Skip to main content
v2026.11,986 entries · CC-BY 4.0

Editorial · CASRAI · AI and ML research outputs

Multimodal Voice and Clinical Biomarker Model Detects Asthma in npj Digital Medicine Study

A multimodal Mixture-of-Experts model combining voice audio and clinical factors achieved an AUROC of 0.85 in detecting asthma, researchers reported.

Published 22 Nov 2026· 7 minute read
Medical acoustic stethoscope on a clean surface in a clinical diagnostic setting
Photo: Jacek Halicki / Wikimedia Commons, CC BY-SA 4.0 (cropped). Source · Licence

On November 22, 2026, researchers from the Luxembourg Institute of Health published a peer-reviewed study in npj Digital Medicine demonstrating that combining smartphone-recorded voice acoustics with structured clinical and demographic data significantly improves machine-learning detection of asthma. The multimodal framework addresses longstanding accessibility barriers associated with conventional pulmonary function testing, providing university health researchers and digital health investigators with an empirical blueprint for remote respiratory screening architectures.

Study Design and Multimodal Machine-Learning Architecture

Diagnosing asthma globally remains constrained by the equipment and staffing requirements of conventional pulmonary function testing. To determine whether non-invasive digital screening could bridge this gap, an international team led by Dr. Vladimir Despotovic, Milena Despotovic, Abir Elbeji, Dr. Petr V. Nazarov, and Dr. Guy Fagherazzi evaluated a hierarchical Mixture-of-Experts (MoE) machine-learning architecture using participant data from the Colive Voice digital health initiative. Their investigation, published in npj Digital Medicine, examined how vocal features recorded on personal smartphones interact with baseline clinical variables to discriminate individuals with asthma from healthy individuals.

The study cohort comprised 1,218 adult participants matched 1:1 into 609 self-reported asthma cases and 609 healthy controls. Matching was strictly enforced across biological sex (419 females, 68.8%; 190 males, 31.2% per cohort), age (female mean 43.1 ± 14.6 years; male mean 40.0 ± 14.3 years for asthma, 39.8 ± 13.7 years for controls), and primary recording language (371 English recordings, 60.9%; 238 French recordings, 39.1% per group). Control participants had no history of chronic respiratory or neurological disorders affecting phonation.

Audio recordings were collected asynchronously through smartphone web interfaces under real-world acoustic conditions. Each participant completed two distinct tasks: a sustained vowel phonation (/a/) held for at least five seconds, and a standardised phonetically balanced reading passage. Audio recordings were converted into 96 × 64 log-Mel spectrograms, from which transferable acoustic representations were extracted using Bootstrap Your Own Latent for Audio (BYOL-A). Alongside acoustic inputs, the researchers integrated ten structured clinical and demographic parameters: age, sex, body mass index (BMI) category, smoking frequency, alcohol consumption, antihistamine medication use, self-reported stress, respiratory quality of life (VQ11 score, range 11–55), and fatigue severity (Fatigue Severity Scale, range 9–63).

Diagnostic Ablation Findings and Model Performance Metrics

To isolate the diagnostic contribution of each data stream, the investigators executed an ablation study comparing unimodal models against bimodal and tripartite multimodal configurations across repeated 10 × 10 stratified cross-validation folds. The evaluation demonstrated that unimodal voice models possessed modest standalone discriminatory capability, but their fusion with clinical and demographic factors yielded substantial performance gains.

Among unimodal baselines, structured clinical and demographic variables alone achieved an Area Under the Receiver Operating Characteristic curve (AUROC) of 0.75 (95% CI 0.75–0.75) and an accuracy of 0.69 (95% CI 0.68–0.69). Sustained vowel phonation alone produced an AUROC of 0.69 (accuracy 0.63), whereas reading passages alone achieved an AUROC of 0.65 (accuracy 0.61). Bimodal fusion of both audio modalities improved AUROC to 0.75. The tripartite full MoE model—combining sustained vowel phonation, reading passage acoustics, and clinical-demographic variables—delivered the strongest performance: AUROC of 0.85 (95% CI 0.83–0.86), accuracy of 0.77 (95% CI 0.77–0.78), sensitivity of 0.75 (95% CI 0.73–0.77), specificity of 0.79 (95% CI 0.77–0.81), precision of 0.79, and an optimal Brier score of 0.17 (95% CI 0.16–0.17).

The table below summarizes the diagnostic metrics recorded across all evaluated configurations in the cross-validation ablation study:

Model Configuration AUROC (95% CI) Accuracy (95% CI) Sensitivity (95% CI) Specificity (95% CI) Brier Score
Reading Passage (RP) Unimodal 0.65 (0.65–0.65) 0.61 (0.60–0.61) 0.57 (0.56–0.58) 0.64 (0.64–0.65) 0.23
Sustained Vowel (SV) Unimodal 0.69 (0.69–0.69) 0.63 (0.63–0.63) 0.66 (0.65–0.66) 0.61 (0.59–0.62) 0.22
Clinical and Demographic (CD) Unimodal 0.75 (0.75–0.75) 0.69 (0.68–0.69) 0.66 (0.63–0.68) 0.72 (0.70–0.74) 0.20
MoE Audio Bimodal (SV + RP) 0.75 (0.74–0.76) 0.69 (0.68–0.70) 0.72 (0.66–0.77) 0.66 (0.61–0.70) 0.21
MoE Hybrid Bimodal (SV + CD) 0.83 (0.82–0.84) 0.76 (0.76–0.77) 0.76 (0.74–0.78) 0.76 (0.74–0.78) 0.18
MoE Hybrid Bimodal (RP + CD) 0.84 (0.83–0.86) 0.76 (0.74–0.77) 0.67 (0.64–0.71) 0.84 (0.82–0.86) 0.18
Full MoE Multimodal (SV + RP + CD) 0.85 (0.83–0.86) 0.77 (0.77–0.78) 0.75 (0.73–0.77) 0.79 (0.77–0.81) 0.17

Evaluation of the hierarchical gating mechanism uncovered distinct operational dynamics between modalities. In participants diagnosed with asthma, greater respiratory symptom burden (elevated VQ11 scores) was significantly correlated with increased gating weights assigned to reading-passage acoustics and reduced weighting of sustained vowel phonation. This pattern indicates that continuous speech tasks provide enhanced discriminative signal during symptomatic respiratory distress, whereas static vowel phonation contributes more steadily during lower symptom severity.

Diagnostic Challenges in Spirometry and Respiratory Clinical Practice

The motivation for exploring acoustic biomarkers stems directly from operational bottlenecks in standard pulmonary assessment. Clinical practice parameters established by the Global Initiative for Asthma (GINA) and diagnostic criteria published by the National Heart, Lung, and Blood Institute (NHLBI) define asthma as a chronic inflammatory disorder characterized by variable respiratory symptoms and documented expiratory airflow limitation. Objective confirmation requires spirometry with bronchodilator reversibility testing, defined in adults as a post-bronchodilator increase in forced expiratory volume in one second (FEV1) of at least 12% and 200 mL relative to baseline values.

Despite its diagnostic authority, spirometry remains underutilized in decentralized research and routine primary care due to equipment costs, operator training requirements, and patient burden during forced exhalations. As hybrid clinical trials expand, academic sponsors increasingly explore objective, non-invasive tools such as wearable devices and digital health technologies to monitor respiratory variation without mandating frequent in-clinic visits.

Research Community and Funder Responses

The findings have generated interest across digital health researchers and clinical informatics departments. Dr. Guy Fagherazzi, director of the Department of Precision Health at the Luxembourg Institute of Health and senior investigator on the study, emphasized that vocal biomarkers should not be conceptualized as stand-alone tests replacing clinical assessment. Instead, the adaptive gating network synthesizes historical demographic risks, current symptom burden, and phonatory acoustic properties into a unified prediction score.

The Colive Voice project operates with internal institutional funding from the Luxembourg Institute of Health, without commercial sponsor involvement in protocol design or reporting. Independent biostatisticians noted that the use of BYOL-A representations avoids manual acoustic feature engineering, mitigating potential algorithmic bias related to native language differences across English and French speakers.

Implications for Clinical Research Operations and Trial Protocols

For university research offices and clinical trial sponsors, smartphone-based vocal biomarkers introduce specific operational considerations:

  • Protocol Design and Decentralized Trials: Capturing sustained vowels and reading passages via web interfaces enables low-burden longitudinal respiratory monitoring in decentralized trials, recording symptom fluctuations between scheduled site visits.
  • Data Governance and Biometric Privacy: Voice recordings constitute biometric data. Research ethics committees and institutional review boards must enforce encryption standards and secure storage architectures that segregate raw audio files from clinical research records.
  • Regulatory Validation Pathways: Applying digital biomarkers as clinical trial endpoints requires formal qualification. Academic consortia pursuing validation must engage structured programs, such as the FDA Biomarker Qualification Program, to establish analytical and clinical validity before tools can support diagnostic stratification.
  • Algorithmic Reproducibility and Transparency: Developing reproducible machine learning workflows requires adherence to international reporting standards, including TRIPOD+AI and STARD-AI frameworks, supported by rigorous prospective clinical trial registration for validation cohorts.

Next Steps for Prospective Clinical Validation and Regulatory Clearance

The research team outlined several critical milestones required before voice-based models can be implemented in clinical software:

  • Prospective Spirometric Validation: Because the Colive Voice database relied on self-reported asthma diagnoses, subsequent prospective studies must benchmark vocal biomarkers directly against clinical spirometry and bronchodilator reversibility testing to confirm ground truth diagnostic labels.
  • Pediatric Cohort Expansion: The current model was evaluated exclusively in adult participants aged 18 to 69. Because asthma represents the most prevalent chronic pediatric disease, evaluating acoustic models in pediatric cohorts represents an urgent research priority.
  • External Hardware Generalizability: Prospective trials must evaluate model robustness across varied smartphone microphones, operating systems, ambient noise profiles, and additional global languages to ensure consistent calibration across diverse socioeconomic settings.
  • Regulatory Clearance Requirements: The multimodal framework currently exists as an investigational research prototype. Deployment as clinical decision support software or software as a medical device (SaMD) requires formal clearance from regulatory authorities, such as FDA 510(k) or De Novo clearance in the United States and CE-mark certification under the European Union Medical Device Regulation (EU MDR).

Sources

Ask CASRAI · full access

Ask about this story

Full access: large monthly usage, up to 10 source passages behind each answer, and MCP and API access for Claude Code, Codex and other CLIs. Cancel any time.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →