Zhijun Yin
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Zhijun Yin has a total of 174 co-authors. Here are the top co-authors by number of shared publications:
- Bradley Malin
Department of Biomedical Informatics, Vanderbilt University Medical Center, Nashville, USA
28 shared publications - Congning Ni
Vanderbilt University
19 shared publications - Ellen Wright Clayton
Center for Biomedical Ethics and Society, Vanderbilt University Medical Center , Nashville, TN 37203
11 shared publications - Qingyuan Song
Vanderbilt University, Nashville, TN, USA
11 shared publications - Lijun Song
Vanderbilt University
9 shared publications - Chao Yan
Department of Biomedical Informatics, Vanderbilt University Medical Center , Nashville, TN 37203
7 shared publications
Latest Publications
- Open Access
CARE-MVLM: Counterfactual Abstention and Region-Grounded Evidence in Mammography Vision-Language Models
Mammography interpretation requires tying each finding to supporting evidence and withholding judgment when that evidence is unavailable. Vision-language models are increasingly used for image-based medical question answering, yet existing benchmarks suggest that they struggle to jointly support accurate answers, evidence grounding, and appropriate abstention. We propose CARE-MVLM, built on Qwen2.5-VL-7B-Instruct, to address this gap through separate modules for answer prediction, answer-conditioned evidence generation, and answerability selection. On 7,358 triplets curated from CBIS-DDSM, MIAS, and VinDr-Mammo, CARE-MVLM substantially improves lesion-sensitive selective behavior over same VLM backbone baselines and exhibits markedly stronger lesion-sensitive abstention than GPT-5.6 and Gemini-3.5-flash while attaining the highest joint answer, grounding, and abstention reliability in the comparison.
Open AccessCARE-MVLM: Counterfactual Abstention and Region-Grounded Evidence in Mammography Vision-Language Models
Mammography interpretation requires tying each finding to supporting evidence and withholding judgment when that evidence is unavailable. Vision-language models are increasingly used for image-based medical question answering, yet existing benchmarks suggest that they struggle to jointly support accurate answers, evidence grounding, and appropriate abstention. We propose CARE-MVLM, built on Qwen2.5-VL-7B-Instruct, to address this gap through separate modules for answer prediction, answer-conditioned evidence generation, and answerability selection. On 7,358 triplets curated from CBIS-DDSM, MIAS, and VinDr-Mammo, CARE-MVLM substantially improves lesion-sensitive selective behavior over same VLM backbone baselines and exhibits markedly stronger lesion-sensitive abstention than GPT-5.6 and Gemini-3.5-flash while attaining the highest joint answer, grounding, and abstention reliability in the comparison.
- Open Access
- 23 September 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Human-Computer Interaction
- Vol. 10 (6) , 1-35
One Question, Four Voices: How Advice for Alzheimer's Caregiving Differs Between Caregivers, Clinicians, and Large Language Models
Online caregiver forums are a critical support infrastructure for Alzheimer’s disease and related dementias (ADRD), where people seek both emotional reassurance and practical guidance. Prior CSCW and HCI work has examined peer and clinician support in these communities and, separately, explored support-oriented chatbots and AI-assisted helping; however, we still lack systematic evidence on how AI responses compare with peer and clinician roles when answering the same caregiver questions in situ. We address this gap through a systematic, side-by-side comparison of responses to 85 real caregiver questions from ALZConnected across four responder groups: peer caregivers , physicians , ChatGPT , and a retrieval-augmented CareGPT grounded in peer discussions. We evaluate responses using 28 measures spanning linguistic form (psycholinguistic cues, structure, and style), emotional support, informational support, and topic scope, including LLM-based scoring for selected measures that we audit and calibrate against double-coded human ratings. Across analyses, we observe a consistent division of communicative labor: peers provided time-anchored narratives and community grounding; clinicians provided concise, advice-forward, clinically bounded guidance; and LLMs produced a formal, verbose, highly structured “consultation” voice. Notably, CareGPT shows minimal measurable separation from ChatGPT, indicating that retrieval-plus-prompt-injection alone is insufficient to shift the model’s communicative stance. Quantitatively, LLM outputs score higher than peers on politeness and emotion-support markers, and are comparable to clinicians on informativeness and physician-alignment measures; yet, they are less narrative and less contextually situated. Topic modeling further shows that LLM responses concentrate narrowly on advice-oriented themes, while peers cover a broader range (e.g., emotional processing, legal/financial concerns, and daily cares). These insights show that AI systems can complement, but not replace, human contributions to ADRD caregiver support. We advocate for hybrid systems that combine AI’s structured delivery with the distinct contributions of human voices to strengthen emotional and informational caregiving support infrastructures, while future work should examine whether similar patterns hold in other caregiving domains.
- 23 September 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Human-Computer Interaction
- Vol. 10 (6) , 1-35
Open AccessOne Question, Four Voices: How Advice for Alzheimer's Caregiving Differs Between Caregivers, Clinicians, and Large Language Models
Online caregiver forums are a critical support infrastructure for Alzheimer’s disease and related dementias (ADRD), where people seek both emotional reassurance and practical guidance. Prior CSCW and HCI work has examined peer and clinician support in these communities and, separately, explored support-oriented chatbots and AI-assisted helping; however, we still lack systematic evidence on how AI responses compare with peer and clinician roles when answering the same caregiver questions in situ. We address this gap through a systematic, side-by-side comparison of responses to 85 real caregiver questions from ALZConnected across four responder groups: peer caregivers , physicians , ChatGPT , and a retrieval-augmented CareGPT grounded in peer discussions. We evaluate responses using 28 measures spanning linguistic form (psycholinguistic cues, structure, and style), emotional support, informational support, and topic scope, including LLM-based scoring for selected measures that we audit and calibrate against double-coded human ratings. Across analyses, we observe a consistent division of communicative labor: peers provided time-anchored narratives and community grounding; clinicians provided concise, advice-forward, clinically bounded guidance; and LLMs produced a formal, verbose, highly structured “consultation” voice. Notably, CareGPT shows minimal measurable separation from ChatGPT, indicating that retrieval-plus-prompt-injection alone is insufficient to shift the model’s communicative stance. Quantitatively, LLM outputs score higher than peers on politeness and emotion-support markers, and are comparable to clinicians on informativeness and physician-alignment measures; yet, they are less narrative and less contextually situated. Topic modeling further shows that LLM responses concentrate narrowly on advice-oriented themes, while peers cover a broader range (e.g., emotional processing, legal/financial concerns, and daily cares). These insights show that AI systems can complement, but not replace, human contributions to ADRD caregiver support. We advocate for hybrid systems that combine AI’s structured delivery with the distinct contributions of human voices to strengthen emotional and informational caregiving support infrastructures, while future work should examine whether similar patterns hold in other caregiving domains.
- Open Access
Accuracy Overstates Evidence Grounding and Abstention Reliability in Mammography Vision-Language Models
Answer accuracy alone cannot determine whether a Vision-Language Model (VLM) relies on clinically relevant mammographic evidence or recognizes when that evidence is unavailable. We introduce an evidence-grounded selective evaluation benchmark that evaluates Pathology classification and Abnormality identification together with label-aware lesion localization. Each original image--task instance is paired with a lesion-removal counterfactual and, when feasible, a size-matched non-lesion random-removal control. Lesion evidence is removed by local tissue reconstruction, while the random-removal view provides a task-irrelevant regional perturbation baseline of comparable spatial extent. We evaluate 16 general-purpose and medically specialized VLMs, including open-source and proprietary models, under a common protocol with the same prompt, image conditions, label spaces, and structured-output requirements. The results show substantial gaps between classification performance and evidence-grounded reliability. InternVL3.5-8B achieved the highest Pathology Macro-F1 at 53.75\%, Huatuo Vision-34B achieved the highest Abnormality Exact Match at 52.44\%, and GPT-5.6 Sol achieved the highest Grounded Answer Success at 21.99\%. Lingshu-7B achieved the highest Counterfactual Specificity at 35.19\%, whereas GPT-5.6 Sol achieved the highest Joint Reliability at 6.47\%. These findings suggest that answer accuracy and abstention behavior can substantially overstate the reliability of current VLMs when predictions are not verified against the visual evidence on which they should depend.
Open AccessAccuracy Overstates Evidence Grounding and Abstention Reliability in Mammography Vision-Language Models
Answer accuracy alone cannot determine whether a Vision-Language Model (VLM) relies on clinically relevant mammographic evidence or recognizes when that evidence is unavailable. We introduce an evidence-grounded selective evaluation benchmark that evaluates Pathology classification and Abnormality identification together with label-aware lesion localization. Each original image--task instance is paired with a lesion-removal counterfactual and, when feasible, a size-matched non-lesion random-removal control. Lesion evidence is removed by local tissue reconstruction, while the random-removal view provides a task-irrelevant regional perturbation baseline of comparable spatial extent. We evaluate 16 general-purpose and medically specialized VLMs, including open-source and proprietary models, under a common protocol with the same prompt, image conditions, label spaces, and structured-output requirements. The results show substantial gaps between classification performance and evidence-grounded reliability. InternVL3.5-8B achieved the highest Pathology Macro-F1 at 53.75\%, Huatuo Vision-34B achieved the highest Abnormality Exact Match at 52.44\%, and GPT-5.6 Sol achieved the highest Grounded Answer Success at 21.99\%. Lingshu-7B achieved the highest Counterfactual Specificity at 35.19\%, whereas GPT-5.6 Sol achieved the highest Joint Reliability at 6.47\%. These findings suggest that answer accuracy and abstention behavior can substantially overstate the reliability of current VLMs when predictions are not verified against the visual evidence on which they should depend.
- Open Access
Online Peer Support Use Among Informal Caregivers of People Living With Alzheimer Disease or Related Dementias: National Cross-Sectional Web-Based Survey
Structured Background: Informal caregivers of persons living with Alzheimer’s disease or related dementias (ADRD) increasingly seek support in online environments. Prior work suggests that perceived value may shape online support use, yet practical and trust-related barriers may still impede uptake even when beliefs are favorable. Objective: To characterize patterns of online peer support use among informal ADRD caregivers and identify and analyze the predisposing, enabling, and need factors under the Andersen and Newman Framework of Health Services Utilization (ANFHSU) framework that are associated with use and nonuse behaviors. Methods: We conducted a cross-sectional, web-based survey that recruited a national, nonprobability sample of US adults through Qualtrics online panels from July 15 to September 16, 2023. Online peer support use was defined as searching for information or discussing caregiving topics in online communities during the past 3 months. Guided by ANFHSU, we measured predisposing characteristics, enabling resources, and need-related factors, along with additional technology-related preference measures, including willingness to use an experience-search tool and a caregiving-education tool. We summarized patterns in platform ecosystem and engagement modes and fitted 5 nested multivariable logistic regression models to identify factors associated with recent online peer support use. Among non-users, we evaluated multiple thresholds to identify a high-belief subgroup, summarized reported barriers, and modeled intention to use online peer support in the next 3 months. Results: Of 18,245 invited panel participants, 12,072 completed the survey (participation rate, 66.2%). After eligibility screening and data-quality exclusions, 1,113 unpaid informal ADRD caregivers comprised the analytic sample. Overall, 740/1,113 (66.5%) caregivers reported online peer support use in the past 3 months. Use was distributed across a fragmented platform ecosystem, with a long tail of infrequently named sources. More than half of users (402/740, 54.3%) primarily read rather than wrote posts, indicating substantial passive participation. In the fully adjusted model, higher belief in the value of online peer support was associated with greater odds of use (OR 1.04, 95% CI 1.01-1.08), along with higher eHealth literacy (OR 1.06, 95% CI 1.02-1.10), greater willingness to use an experience-search tool (OR 1.68, 95% CI 1.39-2.04), and higher caregiving stress (OR 1.03, 95% CI 1.01-1.05). Among non-users (373/1,113, 33.5%), 108/373 (28.9%) met the retained high-belief threshold despite reporting no recent use, indicating a belief-behavior gap. Among non-users, the most commonly reported barriers were sufficient offline support (118/373, 31.6%), lack of time (102/373, 27.3%), dislike of online communities (79/373, 21.2%), limited opportunity to join despite awareness (74/373, 19.8%), and security or trust concerns (67/373, 18.0%). Among non-users, 161/373 (43.2%) reported intending to use online peer support in the next 3 months. Conclusions: Our findings suggest that uptake of online peer support reflects perceived value, digital readiness, and caregiving-related need rather than need alone. A meaningful belief-behavior gap indicates that favorable beliefs do not ensure use when time, trust, and access barriers persist. These findings suggest that efforts to support ADRD caregivers should focus on reducing practical and trust-related barriers to online peer support and should recognize that engagement often occurs through reading rather than posting.
Open AccessOnline Peer Support Use Among Informal Caregivers of People Living With Alzheimer Disease or Related Dementias: National Cross-Sectional Web-Based Survey
Structured Background: Informal caregivers of persons living with Alzheimer’s disease or related dementias (ADRD) increasingly seek support in online environments. Prior work suggests that perceived value may shape online support use, yet practical and trust-related barriers may still impede uptake even when beliefs are favorable. Objective: To characterize patterns of online peer support use among informal ADRD caregivers and identify and analyze the predisposing, enabling, and need factors under the Andersen and Newman Framework of Health Services Utilization (ANFHSU) framework that are associated with use and nonuse behaviors. Methods: We conducted a cross-sectional, web-based survey that recruited a national, nonprobability sample of US adults through Qualtrics online panels from July 15 to September 16, 2023. Online peer support use was defined as searching for information or discussing caregiving topics in online communities during the past 3 months. Guided by ANFHSU, we measured predisposing characteristics, enabling resources, and need-related factors, along with additional technology-related preference measures, including willingness to use an experience-search tool and a caregiving-education tool. We summarized patterns in platform ecosystem and engagement modes and fitted 5 nested multivariable logistic regression models to identify factors associated with recent online peer support use. Among non-users, we evaluated multiple thresholds to identify a high-belief subgroup, summarized reported barriers, and modeled intention to use online peer support in the next 3 months. Results: Of 18,245 invited panel participants, 12,072 completed the survey (participation rate, 66.2%). After eligibility screening and data-quality exclusions, 1,113 unpaid informal ADRD caregivers comprised the analytic sample. Overall, 740/1,113 (66.5%) caregivers reported online peer support use in the past 3 months. Use was distributed across a fragmented platform ecosystem, with a long tail of infrequently named sources. More than half of users (402/740, 54.3%) primarily read rather than wrote posts, indicating substantial passive participation. In the fully adjusted model, higher belief in the value of online peer support was associated with greater odds of use (OR 1.04, 95% CI 1.01-1.08), along with higher eHealth literacy (OR 1.06, 95% CI 1.02-1.10), greater willingness to use an experience-search tool (OR 1.68, 95% CI 1.39-2.04), and higher caregiving stress (OR 1.03, 95% CI 1.01-1.05). Among non-users (373/1,113, 33.5%), 108/373 (28.9%) met the retained high-belief threshold despite reporting no recent use, indicating a belief-behavior gap. Among non-users, the most commonly reported barriers were sufficient offline support (118/373, 31.6%), lack of time (102/373, 27.3%), dislike of online communities (79/373, 21.2%), limited opportunity to join despite awareness (74/373, 19.8%), and security or trust concerns (67/373, 18.0%). Among non-users, 161/373 (43.2%) reported intending to use online peer support in the next 3 months. Conclusions: Our findings suggest that uptake of online peer support reflects perceived value, digital readiness, and caregiving-related need rather than need alone. A meaningful belief-behavior gap indicates that favorable beliefs do not ensure use when time, trust, and access barriers persist. These findings suggest that efforts to support ADRD caregivers should focus on reducing practical and trust-related barriers to online peer support and should recognize that engagement often occurs through reading rather than posting.
- Open Access
From Clinical Free Text to Auditable Concepts: An Agentic Framework for Interpretable Prediction
Across application domains, predictive signals often sit in unstructured free text rather than structured fields, yet turning that text into useful and interpretable features is difficult. Running large language models (LLMs) over an entire corpus is costly and hard to reproduce, while end-to-end text representations can rely on surface cues that are difficult to inspect. We present an agentic workflow that takes a prediction task and a raw text corpus as input and produces an auditable feature layer. The first two agents use an LLM to derive a task-specific predictor taxonomy and weakly label a bounded text sample; routed local extractors then process the corpus, and a deterministic builder aggregates the evidence into a dynamic, longitudinal concept bottleneck. We evaluate the framework on medication discontinuation in a longitudinal oncology cohort and 30-day readmission in MIMIC-IV. With gradient boosting, the longitudinal bottleneck increases area under the receiver operating characteristic curve (AUROC) over coarse concept buckets from 0.700 to 0.761 for medication discontinuation and from 0.576 to 0.609 for readmission. The proposed framework achieves predictive performance comparable to direct BioClinicalBERT prediction on both tasks while additionally providing explicit, interpretable, and traceable task-specific concepts. LLM use is confined to a bounded weak-labeling stage costing $24.00 and $23.39, respectively, compared with projected costs of $10,648 and $11,519 for exhaustive sentence-level LLM processing of the full corpora, demonstrating the substantial cost efficiency of the proposed agentic system.
Open AccessFrom Clinical Free Text to Auditable Concepts: An Agentic Framework for Interpretable Prediction
Across application domains, predictive signals often sit in unstructured free text rather than structured fields, yet turning that text into useful and interpretable features is difficult. Running large language models (LLMs) over an entire corpus is costly and hard to reproduce, while end-to-end text representations can rely on surface cues that are difficult to inspect. We present an agentic workflow that takes a prediction task and a raw text corpus as input and produces an auditable feature layer. The first two agents use an LLM to derive a task-specific predictor taxonomy and weakly label a bounded text sample; routed local extractors then process the corpus, and a deterministic builder aggregates the evidence into a dynamic, longitudinal concept bottleneck. We evaluate the framework on medication discontinuation in a longitudinal oncology cohort and 30-day readmission in MIMIC-IV. With gradient boosting, the longitudinal bottleneck increases area under the receiver operating characteristic curve (AUROC) over coarse concept buckets from 0.700 to 0.761 for medication discontinuation and from 0.576 to 0.609 for readmission. The proposed framework achieves predictive performance comparable to direct BioClinicalBERT prediction on both tasks while additionally providing explicit, interpretable, and traceable task-specific concepts. LLM use is confined to a bounded weak-labeling stage costing $24.00 and $23.39, respectively, compared with projected costs of $10,648 and $11,519 for exhaustive sentence-level LLM processing of the full corpora, demonstrating the substantial cost efficiency of the proposed agentic system.
- Open Access
CPT/HCPCS Code Recommendation from Clinical Notes: A Comparative Evaluation of AI Methods
Automatic coding from clinical notes has been studied extensively for International Classification of Diseases (ICD) codes, yet broad Current Procedural Terminology (CPT) and Healthcare Common Procedure Coding System (HCPCS) recommendation remains comparatively underexplored. Existing studies often focus on one specialty, a limited code vocabulary, or a single model family, leaving it unclear how different artificial intelligence (AI) paradigms perform under a common, clinically meaningful evaluation. We formulate CPT and HCPCS coding as an AI-assisted recommendation task in which a physician or professional coder reviews a short, ranked list of candidate codes supported by the clinical note. Using operative notes from Vanderbilt University Medical Center (VUMC) and discharge summaries from Medical Information Mart for Intensive Care IV (MIMIC-IV), we compare lexical retrieval, Clinical-Longformer, GPT-5.6-Sol, MedGemma-27B, and an inspectable agentic-style retrieve-and-verify system under a controlled review budget. Micro-averaged recall within a fixed number of recommendations measures whether reference codes reach the reviewable list; micro-F1 is reported only where reference labels are sufficiently complete. Zero-shot GPT-5.6-Sol achieves the highest recall within five and ten candidates: 0.717 and 0.800 on VUMC and lower-bound values of 0.689 and 0.738 on MIMIC-IV. The retrieve-and-verify system reaches 0.695 and 0.784 on VUMC and lower-bound values of 0.575 and 0.657 on MIMIC-IV, with a candidate-linked evidence window attached to each retained recommendation. Diagnostic analyses reveal distinct failure sources, including output-length underfilling, confusion among closely related codes, out-of-knowledge-base generation, and incomplete evidence support. These findings establish a systematic evaluation framework for procedure-code recommendation and identify practical requirements for future systems that are accurate, review-efficient, and grounded in clinical evidence.
Open AccessCPT/HCPCS Code Recommendation from Clinical Notes: A Comparative Evaluation of AI Methods
Automatic coding from clinical notes has been studied extensively for International Classification of Diseases (ICD) codes, yet broad Current Procedural Terminology (CPT) and Healthcare Common Procedure Coding System (HCPCS) recommendation remains comparatively underexplored. Existing studies often focus on one specialty, a limited code vocabulary, or a single model family, leaving it unclear how different artificial intelligence (AI) paradigms perform under a common, clinically meaningful evaluation. We formulate CPT and HCPCS coding as an AI-assisted recommendation task in which a physician or professional coder reviews a short, ranked list of candidate codes supported by the clinical note. Using operative notes from Vanderbilt University Medical Center (VUMC) and discharge summaries from Medical Information Mart for Intensive Care IV (MIMIC-IV), we compare lexical retrieval, Clinical-Longformer, GPT-5.6-Sol, MedGemma-27B, and an inspectable agentic-style retrieve-and-verify system under a controlled review budget. Micro-averaged recall within a fixed number of recommendations measures whether reference codes reach the reviewable list; micro-F1 is reported only where reference labels are sufficiently complete. Zero-shot GPT-5.6-Sol achieves the highest recall within five and ten candidates: 0.717 and 0.800 on VUMC and lower-bound values of 0.689 and 0.738 on MIMIC-IV. The retrieve-and-verify system reaches 0.695 and 0.784 on VUMC and lower-bound values of 0.575 and 0.657 on MIMIC-IV, with a candidate-linked evidence window attached to each retained recommendation. Diagnostic analyses reveal distinct failure sources, including output-length underfilling, confusion among closely related codes, out-of-knowledge-base generation, and incomplete evidence support. These findings establish a systematic evaluation framework for procedure-code recommendation and identify practical requirements for future systems that are accurate, review-efficient, and grounded in clinical evidence.
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 11363-11374
scCotag: Prior-Informed Co-Optimal Transport for Robust Single-Cell Multi-Omics Integration
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 11363-11374
Open AccessscCotag: Prior-Informed Co-Optimal Transport for Robust Single-Cell Multi-Omics Integration
- Open Access
- 16 July 2026
- Published by Frontiers Media SA in Frontiers in Digital Health
- Vol. 8, 1833087
Robust nasality representation learning for cleft palate-related velopharyngeal dysfunction screening in real-world settings
Background: Velopharyngeal dysfunction (VPD) is an impaired ability to achieve adequate velopharyngeal closure during speech, often resulting in hypernasality and reduced intelligibility. VPD screening and diagnosis require specialized expertise and controlled recording conditions, limiting scalable access outside high-income countries. Key challenge: Speech-based machine learning models can perform extremely well under standardized clinical recording conditions. However, performance often deteriorates when deployed on consumer devices (e.g., phones or tablets) and in uncontrolled acoustic environments. This degradation is largely driven by domain shift arising from differences in recording conditions (e.g., device and channel characteristics, background noise, and room acoustics), which can cause models to rely on spurious recording artifacts rather than pathology-relevant cues. Methods: This study introduces a two-stage framework to improve robustness under realistic recording scenarios. Nasality representation pre-training employs a nasality-focused representation via supervised contrastive learning (SupCon) using an auxiliary dataset with phoneme alignments to form oral-context versus nasal-context supervision. During Frozen-encoder VPD screening , the encoder is frozen to perform VPD screening using lightweight classifiers on 0.5-second chunks with probability aggregation to produce recording-level decisions using a fixed decision threshold. Here, in-domain refers to standardized clinical recordings used for model development, and out-of-domain refers to heterogeneous public Internet recordings collected under uncontrolled conditions and evaluated without any adaptation. The proposed method is then compared against prior-study baselines, including MFCC features and large pretrained speech representations, using the same evaluation protocol. Results: On the primary in-domain subject-disjoint held-out split of 82 subjects (60 train/22 test; 345 training recordings; 131 test recordings; multiple recordings per subject), the proposed approach reached ceiling recording-level screening performance under this standardized clinical protocol (macro-F1 = 1.000, accuracy = 1.000). To assess sensitivity to this fixed split, an additional subject-level nested 5-fold cross-validation analysis was performed on the full in-domain cohort (82 subjects, 476 recordings), with the encoder frozen and only the second-stage classifiers retrained; the best mean performance was obtained with SVM (macro-F1 = 0 . 981 ± 0 . 022, accuracy = 0 . 985 ± 0 . 016). On a separate out-of-domain set of 131 public Internet recordings, large pretrained speech representations degrade substantially, and MFCC is the strongest baseline (macro-F1 = 0.612, accuracy = 0.641). The proposed method achieves the best overall out-of-domain performance (macro-F1 = 0.679, accuracy = 0.695), improving over the strongest baseline by +0.067 macro-F1 and +0.054 accuracy (point-estimate improvements) under the same evaluation protocol and fixed threshold. Conclusion: Learning a nasality-focused representation prior to clinical classification can reduce sensitivity to recording artifacts and improve robustness when moving from the laboratory to real-world audio recording scenarios. This design supports practical deployment of VPD screening and motivates domain-robust evaluation protocols for deployable speech-based digital health tools.
- 16 July 2026
- Published by Frontiers Media SA in Frontiers in Digital Health
- Vol. 8, 1833087
Open AccessRobust nasality representation learning for cleft palate-related velopharyngeal dysfunction screening in real-world settings
Background: Velopharyngeal dysfunction (VPD) is an impaired ability to achieve adequate velopharyngeal closure during speech, often resulting in hypernasality and reduced intelligibility. VPD screening and diagnosis require specialized expertise and controlled recording conditions, limiting scalable access outside high-income countries. Key challenge: Speech-based machine learning models can perform extremely well under standardized clinical recording conditions. However, performance often deteriorates when deployed on consumer devices (e.g., phones or tablets) and in uncontrolled acoustic environments. This degradation is largely driven by domain shift arising from differences in recording conditions (e.g., device and channel characteristics, background noise, and room acoustics), which can cause models to rely on spurious recording artifacts rather than pathology-relevant cues. Methods: This study introduces a two-stage framework to improve robustness under realistic recording scenarios. Nasality representation pre-training employs a nasality-focused representation via supervised contrastive learning (SupCon) using an auxiliary dataset with phoneme alignments to form oral-context versus nasal-context supervision. During Frozen-encoder VPD screening , the encoder is frozen to perform VPD screening using lightweight classifiers on 0.5-second chunks with probability aggregation to produce recording-level decisions using a fixed decision threshold. Here, in-domain refers to standardized clinical recordings used for model development, and out-of-domain refers to heterogeneous public Internet recordings collected under uncontrolled conditions and evaluated without any adaptation. The proposed method is then compared against prior-study baselines, including MFCC features and large pretrained speech representations, using the same evaluation protocol. Results: On the primary in-domain subject-disjoint held-out split of 82 subjects (60 train/22 test; 345 training recordings; 131 test recordings; multiple recordings per subject), the proposed approach reached ceiling recording-level screening performance under this standardized clinical protocol (macro-F1 = 1.000, accuracy = 1.000). To assess sensitivity to this fixed split, an additional subject-level nested 5-fold cross-validation analysis was performed on the full in-domain cohort (82 subjects, 476 recordings), with the encoder frozen and only the second-stage classifiers retrained; the best mean performance was obtained with SVM (macro-F1 = 0 . 981 ± 0 . 022, accuracy = 0 . 985 ± 0 . 016). On a separate out-of-domain set of 131 public Internet recordings, large pretrained speech representations degrade substantially, and MFCC is the strongest baseline (macro-F1 = 0.612, accuracy = 0.641). The proposed method achieves the best overall out-of-domain performance (macro-F1 = 0.679, accuracy = 0.695), improving over the strongest baseline by +0.067 macro-F1 and +0.054 accuracy (point-estimate improvements) under the same evaluation protocol and fixed threshold. Conclusion: Learning a nasality-focused representation prior to clinical classification can reduce sensitivity to recording artifacts and improve robustness when moving from the laboratory to real-world audio recording scenarios. This design supports practical deployment of VPD screening and motivates domain-robust evaluation protocols for deployable speech-based digital health tools.
- Open Access
- 25 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 6702-6722
Beyond the Response: Examining Reasoning and Execution Fidelity in Large Language Models for Mental Health
- 25 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 6702-6722
Open AccessBeyond the Response: Examining Reasoning and Execution Fidelity in Large Language Models for Mental Health
- Open Access
- 13 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-21
Leveraging Affordances as a Lens: A Systematic Review of Social Media Benefits and Risks for Adolescents
- 13 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-21
Open AccessLeveraging Affordances as a Lens: A Systematic Review of Social Media Benefits and Risks for Adolescents