reference point for ADReSS20 and ADReSSo21, roughly 7.5 accuracy points above the strongest baseline cited in the paper.
  • The cross-dataset numbers imply a model trained on one recording protocol can be applied to another without per-site fine-tuning, a practical requirement for deployment across clinics with different equipment.
  • The ablation results imply that both modalities and the normalization step are each necessary; removing any one component costs between 6.7 and 17.6 accuracy points, so future systems should not drop either modality for convenience.
  • Because all components are locally runnable open-source models, the pipeline offers a concrete path to privacy-compliant cognitive screening without sending identifiable speech or transcripts to external servers.

Reading between the lines

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

  • A de-duplication test the paper leaves open: because the two benchmarks derive from a shared recruitment source, matching participants across datasets by speaker identity would tell whether the cross-dataset transfer is true generalization or partly speaker overlap.
  • The same concatenate-normalize-classify recipe could transfer to other audio-plus-text clinical tasks, such as depression, Parkinson's, or aphasia assessment, where sound and content carry complementary signals.
  • The MMSE-stratified error distribution suggests a practical triage tool: flag borderline cases in the 24–27 MMSE range for in-person follow-up rather than treating the model output as a hard diagnosis.
  • Since the paper only evaluates frozen embeddings, a natural next step is to fine-tune or adapt the audio and text encoders on unlabeled speech from the target clinic; the current gains may be a floor, not a ceiling.
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

5 major / 5 minor

Summary. The paper proposes a multimodal framework for cognitive-impairment (CI) detection from speech, using Qwen2-Audio to extract acoustic embeddings and Qwen3 to extract linguistic embeddings from ASR transcripts, then concatenating, normalizing, and classifying with standard machine-learning classifiers (SVC, XGBoost, logistic regression, neural network). The method is evaluated on the ADReSS20 and ADReSSo21 benchmarks in two settings: a combined training/test setup and a cross-dataset transfer setup. The authors report a held-out accuracy of 92.4% for the neural-network classifier, claim consistent improvements over single-modality baselines and prior work, and report cross-dataset accuracies of 91.5–91.7%, concluding that the method establishes a new state of the art while being privacy-preserving and locally deployable.

Significance. If the results were valid, the paper would make a useful contribution: an open-source, locally deployable LLM-based pipeline that matches or exceeds prior speech-based CI detection on standard benchmarks and transfers across datasets would be of practical interest for scalable cognitive screening. The strengths are the simple and reproducible pipeline design, the use of open models, and the explicit privacy-preserving goal. However, the central quantitative claims currently rest on unverified assumptions about dataset independence and on a test-set-based model selection procedure, so the significance of the contribution cannot be assessed from the present manuscript.

major comments (5)
  1. [Section 2.1 and Section 3.5 (Tables 2, 6, 7)] The paper asserts in Section 2.1 that ADReSS20 and ADReSSo21 were collected from 'distinct speaker cohorts and recording sessions,' but it provides no participant-ID or overlap analysis. Both datasets are derived from the Pitt/DementiaBank corpus, as the paper's own references [7], [13], [45] indicate. If the same speakers appear in both datasets—for example, in the training split of one and the test split of the other—then the combined evaluation in Table 2 and the cross-dataset results in Tables 6–7 are not independent and the 92.4% and 91.5–91.7% accuracies are inflated by speaker leakage. This is a load-bearing issue for the headline claims. The authors must either verify non-overlap with participant identifiers or re-run all experiments on verifiably disjoint participant sets.
  2. [Section 3.1 and Table 4] The reported 92.4% is the best test-set result among four classifiers, and Section 3.1 explicitly discusses the discrepancy between cross-validation and test rankings. Because the neural network was selected after seeing its performance on the held-out test set (Table 2 vs. Table 4), the headline number is an optimistically selected result, not a pre-registered comparison. This selection bias undermines the claim that the neural network 'consistently outperforms' other classifiers and inflates the apparent gain of the multimodal framework. A proper protocol would use a separate validation set or nested cross-validation for model selection and report test performance only for the selected model.
  3. [Section 3.3 and Table 3] The comparison with prior work in Table 3 is not controlled. The cited results (Mortensen et al., Bang et al., Agbavor et al., Luz et al.) come from different experimental setups: they may use different data splits, manual rather than ASR transcripts, different audio preprocessing, or different subsets of ADReSS20/ADReSSo21. In particular, the use of automatic transcripts for all methods in this paper is a deliberate design choice, but the same transcripts are not used for the baselines. The 7.5-point accuracy difference over Mortensen et al. is therefore not interpretable as a fair or meaningful comparison, and the claim of a 'clear improvement over existing methods' is not supported by the presented evidence.
  4. [Section 3.4 and Table 5] The ablation study in Table 5 is performed on the same test set that was used to select the best classifier. The reported contribution of normalization and of each modality is conditional on the neural-network classifier whose hyperparameters and architecture were chosen with knowledge of test-set performance. This makes the ablation results subject to the same selection-bias concern as the headline accuracy. The authors should either present the ablation for a pre-specified classifier or use a nested validation scheme.
  5. [Section 3.5 and Tables 6–7] The cross-dataset generalization claim depends directly on the unverified disjointness of the two datasets. Even if the datasets are disjoint, the results are reported without any statistical significance testing, and the direction-dependent differences (91.5% vs. 91.7%) are small. The paper should provide confidence intervals or significance tests, and, more fundamentally, should first establish that no speaker appears in both datasets before interpreting these numbers as evidence of generalization to 'diverse populations.'
minor comments (5)
  1. [Section 2.2.5, Eq. (14)] Min–Max normalization is defined using z_min and z_max computed from the training set; the manuscript should state explicitly that the same scaler is then applied to the test set, and clarify how it is applied in cross-dataset transfer.
  2. [Section 2.2.1, Eq. (5)] The segmentation length L is not specified. Since L affects the number of segments M_i and the temporal aggregation in Eq. (7), the value should be reported for reproducibility.
  3. [Global] In the Introduction the code is said to be 'made publicly available at https://github.com/kelci2017/CI_Multimodal', but the Data Availability section says 'Code will be shared upon request.' This inconsistency should be resolved.
  4. [Abstract] The phrase 'an CI classification accuracy' should be 'a CI classification accuracy.' Similar grammatical issues occur elsewhere (e.g., 'the neural networks substantially outperforms').
  5. [Figure captions and references] Figure 1 is referenced but not visible in the manuscript text; figures should be embedded. Also, references [50] and [63] appear to be the same paper (Mortensen and Zhu, AMIA 2025); they should be consolidated.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the pipeline is an empirical embedding+classifier system; the headline results rest on untested dataset-disjointness and test-set model selection, but these are validity threats, not derivation-by-construction.

full rationale

This paper is an empirical multimodal classification study, not a theoretical derivation. Its processing chain (Eqs. 1–15) takes frozen Qwen-based audio and text embeddings, concatenates them, applies min-max normalization, and fits a supervised classifier on training splits. The only fitted quantities are classifier parameters and hyperparameters selected through five-fold cross-validation; the reported test numbers are then obtained on held-out splits. No equation makes the predicted label definitionally equal to an input, and no parameter fitted to the test set is renamed as a prediction. The central advertised results (92.4% in Table 2; cross-dataset 91.5%/91.7% in Tables 6–7) depend on an assumption stated in Section 2.1 that ADReSS20 and ADReSSo21 were 'collected from distinct speaker cohorts and recording sessions' despite both deriving from the same Pitt corpus, and Section 3.5 claims generalization 'without any fine-tuning or overlap between datasets.' If participants overlap, the numbers would be inflated by leakage. That is a serious validity and reproducibility concern, but it is not circularity: the results are not equivalent to their inputs by construction. The single self-citation ([28], a generic multimodal-fusion support sentence) is not load-bearing. The practice of choosing the neural-network classifier after inspecting test accuracy is a soundness/selection-bias concern, not circularity. Thus no significant circularity is present; the main risks are external-validity and evaluation-protocol issues, which belong in a correctness review rather than a circularity verdict.

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

The framework introduces no new entities. The main un-validated assumptions are about dataset disjointness, ASR transcript adequacy, and transfer of Qwen embeddings to clinical elderly speech. The free parameters (L, ASR model, classifier hyperparameters) are unspecified, further hurting reproducibility.

free parameters (3)
  • Audio segmentation length L
    Equation (5) defines segmentation into L-second chunks, but L is never specified. It affects both ASR and acoustic embedding quality and is a tuning choice with no justification.
  • Classifier hyperparameters
    Section 3.1 says classifiers were optimized by grid search within five-fold cross-validation, but the grids and selected values for SVC (C, kernel), logistic regression (C), XGBoost (n_estimators, depth), and neural networks (layers, learning rate) are not reported, making the exact pipeline non-reproducible.
  • ASR model choice
    The ASR model Ψ_ASR is never named; the resulting transcripts are the sole text source. Different ASR error rates could change linguistic embeddings and all downstream results.
assumptions (3)
  • domain assumption ADReSS20 and ADReSSo21 are disjoint participant cohorts, so no speaker appears in both datasets.
    Both datasets derive from the Pitt corpus. The entire cross-dataset generalization evaluation (Section 3.5) and the combined train/test split (Section 2.1) depend on this assumption. The paper does not check for subject overlap.
  • domain assumption Automatic transcripts generated by a single ASR pipeline preserve all diagnostically relevant linguistic information compared to manual transcripts.
    The method discards the official manual transcripts and uses ASR output for all text embeddings. If ASR errors have systematic bias toward CI speech, the text embeddings could carry acoustic artifacts rather than genuine linguistic markers.
  • domain assumption Pre-trained Qwen2-Audio and Qwen3 embeddings transfer to deaf/older-adult speech populations.
    The paper itself criticizes wav2vec2.0 and VGGish for being trained on general populations that may not match elderly speakers (Section 1). The same concern applies to Qwen models, but the paper does not address how Qwen embeddings handle age-related acoustic changes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Toward Generalizable Cognitive Impairment Detection with Speech-Based Multimodal Large Language Models." pith.science (2026). https://pith.science/paper/3TDKJ6B3

@misc{pith2026260721496,
  author       = {Pith},
  title        = {Pith review of: Toward Generalizable Cognitive Impairment Detection with Speech-Based Multimodal Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3TDKJ6B3}},
  note         = {Machine review of arXiv:2607.21496}
}
read the original abstract

Cognitive impairment (CI) is a growing public health concern. Early and accurate diagnosis is critical for enabling timely intervention and improving patient outcomes. Speech-based CI detection has emerged as a promising non-invasive approach, as speech signals encode both linguistic and acoustic markers associated with cognitive decline. Recent advances in large language models (LLMs) further strengthen the potential of speech-based assessment by enabling more expressive representation learning and improved generalization across diverse speakers, recording devices, and clinical environments. Moreover, multimodal learning by jointly modeling linguistic and acoustic features allows for a more comprehensive characterization of cognitive and behavioral changes related to CI, leading to more reliable detection. In this work, we propose a multimodal CI detection framework based on open-source LLMs that integrates speech audio and corresponding transcripts while preserving patient privacy. Acoustic embeddings are extracted directly from speech signals, while textual embeddings are generated from automatically transcribed speech. These modality-specific embeddings are then concatenated to create a combined feature vector and used for downstream classification, without requiring access to raw or sensitive patient data. The proposed approach is evaluated on the ADReSS20 and ADReSSo21 benchmark datasets. Experimental results show that the proposed multimodal framework achieves an CI classification accuracy of 92.4% and consistently outperforms single-modality baselines. Our work establishes a new state-of-the-art for CI identification, with the proposed method demonstrating superior cross-dataset generalization. This advance highlights the power of an LLM-based multimodal framework that fuses linguistic and acoustic data to enable robust, scalable, and non-invasive screening.

Figures

Figures reproduced from arXiv: 2607.21496 by the authors.

Figure 1
Figure 1. The workflow of the proposed framework. 2.2.1. Audio Preprocessing The raw continuous-time speech signal is first subjected to preprocessing to reduce variability introduced by recording conditions, speaker characteristics, and dataset-specific acquisition protocols, and to ensure robust feature extraction for both ASR and audio embedding generation. This preprocessing stage standardizes the audio input across subje… view at source ↗
Figure 2
Figure 2. Histogram of correct and incorrect CI classifications with thresholds indicating MMSE cognitive impairment level. Means and standard [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. ROC curve comparing multimodal and single modalities with 5-fold cross validation derived AUC standard deviations. [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

64 extracted references · 12 canonical work pages

  1. [7]

    J. T. Becker, F. Boller, O. L. Lopez, J. Saxton, K. L. McGonigle, The natural history of alzheimer’s disease: Description of study cohort and accuracy of diagnosis, Archives of Neurology 51 (6) (1994) 585–594.doi: 10.1001/archneur.1994.00540180063015

    arXiv 1994

  2. [13]

    S. Luz, F. Haider, S. De La Fuente, D. Fromm, B. Macwhinney, Alzheimer’s dementia recognition through spon- taneous speech: The adress challenge, in: Proceedings of Interspeech, 2020.doi:10.21437/Interspeech. 2020-2571. URLhttp://dx.doi.org/10.21437/Interspeech.2020-2571

    doi:10.21437/interspeech 2020

  3. [45]

    S. Luz, F. Haider, S. de la Fuente, D. Fromm, B. MacWhinney, Detecting cognitive decline using speech only: The adresso challenge, arXiv preprint arXiv:2104.09356 (2021). URLhttps://arxiv.org/abs/2104.09356

    arXiv 2021

  4. [1]

    URLhttps://alz-journals.onlinelibrary.wiley.com/doi/abs/10.1002/alz.13016

    2023 alzheimer’s disease facts and figures, Alzheimer’s & Dementia 19 (4) (2023) 1598–1695.arXiv:https: //alz-journals.onlinelibrary.wiley.com/doi/pdf/10.1002/alz.13016,doi:https://doi.org/ 10.1002/alz.13016. URLhttps://alz-journals.onlinelibrary.wiley.com/doi/abs/10.1002/alz.13016

    work page doi:10.1002/alz.13016 2023

  5. [2]

    Crous-Bou, C

    M. Crous-Bou, C. Minguillón, N. Gramunt, J. L. Molinuevo, Alzheimer’s disease prevention: from risk factors to early intervention, Alzheimer’s Research & Therapy 9 (1) (2017) 71.doi:10.1186/s13195-017-0297-z. URLhttps://doi.org/10.1186/s13195-017-0297-z

    work page doi:10.1186/s13195-017-0297-z 2017

  6. [3]

    M. A. Nowrangi, P. B. Rosenberg, J. M. S. Leoutsakos, Subtle changes in daily functioning predict conversion from normal to mild cognitive impairment or dementia: An analysis of the nacc database, International Psy- chogeriatrics 28 (12) (2016) 2009–2018. URLhttps://pmc.ncbi.nlm.nih.gov/articles/PMC5628501/

    2016

  7. [4]

    Shahin, B

    M. Shahin, B. Ahmed, J. Epps, Zero-shot cognitive impairment detection from speech using audiollm (2025). arXiv:2506.17351. URLhttps://arxiv.org/abs/2506.17351

    arXiv 2025

  8. [5]

    Zolnoori, A

    M. Zolnoori, A. Zolnour, M. Topaz, Adscreen: A speech processing-based screening system for automatic identification of patients with alzheimer’s disease and related dementia, Artificial Intelligence in Medicine 143 (2023) 102624.doi:https://doi.org/10.1016/j.artmed.2023.102624. URLhttps://www.sciencedirect.com/science/article/pii/S0933365723001380 15

    arXiv 2023

Show all 64 references
  1. [6]

    K. Ding, M. Chetty, Azadeh, N. Hoshyar, T. Bhattacharya, B. Klein, Speech based detection of alzheimer’s disease: A survey of ai techniques, datasets and challenges, Artificial Intelligence Review 57 (12) (2024) 1–43. URLhttps://link.springer.com/article/10.1007/s10462-024-10961-6

    2024 doi

  2. [8]

    De La Fuente Garcia, C

    S. De La Fuente Garcia, C. W. Ritchie, S. Luz, Artificial intelligence, speech, and language processing ap- proaches to monitoring alzheimer’s disease: A systematic review, Journal of Alzheimer’s Disease 78 (4) (2020) 1547–1574

    2020

  3. [9]

    Shadle, Phonetics, acoustic, in: K

    C. Shadle, Phonetics, acoustic, in: K. Brown (Ed.), Encyclopedia of Language & Linguistics (Second Edi- tion), second edition Edition, Elsevier, Oxford, 2006, pp. 442–460.doi:https://doi.org/10.1016/ B0-08-044854-2/00001-8. URLhttps://www.sciencedirect.com/science/article/pii/...

    2006

  4. [10]

    Martinc, S

    M. Martinc, S. Pollak, Tackling the adress challenge: A multimodal approach to the automated recognition of alzheimer’s dementia, in: Proceedings of Interspeech, 2020.doi:10.21437/Interspeech.2020-2202. URLhttp://dx.doi.org/10.21437/Interspeech.2020-2202

    2020 doi

  5. [11]

    Edwards, C

    E. Edwards, C. Dognin, B. Bollepalli, M. Singh, Multiscale system for alzheimer’s dementia recognition through spontaneous speech, in: Proceedings of Interspeech, 2020.doi:10.21437/Interspeech.2020-2781. URLhttp://dx.doi.org/10.21437/Interspeech.2020-2781

    2020 doi

  6. [12]

    S. Luz, S. De La Fuente Garcia, F. Haider, D. Fromm, B. Macwhinney, A. Lanzi, Y .-N. Chang, C.-J. Chou, Y .-C. Liu, Connected speech-based cognitive assessment in chinese and english, in: Proceedings of Interspeech, 2024. URLhttp://luzs.gitlab.io/taukadial/

    2024

  7. [14]

    Ortiz-Perez, J

    D. Ortiz-Perez, J. Garcia-Rodriguez, D. Tomás, Cognitive insights across languages: Enhancing multimodal interview analysis, in: Proceedings of Interspeech, International Speech Communication Association, 2024, pp. 952–956

    2024

  8. [15]

    A. P. Perez-Toro, T. Arias-Vergara, P. Klumpp, T. Weise, M. Schuster, E. Noth, J. R. Orozco-Arroyave, A. Maier, Multilingual speech and language analysis for the assessment of mild cognitive impairment: Outcomes from the taukadial challenge, in: Proceedings of Interspeech, 202...

    2024

  9. [16]

    Zhang, Q

    M. Zhang, Q. Cui, W. Li, W. Yu, L. Chen, W. Li, C. Zhu, Y . Lü, Augmented dialectal speech recognition for ai-based neuropsychological scale assessment in alzheimer’s disease, Biomedical Signal Processing and Control 99 (2025) 106821.doi:https://doi.org/10.1016/j.bspc.2024.106...

    2025

  10. [17]

    Zolnour, H

    A. Zolnour, H. Azadmaleki, Y . Haghbin, F. Taherinezhad, M. J. M. Nezhad, S. Rashidi, M. Khani, A. Taleban, S. M. Sani, M. Dadkhah, J. M. Noble, S. Bakken, Y . Yaghoobzadeh, A. H. Vahabie, M. Rouhizadeh, M. Zolnoori, Llmcare: Early detection of cognitive impairment via transfo...

    2025

  11. [18]

    V . D. Badal, J. M. Reinen, E. W. Twamley, E. E. Lee, R. P. Fellows, E. Bilal, C. A. Depp, Investigating acoustic and psycholinguistic predictors of cognitive impairment in older adults: Modeling study, JMIR Aging 7 (2024) e54655.doi:10.2196/54655. URLhttps://doi.org/10.2196/54655 16

    2024 doi

  12. [19]

    Mahon, M

    E. Mahon, M. E. Lachman, V oice biomarkers as indicators of cognitive changes in middle and later adulthood, Neurobiology of Aging 119 (2022) 22–35.doi:https://doi.org/10.1016/j.neurobiolaging.2022. 06.010. URLhttps://www.sciencedirect.com/science/article/pii/S0197458022001415

    2022 doi

  13. [20]

    Galluzzi, W

    F. Galluzzi, W. Garavello, The aging voice: A systematic review of presbyphonia, European Geriatric Medicine 9 (2018) 559–570.doi:10.1007/s41999-018-0095-6. URLhttps://doi.org/10.1007/s41999-018-0095-6

    2018 doi

  14. [21]

    M. Vaca, E. Mora, I. Cobeta, The aging voice, Otolaryngology–Head and Neck Surgery 153 (3) (2015) 409–413.arXiv:https://aao-hnsfjournals.onlinelibrary.wiley.com/doi/pdf/10.1177/ 0194599815592373,doi:https://doi.org/10.1177/0194599815592373. URLhttps://aao-hnsfjournals.onlineli...

    2015 doi

  15. [22]

    Baevski, H

    A. Baevski, H. Zhou, A. Mohamed, M. Auli, wav2vec 2.0: A framework for self-supervised learning of speech representations (2020).arXiv:2006.11477. URLhttps://arxiv.org/abs/2006.11477

    2020 arXiv

  16. [23]

    Hershey, S

    S. Hershey, S. Chaudhuri, D. P. W. Ellis, J. F. Gemmeke, A. Jansen, R. C. Moore, M. Plakal, D. Platt, R. A. Saurous, B. Seybold, M. Slaney, R. J. Weiss, K. Wilson, Cnn architectures for large-scale audio classification, in: Proceedings of the IEEE International Conference on A...

    2017

  17. [24]

    F. F. Poor, H. H. Dodge, M. H. Mahoor, A multimodal cross-transformer-based model to predict mild cognitive impairment using speech, language and vision, Computers in Biology and Medicine 182 (2024) 109199.doi: https://doi.org/10.1016/j.compbiomed.2024.109199. URLhttps://www.s...

    2024

  18. [25]

    J. Gao, P. Li, Z. Chen, J. Zhang, A survey on deep learning for multimodal data fusion, Neural Computa- tion 32 (5) (2020) 829–864.arXiv:https://direct.mit.edu/neco/article-pdf/32/5/829/1865303/ neco_a_01273.pdf,doi:10.1162/neco_a_01273. URLhttps://doi.org/10.1162/neco_a_01273

    2020 doi

  19. [26]

    Lazli, F

    L. Lazli, F. Cheriet, M. Boukadoum, Multiclass prediction of alzheimer’s disease using balanced multimodal data and deep ensemble learning, Biomedical Signal Processing and Control 114 (2026) 109026.doi:https: //doi.org/10.1016/j.bspc.2025.109026. URLhttps://www.sciencedirect....

    2026

  20. [27]

    P. Lu, L. Hu, A. Mitelpunkt, S. Bhatnagar, L. Lu, H. Liang, A hierarchical attention-based multimodal fusion framework for predicting the progression of alzheimer’s disease, Biomedical Signal Processing and Control 88 (2024) 105669.doi:https://doi.org/10.1016/j.bspc.2023.10566...

    2024

  21. [28]

    Huang, B

    Y . Huang, B. Duan, Z. Li, Q. Zhang, Mssafn: A multiscale structure-aware spatial fusion network for alzheimer’s disease classification, Biomedical Signal Processing and Control 113 (2026) 108889.doi:https://doi.org/ 10.1016/j.bspc.2025.108889. URLhttps://www.sciencedirect.com...

    2026

  22. [29]

    Y . Song, L. Xu, An integrated multi-modal and multi-task learning method for alzheimer’s disease diagnosis, Biomedical Signal Processing and Control 113 (2026) 108811.doi:https://doi.org/10.1016/j.bspc. 2025.108811. URLhttps://www.sciencedirect.com/science/article/pii/S174680...

    2026

  23. [30]

    Rohanian, J

    M. Rohanian, J. Hough, M. Purver, Multi-modal fusion with gating using audio, lexical and disfluency features for alzheimer’s dementia recognition from spontaneous speech, in: Proceedings of Interspeech, 2020, pp. 2187– 2191.doi:10.21437/Interspeech.2020-2721

    2020 doi

  24. [31]

    N. Wang, Y . Cao, S. Hao, Z. Shao, K. P. Subbalakshmi, Modular multi-modal attention network for alzheimer’s disease detection using patient audio and language data, in: Proceedings of Interspeech, 2021, pp. 3835–3839

    2021

  25. [32]

    J. Koo, J. H. Lee, J. Pyo, Y . Jo, K. Lee, Exploiting multimodal features from pre-trained networks for alzheimer’s dementia recognition, in: Proceedings of Interspeech, 2020, pp. 2217–2221

    2020

  26. [33]

    S. Li, H. Tang, Multimodal alignment and fusion: A survey (2025).arXiv:2411.17040. URLhttps://arxiv.org/abs/2411.17040

    2025

  27. [34]

    T. Mo, J. C. K. Lam, V . O. K. Li, L. Y . L. Cheung, Leveraging large language models for identifying interpretable linguistic markers and enhancing alzheimer’s disease diagnostics, medRxiv (2024).arXiv: https://www.medrxiv.org/content/early/2024/08/23/2024.08.22.24312463.full...

    2024

  28. [35]

    J. Peng, Y . Wang, Y . Xi, X. Li, X. Zhang, K. Yu, A survey on speech large language models, arXiv preprint (2024)

    2024

  29. [36]

    S. Hu, L. Zhou, S. Liu, S. Chen, L. Meng, H. Hao, J. Pan, X. Liu, J. Li, S. Sivasankaran, L. Liu, F. Wei, Wavllm: Towards robust and adaptive speech large language model, arXiv preprint (Mar. 2024).arXiv:2404.00656. URLhttp://arxiv.org/abs/2404.00656

    2024 arXiv

  30. [37]

    C. Tang, W. Yu, G. Sun, X. Chen, T. Tan, W. Li, L. Lu, Z. Ma, C. Zhang, Salmonn: Towards generic hearing abilities for large language models, in: Proceedings of the International Conference on Learning Representations (ICLR), 2024. URLhttps://github.com/bytedance/SALMONN

    2024

  31. [38]

    Y . Chu, J. Xu, X. Zhou, Q. Yang, S. Zhang, Z. Yan, C. Zhou, J. Zhou, Qwen-audio: Advancing universal audio understanding via unified large-scale audio-language models, arXiv preprint (2023). URLhttps://qwen-audio.github.io/

    2023

  32. [39]

    Y . Chu, J. Xu, Q. Yang, H. Wei, X. Wei, Z. Guo, Y . Leng, Y . Lv, J. He, J. Lin, C. Zhou, J. Zhou, Qwen2-audio technical report, arXiv preprint (2024). URLhttps://github.com/QwenLM/Qwen2-Audio

    2024

  33. [40]

    J. Bai, S. Bai, Y . Chu, Z. Cui, K. Dang, X. Deng, Y . Fan, W. Ge, Y . Han, F. Huang, B. Hui, L. Ji, M. Li, J. Lin, R. Lin, D. Liu, G. Liu, C. Lu, K. Lu, J. Ma, R. Men, X. Ren, X. Ren, C. Tan, S. Tan, J. Tu, P. Wang, S. Wang, W. Wang, S. Wu, B. Xu, J. Xu, A. Yang, H. Yang, J. ...

    2023 arXiv

  34. [41]

    Alsuhaibani, A

    M. Alsuhaibani, A. Pourramezan Fard, J. Sun, F. Far Poor, P. S. Pressman, M. H. Mahoor, A review of machine learning approaches for non-invasive cognitive impairment detection, IEEE access 13 (2025) 56355–56384

    2025

  35. [42]

    W. N. Price, I. G. Cohen, Privacy in the age of medical big data, Nature Medicine 25 (1) (2019) 37–43.doi: 10.1038/s41591-018-0272-7

    2019 doi

  36. [43]

    Yadav, S

    N. Yadav, S. Pandey, A. Gupta, P. Dudani, S. Gupta, K. Rangarajan, Data privacy in healthcare: In the era of artificial intelligence, Indian Dermatology Online Journal 14 (6) (2023) 788–792.doi:10.4103/idoj.idoj_ 543_23. 18

    2023 doi

  37. [44]

    S. Luz, F. Haider, S. de la Fuente, D. Fromm, B. MacWhinney, Alzheimer’s dementia recognition through spontaneous speech: The adress challenge (2020).arXiv:2004.06833. URLhttps://arxiv.org/abs/2004.06833

    2020 arXiv

  38. [46]

    Goodglass, E

    H. Goodglass, E. Kaplan, B. Barresi, Boston Diagnostic Aphasia Examination, 3rd Edition, Lippincott Williams & Wilkins, Philadelphia, 2001

    2001

  39. [47]

    Teipel, D

    S. Teipel, D. Gustafson, R. Ossenkoppele, O. Hansson, C. Babiloni, M. Wagner, S. G. Riedel-Heller, I. Kilimann, Y . Tang, Alzheimer disease: Standard of diagnosis, treatment, care, and prevention, Journal of Nuclear Medicine 63 (7) (2022) 981–985.arXiv:https://jnm.snmjournals....

    2022

  40. [48]

    Sharma, S

    A. Sharma, S. Kaur, N. Memon, A. Jainul Fathima, S. Ray, M. W. Bhatt, Alzheimer’s patients detection using support vector machine (svm) with quantitative analysis, Neuroscience Informatics 1 (3) (2021) 100012.doi: https://doi.org/10.1016/j.neuri.2021.100012. URLhttps://www.sci...

    2021

  41. [49]

    B. A. Llaca-Sánchez, L. R. García-Noguez, M. A. Aceves-Fernández, A. Takacs, S. Tovar-Arriaga, Exploring llm embedding potential for dementia detection using audio transcripts, Eng 6 (7) (2025).doi:10.3390/ eng6070163. URLhttps://www.mdpi.com/2673-4117/6/7/163

    2025

  42. [51]

    A. M. Kashyap, D. Rao, M. R. Boland, L. Shen, C. Callison-Burch, Predicting explainable dementia types with llm-aided feature engineering, Bioinformatics 41 (4) (2025) btaf156.doi:10.1093/bioinformatics/ btaf156

    2025 doi

  43. [52]

    Liashchynskyi, P

    P. Liashchynskyi, P. Liashchynskyi, Grid search, random search, genetic algorithm: A big comparison for nas (2019).arXiv:1912.06059. URLhttps://arxiv.org/abs/1912.06059

    2019 arXiv

  44. [53]

    R. Xiao, X. Cui, H. Qiao, X. Zheng, Y . Zhang, C. Zhang, X. Liu, Early diagnosis model of alzheimer’s disease based on sparse logistic regression with the generalized elastic net, Biomedical Signal Processing and Control 66 (2021) 102362.doi:https://doi.org/10.1016/j.bspc.2020...

    2021

  45. [54]

    J. V . Shanmugam, B. Duraisamy, B. C. Simon, P. Bhaskaran, Alzheimer’s disease classification using pre-trained deep networks, Biomedical Signal Processing and Control 71 (2022) 103217.doi:https://doi.org/10. 1016/j.bspc.2021.103217. URLhttps://www.sciencedirect.com/science/ar...

    2022

  46. [55]

    Botros, F

    J. Botros, F. Mourad-Chehade, D. Laplanche, Explainable multimodal data fusion framework for heart failure detection: Integrating cnn and xgboost, Biomedical Signal Processing and Control 100 (2025) 106997.doi: https://doi.org/10.1016/j.bspc.2024.106997. URLhttps://www.science...

    2025

  47. [56]

    Goodfellow, Y

    I. Goodfellow, Y . Bengio, A. Courville, Deep Learning, MIT Press, Cambridge, MA, 2016. 19

    2016

  48. [57]

    Hastie, R

    T. Hastie, R. Tibshirani, J. Friedman, The Elements of Statistical Learning: Data Mining, Inference, and Predic- tion, 2nd Edition, Springer, New York, NY , 2009

    2009

  49. [58]

    LeCun, Y

    Y . LeCun, Y . Bengio, G. Hinton, Deep learning, Nature 521 (7553) (2015) 436–444.doi:10.1038/ nature14539

    2015

  50. [59]

    Bengio, A

    Y . Bengio, A. Courville, P. Vincent, Representation learning: A review and new perspectives, IEEE Transactions on Pattern Analysis and Machine Intelligence 35 (8) (2013) 1798–1828.doi:10.1109/TPAMI.2013.50

    2013 doi

  51. [60]

    T. Chen, C. Guestrin, Xgboost: A scalable tree boosting system, in: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, ACM, 2016, pp. 785–794.doi:10.1145/ 2939672.2939785

    2016

  52. [61]

    Cortes, V

    C. Cortes, V . Vapnik, Support-vector networks, Machine Learning 20 (3) (1995) 273–297.doi:10.1007/ BF00994018

    1995

  53. [62]

    Kurlowicz, M

    L. Kurlowicz, M. Wallace, The mini-mental state examination (mmse), Journal of Gerontological Nursing 25 (5) (1999) 8–9.doi:10.3928/0098-9134-19990501-08. URLhttps://doi.org/10.3928/0098-9134-19990501-08

    1999 doi

  54. [63]

    G. A. Mortensen, R. Zhu, Early alzheimer’s detection through voice analysis: Harnessing locally deployable llms via adetectolocum, a privacy-preserving diagnostic system, AMIA Joint Summits on Translational Science Proceedings 2025 (2025) 365–374

    2025

  55. [64]

    Agbavor, H

    F. Agbavor, H. Liang, Artificial intelligence-enabled end-to-end detection and assessment of alzheimer’s disease using voice, Brain Sciences 13 (1) (2023) 28.doi:10.3390/brainsci13010028

    2023 doi

  56. [65]

    Bang, S.-H

    J.-U. Bang, S.-H. Han, B.-O. Kang, Alzheimer’s disease recognition from spontaneous speech using large lan- guage models, ETRI Journal 46 (1) (2024) 96–105.arXiv:https://onlinelibrary.wiley.com/doi/pdf/ 10.4218/etrij.2023-0356,doi:https://doi.org/10.4218/etrij.2023-0356. URLht...

    2024 doi

Pith tools

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