The DOI matches its Crossref record; 1 of the 128 references it cites were found in a registry or an index, and 127 are not checked yet; 14 of the 73 sources its claims cite were checked for support.
Explainable and Interpretable AI for Voice and Speech Analysis in Clinical Care: Systematic Review
Open accessAbstract
Background: Driven by recent advances in artificial intelligence (AI), particularly in medicine, audio-based voice and speech biomarkers are increasingly investigated for various medical applications as a complementary or even alternative modality to traditional medical devices. The adoption of deep learning techniques in recent literature is motivated by their superior performance compared to classical machine learning methods. However, ethical and regulatory concerns regarding the black-box nature of these models have limited their integration into clinical workflows. Consequently, explainable artificial intelligence (XAI) has recently been used to address this issue by generating explanations for opaque model outputs. Ideally, medical XAI systems aim to provide human-understandable, clinically grounded explanations essential for enhanced AI trustworthiness and, thereby, facilitate adoption into real-world clinical settings. Objective: We conduct a systematic literature review of XAI methods applied for explaining deep learning techniques in audio-based voice and speech clinical applications. We aim to identify what XAI methods have been used to explain the decisions of deep learning voice and speech AI systems in health care, as well as XAI-informed insights. Additionally, we aim to contextualize these findings with respect to clinical applicability and stakeholder relevance. Lastly, we identify opportunities and recommendations for future clinical audio XAI design. Methods: We used PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses). Six electronic databases (IEEE Xplore, ACM Digital Library, Scopus, PubMed, Web of Science, and Nature) were searched for papers published between January 2015 and February 2025. Eligible studies applied explainability or interpretability methods to deep learning models for voice or speech audio in health care contexts. Risk of bias was assessed using PROBAST+AI (Prediction Model Risk of Bias Assessment Tool). The results were thematically synthesized across explainability categories, input representations, clinical domains, validation strategies, and stakeholder considerations. Results: A total of 30 studies met the inclusion criteria. These studies used a range of explainability approaches, including gradient-based methods, perturbation-based techniques, surrogate model–based methods, model-internal representation analyses, concept-based detectors, and attention-based explanations. Applications spanned diverse clinical domains, including voice disorders, neurodegenerative diseases, psychiatric conditions, and traumatic brain injury. Overall, results indicate that most studies relied primarily on qualitative interpretation of explainability outputs, with limited quantitative validation of explanation consistency across external datasets. Furthermore, none of the included studies explicitly conducted human-in-the-loop evaluations with relevant stakeholders, highlighting a substantial gap in stakeholder alignment. Conclusions: Current XAI practices in clinical voice and speech analysis are limited by insufficient validation, lack of domain-specific design, and misalignment with clinical stakeholder needs. This review highlights opportunities for developing validated, audio-aware, and stakeholder-centered XAI approaches to support trustworthy clinical deployment. Interpretation of these findings should consider limitations related to single-reviewer study selection, potential high-risk of bias, and the repeated use of benchmark datasets.
Verified 2026-10-11 · 127 references not checked yet
Overview
CoChat Summary
TL;DR
This systematic review found 30 studies on explainable/interpretable AI for clinical voice and speech analysis, covering a wide range of XAI methods and clinical domains. Overall, the literature used explanations mostly qualitatively, with limited quantitative validation and no explicit human-in-the-loop evaluation with relevant stakeholders.
Key findings
- The review included 30 studies; all but 1 were published from 2020 onward. The studies spanned voice disorders, Parkinson disease, dementia/Alzheimer disease, psychiatric conditions, traumatic brain injury, aphasia, cleft palate hypernasality, head and neck cancer intelligibility, and autism spectrum disorder.
- The most frequently used explainability approaches were gradient-based saliency methods, input perturbation techniques, and model-internal representation analysis; specifically, Grad-CAM appeared in 5 studies, gradient-based saliency methods in 7 studies, t-SNE in 9 studies, and ablation-based analysis in 9 studies.
- Only 6/30 studies performed model-centric explainability validation using external data, and only 1 study used ground-truth/annotation-based verification; 10 studies relied only on qualitative interpretation and prior literature, and 12 studies reported explanations without quantitative analysis or qualitative literature comparison.
- Risk-of-bias assessment found 20/25 (80%) of model-development studies had high quality concern, and 23/25 (92%) of model-evaluation studies had high risk of bias; 6 studies had data leakage, and none explicitly reported formal human-centered evaluation of explanations with clinicians, speech-language pathologists, or regulators.
Methods
- Systematic review conducted according to PRISMA; a protocol was prepared but not registered.
- Searched 6 databases (IEEE Xplore, ACM Digital Library, Scopus, PubMed, Web of Science, Nature) for papers published January 2015 to February 2025; retrieved 1426 records, removed 78 duplicates, and included 30 studies after full-text screening.
- Eligible studies applied explainability/interpretability methods to deep learning models for voice or speech audio in health care; studies were thematically synthesized by explainability type, input/output representation, validation strategy, domain, and stakeholder considerations. Risk of bias was assessed with PROBAST+AI.
Limitations
- The paper states that study selection and data extraction were conducted by a single reviewer, which may increase the risk of selection bias or missed studies.
- The authors note that conclusions are constrained by the high risk of bias in many underlying models and by repeated use of the same benchmark datasets, raising the risk of circular validation; this is both stated by the paper and evident from the reported dataset reuse.
- The review only covers studies published up to February 2025, so later studies are not included.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Figures



Verification
Claims the paper makes
9 Valid · 3 Not valid · 52 Not checked · 117 Own results and interpretations
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“the intuitiveness of model explanations varies substantially with respect to end users [35-37]”
§ Explainable and Interpretable AI for Voice and Speech Analysis in Clinical Care: Systematic Review > Introduction > Overview
What the sources say
- Supportedcites [35] Wysocki, Assessing the communication gap between AI models and healthcare professionals: explainability, utility and trust in AI-driven clinical decision-making, Artif Intell source
The Discussion section, under “Unmet need to build pragmatically grounded explanatory models,” reports that 78% of healthcare professionals found the feature-contribution explanation easy to interpret, while 17% found the model explanation difficult. This supports the claim that how intuitive an explanation is varies among end users. (LLM, conf=0.89)
- Supportedcites [36] Bienefeld, Solving the explainable AI conundrum by bridging clinicians’ needs and developers’ goals, NPJ Digit Med source
The Abstract reports that developers and clinicians interpreted explanations differently: developers valued model interpretability, while clinicians found Shapley-value explanations unhelpful and sought clinical plausibility instead. This supports the claim that how intuitive or useful an explanation is depends on its audience, in the clinical setting studied. (LLM, conf=0.86)
- Not checked yetcites [37] Kim, Do stakeholder needs differ? - Designing stakeholder-tailored explainable artificial intelligence (XAI) interfaces, Int J Hum Comput Stud source
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Nonetheless, these reviews consistently highlight persistent challenges relating to faithfulness, stability, and standardized evaluation of explanations.”
§ Explainable and Interpretable AI for Voice and Speech Analysis in Clinical Care: Systematic Review > Introduction > XAI in Medicine
What the sources say
- Supportedcites [61] Tjoa, A survey on explainable artificial intelligence (XAI): toward medical XAI, IEEE Trans Neural Netw Learn Syst source
The survey identifies a lack of uniformly adopted interpretability-assessment criteria (Introduction; Types of Interpretability). Its sections on Manipulation of Explanations and Input Invariance discuss unreliable or wrong explanations and the need for explanations to behave consistently under changes to inputs. Together, these support the cited review’s treatment of faithfulness, stability, and evaluation as challenges; this source alone does not establish what other reviews say. (LLM, conf=0.86)
- Partly supportedcites [62] Chaddad, Survey of explainable AI techniques in healthcare, Sensors (Basel) sourceExperimental — accuracy not yet measured
Section 6.3 (Quality of Explanation; page unavailable) identifies unstable LIME and SHAP explanations, inaccurate CAM localization, and evaluation of explanations as an ongoing research challenge. These support concerns about stability and explanation accuracy, but the paper does not specifically identify a lack of standardized evaluation. (LLM, conf=0.84)
- Partly supportedcites [63] Sadeghi, A review of explainable artificial intelligence in healthcare, Comput Electr Eng sourceExperimental — accuracy not yet measured
Section 6.1 (p. 14) says that evaluating explanations remains an open issue and that some researchers question the reliability of XAI methods. Section 6.3 (p. 15) also identifies soundness as a challenge. These passages support concerns related to evaluation and faithfulness, but they do not establish a persistent challenge concerning the stability of explanations or specifically call for standardized evaluation. (LLM, conf=0.84)
- Not checked yetcites [64] Mienye, A survey of explainable artificial intelligence in healthcare: concepts, applications, and challenges, Inf Med Unlocked source
- Supportedcites [65] Sheu, A survey on medical explainable AI (XAI): recent progress, explainability approach, human interaction and scoring system, Sensors (Basel) source
Section 5.5 (Fidelity) says interpretations must align with model predictions, while Section 4.2 describes instability from randomness in LIME explanations. Sections 9.1–9.2 identify the absence of a rating standard and difficulty comparing XAI evaluation results. Together, these support the challenges attributed to this review; the paper does not establish what other reviews say. (LLM, conf=0.87)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“with only a single study [52] relying on remotely collected speech data”
§ Explainable and Interpretable AI for Voice and Speech Analysis in Clinical Care: Systematic Review > Results > Overview of Included Studies
What the sources say
- Supportedcites [52] Wang, The power of speech in the wild: discriminative power of daily voice diaries in understanding auditory verbal hallucinations using deep learning, Proc ACM Interact Mob Wearable Ubiquitous Technol source
The cited work is itself a study whose speech data were collected remotely: Section 3.1 (Procedures and Participants) states participants were 'recruited remotely over the Internet,' carried a smartphone and responded to voice-diary EMAs for 30 days, with data 'firstly stored locally on the phone and transferred to secure research servers,' and Section 1/3.3 describes 4809 (3033 analysed) voice diaries recorded in the naturalistic environment. Its Related Work explicitly frames this against prior speech research 'in controlled laboratory settings,' confirming the speech was gathered remotely rather than in-clinic. The 'only a single included study' quantification is the citing paper's own characterization of its review set, not a claim attributed to this source, so nothing here is unsupported. (LLM, conf=0.86)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“The concentration of quality concerns within the analysis domain, as identified by PROBAST+AI [80], has important implications for XAI applications and the interpretability of explanations.”
§ Explainable and Interpretable AI for Voice and Speech Analysis in Clinical Care: Systematic Review > Discussion > Methodological Quality of Underlying Prediction Models
What the sources say
- Could not verifycites [80] Moons, PROBAST+AI: an updated quality, risk of bias, and applicability assessment tool for prediction models using regression or artificial intelligence methods, BMJ sourceExperimental — accuracy not yet measured
Our copy of PROBAST+AI is a tool-development paper: it describes domain 4 (analysis) as covering choice of analysis method and key statistical considerations with five signalling questions for model development quality and seven for evaluation risk of bias (PROBAST+AI section), and explains how domain ratings are formed (Step 3, Step 4), but it reports no applied assessments and no per-domain distribution of concerns, so nothing states that quality concerns 'concentrate' in the analysis domain. The closest text — that prior reviews showed most AI-based prediction model studies were poorly reported and conducted, with factors such as small sample sizes, poor handling of missing data, failure to deal with overfitting and inadequate performance assessment (Development of PROBAST+AI section) — is credited to other systematic reviews (refs 6, 11, 12, 45, 46) rather than presented as a PROBAST+AI finding, and is not framed as a concentration in one domain. Searched: 'analysis domain high concern for quality', 'worked example / ratings per domain', 'explainable AI interpretability black box' — no passage addresses the claim's point (the XAI/interpretability implication is the citing authors' own inference). (LLM, conf=0.72)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Could not verify
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“SHAP was implemented in 4 studies spanning dementia detection [9], auditory verbal hallucination assessment [52], psychiatric disorder classification [51], and TBI [53].”
§ Explainable and Interpretable AI for Voice and Speech Analysis in Clinical Care: Systematic Review > Results > Explainability Methods
What the sources say
- Not checked yetcites [9] Zhang, DEMENTIA: a hybrid attention-based multimodal and multi-task learning framework with expert knowledge for Alzheimer’s disease assessment from speech, IEEE J Biomed Health Inform source
- Supportedcites [52] Wang, The power of speech in the wild: discriminative power of daily voice diaries in understanding auditory verbal hallucinations using deep learning, Proc ACM Interact Mob Wearable Ubiquitous Technol source
Section 6.1, “Important Interpretable Features,” says the study plotted SHAP values to interpret its predictions of auditory verbal hallucination (AVH) severity. This supports the AVH-assessment portion attached to citation [52]; the other three applications have separate citations. (LLM, conf=0.94)
- Not checked yetcites [51] Schultebraucks, Deep learning-based classification of posttraumatic stress disorder and depression following trauma utilizing visual and auditory markers of arousal and mood, Psychol Med source
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
103 more claims · 4 with their cited sources checked
Create a free account to see all 181 claims and to verify the 59 references still open.
- Probably not a substantial claim. 42
- It may state two separate claims as one. 43
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ Explainable and Interpretable AI for Voice and Speech Analysis in Clinical Care: Systematic Review > Discussion > Limitations
“Study selection and data extraction were conducted by a single reviewer.”
§ Explainable and Interpretable AI for Voice and Speech Analysis in Clinical Care: Systematic Review > Methods > Study Selection
“Full-text screening was performed by two reviewers, with deliberation and discussion used to resolve disagreements.”
Citations
Cites
Rayyan-a web and mobile app for systematic reviews.
Ouzzani, M., Hammady, H., Fedorowicz, Z., & Elmagarmid, A. (2016) · Systematic Reviews · doi:10.1186/s13643-016-0384-4 · Cited by 26,092
"Why Should I Trust You?": Explaining the Predictions of Any Classifier
Ribeiro, M. T., Singh, S., & Guestrin, C. (2016) · Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining · doi:10.1145/2939672.2939778 · Cited by 17,555
Rudin, C. (2019) · Nature Machine Intelligence · doi:10.1038/s42256-019-0048-x · Cited by 10,561
Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization
Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D. (2020) · International Journal of Computer Vision · doi:10.1007/s11263-019-01228-7 · Cited by 6,909
Rajpurkar, P., Chen, E., Banerjee, O., & Topol, E. J. (2022) · Nature Medicine · doi:10.1038/s41591-021-01614-0 · Cited by 3,212
PROBAST: A Tool to Assess the Risk of Bias and Applicability of Prediction Model Studies
Wolff, R. F., Moons, K. G. M., Riley, R. D., Whiting, P. F., Westwood, M., Collins, G. S., Reitsma, J. B., Kleijnen, J., Mallett, S., & for the PROBAST Group† (2019) · Annals of Internal Medicine · doi:10.7326/m18-1376 · Cited by 2,984
Explainable AI: a review of machine learning interpretability methods
Linardatos, P., Papastefanopoulos, V., & Kotsiantis, S. (2020) · Entropy · doi:10.3390/e23010018 · Cited by 2,922
A Survey on Explainable Artificial Intelligence (XAI): Toward Medical XAI
Tjoa, E., & Guan, C. (2021) · IEEE Transactions on Neural Networks and Learning Systems · doi:10.1109/tnnls.2020.3027314 · Cited by 2,524
Explaining Explanations: An Overview of Interpretability of Machine Learning
Gilpin, L. H., Bau, D., Yuan, B. Z., Bajwa, A., Specter, M., & Kagal, L. (2018) · 2018 IEEE 5th International Conference on Data Science and Advanced Analytics (DSAA) · doi:10.1109/dsaa.2018.00018 · Cited by 2,507
Explaining Deep Neural Networks and Beyond: A Review of Methods and Applications
Samek, W., Montavon, G., Lapuschkin, S., Anders, C. J., & Muller, K.-R. (2021) · Proceedings of the IEEE · doi:10.1109/jproc.2021.3060483 · Cited by 1,477
Cited by
Still being checked.
Metadata
- Authors
- Ebraheem, M., Toghranegar, J., Bridge2AI-Voice Consortium, Bensoussan, Y., & Templeton, J. M.
- Affiliations
- University of South Florida (US)
- Corresponding author
- Ebraheem, Mohamed
- Venue
- Journal of Medical Internet Research
- Publisher
- JMIR Publications Inc.
- Published
- 2026-06-24
- Volume
- 28
- Pages
- e83790-e83790
- ISSN
- 1438-8871 (ISSN-L)
- PMC
- PMC13293602
- OpenAlex
- W4414100850
- Cited by
- 0
- References
- 126
- FWCI
- 0.00
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- cc-by
- Locations
- Journal of Medical Internet Research (publishedVersion, cc-by)
- Funders
- National Institutes of Health (Bridge2AI)
- Common Fund
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
- Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
- Health Sciences › Health Professions › Speech and Hearing › Voice and Speech Disorders
- Keywords
- explainable AI, speech biomarkers, gradient-based explanation, attention-based explanations, surrogate models, neurodegenerative diseases, voice disorders, traumatic brain injury, psychiatric conditions, stakeholder alignment, clinical adoption of AI, human-in-the-loop evaluation, clinical workflow integration
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-11, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-11 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Explainable and Interpretable AI for Voice and Speech Analysis in Clinical Care: Systematic Review. https://app.cochat.ai/papers/10.2196/83790
- The paper
- Ebraheem, M., Toghranegar, J., Bridge2AI-Voice Consortium, Bensoussan, Y., & Templeton, J. M. (2026). Explainable and Interpretable AI for Voice and Speech Analysis in Clinical Care: Systematic Review. Journal of Medical Internet Research. https://doi.org/10.2196/83790
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.